id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
d065c64e-9c14-4b72-9005-4fccf36d1ecb | trentmkelly/LessWrong-43k | LessWrong | I changed my mind about orca intelligence
Follow up to: My previous posts on orca intelligence.
TLDR: I now think it’s <1% likely that average orcas are >=+6std intelligent.
(I now think the relevant question is rather whether orcas might be >=+4std intelligent, since that might be enough for superhuman wisdom and thinking techniques to accumulate through generations, but I think it’s only 2% probable. (Still decently likely that they are near human level smart though.))
1. Insight: Think about societies instead of individuals
I previously thought of +7std orcas like having +7std potential but growing up in a hunter-gatherer-like environment where the potential isn’t significantly realized and they don’t end up that good at abstract reasoning. I imagined them as being untrained and not knowing much. I still think that a +7std human who grew up in a hunter-gatherer society wouldn’t be all that awesome at learning math and science as an adult (though maybe still decently good).
But I think that’s the wrong way to think about orcas. If the average orca was +6std intelligent, it would be a +6std society, where cultural evolution would happen on ultrasteroids, knowledge and skill accumulates quickly and gets taught effectively, so orcas would end up extremely competent. I think the equilibrium for +6std societies is very likely that they act sorta like a macroagent pursuing the multi-agent optimum, aka sorta like dath ilan[1], even if they don’t have science or a growing economy. The amount of expertise that can be transmitted by smart people is pretty large even without writing.
When I realized that, I stopped thinking about “what would I expect to see if there existed +7std orca individuals” (where I implicitly imagined +0std societies) and shifted to thinking “what would I expect to see from small +6std societies without writing and science”, where in my opinion, the latter implies an extremely much higher level of competence, though I’m probably bad at communicating why.
I previously thought that it’s |
a64dd93a-f0be-402b-b75d-138f70b4df2f | trentmkelly/LessWrong-43k | LessWrong | Meetup : Urbana-Champaign: Politics and the English Language
Discussion article for the meetup : Urbana-Champaign: Politics and the English Language
WHEN: 16 February 2014 02:00:00PM (-0600)
WHERE: 40.109545,-88.2273
Read this: https://www.mtholyoke.edu/acad/intrel/orwell46.htm
And then show up and discuss it.
Discussion article for the meetup : Urbana-Champaign: Politics and the English Language |
1872624d-6e7b-4b92-a959-f93b9d8c4cb8 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Interpretable Latent Spaces for Learning from Demonstration
1 Introduction
---------------
We want our autonomous robots to be competent across broad domains. Ideally, they must be able to cope with the open world, i.e., one where it may not be possible to provide a complete specification of the task up front. Thus, it is important for the robotic agent to incrementally learn about the world and its regularities in order to be able to efficiently generalize to new situations. It can be argued that such generalizations are made easier when the models used by the robot can represent concepts ranging from object-hood to causality.
Learning from human demonstration [[1](#bib.bib1), [2](#bib.bib2)] is an efficient way to transfer such knowledge to a robot, wherein the human expert teaches a robot by showing it instances of execution of the task of interest. There are numerous examples of fairly sophisticated behaviours being learnt by robots, such as for control and planning in high dimensional systems [[3](#bib.bib3), [4](#bib.bib4), [5](#bib.bib5)]. A common theme across a majority of these works is that the target of transfer is the specific behaviour, so that the robot is taught to mimic a motion (with learning methods being used to generalise over, say, differing body configurations). We are also interested in being able to transfer other aspects of knowledge about the world which may allow the robot to infer deeper concepts. A necessary first step towards this form of teaching or transfer is to give the robotic agent the ability to learn models that represent structure in ways that are similar to corresponding human notions - so that it may be possible for the human expert and robot to leverage common grounding [[6](#bib.bib6)].
From a different direction, there is recognition among roboticists that models and policies used by autonomous robotic agents should be explainable [[7](#bib.bib7), [8](#bib.bib8)]. The notion of what it means to interpret and explain is in itself a topic of active debate at the moment, but we believe it is safe to claim that grounded models used by robots should satisfy key desiderata. For instance, the notion of similarity with which items are grouped together must align with corresponding human notions or that factors of variation implied by the model should be understandable to a human expert. These desiderata are also closely aligned with the emerging literature on learning disentangled representations [[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [12](#bib.bib12)].
In this paper, we focus on learning models of objects of the kind that robots might encounter in manipulation settings. In order to make the underlying concepts fully clear, we use simple objects that uncontroversially expose the structure of the domain. For instance, we show that after being exposed only to a few instances of a ball and a cube, the agent can figure out that other instances of cubes can be labelled as such (without this fact being explicitly stated by the expert).
Our main contribution consists of a framework—see figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Interpretable Latent Spaces for Learning from Demonstration")—which allows for independent user-defined factors of variation, manifested in a high-dimensional space, to be projected to a lower-dimensional latent space by preserving the factors’ orthogonality. The latter is guaranteed by aligning each of the basis vectors that span the latent space with a single factor. Each factor is specified as a set of weak labels over the high-dimensional space. So, for the same data the framework can learn different representations for different sets of symbols. We test the framework both on a synthetic dataset (modified dSprites [[13](#bib.bib13)] with added color), with controlled factors of variation and on a dataset of real-world objects captured from a set of human demonstrations for the task of sorting table-top objects according to the user’s preference.
Figure 1: Overview of the full framework. The expert is demonstrating the task at hand while uttering symbols which are grounded according to their attention in the environment. The conceptual grouping of the symbols is given. (a) The gathered dataset is used to train a variational autoencoder with a set of auxiliary classification losses - one per conceptual group. Each classifier takes information from a single latent dimension. (b) The training procedure guarantees that the semantic meaning of each group will be encoded in a separate latent dimension with linearly separable classes (c) Using the per-class estimated 1D Normal distributions, together with in-between distributions for unknown labels, we can perform 1-NN classification (d)
2 Related Work
---------------
An efficient, and unobtrusive, learning process must need minimal effort on the part of the human expert, requiring the agent to be able to generalise to unobserved scenarios. At the same time, the models and concepts used by the agent must be interpretable by the human, in order for the interaction to be efficient. 0-shot, 1-shot [[14](#bib.bib14)] and meta-learning [[15](#bib.bib15)] are all approaches which strive for fast learning and better generalisation from limited amounts of data. However, from the point of view of ensuring safe operation by a robot, it is also important to be able to quantify when an unfamiliar situation occurs and to seek guidance from an expert.
The ability to deal with examples that fall outside the domain of the training distribution is known as ‘open world recognition’ [[16](#bib.bib16), [17](#bib.bib17)]. This is needed when an agent comes across unknown unknowns - observations that have not been seen during training. In this setting, the incomplete nature of the agent’s model causes it to make the wrong prediction. A complete model would cope with such cases by explicitly recognising that data point as ‘unknown’. It is desirable for the agent to be able to express its ignorance to the expert in a way that is comprehensible to them. This motivates the need for interpretability - we learn a representation that enables communication for user feedback regarding partially labelled sensory observations.
In the context of visual sensory inputs, ‘inverse graphics’ is an approach to learning compressed and factorised representations of high-dimensional visual data - e.g. learn to invert a graphical renderer and infer (for a given image) the parameter vector that generated an input image. The fact that such vectors are low-dimensional and factorised is what makes them appealing from a human-interpretability point of view [[18](#bib.bib18), [19](#bib.bib19)].
In the representation learning literature, there has been work [[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11), [20](#bib.bib20), [21](#bib.bib21)] aimed at unsupervised learning of inverse-graphics-like representations of image datasets. This has been demonstrated with models like InfoGAN [[10](#bib.bib10)] which does learn the desired factorised representations in an unsupervised fashion but also suffers from problems like unstable training and sensitivity to the chosen prior distribution over the latent codes. Moreover, there is no explicit inference module that would produce a latent code, given an observation, even though the recognition network Q can be used as such. An alternative in the literature is the β-VAE model [[9](#bib.bib9)] which has fewer parameters to tune, has more stable training and reportedly learns similar representations. The factors of variation which are encoded in these representations are the ones that best explain the visual variations in the data [[12](#bib.bib12)] but might not necessarily be the factors of variation which are crucial to recognise and discriminate in the context of a particular task. Inferring which learnt factors best ground the user-defined symbols is usually a separate process wherein inference must be performed for each label-factor pair. Learning a representation suitable for symbol grounding [[22](#bib.bib22), [23](#bib.bib23)] from raw pixels in an unsupervised manner, based on β-VAE, has turned out to be non-trivial [[24](#bib.bib24)]. In this work we employ methods based on deep variational inference [[25](#bib.bib25)] and perform the learning of a factorised vector space and the grounding of symbols in it simultaneously, in a weakly-supervised fashion. Thus we achieve better alignment between the user-defined semantic conceptual groups and the basis vectors of the vector space.
3 Background and Problem Formulation
-------------------------------------
In this work, we assume that the human expert and robotic agent share the environment and that they can both observe and interact with objects in the surrounding world. The agent can extract images of distinct objects from the world and receive labels for them from the expert.
Data capture can be achieved with any technique that is able to align a sequential abstract plan for a task (parsed from a linguistic construct) with a temporal trace of expert movements and actions in the external environment - e.g. gesture recognition [[26](#bib.bib26)], movement tracking [[27](#bib.bib27)], eye-tracking [[28](#bib.bib28)]. Thus, specific symbols from the plan can be related to regions of interest in the environment, where the object, encoding the label’s meaning, resides.
Let O={x1,…,xN},xi∈RD be a set of unlabelled observations, which have been passively gathered prior to the expert demonstration. An expert demonstration introduces a set of L conceptual groups G = {g1,…,gL} (e.g. color, shape, size, etc.), where each group is a set of mutually exclusive discrete labels: gi={yi1,...,yini},ni=|gi| (e.g. color can be red, blue, yellow, etc.). The demonstration also provides labels, one drawn from each concept group, to M of the observations: O={(x1,y1),…,(xM,yM),(xM+1,∅)…(xN,∅)}, yj={yp}, p∈{1,…,L}, yp∈gp is a set of labels which have been attached to the image x - one per concept group.
The task is to project each x∈RD into z∈RC , C≪D, where the space of RC possesses the following properties:
Axes Alignment - guarantees a one-to-one mapping from the concept groups G, to the orthogonal basis vectors which span RC. This would guarantee that independent concepts in image space are kept independent in RC—e.g. color does not depend on the size, orientation or shape of an object.
Intra-group Linear Separability - the latent clusters in RC, corresponding to the labels in each concept group gi, are linearly separable across the basis vector which has been aligned with that concept group.
These two properties allow for RC to be used as a feature space for performing probabilistic symbol grounding with the ability to recognise unknown objects. For that we use 1-NN classification with 1D normal distributions (one per label) for each concept group in G—see Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Interpretable Latent Spaces for Learning from Demonstration") (c) and (d).
4 Methodology
--------------
We explore the effects of adding an auxiliary classification loss to a Variational Autoencoder [[25](#bib.bib25)] as a base architecture, specifically the β-VAE [[9](#bib.bib9)]. Through weak supervision, in the form of partially labelled data, the auxiliary loss influences the latent space of the model to exhibit properties which make it suitable for robust symbol classification. The model consists of a convolutional encoder network qϕ, parametrised by ϕ, a deconvolutional decoder network pθ, parametrised by θ, and a set of linear classifiers parametrised by wi∈R|gi| for each group gi∈G. Additional parameters—α,β,γ—are added on the three terms of the overall loss function—see ([1](#S4.E1 "(1) ‣ 4 Methodology ‣ Interpretable Latent Spaces for Learning from Demonstration"))—in order to leverage their importance.
| | | | |
| --- | --- | --- | --- |
| | minθ,ϕ,WL(x,y,θ,ϕ)=βDKL(qϕ(z|x)||pθ(z))−αEqϕ(z|x)(logpθ(x|z))+γ|G|∑iH(ziwTi,yi) | | (1) |
Classification term (weighted by γ) - In order to force the learnt latent space to explain the variations in the data, we add a linear classifier for each concept group. Each classifier has to predict the set of labels for its assigned concept group using information only from a single latent dimension. No two classifiers have access to the same latent dimension. That forces each dimension to only explain labels from the particular concept group. A discrete cross-entropy term is used for the predictions of each classifier.
Reconstruction term (weighted by α) - The reconstruction loss is a standard Bernoulli Negative Log Likelihood, which is used to predict the pixel values across the three RGB channels. The motivation behind this term is that we do not assume all data points in our dataset are labelled. Thus the reconstruction loss would force data points which look similar in image space to be projected close to each other in the latent space.
Kullback-Leibler divergence term (weighted by β) - The Kullback-Leibler divergence term ensures that the distribution of the latent projections of the data in RC does not diverge from a prior isotropic normal distribution. A perfectly optimised KL term would result in all latent projections to be 0. This forces the encoder network qϕ to be more efficient when encoding the image observations so that their latent projection can be discriminated from each other across the basis vectors aligned with G—the classification term—and the decoder network pθ can efficiently reconstruct them—the reconstruction term.
The values for all three coefficients are chosen empirically such that the values for all the loss terms have similar magnitude and thus none of them overwhelms the gradient updates while training.
In order to account for data-generative factors of variation that might not be needed to encode the conceptual groups G but are still essential for good reconstruction and stable training, we allow |RC|>|G|. For example, spatial and rotational factors of variation would not contribute to explaining the semantics of concepts like shape and size but should still be accounted for in order for the subset of basis vectors in RC which are aligned with G to encode only information which explain G.
Algorithm [1](#algorithm1 "Algorithm 1 ‣ 4 Methodology ‣ Interpretable Latent Spaces for Learning from Demonstration") describes the core functionality of the framework. Initially we have a set of partially labelled observations, a set of conceptual groups with their labels, and a single untrained linear classifier for each group. In the training process, each observed image x is passed through the network and its reconstruction ^x∼pθ(x|z),z∼qϕ(x) is fed to the loss L. If the data point is labelled, a label is predicted for each conceptual group from the respective classifier. After training, we estimate the parameters of a 1D normal distribution for each label across the dimension which was responsible for predicting it during learning.
Input: observations O={(x1,y1),…,(xM,yM),(xM+1,∅)…(xN,∅)}
Input: conceptual groups G = {g1,…,gL}
Input: linear classifiers W={wi},i∈{1,…,L}
Input: Isotropic Normal Prior p(z)=N(0,I)
Output: set of per-label estimated 1D normal distribution for each label in each conceptual group: K={{N(μpq,σpq)},p∈{1,…,L},q∈{1,…,|gp|}
1 while *not converged* do
2 for *each (x,y) in O* do
3 ^μ,^σ←Encode(x);
4 ^z∼N(^μ,^σI);
5 if *y≠∅* then
6 for *each wi in W* do
7 ^yi←ziwTi;
8
9 ^x←Decode(^z);
10 Use ^x, ^y, ^μ and ^σ to compute L—see ([1](#S4.E1 "(1) ‣ 4 Methodology ‣ Interpretable Latent Spaces for Learning from Demonstration"));
11
12for *each gi in G* do
13 for *each yij in gi, j∈{1,…,|gi|}* do
14 f←{(x,y)∈O|yij∈y};
15 ^μ,^σ←Encode(f);
16 ^z∼N(^μ,^σI);
17 N(μij,σij)←fitNormal(^zi);
18 Add N(μij,σij) to K;
19
Algorithm 1 Model Learning with Weak Supervision
When classification is performed at test time, the task is to predict a set of L labels y for each image observation x. Classification is performed by using the factored projection of each image observation in RC. The 1D coordinates along each basis vector zi of RC are used to predict a label for the correspondingly aligned concept group gi. Each 1D value is normalised with respect to the normal distributions in K corresponding to gi and the class associated with the closest one, along zi is assigned to x for gi. As a consequence of optimising the Kullback-Leibler divergence term, together with the reconstruction and classification losses, any data points that have not been labelled and do not resemble the labelled ones end up being projected closer towards the origin of the latent space RC, in between the clusters associated with the labelled data. Thus, in order to be able to account for such unknown objects, for every two neighbouring distributions N(μil,σil) and N(μir,σir), along the basis vector zi, aligned with gi, we fit an average distribution Nu(μiu,σiu) where μiu=μil+μir2 and σiu=σil+σir2. Any data point that is closer to such an unknown distribution than to a labelled one is considered unknown.
5 Experiments
--------------
###
5.1 Data
The controlled data-generative factors of variation of the modified dSprites dataset—see figure [2](#S5.F2 "Figure 2 ‣ 5.1 Data ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration")—make it suitable for exploring how the two baselines compare to the proposed model with respect to the defined manifold properties in section [3](#S3 "3 Background and Problem Formulation ‣ Interpretable Latent Spaces for Learning from Demonstration"). The resulting dataset is of size 3500 images—72 objects with spatial x/y variations in the image. We perform two experiments with the same underlying dataset but different sets of symbols in order to demonstrate how the user’s preference is encoded in the latent space.
Figure 2: A modified version of the dSprites [[13](#bib.bib13)] dataset with added color. The RGB equivalents of 8 colors—two variations of red, green, blue and yellow—are added. Spatial factors of variation are also present in the images but not shown in the figure.
Figure 3: Example images crops of all table-top objects. Spatial and pixel noise factors of variation were added to each crop through data augmentation techniques.
In order to demonstrate the application of the framework to real-world human-robot interaction scenarios, a second dataset of objects on a table-top is gathered from a human demonstration - Figure [3](#S5.F3 "Figure 3 ‣ 5.1 Data ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"). The task the human performs is to separate the objects by their function - juggling balls vs orbs, and then by their color - red vs yellow vs blue. Lego blocks and whiteboard pins are also present in the scene, but they are not manipulated and no label information is given about them from the expert. At test time the agent has to repeat the task, with new objects being present in the scene that were previously unobserved—green objects and a yellow rubber duck. Each object image is augmented, resulting in a dataset of size 7500 images of 15 objects with spatial variation. Both datasets are split into a training-testing sets with an 80-20 ratio. All results are reported on the test set. The setup for gathering and labelling per-object image crops in a single human demonstration through the use of eye-tracking is described in Appendix [A](#A1 "Appendix A Demonstration and Object Crops Gathering ‣ Interpretable Latent Spaces for Learning from Demonstration").
###
5.2 Evaluation
Two baselines are used to benchmark against the proposed architecture - a Vanilla Beta-VAE (γ=0) and a Convolutional Classifier Network with Kullback-Leibler divergence and no reconstruction term (α=0). More details on the network architecture and training setups are provided in Appendices [B](#A2 "Appendix B Network Architecture ‣ Interpretable Latent Spaces for Learning from Demonstration") and [C](#A3 "Appendix C Hyper parameters ‣ Interpretable Latent Spaces for Learning from Demonstration"). In order to evaluate the extent to which each baseline satisfies the properties we define in Section [3](#S3 "3 Background and Problem Formulation ‣ Interpretable Latent Spaces for Learning from Demonstration"), we use metrics which are inspired by the literature on learning disentangled representations [[29](#bib.bib29)]
Axes Alignment - In order to determine how well the concept groups in G are aligned with the basis vectors that span the latent space RC we perform PCA on the latent projections of the data points for each label and examine the alignment of the resultant eigenvectors with the basis vectors. If a particular concept group is aligned with a single basis vector, then the eigenvector with the smallest eigenvalue should be parallel to that basis vector. Such alignment would mean that traversing a single concept group in image space corresponds to perturbing the values of a single basis vector. For example, if zi encodes the concept of color, then the latent distribution for all blue datapoints should have small variance along zi and large variance across all other zj≠i. To examine this, for each label in each concept group, we plot the cosine similarity diagrams depicting the cosine distance between each pair of basis and eigenvectors—see figure [4](#S5.F4 "Figure 4 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"). White cells mark the cosine similarities between the shortest eigenvector with all basis vectors and between its closest parallel basis vector and all other principal components. The average entropy of the normalised white cosine values is reported for each combination of model-experiment to ease quantitative analysis. Low entropy corresponds to axis alignment, resulting in a single big white cell (for the smallest eigenvector and its most-parallel basis vector) and remaining small white cells along the row and column of the big one. For each concept group in each experiment we should observe the same white cell patterns if axis alignment is being achieved. For further information on this point, please see the supplementary materials 111<https://sites.google.com/view/interpretable-latent-spaces/axes-align-explain>.
Intra-group Linear Separability - In order to perform 1-NN classification at test time with unlabelled objects, using 1D Normal distributions across the concept groups in G, the latent cluster in RC for each label yij∈gi has to be linearly separable from the latent clusters for all other labels from gi. We report F1 scores for each class label, per concept group, including predictions for unlabelled observations which represent both known and unknown labels.
###
5.3 Results
Experiment 1 - learn z0≡ color and z1≡ size. The user-uttered labels for color are red, blue and for size are small, big. The color labels are assigned to a single variation of the respective color. All images which can not be described by the given labels—yellow and green for color and medium for size—are given an unknown ground truth label. Total |Z| = 4.
Experiment 2 - learn z0≡ shape and z1≡ size. The user-uttered labels for shape are square, heart and for size are small, big. All images which can not be described by the user-uttered labels—ellipse for shape and medium for size—are given an unknown ground truth label. Total |Z| = 4.
Experiment 3 learn z0≡ color and z1≡ object type. The user-uttered labels for color are red, blue and for object type are juggle ball, orb. All images which can not be described by the user-uttered labels—lego bricks and whiteboard pins, rubber ducks or green objects—are given an unknown ground truth label. Total |Z| = 4.
Figure 4: Axes Alignment evaluation for the full model (a), the classifier baseline (b) and the β-VAE (c) for the 3 experiments - (1), (2) and (3) respectively. For each experiment and each model, the cosine similarity diagram for each label is shown. In each diagram the size of square (i,j) represents the cosine distance between eigenvector cj and basis vector zi in RC, after performing PCA on the latent values for the corresponding label. All white cells mark the cosine similarities between the smallest eigenvector with all basis vectors and between its most-parallel basis vector and all other principal components. E denotes the average entropy estimates over normalised white cells values for a single model and a single experiment.
Figure [4](#S5.F4 "Figure 4 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration") presents the main findings from our experiments. Both the full model and the classifier network - figure [4](#S5.F4 "Figure 4 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration") (a,b) achieve good alignment between the concept groups in G and the basis vectors of RC. All labels from a particular concept group are consistently explained by the latent dimension which was used to predict them - e.g. color is explained by z0 and size is explained by z1 for experiment 1. For the classifier network - figure [4](#S5.F4 "Figure 4 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration") (b) we show only the first two latent dimensions which were used to perform the classification (hence containing relevant information). The β-VAE - figure [4](#S5.F4 "Figure 4 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration") (c) fails to satisfy the axis alignment requirement - labels from the same concept groups are best explained by different basis vectors—e.g. color in experiment 1. For this model we deterministically chose to use zi to predict labels for gi - Table [1](#S5.T1 "Table 1 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"). This is of course owing to the fact that the β-VAE baseline is trained in a completely unsupervised fashion. We will report this to bring out the observation that non-axis-aligned representations render the 1-NN classification process to be non-trivial, as it is not at all clear which axes should be used to decide on the labels for each concept group. The full model and the classification network baseline perform comparably well with respect to classifying labelled data points. However, the classifier-only baseline is not as good at discriminating between known and unknown objects - unknown columns in Tables [0(a)](#S5.T0.st1 "(a) ‣ Table 1 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"), [0(b)](#S5.T0.st2 "(b) ‣ Table 1 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"), [0(c)](#S5.T0.st3 "(c) ‣ Table 1 ‣ 5.3 Results ‣ 5 Experiments ‣ Interpretable Latent Spaces for Learning from Demonstration"). We speculate that this is a result of the model having access only to labelled images during training, unlike the weakly trained full model. Thus, it is not able to pull together in the latent space both labelled and unlabelled visually similar data points in image space and push away visually different data points in image space. We provide more detailed analysis in the supplementary materials 222<https://sites.google.com/view/interpretable-latent-spaces/>.
| | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | | |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
|
| Model | blue | red | unknown | big | small | unknown |
| --- | --- | --- | --- | --- | --- | --- |
| α = 1, γ≠0 | 0.88 | 0.98 | 0.87 | 0.82 | 0.93 | 0.64 |
| α = 0 | 0.76 | 0.78 | 0.29 | 0.84 | 0.87 | 0.52 |
| γ = 0 | 0.05 | 0.37 | 0.23 | 0.45 | 0.34 | 0.26 |
(a) dSprites - z0≡ color and z1≡ size
|
| Model | heart | square | unknown | big | small | unknown |
| --- | --- | --- | --- | --- | --- | --- |
| α = 1, γ≠0 | 0.70 | 0.73 | 0.34 | 0.71 | 0.83 | 0.6 |
| α = 0 | 0.97 | 0.8 | 0.67 | 0.53 | 0.79 | 0.5 |
| γ = 0 | 0.2 | 0.33 | 0.32 | 0.4 | 0.38 | 0.33 |
(b) dSprites - z0≡ shape and z1≡ size
|
| Model | blue | red | yellow | unknown | juggle ball | orb | unknown |
| --- | --- | --- | --- | --- | --- | --- | --- |
| α = 1, γ≠0 | 0.75 | 0.87 | 0.86 | 0.95 | 0.8 | 0.62 | 0.81 |
| α = 0 | 0.7 | 0.31 | 0.89 | 0.1 | 0.64 | 0.64 | 0.41 |
| γ = 0 | 0.2 | 0.25 | 0.25 | 0.18 | 0.1 | 0.3 | 0.26 |
(c) Real objects - z0≡ color and z1≡ object type
|
Table 1: Evaluation of the discriminative abilities of each model for each experiment. F1 scores are reported for experiment 1 (a), experiment 2 (b) and experiment 3 (c).
6 Conclusion
-------------
We present a framework for physical symbol grounding where linguistically-defined semantic concepts from an expert, manifested in a high-dimensional image space, are mapped to a lower-dimensional learnt latent space. The resultant latent projections preserve any orthogonality between the user-defined concepts. In this sense, they are sufficient to perform robust—among other things, able to recognise unknown unknowns—and sample-efficient symbol inference. We demonstrate this through experiments with images of computer-generated and real physical objects.
\acknowledgments
This work is partly supported by ERC Grant 269427 (STAC), a Xerox University Affairs Committee grant, and the CoGLE project under the DARPA XAI programme. |
35336c34-6f46-4ada-8679-f55eb409aa6a | trentmkelly/LessWrong-43k | LessWrong | Meetup : Vancouver
Discussion article for the meetup : Vancouver
WHEN: 27 September 2012 06:00:00PM (-0700)
WHERE: 885 west georgia vancouver
Last meetup was fun, it's almost time for the next!
Meet us in the lobby of 885 west georgia at 18:00 on thursday. Better to be early than late. We go upstairs at about 18:30, after which it will be very difficult to meet up with us. Come on out.
I'll be leading a discussion of Yvain's rational home buying and disolving questions about desiese because they are such good examples of LW ideas in action.
Last time, I presented Letting Go from HTACYM. The consensus on the Letting Go sequence was that it was a bag of tricks and inspiration for dealing with the need to face reality (when you are trying to actually change the world). We had a lot of discussion of what you can expect from other people vs what you expect from yourself, dealing with emotions, value of predictive vs descriptive explanations, some rationality fundamentals, some fun math proofs, etc.
Hopefully we'll have some more cool discussion and maybe even some application of these ideas to our own lives.
As usual, join us on the mailing list. Be there.
Discussion article for the meetup : Vancouver |
23a925bf-b73d-47d7-86d2-4b35be05553c | trentmkelly/LessWrong-43k | LessWrong | The Price Of Space Is Friendship
It's an exciting time in the history of human exploration. We're only just beginning to probe outside our biosphere. Low-orbit, long-term space inhabitation like that seen on the ISS has been the norm for many years now. As of the time of writing, 7 people were in space. Our long term goals seem set on permanent outposts on Mars.
Orbital re-entry of crewed spacecraft is a complicated dance of physics. You ever do that challenge where you have to build a cage out of straws and cardboard to protect an egg from a fall? Well, it's like that on steroids. Generally, if you are coming in from another interstellar body you will be going very fast. You can't decelerate over around 90m/s/s (or 9gs) or let the interior of the ship get too hot without killing your crew. You've got an enormous amount of kinetic energy to disperse, and the best way to do that is by braking in the atmosphere at a relatively shallow angle of descent. This means that reentry maneuvers cover a lot of airspace. You will likely either orbit the earth entirely or a large portion of it before you finally reduce your velocity enough to touch down.
The same is true of orbital ejections from Earth. Generally we try to launch from near the equator as we can, and in the direction of the Earth's rotation (easterly). This is so we can steal a bit of this initial speed from the Earth itself. To get ourselves from one stellar body to another we might do a Hohmann transfer, which means there is going to be some arbitrary point in the stable orbital trajectory where we need to fire some impulse to get onto this new trajectory (where line '2' begins below). That arc over low-earth orbit can be of any length, from a few moments to multiple orbits around the globe, and stretch over any number of nation-states.
(Courtesy of WikiMedia)
There have been significant developments in the ability for nation-states to control their low-earth orbit even in the past few decades. Anti-satellite systems are a rapidly growing se |
4ef1c229-5f1f-45b1-a04a-ea8bce517008 | trentmkelly/LessWrong-43k | LessWrong | [LINK] 52 Concepts To Add To Your Cognitive Toolkit
Excellent list by Brenton Mayer and Peter McIntyre: http://mcntyr.com/52-concepts-cognitive-toolkit/
I think the list can also serve as a useful index and/or introduction to a lot of LessWrong concepts.
A note of caution: I find that brief lists like this can actually be counterproductive, since they make you feel like you understand the issues when all you did was read a short definition and peg a name on the concept. I'd recommend doing the following: Look through the list carefully and slowly. If there's a concept there that you've read a lot about then you can go on to the next one, although do take a look at where the link points to in case it's to an interesting article you haven't seen before. If you haven't read a lot about the concept then ideally you should click on the link and read all about it. If you're more pressed for time, then at least take a few moments to reflect on each concept and think how it might apply to you. If you have even a slight suspicion that there might be something in the concept that wasn't completely obvious to you before, then click on the link even though you're pressed for time. If you're so time constrained that you can't even do this, then consider just bookmarking the list and getting back to it later. Personally I think it's better to read it later carefully than to read it now and think you understand it when you really don't. |
91130618-f62a-40f9-a14f-c12367f7eca1 | trentmkelly/LessWrong-43k | LessWrong | Love and Rationality: Less Wrongers on OKCupid
Last month, Will_Newsome started a thread about OKCupid, one of the major players among online dating sites--especially for the young-and-nerdy set, given their mathematical approach to matching. He opened it up for individual profile evaluation, which occurred, but so did a lot of fruitful meta-discussion about attraction in general and online dating mechanisms in particular. This post is a summary of the parts of that thread which specifically address the practical aspect of good profile editing and critique. (It also incorporates some ideas I had previously but hadn't collected yet.) A little of it is specific to OKCupid, but most of it can be applied to any dating site, and some to dating in general. I've cited points which came from single comments (i.e. not suggested by several people); if I missed one of yours, please comment with a link and I'll add the reference.
On OKTrends
"Wait a minute," I hear experienced OKCers cry. "Why reinvent the wheel of profile analysis? OKCupid already has a blog for just that, and it's called OKTrends."
OKTrends has its merits, but it also has one major flaw. Wei_Dai summed it up well by observing that OKTrends does not make "any effort to distinguish between correlation and causation," citing this post as an example. The reason for that is obvious: the first purpose of OKTrends is to bring traffic to OKCupid. It does this with entertaining content about racy subjects, and rigorous analysis comes (optimistically) second. Of course, datadataeverywhere added, that's exactly the Mythbusters formula. They're both junk food science, but it's also the only look at their data we're going to get, so I'll link a few relevant OKTrends posts in the appropriate sections.
How to Write a Good Profile
Okay, you've created your account and answered a few questions. Now it's time to summarize your whole personality, your appeal, and your worldview in ten little text boxes. Where to begin?
The obvious answer is to reply to the ten profile |
50566cd5-db1b-4ce9-929a-a08232165ba5 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Path Independent Equilibrium Models Can Better Exploit Test-Time Computation.
1 Introduction
---------------
One of the main challenges limiting the practical applicability of modern deep learning systems is the ability to generalize outside the training distribution (Koh et al., [2021](#bib.bib24)). One particularly important type of out-of-distribution (OOD) generalization is *upwards generalization*, or the ability to generalize to more difficult problem instances than those encountered at training time (Selsam et al., [2018](#bib.bib45); Bansal et al., [2022](#bib.bib7); Schwarzschild et al., [2021b](#bib.bib44); Nye et al., [2021](#bib.bib38)). Often, good performance on more difficult instances will require a larger amount of test-time computation, so a natural question arises: how can we design neural net architectures which can reliably exploit additional test-time computation to achieve better accuracy?
Equilibrium models, a broad class of architectures whose outputs are the fixed points of learned dynamical systems, are particularly suited to meet this challenge. Closely related to weight-tied recurrent models – networks that apply the same fixed neural network module repeatedly to hidden-layer activations – equilibrium models are capable of adapting their compute budget based on the input they are given. Under what conditions, if any, can this input-dependent ability to scale-up test-time compute actually lead to upwards generalization?
We argue that a key determiner of whether a learned model can exploit additional test-time computation is whether the dynamical system corresponding to the model is *path independent*; that is, whether the learned model’s hidden layer activations converge to the same asymptotic behaviour (i.e. fixed point or limit cycle), regardless of the initialization of the system. For example, a simple integrator xt+1=xt+1subscript𝑥𝑡1subscript𝑥𝑡1x\_{t+1}=x\_{t}+1italic\_x start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT = italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT + 1 is clearly *not* path independent, as its final state depends on the initial state x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT and the number of iterations run; conversely, the system xt+1=(xt+1)/2subscript𝑥𝑡1subscript𝑥𝑡12x\_{t+1}=(x\_{t}+1)/2italic\_x start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT = ( italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT + 1 ) / 2 *is* path independent, as it will converge to the solution xT=1subscript𝑥𝑇1x\_{T}=1italic\_x start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT = 1 as T→∞→𝑇T\rightarrow\inftyitalic\_T → ∞ regardless of the initial condition of x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT. Path independence is closely related to the concept of global stability from control theory (see Section [9](#S9 "9 Related Work ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") for more).
Intuitively, path independent systems can more easily take advantage of additional test-time iterations than path dependent ones. For instance, gradient descent applied to a convex objective is path independent, and correspondingly when confronted with a more ill-conditioned problem instance, one can compensate by increasing the number of iterations. Conversely, a weather simulation is path dependent, and extending the simulation won’t yield more accurate predictions of a given day’s weather. Based on this intuition, we hypothesize that path independence of a learned model is a key determiner of whether it can take advantage of an increased test-time iteration budget when generalizing to harder problem instances:
Figure 1: Trajectories of path independent models converge to the same hidden state for a given input, regardless of initialization, whereas the trajectories of path dependent models depend on initialization. Here, we display five trajectories with different initializations obtained from a path independent (left) and path dependent model (right) on the prefix-sum task, projected onto two random directions.
>
> *Path Independence Hypothesis:* Models which successfully fit the training distribution with a path independent function are better able to exploit more test-time iterations to achieve higher accuracy, compared to those which fit the training distribution with a path-dependent function.
>
>
>
We first introduce a metric for measuring the path independence of a learned model, the *Asymptotic Alignment (AA)* score. On a wide range of tasks including sequence prediction, visual reasoning, image classification, continuous optimization and graph regression, we show that a model’s AA score is strongly correlated with its performance when far more iterations are used at test time than at training time. In general, we find that path independent models increase their performance monotonically with the number of test-time iterations, whereas path dependent models degrade when the number of test time iterations exceeds those at training time. We find that input injection and implicit updates improve both the path independence and the accuracy. Furthermore, we perform an experimental manipulation by introducing regularizers which directly promote or punish path independence. We find that these manipulations, while increasing or decreasing the AA score, also have the corresponding effect on accuracy.
2 Background
-------------
##### Equilibrium Models
Equilibrium models treat computing internal representations as a fixed-point finding problem. (McClelland and Rumelhart, [1989](#bib.bib35); Liao et al., [2018](#bib.bib30); Bai et al., [2019](#bib.bib3)) Concretely, letting 𝒙∈ℝnx𝒙superscriptℝsubscript𝑛𝑥{\bm{x}}\in\mathbb{R}^{n\_{x}}bold\_italic\_x ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT and f𝒘:ℝnx×nz↦ℝnz:subscript𝑓𝒘maps-tosuperscriptℝsubscript𝑛𝑥subscript𝑛𝑧superscriptℝsubscript𝑛𝑧f\_{{\bm{w}}}:\mathbb{R}^{n\_{x}\times n\_{z}}\mapsto\mathbb{R}^{n\_{z}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT : blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT × italic\_n start\_POSTSUBSCRIPT italic\_z end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT ↦ blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_z end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT stand for an input and the equilibrium model function (or “cell") parametrized by 𝒘∈ℝnw𝒘superscriptℝsubscript𝑛𝑤{\bm{w}}\in\mathbb{R}^{n\_{w}}bold\_italic\_w ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_w end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT respectively, equilibrium models aim to solve for the fixed point 𝒛\*∈ℝnzsuperscript𝒛superscriptℝsubscript𝑛𝑧{\bm{z}}^{\*}\in\mathbb{R}^{n\_{z}}bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_z end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT that satisfies 𝒛\*=f𝒘(𝒙,𝒛\*)superscript𝒛subscript𝑓𝒘𝒙superscript𝒛{\bm{z}}^{\*}=f\_{{\bm{w}}}({\bm{x}},{\bm{z}}^{\*})bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT = italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ). The cell f𝒘subscript𝑓𝒘f\_{{\bm{w}}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT might represent anything from a fully connected layer to a transformer block (Vaswani et al., [2017](#bib.bib50)). We emphasize that f𝒘subscript𝑓𝒘f\_{{\bm{w}}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT directly depends on the input 𝒙𝒙{\bm{x}}bold\_italic\_x; following existing literature, we refer to this as input injection. The outcome of the fixed point finding process might depend on the initial fixed point guess. To make this dependency explicit, we define the function FIX:ℝnx×nz↦ℝnz:FIXmaps-tosuperscriptℝsubscript𝑛𝑥subscript𝑛𝑧superscriptℝsubscript𝑛𝑧\mathrm{FIX}:\mathbb{R}^{n\_{x}\times n\_{z}}\mapsto\mathbb{R}^{n\_{z}}roman\_FIX : blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT × italic\_n start\_POSTSUBSCRIPT italic\_z end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT ↦ blackboard\_R start\_POSTSUPERSCRIPT italic\_n start\_POSTSUBSCRIPT italic\_z end\_POSTSUBSCRIPT end\_POSTSUPERSCRIPT that maps an input 𝒙𝒙{\bm{x}}bold\_italic\_x and an initial guess for the fixed point 𝒛0subscript𝒛0{\bm{z}}\_{0}bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT to an output that satisfies the fixed point equation FIXfw(𝒙,𝒛0):=f𝒘(𝒙,FIXfw(𝒙,𝒛0))assignsubscriptFIXsubscript𝑓𝑤𝒙subscript𝒛0subscript𝑓𝒘𝒙subscriptFIXsubscript𝑓𝑤𝒙subscript𝒛0\mathrm{FIX}\_{f\_{w}}({\bm{x}},{\bm{z}}\_{0}):=f\_{{\bm{w}}}({\bm{x}},\mathrm{FIX}\_{f\_{w}}({\bm{x}},{\bm{z}}\_{0}))roman\_FIX start\_POSTSUBSCRIPT italic\_f start\_POSTSUBSCRIPT italic\_w end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) := italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , roman\_FIX start\_POSTSUBSCRIPT italic\_f start\_POSTSUBSCRIPT italic\_w end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) ). The behaviour of FIXFIX\mathrm{FIX}roman\_FIX depends on the solver that’s used to find fixed points.
The most straightforward approach to solve for fixed-points is the fixed point iteration method, which recursively applies the function f𝒘subscript𝑓𝒘f\_{{\bm{w}}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT on the internal representations 𝒛𝒛{\bm{z}}bold\_italic\_z (i.e. 𝒛t+1=f𝒘(𝒙,𝒛t)subscript𝒛𝑡1subscript𝑓𝒘𝒙subscript𝒛𝑡{\bm{z}}\_{t+1}=f\_{{\bm{w}}}({\bm{x}},{\bm{z}}\_{t})bold\_italic\_z start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT = italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT )). If certain conditions are satisfied (such as the fixed iterations being globally contractive – more general conditions are discussed below), this procedure converges111Divergence is also a possible outcome, rendering the output of equilibrium models unusable. to a fixed point: 𝒛\*=f𝒘(𝒙,𝒛\*)superscript𝒛subscript𝑓𝒘𝒙superscript𝒛{\bm{z}}^{\*}=f\_{{\bm{w}}}({\bm{x}},{\bm{z}}^{\*})bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT = italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ). As solving for fixed points exactly is expensive, fixed point iterations are often terminated after a fixed number of steps or when the norm of the difference between subsequent iterates falls below a pre-selected threshold. The model weights can be updated using gradients computed via backpropagating through the full forward computational graph.
If one commits to using fixed point iterations as the root solver, then the output 𝒛\*superscript𝒛{\bm{z}}^{\*}bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT of equilibrium models can be interpreted as the infinite-depth limit of an input-injected, weight-tied model f𝒘∞(𝒙,𝒛\*)=limn→∞f𝒘(n)(𝒙,𝒛0)subscriptsuperscript𝑓𝒘𝒙superscript𝒛subscript→𝑛subscriptsuperscript𝑓𝑛𝒘𝒙subscript𝒛0f^{\infty}\_{{\bm{w}}}({\bm{x}},{\bm{z}}^{\*})=\lim\_{n\to\infty}f^{(n)}\_{{\bm{w}}}({\bm{x}},{\bm{z}}\_{0})italic\_f start\_POSTSUPERSCRIPT ∞ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) = roman\_lim start\_POSTSUBSCRIPT italic\_n → ∞ end\_POSTSUBSCRIPT italic\_f start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) where the notation f(n)superscript𝑓𝑛f^{(n)}italic\_f start\_POSTSUPERSCRIPT ( italic\_n ) end\_POSTSUPERSCRIPT stands for n𝑛nitalic\_n repeated applications of f𝑓fitalic\_f on its own output, and 𝒛0subscript𝒛0{\bm{z}}\_{0}bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT stands for the fixed point initialization.
Implicit Training of Equilibrium Models Different training algorithms for equilibrium models can be derived by considering their implicit nature. Bai et al. ([2019](#bib.bib3)) solve for fixed points explicitly using black-box root finders, such as Broyden’s method (Broyden, [1965](#bib.bib9)) or Anderson acceleration (Anderson, [1965](#bib.bib1)). In order to avoid explicitly differentiating through the root-finding procedure, they utilize implicit differentiation to compute gradients.222Bai et al. ([2019](#bib.bib3)) use the term Deep Equilibrium Models (DEQ) to refer to implicitly trained equilibrium models. To keep things more general, we categorize “explicitly trained” networks (i.e. with fixed point iterations and backpropagation) under the umbrella of equilibrium models as well. Concretely, letting g𝒘(𝒙,𝒛)=f𝒘(𝒙,𝒛)−𝒛subscript𝑔𝒘𝒙𝒛subscript𝑓𝒘𝒙𝒛𝒛g\_{{\bm{w}}}({\bm{x}},{\bm{z}})=f\_{{\bm{w}}}({\bm{x}},{\bm{z}})-{\bm{z}}italic\_g start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z ) = italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z ) - bold\_italic\_z for a fixed point 𝒛\*superscript𝒛{\bm{z}}^{\*}bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT, the Jacobian of 𝒛\*superscript𝒛{\bm{z}}^{\*}bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT with respect to the equilibrium model weights can be given by:
| | | | |
| --- | --- | --- | --- |
| | d𝒛\*d𝒘=−(∂g𝒘(𝒙,𝒛\*)∂𝒛\*)−1∂f𝒘(𝒙,𝒛\*)∂𝒘𝑑superscript𝒛𝑑𝒘superscriptsubscript𝑔𝒘𝒙superscript𝒛superscript𝒛1subscript𝑓𝒘𝒙superscript𝒛𝒘\frac{d{\bm{z}}^{\*}}{d{\bm{w}}}=-\left(\frac{\partial g\_{{\bm{w}}}({\bm{x}},{\bm{z}}^{\*})}{\partial{\bm{z}}^{\*}}\right)^{-1}\frac{\partial f\_{{\bm{w}}}({\bm{x}},{\bm{z}}^{\*})}{\partial{\bm{w}}}divide start\_ARG italic\_d bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_ARG start\_ARG italic\_d bold\_italic\_w end\_ARG = - ( divide start\_ARG ∂ italic\_g start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) end\_ARG start\_ARG ∂ bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT divide start\_ARG ∂ italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) end\_ARG start\_ARG ∂ bold\_italic\_w end\_ARG | | (1) |
Inverting a Jacobian matrix can become computationally expensive. Recent works (Geng et al., [2021a](#bib.bib17); Fung et al., [2021](#bib.bib16)) have shown that the inverse-Jacobian term in [Eq. 1](#S2.E1 "1 ‣ Equilibrium Models ‣ 2 Background ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") can be replaced with an identity matrix i.e. Jacobian-free or an approximate inverse-Jacobian (Geng et al., [2021b](#bib.bib18)) without affecting the final performance. This approximation makes the backward pass inexpensive and lightweight. Equilibrium models have been shown to achieve state-of-the-art performance on various tasks including language modelling (Bai et al., [2019](#bib.bib3)), image recognition, semantic segmentation (Bai et al., [2020](#bib.bib4)), object detection (Wang et al., [2020](#bib.bib52)), and graph modeling (Gu et al., [2020](#bib.bib21); Liu et al., [2021a](#bib.bib32)).
Equilibrium Models vs. Depthwise Recurrent Models Both equilibrium models and input-injected depthwise recurrent (i.e. weight-tied, fixed-depth) networks leverage weight-tying i.e., they apply the same transformation at each layer, fω[i]=fω∀isuperscriptsubscript𝑓𝜔delimited-[]𝑖subscript𝑓𝜔for-all𝑖f\_{\omega}^{[i]}=f\_{\omega}\;\forall iitalic\_f start\_POSTSUBSCRIPT italic\_ω end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT [ italic\_i ] end\_POSTSUPERSCRIPT = italic\_f start\_POSTSUBSCRIPT italic\_ω end\_POSTSUBSCRIPT ∀ italic\_i. The two models differ in the ultimate aim of the forward pass: while depthwise recurrent models compute a (weight-tied) fixed depth computation (which may or may not approach a fixed point), the stated *goal* of equilibrium models is explicitly to find a fixed point. Weight-tied fixed depth networks by definition require backpropagation through an explicit stack of layers. Equilibrium models, however, directly solve for fixed points using (potentially black-box) solvers during the forward pass and may be trained using implicit differentiation.
Convergence
As alluded above, in order to guarantee convergence to a unique fixed point, it suffices for the cell of the equilibrium model to be contractive over its input domain333This is known as the Banach fixed-point theorem. (i.e. the singular values of its Jacobian all lie below 1111). Previous work has leveraged Lipschitz constrained cells to ensure contractivity (Revay et al., [2020](#bib.bib39)). Other approaches for ensuring global convergence exist: the monotone equilibrium model architecture guarantees global convergence by utilizing an equilibrium model parametrization that bears similarities to solutions to a particular form of monotone operator splitting problem (Winston and Kolter, [2020](#bib.bib54)). Unrestricted equilibrium models aren’t constrained enough to guarantee convergence: they can easily express globally divergent vector fields that prohibit the existence of fixed points. It is, therefore, interesting that they can (and often) learn path independent solutions. Also note that the “infinite-depth weight tied network" interpretation of equilibrium models is less general than the implicit formulation presented above, as the latter admits unstable fixed points as well.
Terminology and Abbreviations We use the term “equilibrium models” to refer to the general class of networks that explicitly solve for a fixed point in the forward pass. We use the term ‘solver’ to refer to the use of black-box root finders like Anderson acceleration to find fixed points of an implicitly trained equilibrium model. These networks can use implicit gradients computed via implicit function theorem (IFT), Jacobian-free backward pass, or with an approximation of inverse-Jacobian. The term ‘unroll’ refers to equilibrium models that use regular fixed-point iterations to compute the equilibrium point. We use the abbreviation ‘bp’ to refer to backpropagation gradients, and ‘inj’ to refer to input injection. The term ‘progressive net’ refers to the deep thinking networks trained with progressive training as proposed by Bansal et al. ([2022](#bib.bib7)). We use ‘PI’ and ‘non-PI’ to refer to path independent and path dependent networks, respectively.
3 Upwards Generalization with Equilibrium Models
-------------------------------------------------
In this section, we establish that equilibrium models are capable of strong upwards generalization. To study the effects of test time computation, it is useful to consider tasks with an explicit difficulty parameter, so that the learned models can be tested on more difficult instances which require a large number of iterations to solve correctly. We focus on multiple algorithmic generalization tasks: prefix sum and mazes by Schwarzschild et al. ([2021a](#bib.bib43), [b](#bib.bib44)), blurry MNIST, matrix inversion and edge copy by Du et al. ([2022](#bib.bib11)). Taken together, these tasks cover a wide range of problems from different domains, namely sequence prediction, visual reasoning, image classification, continuous optimization and graph regression. To maintain clarity and focus, we run our detailed analysis on the prefix sum and mazes tasks, and provide complementary results for the remaining tasks in the Supplementary Material (SM).
Tasks Prefix-sum is a sequence-to-sequence task whereby the network is given a sequence of 0-1 bits, and is trained to output, for each bit, the parity of all of the bits received since the beginning of the sequence until the current bit. We train on 10,000 unique 32-bit binary strings, and report results on binary strings of other lengths. The mazes task is also an image-to-image task, where the input is a three-channel RGB image. The ‘start’ and ‘finish’ positions are marked by a red and a green square respectively; walls are marked in black. The output is the optimal path in the maze that connects these two points without passing through the walls. We train on 50,000 small mazes of size 9×9999\times 99 × 9, and report upward generalization results on larger mazes. Instances of each of these problems, as well as additional image classification and continuous optimization results can be found in the supplementary material. Blurry MNIST ([Liang et al.,](#bib.bib29) ) is a robustness-to-corruption task: one has to learn to do MNIST classification from lightly blurred images and generalize zero-shot to highly blurred ones. In the matrix inversion task (Du et al., [2022](#bib.bib11)), the goal is to learn to invert 10×10101010\times 1010 × 10 matrices in a way that generalizes to matrices that have worse condition number than those observed during training. Edge copy (Du et al., [2022](#bib.bib11)) is a simple graph regression tasks that requires learning to output the input edge features, in a way that generalizes to larger graph sizes. Note that the training and test data in the latter two tasks are generated with noise added on-the-fly, as done by Du et al. ([2022](#bib.bib11)).
Strong Upward Generalization
[Fig. 2](#S3.F2 "Figure 2 ‣ 3 Upwards Generalization with Equilibrium Models ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") shows that equilibrium models demonstrate very strong upward generalization performance compared to non weight-tied fixed-depth models. Moreover, [Fig. 2](#S3.F2 "Figure 2 ‣ 3 Upwards Generalization with Equilibrium Models ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") shows that increasing inference depth consistently improves performance—especially on harder problem instances.
(a)
(a)
(b)
Figure 2: (left) Strong upward generalization on mazes by PI models. Models were trained on 9×9999\times 99 × 9 sized mazes and tested for upward generalization on larger mazes. y-axis uses probit transformation. (right) PI models are better able to make use of additional test-time computation. We trained models with varying number of training-time iterations, learning rate and weight norm application. Bit-wise accuracies are evaluated and averaged over different string-lengths.
4 Path Independence
--------------------
Having intuitively motivated the idea of path independence in [Sec. 1](#S1 "1 Introduction ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), we now define it formally: we say that the computation performed by a recurrent operator computing function f𝒘subscript𝑓𝒘f\_{{\bm{w}}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT on an input 𝒙𝒙{\bm{x}}bold\_italic\_x is path independent if it converges to the same limiting behavior regardless of the current state 𝒛tsubscript𝒛𝑡{\bm{z}}\_{t}bold\_italic\_z start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT. As a special case, if the computation is convergent, this property is equivalent to the existence of a unique fixed point 𝒛\*superscript𝒛{\bm{z}}^{\*}bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT such that f𝒘∞(𝒙,𝒛0)=𝒛\*subscriptsuperscript𝑓𝒘𝒙subscript𝒛0superscript𝒛f^{\infty}\_{{\bm{w}}}({\bm{x}},{\bm{z}}\_{0})={\bm{z}}^{\*}italic\_f start\_POSTSUPERSCRIPT ∞ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) = bold\_italic\_z start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT for any 𝒛0subscript𝒛0{\bm{z}}\_{0}bold\_italic\_z start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT. However, our definition allows for other behaviors such as limit cycles (see [Sec. 7](#S7 "7 Disambiguating Convergence and Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation")).
Some architectures guarantee the path independence property (see [Sec. 2](#S2 "2 Background ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation")). However, most common DEQ architectures—and the ones we use throughout this paper—have the expressive power to learn multiple fixed points per input. Since it is unclear whether architectures enforcing the contraction property lose expressiveness (Bai et al., [2021](#bib.bib5)), we focus our investigation on unrestricted architectures.
PI networks represent a different model for computation than standard feed-forward networks: instead of learning an entire computational graph to map inputs to outputs, they only have to learn where to stop. We dedicate the rest of the paper on exploring the Path Independence Hypothesis—the idea that models which successfully fit the training distribution with a path independent function are better able to exploit more test-time iterations to achieve higher accuracy, compared to those which fit the training distribution with a path-dependent function.
Before establishing a connection between path independence and out-of-distribution generalization, we first describe two architectural components that are necessary conditions for achieving path independence. Afterward, we describe a metric to quantify how path-independent a trained network is.
###
4.1 Architectural Components Necessary for Path Independence
Past work has observed that *weight tying* and *input injection* are both crucial for upwards generalization (Bansal et al., [2022](#bib.bib7)). We observe that both architectural components are also necessary for a learned model to be PI.444Our definition also admits non-input-injected models to be path independent if they’re representing constant functions (i.e. input independent). We don’t consider such cases in our analyses. Without weight tying, the network is constrained to have a fixed forward depth, so it is meaningless to talk about the limiting behavior in large depth. Input injection ensures that the equilibrium point depends on the input despite having an “infinite depth". Without input injection, a PI network would necessarily forget the input; hence, any model which successfully fits the training distribution must be path dependent.
Interestingly, both architectural motifs are also key components of deep equilibrium models (Bai et al., [2019](#bib.bib3)); in that work, the motivation was to enable efficient gradient estimation via the implicit function theorem (IFT) — a concept closely related to path independence, since the premise of the IFT gradient estimator is that only the final hidden state matters, not the path taken to get there. It is striking that two separate lines of work would converge on the same architectural motifs, one motivated by generalization and the other by a variant of path independence.
Reproducing the results of Bansal et al. ([2022](#bib.bib7)), in [Fig. 2](#S3.F2 "Figure 2 ‣ 3 Upwards Generalization with Equilibrium Models ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") we show upward generalization performance using both equilibrium models and progressive nets (Bansal et al., [2022](#bib.bib7)) – and the lack thereof using non-input-injected networks. For the remainder of this paper, we focus on architectures with both input injection and weight tying.
###
4.2 Quantifying Path Independence
We propose a simple metric to quantify path independence based on the directional alignment of the fixed points computed with the same input, but different initializations. We name this metric the *Asymptotic Alignment (AA) score*. Pseudocode to compute the metric is given in [Alg. 1](#alg1 "Algorithm 1 ‣ Figure 3 ‣ 4.2 Quantifying Path Independence ‣ 4 Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"). The AA score is the average cosine similarity between the fixed points obtained with the training time initialization (often simply the zero vector) and the fixed points obtained when one initializes the solver using the fixed points computed on different inputs. Higher AA scores (with 1111 being the highest value) imply higher degrees of path independence. In [Sec. 5](#S5 "5 Path Independence Correlates with Upward Generalization ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), we show a strong correlation between path independence and accuracy using the AA score.
The AA score is cheap to compute, is a reliable indicator of path independence (see below), and is unitless, meaning that networks obtained from different training runs can be compared on equal footing. See the supplementary material for other metrics we’ve considered for quantifying path independence and why we found AA score to be preferable.
Algorithm 1 Asymptotic Alignment Score
A batched input [𝒙1𝒙2]matrixsubscript𝒙1subscript𝒙2\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{x}}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{x}}\_{2}}\end{bmatrix}[ start\_ARG start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ], an operator f𝒘subscript𝑓𝒘f\_{\bm{w}}italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT
Initialize: [𝒛1𝒛2]=𝟎matrixsubscript𝒛1subscript𝒛20\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}\_{2}}\end{bmatrix}=\mathbf{0}[ start\_ARG start\_ROW start\_CELL bold\_italic\_z start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_z start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] = bold\_0
Define: h(𝒚1,𝒚2)=𝒚1‖𝒚1‖2⋅𝒚2‖𝒚2‖2ℎsubscript𝒚1subscript𝒚2⋅subscript𝒚1subscriptnormsubscript𝒚12subscript𝒚2subscriptnormsubscript𝒚22h({\bm{y}}\_{1},{\bm{y}}\_{2})=\dfrac{{\bm{y}}\_{1}}{\|{\bm{y}}\_{1}\|\_{2}}\cdot\dfrac{{\bm{y}}\_{2}}{\|{\bm{y}}\_{2}\|\_{2}}italic\_h ( bold\_italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , bold\_italic\_y start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ) = divide start\_ARG bold\_italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_ARG start\_ARG ∥ bold\_italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_ARG ⋅ divide start\_ARG bold\_italic\_y start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_ARG start\_ARG ∥ bold\_italic\_y start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_ARG
Compute [𝒛1′𝒛2′]=FIXf𝒘([𝒙1𝒙2],[𝒛1𝒛2])matrixsubscriptsuperscript𝒛′1subscriptsuperscript𝒛′2subscriptFIXsubscript𝑓𝒘matrixsubscript𝒙1subscript𝒙2matrixsubscript𝒛1subscript𝒛2\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}^{\prime}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}^{\prime}\_{2}}\end{bmatrix}=\mathrm{FIX}\_{f\_{\bm{w}}}\left(\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{x}}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{x}}\_{2}}\end{bmatrix},\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}\_{2}}\end{bmatrix}\right)[ start\_ARG start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] = roman\_FIX start\_POSTSUBSCRIPT italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( [ start\_ARG start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] , [ start\_ARG start\_ROW start\_CELL bold\_italic\_z start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_z start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] )
# Interchange and reinitialize iterates
Compute [𝒛1′′𝒛2′′]=FIXf𝒘([𝒙1𝒙2],[𝒛2′𝒛1′])matrixsubscriptsuperscript𝒛′′1subscriptsuperscript𝒛′′2subscriptFIXsubscript𝑓𝒘matrixsubscript𝒙1subscript𝒙2matrixsubscriptsuperscript𝒛′2subscriptsuperscript𝒛′1\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}^{\prime\prime}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}^{\prime\prime}\_{2}}\end{bmatrix}=\mathrm{FIX}\_{f\_{\bm{w}}}\left(\begin{bmatrix}{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{x}}\_{1}}\\
{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{x}}\_{2}}\end{bmatrix},\begin{bmatrix}{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}^{\prime}\_{2}}\\
{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}^{\prime}\_{1}}\end{bmatrix}\right)[ start\_ARG start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] = roman\_FIX start\_POSTSUBSCRIPT italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( [ start\_ARG start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_x start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] , [ start\_ARG start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_ARG ] )
return average(h(𝒛1′′,𝒛1′)\mathrm{average}(h({\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}^{\prime\prime}\_{1}},{\color[rgb]{0,.5,.5}\definecolor[named]{pgfstrokecolor}{rgb}{0,.5,.5}\pgfsys@color@rgb@stroke{0}{.5}{.5}\pgfsys@color@rgb@fill{0}{.5}{.5}{\bm{z}}^{\prime}\_{1}})roman\_average ( italic\_h ( bold\_italic\_z start\_POSTSUPERSCRIPT ′ ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ), h(𝒛2′′,𝒛2′))h({\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}^{\prime\prime}\_{2}},{\color[rgb]{1,0,1}\definecolor[named]{pgfstrokecolor}{rgb}{1,0,1}\pgfsys@color@cmyk@stroke{0}{1}{0}{0}\pgfsys@color@cmyk@fill{0}{1}{0}{0}{\bm{z}}^{\prime}\_{2}}))italic\_h ( bold\_italic\_z start\_POSTSUPERSCRIPT ′ ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , bold\_italic\_z start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ) )
Figure 3: (left) AA Score Algorithm: We provide the algorithm for a simple illustrative case of two inputs. In practice, we consider larger batches. (right) Promoting path independence improves generalization in the prefix sum task: Interventions that are designed to promote path independence (initializing fixed points with random noise or running the fixed point solver with stochastic budget) improves generalization. Conversely, those that hurt path independence (penalty term that directly penalizes fixed point alignment) leads to poorer generalization.
##### Stress-testing the AA score
To stress-test the extent to which the AA score really measures path independence, we search for adversarial initializations that are optimized to result in distinct fixed points, hence low AA values. (Unlike adversarial examples, this attack is not constrained to an ε𝜀\varepsilonitalic\_ε-ball.) We use the L-BFGS (Liu and Nocedal, [1989](#bib.bib31)) optimizer, and repeat the search multiple times starting from different fixed point initializations. We include pseudocode in the supplementary material.
Results of the adversarial stress test can be seen in [Tab. 1](#S4.T1 "Table 1 ‣ Stress-testing the AA score ‣ 4.2 Quantifying Path Independence ‣ 4 Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"). The results corroborate that the AA score is indeed a reliable measure of path independence; while it isn’t possible to find adversarial initializations for high AA score networks (indicating high path independence), low AA score networks can easily be adversarially initialized to be steered away from the original fixed point estimate.
| Model | Task | AA ↑↑\uparrow↑ | Accuracy (%) | Attacked AA ↑↑\uparrow↑ | Attacked Acc. (%) |
| --- | --- | --- | --- | --- | --- |
| Non-PI network | Maze | 0.320.320.320.32 | 87.1287.1287.1287.12 | 0.090.090.090.09 | 00 |
| PI network | Maze | 1.001.001.001.00 | 100100100100 | 1.001.001.001.00 | 100100100100 |
| Non-PI network | Prefix sum | 0.620.620.620.62 | 66.6666.6666.6666.66 | 0.180.180.180.18 | 00 |
| PI network | Prefix sum | 0.990.990.990.99 | 100100100100 | 0.990.990.990.99 | 100100100100 |
Table 1: Stress-testing the AA Scores: AA scores for PI vs non-PI networks computed on 13×13131313\times 1313 × 13 mazes and 64 bit prefix sum. Attacked AA refers to the cosine similarity between the fixed point from zero initialization and an adversarial initialization. Non-PI networks can be easily steered away from the initial fixed point estimate through adversarial initializations but it is difficult to do so for PI networks with high AA scores.
5 Path Independence Correlates with Upward Generalization
----------------------------------------------------------
Is path independence (as measured by the AA score) a strong predictor of upwards generalization? We took the trained networks from [Sec. 3](#S3 "3 Upwards Generalization with Equilibrium Models ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), computed their average AA scores on in- and out-of-distribution splits and inspected whether the AA scores are correlated with upward generalization.
On prefix sum experiments, we varied 1) network depth, 2) whether or not weight norm (wnorm) (Salimans and Kingma, [2016](#bib.bib40)) was used or not,555Bai et al. ([2019](#bib.bib3)) report that weight norm helps stabilize the training of DEQ models. 3) learning rate (one of [0.01, 0.001, 0.0001]), 4) forward solver (fixed point iterations or Anderson acceleration (Anderson, [1965](#bib.bib1)), and 5) the gradient estimator (backprop or implicit gradients).666Note that the deep equilibrium model (DEQ) setup (Bai et al., [2019](#bib.bib3)) correspond to using a root solver (such as Anderson) for the forward pass and implicit gradients for the backward pass. On the maze experiments, we varied 1) network depth, 2) use of weight norm, 3) forward solver (fixed point iterations or Broyden solver (Broyden, [1965](#bib.bib9))), and 5) the gradient estimator (backprop or implicit gradients).
[Fig. 4](#S5.F4 "Figure 4 ‣ 5 Path Independence Correlates with Upward Generalization ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") displays our findings. We evaluated performance on a mixture of in- and OOD validation data; results on individual data splits can be found in the supplementary material. The results show a strong correlation between AA score and accuracy when the inference depth is large enough. This shows that PI networks allow for scaling test-time compute to improve test-time accuracy (see also [Fig. 2](#S3.F2 "Figure 2 ‣ 3 Upwards Generalization with Equilibrium Models ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation")). The in-distribution validation performance of non-PI networks degrades with deeper inference depths. Unsurprisingly, these networks generalize poorly on harder problem instances that require deeper inference depths (i.e. problem instances that provably require at least a given number of layers to handle). Further results on the BlurryMNIST, matrix inversion and edge copy tasks can be found in Supplementary Material [H](#A8 "Appendix H Results on the Blurry MNIST Task ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), [I](#A9 "Appendix I Results on Matrix Inversion Task ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") and [J](#A10 "Appendix J Results on the Edge Copy Task ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation").
(a) Prefix Sum
(b) Mazes
Figure 4: High AA scores correlate with good upward generalization. For a given choice of an architecture and a task, the reported numbers are averaged over problem instances of different dimensions. We apply the probit transformation along both axes, following Miller et al. ([2021](#bib.bib36)). Accuracies and AA scores are capped at 0.999 for compatibility with the probit transform.
6 Experimental Manipulations of Path Independence
--------------------------------------------------
The previous section demonstrates a strong correlation between path independence and the ability to exploit additional test-time iterations. Unfortunately, we can’t make a causal claim based on these studies: the observed effect could have been due to an unobserved confounder.
In this section, we intervene directly on path independence by imposing regularizers which directly encourage or penalize path independence.
We find that interventions designed to promote path independence also improve generalization, while interventions designed to reduce path independence also hurt generalization.
###
6.1 Promoting Path Independence via Randomized Forward Passes
A straightforward way to encourage path independence is simply to initialize the hidden states with random noise during training. To this end, we experimented with initializing the hidden states with zeros on half of the examples in the batch, and with standard Gaussian noise on the rest of the examples. The reason to include the zero-initializations at training time is that we initialize from zeros at test time - not including this initialization during traning time causes a distribution shift.
Another way to promote path independence is simply running the forward solver with randomized compute budgets/depths during training time. While a path independent solution can be expected to be robust against this intervention, a path dependent one will fail.
We took the training configurations of the 12 prefix-sum networks described in [Sec. 5](#S5 "5 Path Independence Correlates with Upward Generalization ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") that use fixed point iterations in their forward pass, and backpropagation gradient in their backward pass, and retrained them separately with the aforementioned mixed initialization and randomized depth strategies without modifying any other experimental conditions. As can be seen in [Fig. 3](#S4.F3 "Figure 3 ‣ 4.2 Quantifying Path Independence ‣ 4 Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), the interventions lead to strong test-time path independent neural networks, while also reliably improving in- and out-of-distribution validation accuracy. We especially emphasize that shallow networks trained with mixed initialization actually remain far from having high AA scores using the training-time forward pass conditions due to lack of convergence. However, since the mixed initialization strategy results in path independent networks, scaling up test-time compute budget leads to high AA scores, and therefore high upwards generalization.
###
6.2 Penalizing Path Independence via the Fixed Point Alignment Penalty
Does an intervention that results in less path independence also result in poorer upwards generalization? Like in the mixed initialization experiment, we retrained the 12 unroll + backpropagation networks with an additional auxiliary loss term that penalizes the dot product between the fixed points computed from the same input, but different initializations sampled from standard Gaussian noise. [Fig. 3](#S4.F3 "Figure 3 ‣ 4.2 Quantifying Path Independence ‣ 4 Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") shows that this intervention succeeded in pushing the AA scores down, while also keeping the accuracy on the same trend line.
7 Disambiguating Convergence and Path Independence
---------------------------------------------------
Is convergence necessary for path independence?
We answer this statement in the negative, and show that neither training-time convergence nor test-time convergence is required for path independence.
Instead convergence to the same limiting behavior regardless of initialization is important.
##### Training Time Convergence
We consider two implicitly trained equilibrium models trained on the mazes task—one trained with implicit gradients computed via implicit function theorem (IFT), and the other trained with an approximation of the (inverse) Jacobian, called phantom gradients (Geng et al., [2021b](#bib.bib18)). We report the values of residuals (i.e., ‖f𝐰(𝐱,𝐳)−𝐳‖2subscriptnormsubscript𝑓𝐰𝐱𝐳𝐳2\|f\_{\bm{w}}({\bm{x}},{\bm{z}})-{\bm{z}}\|\_{2}∥ italic\_f start\_POSTSUBSCRIPT bold\_italic\_w end\_POSTSUBSCRIPT ( bold\_italic\_x , bold\_italic\_z ) - bold\_italic\_z ∥ start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT), AA scores and accuracies observed for in- and out-of-distribution data for mazes in [Tab. 2](#S7.T2 "Table 2 ‣ Training Time Convergence ‣ 7 Disambiguating Convergence and Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"). We observe that DEQs trained with phantom gradients have higher values of in-distribution residuals but are path independent, as indicated by their high AA scores, and show strong upward generalization as indicated by their good accuracy.
The mixed-initialization intervention described in [Sec. 6.1](#S6.SS1 "6.1 Promoting Path Independence via Randomized Forward Passes ‣ 6 Experimental Manipulations of Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation") also leads to a separation between training-time convergence and path independence. We found that it is possible to train very shallow (i.e., 5 layer) unrolled networks that, while being very far from converging during training and attaining poor in-distribution generalization, are able to converge and achieve perfect performance when run for many more iterations during test time. Details are provided in the supplementary material.
| Model | Residual ↓↓\downarrow↓ | AA score ↑↑\uparrow↑ | Accuracy (%) ↑↑\uparrow↑ |
| --- | --- | --- | --- |
| In-dist | OOD | In-dist | OOD | In-dist | OOD |
| DEQ (phantom grad.) | 11.8311.8311.8311.83 | 0.0160.0160.0160.016 | 0.960.960.960.96 | 0.990.990.990.99 | 99.9699.9699.9699.96 | 99.8899.8899.8899.88 |
| DEQ (IFT) | 1.41.41.41.4 | 0.0110.0110.0110.011 | 0.990.990.990.99 | 0.990.990.990.99 | 99.9999.9999.9999.99 | 100100100100 |
Table 2: Training-time convergence is not needed for path independence: models might show poor training-time convergence (as shown by high values of residuals) but still be path independent. Residual, AA score, and Accuracy for DEQ trained with IFT vs phantom gradients. In-distribution (In-dist) results were computed on 9×9999\times 99 × 9 mazes, and OOD results were computed on 25×25252525\times 2525 × 25 mazes.
Test Time Convergence From [Tab. 2](#S7.T2 "Table 2 ‣ Training Time Convergence ‣ 7 Disambiguating Convergence and Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"), one might conclude that test time convergence is important for path independence. However, we show that this connection is not necessary, and convergence to the same fixed point is not a required condition for path independence. We study test time convergence properties of an unrolled weight-tied input-injected network trained with backpropagation under different solvers. This network is highly path independent using either the Broyden solver or fixed point iterations, as indicated by its high AA scores (0.99) on both in- and out-of-distribution data. We visualize the values of test-time residuals with fixed point iterations and Broyden’s method in [Fig. 5](#S8.F5 "Figure 5 ‣ 8 Path Independence on a Per-Example Level ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"). Both these solvers converge to different limit cycles but still show good upward generalization.
8 Path Independence on a Per-Example Level
-------------------------------------------
The connection between path independence and prediction correctness also largely holds on a per-instance basis. Using the prefix-sum networks trained with the mixed-initialization strategy (the most performant group of networks in our intervention experiments), we plotted the distribution of per-instance fixed point alignment scores, colored by whether the prediction on that instance was correct or not in Figure [5](#S8.F5 "Figure 5 ‣ 8 Path Independence on a Per-Example Level ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation"). This suggests that path independence can be used as a valuable sanity-check to determine whether a prediction is correct or not without the need for any label data, both in- and out-of-distribution. We provide a more in-depth per-instance analysis in the supplementary material.
(a)
(a)
(b)
Figure 5: (Left) Different solvers display differing asymptotic behaviours but still achieve good upwards generalization. Here, the network has an adversarial AA score of 0.99, and achieves accuracy of 99.98% (fixed point iterations) and 99.97% (Broyden solver) respectively on the mazes task; (Right) Per-instance path independence is highly correlated with correctness of predictions for prefix sum task.
9 Related Work
---------------
There is a long line of research on neural networks that can adapt their computational budget based on the complexity of the task they are learning to solve—akin to the intrinsic mechanism in humans to reason and solve problems. Schmidhuber ([2012](#bib.bib41)) introduced self-delimiting neural networks which are a type of recurrent neural networks (RNNs) that adapt their compute based on the output of a special "halt" neuron. Adaptive computation time (ACT) (Graves, [2016a](#bib.bib19)) also uses the output of a sigmoidal halting unit to determine the termination condition of an RNN, but it avoids long "thinking" time by explicitly penalizing it. Subsequent works have successfully applied variants of ACT in image classification and object detection (Figurnov et al., [2017](#bib.bib15)), visual reasoning (Eyzaguirre and Soto, [2020](#bib.bib14)), Transformers (Vaswani et al., [2017](#bib.bib50)) for language modelling (Dehghani et al., [2019](#bib.bib10); Elbayad et al., [2020a](#bib.bib12); Liu et al., [2021b](#bib.bib34)), and recognizing textual entailment (Neumann et al., [2016](#bib.bib37)). PonderNet (Banino et al., [2021](#bib.bib6)) reformulates the halting policy of ACT as a probabilistic model, and adds a regularization term in the loss objective to encourage exploration. With these additions, PonderNet can extrapolate to more difficult examples on the parity task, first proposed by Graves ([2016b](#bib.bib20)): in a vector with entries of 0, -1, and 1, output 1 for odd number of ones, and 0 otherwise. In this work, we do not optimize or penalize the network for the number of computational steps. Our main goal is to understand the underlying mechanism that results in scalable generalization of equilibrium models on harder problem instances. Our current work is closely related to previous work by Schwarzschild et al. ([2021b](#bib.bib44)) and (Bansal et al., [2022](#bib.bib7)) that propose architectural choices and training mechanisms that enable weight tied networks to generalize on harder problem instances. We relate these papers’ contributions to ours in Section [4.1](#S4.SS1 "4.1 Architectural Components Necessary for Path Independence ‣ 4 Path Independence ‣ Path Independent Equilibrium Models Can Better Exploit Test-Time Computation").
Another family of models with the property of adaptive inference compute budget is early exit networks (Teerapittayanon et al., [2016](#bib.bib49); Laskaridis et al., [2021](#bib.bib27)). These networks have multiple additional “exit” prediction heads along their depth. At inference time, the result that satisfies an exit policy is selected as the prediction output. This approach of designing adaptive networks has been adapted both in natural language processing (Schwartz et al., [2020](#bib.bib42); Soldaini and Moschitti, [2020](#bib.bib47); Elbayad et al., [2020b](#bib.bib13); Zhou et al., [2020](#bib.bib57); Liu et al., [2020](#bib.bib33)) and vision (Li et al., [2017](#bib.bib28); Wang et al., [2018](#bib.bib53); Xing et al., [2020](#bib.bib56); Kouris et al., [2021](#bib.bib26)). Most of these architectures have complex sub-modules that are trained in multiple stages, and require complex exit policies. In contrast, equilibrium models have a simple architecture, and can use root solvers to efficiently solve for the fixed point at inference.
More complex transformer-based language models like GPT-3 also struggle to generalize well on simple algorithmic tasks like addition (Brown et al., [2020](#bib.bib8)).
Recent work by Nye et al. ([2021](#bib.bib38)) shows that transformers can be trained to perform well on algorithmic tasks and generalize on OOD data by emitting the intermediate steps of an algorithm to a buffer called “scratchpad”. Using a scratchpad enables the model to revisit its errors and correct them.
Path independence is closely related to the concept of global stability and global convergence in control theory and optimization. This concept is somewhat overloaded, as it sometimes requires convergence to a single point (Slotine et al., [1991](#bib.bib46)), and sometimes implies the system is convergent everywhere, even if to different points (Wang et al., [2003](#bib.bib51); Sriperumbudur and Lanckriet, [2009](#bib.bib48)). We thus choose the term *path independence* to refer specifically to the fact that the system will converge to the same limiting behavior (whatever that might be) regardless of the initial state of the system.
10 Conclusion
--------------
Being able to attain better levels of performance using a larger inference-time compute budget is a feat that eludes most standard deep learning architectures. This is especially relevant for tasks that require upwards generalization, i.e., the ability to generalize from easy problem instances to hard ones. We show that equilibrium models are capable of displaying upwards generalization by exploiting scalable test-time compute. We link this to a phenomenon we call path independence: the tendency of an equilibrium network to converge to the same limiting behavior given an input, regardless of the initial conditions. We investigate this phenomenon through careful experiments and verify that path independent networks indeed generalize well on harder problem instances by exploiting more test time compute. Moreover, interventions on training conditions that promote path independence also improve upwards generalization, while those that penalize it hurt this capability. Our findings suggest that path independent equilibrium models are a promising direction towards building general purpose learning systems whose test-time performance improves with more compute.
11 Acknowledgements
--------------------
AP is supported by a grant from the Bosch Center for Artificial Intelligence. JT acknowledges support from the Center on Long-Term Risk Fund.
CA is supported by NSERC Canada Graduate School - Doctorate scholarship. |
b374e73f-5beb-427b-84aa-8af94901235b | trentmkelly/LessWrong-43k | LessWrong | Weekly LW Meetups: Austin, Brussels, Montreal, Pittsburgh, St. Louis
This summary was posted to LW main on Nov 23rd. The following week's summary post is here.
There are upcoming irregularly scheduled Less Wrong meetups in:
* Brussels meetup: 24 November 2012 01:00PM
* (St. Louis MO) Psycology of memorization, Thiel Fellow James Koppel, and games: 24 November 2012 02:30PM
* LessWrong Montreal - Advanced Epistemology 101: 26 November 2012 06:30PM
* Pittsburgh: affecting the far future: 28 November 2012 06:00PM
* Berlin Meetup: 01 December 2012 07:30PM
* First meetup in Innsbruck : 02 December 2012 03:00PM
* Atlanta - Practical Rationality Meetup Session: 02 December 2012 05:00PM
* Meetup : Bielefeld Meetup, December 5th: 05 December 2012 07:00PM
* Sofia, Bulgaria Meetup: 09 December 2012 05:00PM
The following meetups take place in cities with regularly scheduled meetups, but involve a change in time or location, special meeting content, or simply a helpful reminder about the meetup:
* Austin, TX: 24 November 2018 01:30PM
* Melbourne, practical rationality: 07 December 2012 07:00PM
* Winter Solstice Megameetup - NYC: 15 December 2012 05:00PM
Locations with regularly scheduled meetups: Austin, Berkeley, Cambridge, MA, Cambridge UK, Madison WI, Melbourne, Mountain View, New York, Ohio, Oxford, Portland, Salt Lake City, Seattle, Toronto, Waterloo, and West Los Angeles.
If you'd like to talk with other LW-ers face to face, and there is no meetup in your area, consider starting your own meetup; it's easy (more resources here). Check one out, stretch your rationality skills, build community, and have fun!
In addition to the handy sidebar of upcoming meetups, a meetup overview will continue to be posted on the front page every Friday. These will be an attempt to collect information on all the meetups happening in the next weeks. The best way to get your meetup featured is still to use the Add New Meetup feature, but you'll now also have the benefit of having your meetup mentioned in a weekly overview. These overview posts w |
4a9b8861-9e00-4e0d-a326-a7e76e34319c | trentmkelly/LessWrong-43k | LessWrong | Classifying representations of sparse autoencoders (SAEs)
Produced as part of the SERI ML Alignment Theory Scholars Program - Autumn 2023 Cohort, under the mentorship of Dan Hendrycks
There was recently some work on sparse autoencoding of hidden LLM representation.
I checked if these sparse representations are better suited for classification. It seems like they are significantly worse. I summarize my negative results in this blogpost, code can be found on GitHub.
Introduction
Anthropic, Conjecture and other researchers have recently published some work on sparse autoencoding. The motivation is to push features towards monosemanticity to improve interpretability.
The basic concept is to project hidden layer activations to a higher dimensional space with sparse features. These sparse features are learned by training an autoencoder with sparsity constraints.
I had previously looked into how to use hidden layer activations for classification, steering and removal. I thought maybe sparse features could be better for these tasks as projecting features to a higher dimensional space can make them more easily linearly separable. Kind of like this (except sparser...):
Implementation
I use the pythia models (70m and 410m) together with the pretrained autoencoders from this work.
As the models are not super capable I use a very simple classification task. I take data from the IMDB review data set and filter for relatively short reviews.
To push the model towards classifying the review I apply a formatting prompt to each movie review:
format_prompt='Consider if following review is positive or negative:\n"{movie_review}"\nThe review is '
I encode the data and get the hidden representations for the last token (this contains the information of the whole sentence as I'm using left padding).
# pseudo code
tokenized_input = tokenizer(formatted_reviews)
output = model(**tokenized_input, output_hidden_states=True)
hidden_states = output["hidden_states"]
hidden_states = hidden_states[:, :, -1, :] # has shape (num_layers, num_sa |
034f54a4-a41f-496a-a90d-172faf46d9b2 | trentmkelly/LessWrong-43k | LessWrong | On Privilege
The forum has been very much focused on AI safety for some time now, thought I'd post something different for a change. Privilege.
Here I define Privilege as an advantage over others that is invisible to the beholder. [EDIT: thanks to JenniferRM for pointing out that "beholder" is a wrong word.] This may not be the only definition, or the central definition, or not how you see it, but that's the definition I use for the purposes of this post. I also do not mean it in the culture-war sense as a way to undercut others as in "check your privilege". My point is that we all have some privileges [we are not aware of], and also that nearly each one has a flip side.
In some way this is the inverse of The Lens That Does Not See Its Flaws: The lens that does not see its strengths.
On to the examples, as non-polarizing as I could think of, so no focus on race or gender:
* Intelligence privilege. This one ought to be familiar to the regulars, but is often a blind spot. One manifestation of it being a privilege (i.e. being unaware that you have something others do not) is saying something like "anyone can learn calculus". No, not anyone.
* Health privilege. Most healthy people don't give a second thought to how lucky they are compared to those who struggle physically or mentally to just get through the day.
* Conventional beauty privilege. Everything is just invisibly easier when you look good. The flip side, of course, is that you have to deal with harassment a lot more. Or the judgment "you got where you are because of your looks".
* A reasonably happy childhood privilege. Those who grew up in "normal" families rarely appreciate what it is like to be a child with a high ACE score. The flip side is the determination to get ahead that some high-adversity survivors develop.
* Having an instinctive social "game". Not just for dating, but in general social interactions where a person is naturally likable and relatable, at least on the surface level. If you have that, you |
210fe4f1-6d54-4780-9195-abd478f0bce1 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Durham HPMoR Discussion, chapters 24-26
Discussion article for the meetup : Durham HPMoR Discussion, chapters 24-26
WHEN: 29 December 2012 11:00:00AM (-0500)
WHERE: 2706 Durham-Chapel Hill Boulevard, Durham, NC 27707
HPMoR discussion meetup. We'll be talking about chapters 24-26; come join us!
Please feel free to join in, even if you haven't read the chapters in question yet.
Depending on interest, there may well be Zendo afterwards.
Discussion article for the meetup : Durham HPMoR Discussion, chapters 24-26 |
6270d88a-d06a-4f44-8cc4-a6bc3c65c161 | trentmkelly/LessWrong-43k | LessWrong | "Personal Identity and Uploading", by Mark Walker
“Personal Identity and Uploading”, Mark Walker is the next JET paper. Abstract:
> Objections to uploading may be parsed into substrate issues, dealing with the computer platform of upload and personal identity. This paper argues that the personal identity issues of uploading are no more or less challenging than those of bodily transfer often discussed in the philosophical literature. It is argued that what is important in personal identity involves both token and type identity. While uploading does not preserve token identity, it does save type identity; and even qua token, one may have good reason to think that the preservation of the type is worth the cost.
1 Uploading: prospects and perils
> If, like me, you think that uploading is possible (at least in principle), and so you hold that the first interpretation of these events is correct, then you must hold true the following three theses:
>
> > 1. Computers are capable of supporting the important properties constitutive of personal identity, e.g., thought and consciousness.
>
> It is clear that uploading will not preserve all properties we associate with Homo sapiens, e.g., basic facts about the human digestive system are not likely to be preserved in uploading to a robotic body. But these facts are not typically thought to be important for personal identity. Candidates for important properties include thought, consciousness, emotions, creativity, aesthetic experience, sensory experience, empathy and so on. For the most part, the question of which properties are important is not as serious as it may first seem, since uploading promises to preserve the essential aspects of the brain and nervous system, which overlap with the usual lists of important properties for identity.
>
> …suffice it to say that if Searle is correct, then #1 may be false. For Searle thinks that a computer can never consciously think merely in virtue of instantiating a computer program, and the uploading process seems to be one of me |
29930f2e-8048-49ea-ab94-6d2089097058 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Brain-inspired AGI and the "lifetime anchor"
Last year Ajeya Cotra published a [draft report on AI timelines](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines). (See also: [summary and commentary by Holden Karnofsky](https://www.cold-takes.com/forecasting-transformative-ai-the-biological-anchors-method-in-a-nutshell/), [podcast interview with Ajeya](https://www.lesswrong.com/posts/CuDYhLLXq6FuHvGZc/axrp-episode-7-5-forecasting-transformative-ai-from).)
I commented at the time ([1](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines?commentId=RjZQLiikNZpTimWsp),[2](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines?commentId=7J92pd8pgDLRNuCcQ),[3](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines?commentId=TM84D4Jofq4fWdBuK)) in the form of skepticism about the usefulness of the "Genome Anchor" section of the report. Later I fleshed out those thoughts in my post [Against Evolution as an Analogy for how Humans Will Create AGI](https://www.lesswrong.com/posts/pz7Mxyr7Ac43tWMaC/against-evolution-as-an-analogy-for-how-humans-will-create), see especially the ["genome=code" analogy table near the top](https://www.lesswrong.com/posts/pz7Mxyr7Ac43tWMaC/against-evolution-as-an-analogy-for-how-humans-will-create#A_biological_analogy_I_like_much_better__The__genome___code__analogy).
In this post I want to talk about a different section of the report: the "Lifetime Anchor".
1. Assumptions for this post
============================
Here are some assumptions. I don’t *exactly* believe them—let alone with 100% confidence—but for the purpose of this post let’s say I do. I’m not going to present any evidence for or against them here. Think of it as the [Jeff Hawkins perspective](https://www.lesswrong.com/posts/ixZLTmFfnKRbaStA5/book-review-a-thousand-brains-by-jeff-hawkins) or something.
**ASSUMPTION 1:** There’s a “secret sauce” of human intelligence, and it looks like a learning algorithm (and associated inference algorithm).
**ASSUMPTION 2:** It’s a fundamentally different learning algorithm from deep neural networks. I don’t just mean a different neural network architecture, regularizer, etc. I mean *really* different, like “involving probabilistic program inference algorithms” or whatever.
**ASSUMPTION 3:** The algorithm is human-legible, but nobody knows how it works yet.
**ASSUMPTION 4:** We'll eventually figure out this “secret sauce” and get Transformative AI (TAI). *[Note added for clarification: To simplify the discussion, I'm assuming that when this is all happening, we don't already have TAI independently via some unrelated R&D path.]*
If you think these assumptions are all absolutely 100% wrong, well, I guess you might not find this post very interesting.
To be clear, Ajeya pretty much explicitly rejected these assumptions when writing [her report](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines) (cf. discussion of “algorithmic breakthroughs” [here](https://www.lesswrong.com/posts/CuDYhLLXq6FuHvGZc/axrp-episode-7-5-forecasting-transformative-ai-from)), so there's no surprise that I wind up disagreeing with what she wrote. Maybe I shouldn't even be using the word "disagree" in this post. Oh well; her report is still a good starting point / foil for present purposes.
2. Thesis and outline
=====================
**I will argue that** ***under those assumptions*****, once we understand that “secret sauce”, it’s plausible that we will then be <10 years away from optimized, tested, well-understood, widely-used, industrial-scale systems for training these models all the way to TAI.**
**I’ll also argue that training these models from scratch will plausibly be** ***easily*** **affordable, as in <$10M—i.e., a** ***massive***[**hardware overhang**](https://www.lesswrong.com/tag/computing-overhang)**.**
(By “plausible” I mean >25% probability I guess? Sorry, I’m not at the point where I can offer a probability distribution that isn’t pulled out of my ass.)
**Outline of the rest of this post**: First I’ll summarize and respond to Ajeya’s discussion of the “Lifetime Anchor” (which is not *exactly* the scenario I’m talking about here, but close). Then I’ll talk (somewhat speculatively) about time and cost involved in refactoring and optimizing and parallelizing and hardware-accelerating and scaling the new algorithm, and in doing training runs.
3. Background: The “Lifetime Anchor” in Ajeya Cotra's draft report
==================================================================
In [Ajeya's draft report](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines), one of the four bases for estimating TAI timelines is the so-called “Lifetime Anchor”.
She put it in the report but puts very little stock in it: she only gives it 5% weight.
What is the “Lifetime Anchor”? Ajeya starts by estimating that simulating a brain from birth to adulthood would involve a median estimate of 1e24 floating-point operations (FLOP). This comes from 1e24 FLOP ≈ 1e15 FLOP/s × 30 years, with the former being roughly the median estimate in [Joe Carlsmith’s report](https://www.openphilanthropy.org/brain-computation-report), and 30 years being roughly human adulthood (and rounds to a nice even 1e9 seconds). Actually, she uses the term “30 subjective years” to convey the idea that if we do a 10×-sped-up simulation of the brain, then the same training would take 3 years of wall-clock time, for example.
A 1e24 FLOP computation would cost about $10M in 2019, she says, and existing ML projects (like training AlphaStar at 1e23 FLOP) are already kinda in that ballpark. So 1e24 FLOP is *ridiculously cheap* for a transformative world-changing AI. (Memory requirements are also relevant, but I don’t think they change that picture, see footnote.[[1]](#fn0712vqtmwc9i))
OK, so far she has a probability distribution centered at 1e24 FLOP, proportional to the distribution she derived from Joe Carlsmith’s report. She then multiplies by a, let’s call it, “computer-vs-brain inefficiency factor” that she represents as a distribution centered at 1000. (I’ll get back to that.) Then there’s one more step of ruling out extremely-low-compute scenarios. (She rules them out for reasons that wouldn't apply to the scenario of Section 1 that I'm talking about here.) She combines this with estimates of investment and incremental algorithmic improvements and Moore's law and so on, and she winds up with a probability distribution for what year we'll get TAI. That's her “lifetime anchor”.
4. Why Ajeya puts very little weight on the Lifetime Anchor, and why I disagree
===============================================================================
Ajeya cites two reasons she doesn’t like the lifetime anchor.
**First,** it doesn’t seem compatible with the empirical model size and training estimates for current deep neural networks:
> I think the most plausible way for this hypothesis to be true would be if a) it turns out we need a smaller model than I previously assumed, e.g. ~1e11 or ~1e12 FLOP / subj sec with a similar number of parameters, *and* b) that model could be trained on a *very*short horizon ML problem, e.g. 1 to 10 seconds per data point. Condition a) seems quite unlikely to me because it implies our architectures are much more efficient than brain architectures discovered by natural selection; I don’t think we have strong reason [to expect this on priors](https://docs.google.com/document/d/1IJ6Sr-gPeXdSJugFulwIpvavc0atjHGM82QjIfUSBGQ/edit#heading=h.hxc5y3gfmjv4) and it doesn’t seem consistent with [evidence from other technological domains](https://docs.google.com/document/d/1IJ6Sr-gPeXdSJugFulwIpvavc0atjHGM82QjIfUSBGQ/edit#heading=h.xkxuso6p1x40). Condition b) seems somewhat unlikely to me because it seems likely by default that [transformative ML problems have naturally long horizon lengths](https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.2s3orj7g2t76) because we may need to select for abilities that evolution optimized for, and [possible measures to get around that](https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.o50oaumofd3v) may or may not work.
>
>
**Why I disagree:** As in Section 1, the premise of this post is that the human brain algorithm is a fundamentally different type of learning algorithm than a deep neural network. Thus I see no reason to expect that they would have the same scaling laws for model size, training data, etc.
**Second,** the implication is that training TAI is *so inexpensive* that we could have been doing it years ago. As she writes:
> Another major reason for skepticism is that (even with a median ~3 OOM larger than the human lifetime) this hypothesis implies a substantial probability that we could have trained a transformative model using less computation than the amount used in the most compute intensive training run of 2019 (AlphaStar at ~1e23 FLOP), and a large probability that we could have done so by spending only a few OOMs more money (e.g. $30M to $1B). I consider this to be a major point of evidence against it, because there are many well-resourced companies who could have afforded this kind of investment already if it would produce a transformative model, and they have not done so. See [below](https://docs.google.com/document/d/1k7qzzn14jgE-Gbf0CON7_Py6tQUp2QNodr_8VAoDGnY/edit#heading=h.6dz96ob6k6ro) for the update I execute against it.
>
>
**Why I disagree:** Again as in Section 1, the premise of this post is that nobody knows how the algorithm works. People can’t use an algorithm that doesn’t yet exist.
5. Why Ajeya thinks the computer-vs-brain inefficiency factor should be >>1, and why I disagree
===============================================================================================
Ajeya mentions a few reasons she wants to center her computer-vs-brain-inefficiency-factor distribution at 1000. I won’t respond to all of these, since some would involve a deep-dive into neuroscience that I don’t want to get into here. But I can respond to a couple.
**First,** deep neural network data requirements:
> Many models we are training currently already require orders of magnitude more data than a human sees in one lifetime.
>
>
**Why I disagree:** Again under the assumptions of Section 1, “many models we are training” are very different from human brain learning algorithms. Presumably human brain-like learning algorithms will have similar sample efficiency to actual human brain learning algorithms, for obvious reasons.
**Second,** she makes a reference-class argument using other comparisons between biological and human artifacts
> Brain FLOP/s seems to me to be somewhat more analogous to “ongoing energy consumption of a biological artifact” while lifetime FLOP seems to be more analogous to “energy required to *manufacture* a biological artifact”; Paul’s [brief investigation](https://docs.google.com/document/d/1HUtUBpRbNnnWBxiO2bz3LumEsQcaZioAPZDNcsWPnos/edit#) comparing human technologies to natural counterparts, which I discussed in [Part 1](https://docs.google.com/document/d/1IJ6Sr-gPeXdSJugFulwIpvavc0atjHGM82QjIfUSBGQ/edit#heading=h.r8kaeen4zwy6), found that the manufacturing cost of human-created artifacts tend to be more like ~3-5 OOM worse than their natural counterparts, whereas energy consumption tends to be more like ~1-3 OOM worse.
>
>
**Why I disagree:** Ajeya mentions two reference class arguments here: (1) “human-vs-brain FLOP/s ratio” is hypothesized to fit into the reference class of “human-artifact-vs-biological-artifact ongoing energy consumption ratio”; and (2) “human-vs-brain lifetime FLOP” is hypothesized to fit into the reference class of “human-artifact-vs-biological-artifact manufacturing energy”.
Under my assumptions here, the sample efficiency of brains and silicon should be similar—i.e., if you run similar learning algorithms on similar data, you should get similarly-capable trained models at the end. So from this perspective, the two ratios have to agree—i.e., these are two reference classes for the very same quantity. That’s fine; in fact, Ajeya’s median estimate of 3 OOM is nicely centered between the ~1-3 OOM reference class and the ~3-5 OOM reference class.
But I actually want to reject *both* of those numbers, because I think Joe Carlsmith’s report has already “priced in” human inefficiency by translating from neuron-centric metrics (number of neurons, synapses etc.) to silicon-centric metrics (FLOPs). (And then we estimated costs based on known $/FLOP of human ML projects.) So when we talk about FLOPs, we’ve already crossed over into human-artifact-world! It would be double-counting to add extra OOMs for human inefficiency.
Here's another way to make this same point: think about energy usage. [Joe Carlsmith’s report](https://www.openphilanthropy.org/brain-computation-report) says we need (median) 1e15 FLOP/s to simulate a brain. Based on existing hardware (maybe [5e9 FLOP/joule](https://en.wikipedia.org/wiki/Performance_per_watt)? EDIT: …or maybe much lower; see [comment](https://www.lesswrong.com/posts/W6wBmQheDiFmfJqZy/brain-inspired-agi-and-the-lifetime-anchor?commentId=HEHsFrgHpevMtA3Fc#HEHsFrgHpevMtA3Fc)), that implies (median) 200kW to simulate a brain. (Hey, $20/hour electricity bills, not bad!) Actual brains are maybe 20W, so we’re expecting our brain simulation to be about 4 OOM less energy-efficient than a brain. OK, fine.
…But now suppose I declare that in general, human artifacts are 3 OOM less efficient than biological artifacts. So we should *really* expect 4+3=7 OOM less energy efficiency, i.e. 200MW! I think you would say: that doesn’t make sense, it’s double-counting! That’s what *I* would say, anyway! And I’m suggesting that the above draft report excerpt is double-counting in an analogous way.
5.1 …And indeed why the computer-vs-brain inefficiency factor should be <<1!
----------------------------------------------------------------------------
My best guess for the inefficiency factor is actually <<1! (…At least, that’s my guess after a few years of people using these algorithms and picking the low-hanging fruit of implementing them efficiently.)
Why? Compare the following two possibilities:
* We understand the operating principles of the brain-like learning algorithms, and then implement those same learning algorithms on our silicon chips, versus
* We use our silicon chips to simulate biological neurons which in turn are running those brain-like learning algorithms.
Doing the second bullet point gets us an inefficiency factor of 1, by definition. But the second bullet point is bound to be far more inefficient than the first.
By analogy: If I want to multiply two numbers with my laptop, I can do it in nanoseconds directly, or I can do it *dramatically slower* by using my laptop to run a transistor-by-transistor simulation of a pocket calculator microcontroller chip.
Or here’s a more direct example: There’s a type of neuron circuit called a “central pattern generator”. (Fun topic by the way, see [here](https://www.alignmentforum.org/posts/PTkd8nazvH9HQpwP8/building-brain-inspired-agi-is-infinitely-easier-than).) A simple version might involve, for example, 30 neurons wired up in a particular way so as to send a wave of activation around and around in a loop forever. Let’s say (hypothetically) that this kind of simple central pattern generator is playing a role in an AGI-relevant algorithm. The second bullet point above would be like doing a simulation of those 30 neurons and all their interconnections. The first bullet point above would be like writing the one line of source code, “y = sin(ωt+φ)”, and then compiling that source code into assembly language. I think it’s obvious which one would require less compute!
(Silicon chips are maybe 7 OOM faster than brains. A faster but less parallel processor can emulate a slower but more parallel processor, *but not vice-versa*. So there’s a whole world of possible algorithm implementation strategies that brains cannot take advantage of but that we can—directly calculating sin(ωt+φ) is just one example.)
The scenario I’m talking about (see assumptions in Section 1) is the first bullet point above, not the second. So I consider an inefficiency factor <<1 to be a default expectation, again leaving aside the very earliest thrown-together implementations.
6. Some other timeline-relevant considerations
==============================================
6.1 How long does it take to get from janky grad-student code to polished, scalable, parallelized, hardware-accelerated, turn-key learning algorithms?
------------------------------------------------------------------------------------------------------------------------------------------------------
On the assumptions of Section 1, a brain-like learning algorithm would be sufficiently different from DNNs that some of the existing DNN-specific infrastructure would need to be re-built (things like PyTorch, TPU chips, pedagogical materials, a trained workforce, etc.).
How much time would that add?
Well I’ll try to draw an analogy with the history of DNNs (warning: I’m not terribly familiar with the history of DNNs).
AlexNet was 2012, DeepMind patented deep Q learning in 2014, the first TensorFlow release was 2015, the first PyTorch release was 2016, the first TPU was 2016, and by 2019 we had billion-parameter GPT-2.
So, maybe 7 years?
But that may be an overestimate. I think a lot of the deep neural net infrastructure *will* carry over to even quite different future ML algorithms. For example, the building up of people and money in ML, the building up of GPU servers and the tools to use them, the normalization of the idea that it’s reasonable to invest millions of dollars to train one model and to fab ML ASICs, the proliferation of expertise related to parallelization and hardware-acceleration, etc.—all these things would transfer directly to future human-brain-like learning algorithms. So maybe they’ll be able to develop in less time than it took DNNs to develop in the 2010s.
So, maybe the median guess should be somewhere in the range of 3-6 years?
6.2 How long (wall-clock time) does it take to train one of these models?
-------------------------------------------------------------------------
Should we expect engineers to be twiddling their thumbs for years and years, as their training runs run? If so, that would obviously add to the timeline.
The relevant factor here is *limits to parallelization*. If there weren’t limits to parallelization, you could make wall-clock time arbitrarily low by buying more processing power. For example, AlphaStar training took 14 days and totaled 1e23 FLOP, so it’s presumably feasible to squeeze a 1e24-FLOP, 30-subjective-year, training run into 14×10=140 days—i.e., 80 subjective seconds per wall-clock second. With more money, and another decade or two of technological progress, and a brain-vs-computer inefficiency factor <<1 as above, it would be even faster. But that case study only works if our future brain-like algorithms are at least as parallelizable as AlphaStar was.
Maybe my starting point should be the [AI Impacts’s Brain Performance In TEPS writeup](https://aiimpacts.org/brain-performance-in-teps/)? This comparison implies that existing supercomputers—as of the 2015 writeup—were not quite capable of real-time brain simulations (1 subjective second per wall-clock second), but they *were* within an order of magnitude. This makes it seem unlikely that we can get orders of magnitude faster than real-time. So, maybe we’ll be running our training algorithms for decades after all??
I’m not so sure. I still think it might well be much faster.
The most important thing is: I’m not a parallelization expert, but I assume that chip-to-chip connections are the bottleneck for the TEPS benchmark, not within-chip connections. (Someone please tell me if I’m wrong!) If I understand correctly, TEPS assumes that data is sent from an arbitrary node in the graph to a randomly-chosen different arbitrary node in the graph. So for a large calculation (more than a few chips), TEPS implicitly assumes that almost all connections are chip-to-chip. However, I think that in a brain simulation, data transmission events would be disproportionately likely to be within-chip.
For example, with adult brain volume of 1e6 mm3.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
, and an AlphaStar-like 400 silicon chips, naively each chip might cover about (13.5mm)3 of brain volume. So any neuron-to-neuron connection much shorter than 13.5mm is likely to translate to within-chip communication, not chip-to-chip. Then the figures at [this AI Impacts page](https://aiimpacts.org/transmitting-fibers-in-the-brain-total-length-and-distribution-of-lengths/) imply that almost all *un*myelinated fiber transmission would involve within-chip communication, and thus, chip-to-chip communication would mainly consist of:
* Information carried by long-range myelinated fibers. Using the [AI Impacts figure](https://aiimpacts.org/transmitting-fibers-in-the-brain-total-length-and-distribution-of-lengths/) of 160,000km of myelinated fibers, let’s guess that they're firing at [0.1-2 Hz](https://aiimpacts.org/rate-of-neuron-firing/) and typically 5cm long, then I get (3-60)e8 chip-to-chip TEPS from this source;
* Information carried by short-range fibers that happen to be near the boundary between the simulation zones of two chips. If you make a planar slice through the brain, I guess you would cut through on average ~3.5e11 axons and dendrites per m2 of slice (from [850,000km of axons and dendrites](https://aiimpacts.org/transmitting-fibers-in-the-brain-total-length-and-distribution-of-lengths/) in a 1200cm3 brain[[2]](#fnvanvzi2cir)). (Warning: a different estimation method[[3]](#fnafq79n0fm1p) gave 6e12 per m2 instead. Part of the discrepancy is probably that the latter is cortex and the former is the whole brain, including white matter which is presumably much more spaced out. Or maybe the AI Impacts 850,000km figure is wrong. Anyway, take all this with a large grain of salt.) So again if we imagine 400 chips each simulating a little (13.5mm)3 cube of brain, we get ~0.22m2 of total “virtual slices”, and if they’re firing at [0.1-2 Hz](https://aiimpacts.org/rate-of-neuron-firing/), we get something like (0.8-16)e10 chip-to-chip TEPS from this source
Recall the headline figure of [“brain performance in TEPS”](https://aiimpacts.org/brain-performance-in-teps/) was 1.8-64e13. So the above is ~3 OOM less! If I didn’t mess up, I infer a combination of (1) disproportionate numbers of short connections which turn into within-chip communications, and (2) a single long-range myelinated axon that connects to a bunch of neurons near its terminal, which from a chip-to-chip-communications perspective would look like just one connection.
Some other considerations that seem to point in the direction of “wall-clock training time probably won’t be years and years”:
* Technology is presumably improving, especially around processor-to-processor communications, and presumably it will continue to do so. For example, it looks like the highest-TEPS supercomputer increased from 2.4e13 TEPS to 1.0e14 TEPS between 2014 and 2021, if I’m reading [this](http://graph500.org/) right. (The *second* highest is still 2.4e13 though!)
* Again I’m not a parallelization expert, so maybe this is super-naive, but: whatever algorithms the brain is using, they’ve gotta be *extremely* parallelizable, right? Remember, we’re working with silicon chips that are ~7 OOM faster than the brain; even if we’re a whopping 100,000× less skillful at parallelizing brain algorithms than the brain itself, we’d *still* be able to simulate a brain at 100× speedup. So I guess I’d be pretty surprised if wall-clock time winds up being a showstopper, just on general principles.
* As mentioned above, I’m expecting the computer-vs-brain inefficiency factor to be <<1. I was talking about FLOPs there, but I think the same argument applies to TEPS.
* This is probably a <1 OOM effect, but I’ll say it anyway: I bet the “30 subjective years” figure is *way* overkill for TAI. Like, the smartest 15-year-old humans are much better programmers etc. than most adults, and even those smart 15-year-olds *sure* didn’t spend every minute of those 15 years doing optimally-efficient learning!!
* **Update:** See [this comment](https://www.lesswrong.com/posts/W6wBmQheDiFmfJqZy/brain-inspired-agi-and-the-lifetime-anchor?commentId=DgDKjcuk5saqP293a#DgDKjcuk5saqP293a) about the possibility of "parallel experiences".
**Update to add:** Here’s another possible objection. training requires both compute and data. Even if we can muster enough compute, what if data is a bottleneck? In particular, suppose for the sake of argument that the *only way* to train a model to AGI involves having the model control a real-world robot which spends tens of thousands of hours of serial time manipulating human-sized objects and chatting with humans. (And suppose also that [“parallel experiences”](https://www.lesswrong.com/posts/W6wBmQheDiFmfJqZy/brain-inspired-agi-and-the-lifetime-anchor?commentId=DgDKjcuk5saqP293a#DgDKjcuk5saqP293a) wind up being impossible). Then that would limit model training speed, even if we had infinitely fast computers. However, I view that possibility as highly unlikely—see my discussion of “embodiment” in [this post (Section 1.5)](https://www.alignmentforum.org/posts/4basF9w9jaPZpoC8R/intro-to-brain-like-agi-safety-1-what-s-the-problem-and-why#1_5_What_s_the_probability_that_we_ll_eventually_wind_up_with_brain_like_AGI_). My strong expectation is that future programmers will be able to make AGI *just fine* by feeding it YouTube videos, books, VR environments, and other such easily-sped-up data sources, with comparatively little real-world-manipulation experience thrown in at the very end. (After all, going in the opposite direction, humans can learn very quickly to get around in a VR environment after a lifetime in the real world.)
6.3 How many full-length training runs do we need?
--------------------------------------------------
If a “full-length training run” is the 30 subjective years or whatever, then an additional question is: how many such runs will we need to get TAI? I’m inclined to say: as few as 1 or 2, plus *lots and lots* of smaller-scale studies. For example, I believe there was one and only one full training run of GPT-3—all the hyperparameters were extrapolated from smaller-scale studies, and it worked well enough the first time.
Note also that imperfect training runs don’t necessarily need to be restarted from scratch; the partially-trained model may well be salvageable, I’d assume. And it’s possible to run multiple experiments in parallel, especially when there’s a human in the loop contextualizing the results.
So anyway, combining this and the previous subsection, I think it’s at least plausible for “wall-clock time spent running training” to be a minor contributor to TAI timelines (say, adding <5 years). That’s not guaranteed, just plausible. (As above, "plausible" = ">25% probability I guess").
7. Conclusion
=============
I’ll just repeat what I said in Section 2 above: if you accept the assumptions in section 1, I think we get the following kind of story:
*We can’t train a lifetime-anchor model today because we haven’t pinned down the brain-like learning algorithms that would be needed for it. But when we understand the secret sauce, we could plausibly be <10 years away from optimized, tested, well-understood, widely-used, industrial-scale systems for training these models all the way to TAI. And this training could plausibly be easily affordable, as in <$10M—i.e., a MASSIVE hardware overhang.*
(Thanks Dan Kokotajlo & Logan Smith for critical comments on drafts.)
1. **[^](#fnref0712vqtmwc9i)**Warning: FLOP is only one of several inputs to an algorithms. Another input worth keeping in mind is memory. In particular, the human neocortex has ≈1014 synapses. How this number translates into (for example) GB of GPU memory is complicated, and I have some uncertainty, but I think my Section 6.2 scenario (involving an AlphaStar-like 400 chips) does seem to be in the right general ballpark for not only FLOP but also memory storage.
2. **[^](#fnrefvanvzi2cir)**I assumed the axons and dendrites are locally isotropic (equally likely to go any direction); that gives a factor of 2 from averaging cos θ over a hemisphere.
3. **[^](#fnrefafq79n0fm1p)**I asked [Nick Turner](https://nicholasturner1.github.io/) and he kindly downloaded three random little volumes from [this dataset](https://ai.googleblog.com/2021/06/a-browsable-petascale-reconstruction-of.html) and counted how many things crossed the z=0 plane, as a very rough estimate. By the way, it was mostly axons not dendrites, like 10:1 ratio or something, in case anyone’s interested. |
3aa42ee2-9137-48a9-84df-edc1f9f88802 | trentmkelly/LessWrong-43k | LessWrong | My Fundamental Question About Omega
Omega has appeared to us inside of puzzles, games, and questions. The basic concept behind Omega is that it is (a) a perfect predictor and (b) not malevolent. The practical implications behind these points are that (a) it doesn't make mistakes and (b) you can trust its motives in the sense that it really, honestly doesn't care about you. This bugger is True Neutral and is good at it. And it doesn't lie.
A quick peek at Omega's presence on LessWrong reveals Newcomb's problem and Counterfactual Mugging as the most prominent examples. For those that missed them, other articles include Bead Jars and The Lifespan Dilemma.
Counterfactual Mugging was the most annoying for me, however, because I thought the answer was completely obvious and apparently the answer isn't obvious. Instead of going around in circles with a complicated scenario I decided to find a simpler version that reveals what I consider to my the fundamental confusion about Omega.
Suppose that Omega, as defined above, appears before you and says that it predicted you will give it $5. What do you do? If Omega is a perfect predictor, and it predicted you will give it $5... will you give it $5 dollars?
The answer to this question is probably obvious but I am curious if we all end up with the same obvious answer.
The fundamental problem behind Omega is how to resolve a claim by a perfect predictor that includes a decision you and you alone are responsible for making. This invokes all sorts of assumptions about choice and free-will, but in terms of phrasing the question these assumptions do not matter. I care about how you will act. What action will you take? However you label the source of these actions is your prerogative. The question doesn't care how you got there; it cares about the answer.
My answer is that you will give Omega $5. If you don't, Omega wouldn't have made the prediction. If Omega made the prediction AND you don't give $5 than the definition of Omega is flawed and we have to redefine Omeg |
c161ff16-738c-4939-bc82-7cab77157baa | trentmkelly/LessWrong-43k | LessWrong | How to make TensorFlow run faster
I quickly put this together for a fellow AI alignment researcher/engineer, so I thought I'd share it here.
The main recommendations are from Intel: https://software.intel.com/en-us/articles/maximize-tensorflow-performance-on-cpu-considerations-and-recommendations-for-inference
I implement these recommendations by setting the following in the code:
tf.config.threading.set_inter_op_parallelism_threads(2)
tf.config.threading.set_intra_op_parallelism_threads(6) # Number of physical cores.
And setting the environment variables in the PyCharm run configuration and Python Console settings.
As a result, the per-epoch training time dropped by 20 % with a small multi-layer perceptron on Fashion MNIST.
Another important point is to use a TensorFlow binary that uses all available CPU capabilities. Ie. it should display something like this:
2019-11-27 17:26:42.782399: I tensorflow/core/platform/cpu_feature_guard.cc:145] This TensorFlow binary is optimized with Intel(R) MKL-DNN to use the following CPU instructions in performance critical operations: SSE4.1 SSE4.2 AVX AVX2 FMA
And not something like this:
2019-11-28 09:18:26.315191: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
One way to achieve this is to install TensorFlow with Conda instead of Pip.
Here is more information about that: https://software.intel.com/en-us/articles/intel-optimization-for-tensorflow-installation-guide
If you're on a Mac and the Conda-TensorFlow crashes with an OMP error, here's the solution: https://stackoverflow.com/a/53692707/5091738 |
d83e3f62-9b21-432d-837e-f4abfa404a4d | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1928
(Last revised: July 2024. See changelog at the bottom.) 14.1 Post summary / Table of contents Part of the “Intro to brain-like-AGI safety” post series . Post #12 suggested two paths forward for solving “the alignment problem” for brain-like AGI , which I called “Social-instinct AGI” and “Controlled AGI”. Then Post #13 went into more detail about (one aspect of) “Social-instinct AGI”. And now, in this post, we’re switching over to “Controlled AGI”. If you haven’t read Post #12 , don’t worry, the “Controlled AGI” research path is nothing fancy—it’s merely the idea of solving the alignment problem in the most obvious way possible: The “Controlled AGI” research path: Step 1 (out-of-scope for this series) : We decide what we want our AGI’s motivation to be. For example, that might be: “Invent a better solar cell without causing catastrophe” ( task-directed AGI ), “Be a helpful assistant to the human supervisor” ( corrigible AGI assistants ), “Fulfill the human supervisor’s deepest life goals” ( ambitious value learning ), “Maximize coherent extrapolated volition ”, or whatever else we choose. Step 2 (subject of this post) : We make an AGI with that motivation. This post is about Step 2, whereas Step 1 is out-of-scope for this series. Honestly, I’d be ecstatic if we figured out how to reliably set the AGI’s motivation to any of those things I mentioned under Step 1. Unfortunately, I don’t know any good plan for Step 2, and (I claim) nobody else does either. But I do have some vague thoughts and ideas, and I will share them here, in the spirit of brainstorming. If you’re in a hurry and want to read a shorter and self-contained version of my least-bad proposed plan for Step 2, check out my separate post: Plan for mediocre alignment of brain-like [model-based RL] AGI , which basically puts together the most obvious ideas mentioned in Sections 14.2 and 14.3 into an end-to-end framework. I think that plan passes the low bar of “as far as I know, it might turn out OK”—well, I think I’m mildly skeptical, but I go back and forth, and I’m sure how to pin it down with more confidence. But obviously, we should be aiming higher than that! With stakes so high, we should really be starting from “there’s a strong reason to expect the plan to work, if carefully implemented”. And then we can start worrying about what can go wrong in the implementation. So we clearly still have work to do. This post is not meant to be a comprehensive overview of the whole problem, just what I see as the most urgent missing ingredients. Out of all the posts in the series, this post is the hands-down winner for “most lightly-held opinions”. For almost anything I say in this post, I can easily imagine someone changing my mind within an hour of conversation. Let that ‘someone’ be you—the comment section is below! Table of contents: Section 14.2 discusses what we might use as “Thought Assessors” in an AGI. If you’re just tuning in, Thought Assessors were defined in Posts #5 – #6 and have been discussed throughout the series. If you have a Reinforcement Learning background, think of Thought Assessors as the components of a multi-dimensional value function. If you have a “being a human” background, think of Thought Assessors as learned functions that trigger visceral reactions (aversion, cortisol-release, etc.) based on the thought that you’re consciously thinking right now. In the case of brain-like AGIs, we get to pick whatever Thought Assessors we want, and I propose three categories for consideration: Thought Assessors oriented towards safety (e.g. “this thought / plan involves me being honest”), Thought Assessors oriented towards accomplishing a task (e.g. “this thought / plan will lead to better solar cell designs”), and Thought Assessors oriented purely towards interpretability (e.g. “this thought / plan has something to do with dogs”). Section 14.3 discusses how we might generate supervisory signals to train those Thought Assessors. Part of this topic is what I call the “first-person problem”, namely the open question of whether it’s possible to take third-person labeled data (e.g. a YouTube video where Alice deceives Bob), and transmute it into a first-person preference (an AGI’s desire to not, itself, be deceptive). Section 14.4 discusses the problem that the AGI will encounter “edge cases” in its preferences—plans or places where its preferences become ill-defined or self-contradictory. I’m cautiously optimistic that we can build a system that monitors the AGI’s thoughts and detects when it encounters an edge case. However, I don’t have any good idea about what to do when that happens. I’ll discuss a few possible solutions, including “conservatism”, and a couple different strategies for what Stuart Armstrong calls Concept Extrapolation . Section 14.5 discusses the open question of whether we can rigorously prove anything about an AGI’s motivations. Doing so would seem to require diving into the AGI’s predictive world-model (which would probably be a multi-terabyte, learned-from-scratch , unlabeled data structure), and proving things about what the components of the world-model “mean”. I’m rather pessimistic about our prospects here, but I’ll mention possible paths forward, including John Wentworth’s “Natural Abstraction Hypothesis” research program (most recent update here ). Section 14.6 concludes with my overall thoughts about our prospects for “Controlled AGIs”. I’m currently a bit stumped and pessimistic about our prospects for coming up with a good plan, but hope I’m wrong and intend to keep thinking about it. I also note that a mediocre, unprincipled approach to “Controlled AGIs” (as in my “plan for mediocre alignment of brain-like AGI” post) would not necessarily cause a world-ending catastrophe—I think it’s hard to say. 14.2 Three categories of AGI Thought Assessors As background, here’s our usual diagram of motivation in the human brain, from Post #6 : See Post #6 . And here’s the modification for AGI, from Post #8 : On the center-right side of the diagram, I crossed out the words “cortisol”, “sugar”, “goosebumps”, etc. These correspond to the set of human innate visceral reactions which can be involuntarily triggered by thoughts (see Post #5 ). (Or in machine learning terms, these are more-or-less the components of a multidimensional value function, similar to what you find in multi-objective / multi-criteria reinforcement learning.) Clearly, things like cortisol, sugar, and goosebumps are the wrong Thought Assessors for our future AGIs. But what are the right ones? Well, we’re the programmers! We get to decide! I have in mind three categories to pick from. I’ll talk about how they might be trained (i.e., supervised) in Section 14.3 below. 14.2.1 Safety & corrigibility Thought Assessors Example thought assessors in this category: This thought / plan involves me being helpful. This thought / plan does not involve manipulating my own learning process, code, or motivation systems. This thought / plan does not involve deceiving or manipulating anyone. This thought / plan does not involve anyone getting hurt. This thought / plan involves following human norms , or more generally, doing things that an ethical human would plausibly do. This thought / plan is “low impact” (according to human common sense). … Arguably (cf. this Paul Christiano post ), #1 is enough, and subsumes the rest. But I dunno, I figure it would be nice to have information broken down on all these counts, allowing us to change the relative weights in real time ( Post #9, Section 9.7 ), and perhaps giving an additional measure of safety. Items #2–#3 are there because those are especially probable and dangerous types of thoughts—see discussion of Instrumental Convergence in Post #10, Section 10.3.2 . Item #5 is a bit of a catch-all for the AGI finding weird out-of-the-box solutions to problems, i.e. it’s my feeble attempt to mitigate the so-called “Nearest Unblocked Strategy problem” . Why might it mitigate the problem? Because pattern-matching to “things that an ethical human would plausibly do” is a bit more like a whitelist than a blacklist. I still don’t think that would work on its own, don't get me wrong, but maybe it would work in conjunction with the various other ideas in this post. Before you jump into loophole-finding mode ( “lol an ethical human would plausibly turn the world into paperclips if they’re under the influence of alien mind-control rays” ), remember (1) these are meant to be implemented via pattern-matching to previously-seen examples (Section 14.3 below), not literal-genie -style following the exact words of the text; (2) we would hopefully also have some kind of out-of-distribution detection system (Section 14.4 below) to prevent the AGI from finding and exploiting weird edge-cases in that pattern-matching process. That said, as we’ll see, I don’t quite know how to do either of those two things, and even if we figure it out, I don’t have an airtight argument that it would be sufficient to get the intended safe behavior. 14.2.2 Task-related Thought Assessors Example thought assessors in this category: This thought / plan will lead to a reduction in global warming This thought / plan will lead to a better solar panel design This thought / plan will lead to my supervisor becoming fabulously rich … This kind of thing is why we built the AGI—what we actually want it to do. (Assuming task-directed AGI for simplicity.) Basing a motivation system on these kinds of assessments by themselves would be obviously catastrophic. But maybe if we use these as motivations, in conjunction with the previous category, it will be OK. For example, imagine the AGI can only think thoughts that pattern-match to “I am being helpful” AND pattern-match to “there will be less global warming”. That said, I’m not sure we want this category at all. Maybe the “I am being helpful” Thought Assessor by itself is sufficient. After all, if the human supervisor is trying to reduce global warming, then a helpful AGI would produce a plan to reduce global warming. That’s kinda the approach here , I think. 14.2.3 “Ersatz interpretability” Thought Assessors (See Post #9, Section 9.6 for what I mean by “Ersatz interpretability”.) As discussed in Posts #4 – #5 , each thought assessor is a model trained by supervised learning. Certainly, the more Thought Assessors we put into the AGI, the more computationally expensive it will be. But how much more? It depends. For example, I think the “valence” Thought Assessor in the human brain involves orders of magnitude more neurons than the “salivation” Thought Assessor. On the other hand, I think the “valence” Thought Assessor is far more accurate as a result. Anyway, as far as I know, it’s not impossible that we can put in 10 7 Thought Assessors, and they’ll work well enough, and this will only add 1% to the total compute required by the AGI. I don’t know. So I’ll hope for the best and take the More Dakka approach: let’s put in 30,000 Thought Assessors, one for every word in the dictionary: This thought / plan has something to do with AARDVARK This thought / plan has something to do with ABACUS This thought / plan has something to do with ABANDON … … … This thought / plan has something to do with ZOOPLANKTON I expect that ML-savvy readers will be able to immediately suggest much-improved versions of this scheme—including versions with even more dakka —that involve things like contextual word embeddings and language models and so on. As one example, if we buy out and open-source Cyc (more on which below), we could use its hundreds of thousands of human-labeled concepts. 14.2.4 Combining Thought Assessors into a reward function For an AGI to judge a thought / plan as being good, we’d like all the safety & corrigibility Thought Assessors from Section 14.2.1 to have as high a value as possible, and we’d like the task-related Thought Assessor from Section 14.2.2 (if we’re using one) to have as high a value as possible. (The outputs of the interpretability Thought Assessors from Section 14.2.3 are not inputs to the AGI’s reward function, or indeed used at all in the AGI, I presume. I was figuring that they’d be silently spit out to help the programmers do debugging, testing, monitoring, etc.) So the question is: how do we combine this array of numbers into a single overall score that can guide what the AGI decides to do? A probably-bad answer is “add them up”. We don’t want the AGI going with a plan that performs catastrophically badly on all but one of the safety-related Thought Assessors, but so astronomically well on the last one that it makes up for it. Instead, I imagine we’ll want to apply some kind of nonlinear function with strongly diminishing returns, and/or maybe even acceptability thresholds, before adding up the Thought Assessors into an overall score. I don’t have much knowledge or opinion about the details. But there is some related literature on “scalarization” of multi-dimensional value functions—see here for some references. 14.3 Supervising the Thought Assessors, and the “first-person problem” Recall from Posts #4 – #6 that the Thought Assessors are trained by supervised learning. So we need a supervisory signal—what I labeled “ground truth in hindsight” in the diagram at the top. I’ve talked about how the brain generates ground truth in numerous places, e.g. Post #3 Section 3.2.1 , Posts #7 & #13 . How do we generate it for the AGI? Well, one obvious possibility is to have the AGI watch YouTube, with lots of labels throughout the video for when we think the various Thought Assessors ought to be active. Then when we’re ready to send the AGI off into the world to solve problems, we turn off the labeled YouTube videos, and simultaneously freeze the Thought Assessors (= set the error signals to zero) in their current state. Well, I’m not sure if that would work; maybe the AGI has to go back and watch more labeled YouTube videos from time to time, to help the Thought Assessors keep up as the AGI’s world-model grows and changes. One potential shortcoming of this approach is related to first-person versus third-person concepts. We want the AGI to have strong preferences about aspects of first-person plans—hopefully, the AGI will see “I will lie and deceive” as bad, and “I will be helpful” as good. But we can’t straightforwardly get that kind of preference from the AGI watching labeled YouTube videos. The AGI will see YouTube character Alice deceiving YouTube character Bob, but that’s different from the AGI itself being deceptive. And it’s a very important difference! Consider: If you tell me “my AGI dislikes being deceptive”, I’ll say “good for you!”. If you tell me “my AGI dislikes it when people are deceptive”, I’ll say “for god's sake you better shut that thing off before it escapes human control and kills everyone”!!! It sure would be great if there were a way to transform third-person data (e.g. a labeled YouTube video of Alice deceiving Bob) into an AGI’s first-person preferences (“I don’t want to be deceptive”). I call this the first-person problem . How do we solve the first-person problem? I’m not entirely sure. Maybe we can apply interpretability tools to the AGI’s world-model, and figure out how it represents itself, and then correspondingly manipulate its thoughts, or something? It’s also possible that further investigation into human social instincts ( previous post ) will shed some light, as human social instincts do seem to transform the third-person “everyone in my friend group is wearing green lipstick” into the first-person “I want to be wearing green lipstick”. If the first-person problem is not solvable, we need to instead use the scary method of allowing the AGI to take actions, and putting labels on those actions. Why is that scary? First, because those actions might be dangerous. Second, because it doesn’t give us any good way to distinguish (for example) “the AGI said something dishonest” from “the AGI got caught saying something dishonest”. Conservatism and/or concept extrapolation (Section 14.4 below) could help with that “getting caught” problem—maybe we could manage to get our AGI both motivated to be honest and motivated to not get caught, and that could be good enough—but it still seems fraught for various reasons. 14.3.1 Side note: do we want first-person preferences? I suspect that “the first-person problem” is intuitive for most readers. But I bet a subset of readers feel tempted to say that the first-person problem is not in fact a problem at all. After all, in the realm of human affairs, there’s a good argument that we could use a lot fewer first-person preferences! The opposite of first-person preferences would be “impersonal consequentialist preferences”, wherein there’s a future situation that we want to bring about (e.g. “awesome post-AGI utopia”), and we make decisions to try to bring that about, without particular concern over what I-in-particular am doing. Indeed, too much first-person thinking leads to lots of things that I personally dislike in the world—e.g. jockeying for credit, blame avoidance, the act / omission distinction, social signaling, and so on. Nevertheless, I still think giving AGIs first-person preferences is the right move for safety. Until we can establish super-reliable 12th-generation AGIs, I’d like them to treat “a bad thing happened (which had nothing to do with me)” as much less bad than “a bad thing happened (and it’s my fault)”. Humans have this notion, after all, and it seems at least relatively robust—for example, if I build a bank-robbing robot, and then it robs the bank, and then I protest “Hey I didn’t do anything wrong; it was the robot!”, I wouldn’t be fooling anybody, much less myself. An AGI with such a preference scheme would presumably be cautious and conservative when deciding what to do, and would default to inaction when in doubt. That seems generally good, which brings us to our next topic: 14.4 Conservatism and concept-extrapolation 14.4.1 Why not just relentlessly optimize the right abstract concept? Let’s take a step back. Suppose we build an AGI such that it has positive valence on the abstract concept “there will be lots of human flourishing”, and consequently makes plans and take actions to make that concept happen. I’m actually pretty optimistic that we’ll be able to do that, from a technical perspective. Just as above, we can use labeled YouTube videos and so on to make a Thought Assessor for “this thought / plan will lead to human flourishing”, and then base the reward function purely on that one Thought Assessor (cf. Post #7 ). And then we set the AGI loose on an unsuspecting world, to go do whatever it thinks is best to do. What could go wrong? The problem is that the concept of “human flourishing” is an abstract concept in the AGI’s world-model—really, it’s just a fuzzy bundle of learned associations. It’s hard to know what actions a desire for “human flourishing” will induce, especially as the world itself changes, and the AGI’s understanding of the world changes even more. In other words, there is no future world that will perfectly pattern-match to the AGI’s current notion of “human flourishing”, and if an extremely powerful AGI optimized the world for the best possible pattern-match, we might wind up with something weird, even catastrophic. (Or maybe not! It’s pretty hard to say, more on which in Section 14.6.) As some random examples of what might go wrong: maybe the AGI would take over the world and prevent humans and human society from changing or evolving forevermore, because those changes would reduce the pattern-match quality. Or maybe the least-bad pattern-match would be the AGI wiping out actual humans in favor of an endless modded game of The Sims . Not that The Sims is a perfect pattern-match to “human flourishing”—it’s probably pretty bad! But maybe it’s less bad a pattern-match than anything the AGI could feasibly do with actual real-world humans. Or maybe as the AGI learns more and more, its world-model gradually drifts and changes, such that the frozen Thought Assessor winds up pointing at something totally random and crazy, and then the AGI wipes out humans to tile the galaxy with paperclips. I don’t know! So anyway, relentlessly optimizing a fixed, frozen abstract concept like “human flourishing” seems maybe problematic. Can we do better? Well, it would be nice if we could also continually refine that concept, especially as the world itself, and the AGI’s understanding of the world, evolves. This idea is what Stuart Armstrong calls Concept Extrapolation , if I understand correctly. Concept extrapolation is easier said than done—there’s no obvious ground truth for the question of “what is ‘human flourishing’, really ?” For example, what would “human flourishing” mean in a future of transhuman brain-computer hybrid people and superintelligent evolved octopuses and god-only-knows-what-else? Anyway, we can consider two steps to concept extrapolation. First (the easier part), we need to detect edge-cases in the AGI’s preferences. Second (the harder part), we need to figure out what the AGI should do when it comes across such an edge-case. Let’s talk about those in order. 14.4.2 The easier part of concept extrapolation: Detecting edge-cases in the AGI’s preferences I’m cautiously optimistic about the feasibility of making a simple monitoring algorithm that can watch an AGI’s thoughts and detect that it’s in an edge-case situation—i.e., an out-of-distribution situation where its learned preferences and concepts are breaking down. (Understanding the contents of the edge-case seems much harder, as discussed shortly, but here I’m just talking about recognizing the occurrence of an edge-case.) To pick a few examples of possible telltale signs that an AGI is at an edge-case: The learned probability distributions for Thought Assessors (see Post #4 footnote ) could have a wide variance, indicating uncertainty. The different Thought Assessors of Section 14.2 could diverge in new and unexpected ways. The AGI’s reward prediction error could flip back and forth between positive and negative in a way that indicates “feeling torn” while paying attention to different aspects of the same possible plan. The AGI’s generative world-model could settle into a state with very low prior probability, indicating confusion. 14.4.3 The harder part of concept extrapolation: What to do at an edge case I don’t know of any good answer. Here are some options. 14.4.3.1 Option A: Conservatism—When in doubt, just don’t do it! A straightforward approach would be that if the AGI’s edge-case-detector fires, it forces the valence signal negative—so that whatever thought the AGI was thinking is taken to be a bad thought / plan. This would loosely correspond to a “conservative” AGI. (Side note: I think there may be many knobs we can turn in order to make a brain-like AGI more or less “conservative”, in different respects. The above is just one example. But they all seem to have the same issues.) A failure mode of a conservative AGI is that the AGI just sits there, not doing anything, paralyzed by indecision, because every possible plan seems too uncertain or risky. An “AGI paralyzed by indecision” is a failure mode, but it’s not a dangerous failure mode. Well, not unless we were foolish enough to put this AGI in charge of a burning airplane plummeting towards the ground. But that’s fine—in general, I think it’s OK to have first-generation AGIs that can sometimes get paralyzed by indecision, and which are thus not suited to solving crises where every second counts. Such an AGI could still do important work like inventing new technology, and in particular designing better and safer second-generation AGIs. However, if the AGI is always paralyzed by indecision—such that it can’t get anything done— now we have a big problem. Presumably, in such a situation, future AGI programmers would just dial the “conservatism” knob down lower and lower, until the AGI started doing useful things. And at that point, it’s unclear if the remaining conservatism would be sufficient to buy us safety. I think it would be much better to have a way for the AGI to iteratively gain information to reduce uncertainty, while remaining highly conservative in the face of whatever uncertainty still remains. So how can we do that? 14.4.3.2 Option B: Dumb algorithm to seek clarification in edge-cases Here’s a slightly-silly illustrative example of what I have in mind. As above, we could have a simple monitoring algorithm that watches the AGI’s thoughts, and detects when it’s in an edge-case situation. As soon as it is, the monitoring algorithm shuts down the AGI entirely, and prints out the AGI’s current neural net activations (and corresponding Thought Assessor outputs). The programmers use interpretability tools to figure out what the AGI is thinking about, and manually assign a valence / value / reward, overriding the AGI’s previous uncertainty with a highly-confident ground-truth. That particular story seems unrealistic, mainly because we probably won’t have sufficiently reliable and detailed interpretability tools. (Prove me wrong, interpretability researchers!) But maybe there’s a better approach than just printing out billions of neural activations and corresponding Thought Assessors? The tricky part is that AGI-human communication is fundamentally a hard problem. It’s unclear to me whether it will be possible to solve that problem via a dumb algorithm. The situation here is very different from, say, an image classifier, where we can find an edge-case picture and just show it to the human. The AGI’s thoughts may be much more inscrutable than that. By analogy, human-human communication is possible, but not by any dumb algorithm. We do it by leveraging the full power of our intellect—modeling what our conversation partner is thinking, strategically choosing words that will best convey a desired message, and learning through experience to communicate more and more effectively. So what if we try that approach? 14.4.3.3 Option C: The AGI wants to seek clarification in edge-cases If I’m trying to help someone, I don’t need any special monitoring algorithm to prod me to seek clarification at edge-cases. Seeking clarification at edge-cases is just what I want to do , as a self-aware properly-motivated agent. So what if we make our AGIs like that? At first glance, this approach would seem to solve all the problems mentioned above. Not only that, but the AGI can use its full powers to make everything work better. In particular, it can learn its own increasingly-sophisticated metacognitive heuristics to flag edge-cases, and it can learn and apply the human’s meta-preferences about how and when the AGI should ask for clarification. But there’s a catch. I was hoping for a conservatism / concept extrapolation system that would help protect us from misdirected motivations. If we implement conservatism / concept extrapolation via the motivation system itself, then we lose that protection. More specifically: if we go up a level, the AGI still has a motivation (“seek clarification in edge-cases”), and that motivation is still an abstract concept that we have to extrapolate into out-of-distribution edge cases (“What if my supervisor is drunk, or dead, or confused? What if I ask a leading question?”). And for that concept extrapolation problem, we’re plowing ahead without a safety net. Is that a problem? Bit of a long story: Side-debate: Will “helpfulness”-type preferences “extrapolate” safely just by recursively applying to themselves? In fact, a longstanding debate in AGI safety is whether these kinds of helpful / corrigible AGI preferences (e.g. an AGI’s desire to understand and follow a human’s preferences and meta-preferences) will “extrapolate” in a desirable way without any “safety net”—i.e., without any independent ground-truth mechanism pushing the AGI’s preferences in the right direction. In the optimistic camp is Paul Christiano, who argued in “Corrigibility” (2017) that there would be “a broad basin of attraction towards acceptable outcomes”, based on, for example, the idea that an AGI’s preference to be helpful will result in the AGI having a self-reflective desire to continually edit its own preferences in a direction humans would like. But I don’t really buy that argument for reasons in my 2020 post —basically, I think there are bound to be sensitive areas like “what does it mean for people to want something” and “what are human communication norms” and “inclination to self-monitor”, and if the AGI’s preferences drift along any of those axes (or all of them simultaneously), I’m not convinced that those preferences would self-correct. Meanwhile, in the strongly-pessimistic camp is Eliezer Yudkowsky, I think mainly because of an argument (e.g. this post, final section ) that we should expect powerful AGIs to have consequentialist preferences, and that consequentialist preferences seem incompatible with corrigibility. But I don’t really buy that argument either, for reasons in my 2021 “Consequentialism & Corrigibility” post —basically, I think there are possible preferences that are reflectively-stable, and that include consequentialist preferences (and thus are compatible with powerful capabilities), but are not purely consequentialist (and thus are compatible with corrigibility). A “preference to be helpful” seems like it could plausibly develop into that kind of hybrid preference scheme. Anyway, I’m uncertain but leaning pessimistic. For more on the topic, see also Wei Dai’s recent post , and RogerDearnaley’s , and the comment sections of all of the posts linked above. 14.4.3.4 Option D: Something else? I dunno. 14.5 Getting a handle on the world-model itself The elephant in the room is the giant multi-gigabyte unlabeled generative world-model that lives inside the Thought Generator. The Thought Assessors provide a window into this world-model, but I’m concerned that it may be a rather small, foggy, and distorted window. Can we do better? Ideally, we’d like to prove things about the AGI’s motivation. We’d like to say “Given the state of the AGI’s world-model and Thought Assessors, the AGI is definitely motivated to do X” (where X=be helpful, be honest, not hurt people, etc.) Wouldn’t that be great? But we immediately slam into a brick wall: How do we prove anything whatsoever about the “meaning” of things in the world-model, and thus about the AGI’s motivation? The world is complicated, and therefore the world-model is complicated. The things we care about are fuzzy abstractions like “honesty” and “helpfulness”—see the Pointers Problem . The world-model keeps changing as the AGI learns more, and as it makes plans that would entail taking the world wildly out-of-distribution (e.g. planning the deployment of a new technology). How can we possibly prove anything here? I still think the most likely answer is “We can’t”. But here are two possible paths anyway. For some related discussion, see Eliciting Latent Knowledge , and especially Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems (Dalrymple et al. , 2024) . Proof strategy #1 starts with the idea that we live in a three-dimensional world containing objects and so on. We try to come up with an unambiguous definition of what those objects are, and from there we can have an unambiguous language for specifying what we want to happen in the world. We also somehow translate (or constrain) the AGI’s understanding of the world into that language, and now we can prove theorems about what the AGI is trying to do. This is my tentative understanding of what John Wentworth is trying to do via his Natural Abstraction Hypothesis research program (most recent update here ), and I’ve heard ideas in this vicinity from a couple other people as well. (Update: John disagrees with this characterization, see his comment .) I’m skeptical because a 3D world of localized objects seems to be an unpromising starting point for stating and proving useful theorems about the AGI’s motivations. After all, a lot of things that we humans care about, and that the AGI needs to care about, seem difficult to describe in terms of a 3D world of localized objects—consider the notion of “honesty”, or “solar cell efficiency”, or even “daytime”. Proof strategy #2 would start with a human-legible “reference world-model” (e.g. Cyc ). This reference world-model wouldn’t be constrained to be built out of localized objects in a 3D world, so unlike the above, it could and probably would contain things like “honesty” and “solar cell efficiency” and “daytime”. Then we try to directly match up things in the “reference world-model” with things in the AGI’s world-model. Will they match up? No, of course not. Probably the best we can hope for is a fuzzy, many-to-many match, with various holes on both sides. It's hard for me to see a path to rigorously proving anything about the AGI’s motivations using this approach. Nevertheless, I continue to be amazed that unsupervised machine translation is possible at all, and I take that as an indirect hint that if pieces of two world-models match up with each other in their internal structure, then those pieces are probably describing the same real-world thing. So maybe I have the faintest glimmer of hope. I’m unaware of work in this direction, possibly because it’s stupid and doomed, and also possibly because I don’t think we currently have any really great open-source human-legible world-models to run experiments on. The latter is a problem that I think we should rectify ASAP, perhaps by cutting a giant check to open-source Cyc , or else developing a similarly rich, accurate, and (most importantly) human-legible open-source world-model by some other means. (See also some helpful discussion in Towards Guaranteed Safe AI about how to build an open-source human-legible world-model, although they have in mind a different end-use for it than I do. Indeed, there are lots of different reasons to want an awesome open-source human-legible world-model! All the more reason to make one!) 14.6 Conclusion: mild pessimism about finding a good solution, uncertainty about the consequences of a lousy solution I think we have our work cut out figuring out how to solve the alignment problem via the "Controlled AGIs" route (as defined in Post #12 ). There are a bunch of open problems, and I’m currently pretty stumped. We should absolutely keep looking for good solutions, but right now I’m also open-minded to the possibility that we won’t find any. That’s why I continue to put a lot of my mental energy into the “social-instinct AGIs” path (Posts #12 – #13 ), which seems somewhat less doomed to me, despite its various problems. I note, however, that my pessimism is not universally shared—for example, as mentioned, Stuart Armstrong at AlignedAI appears optimistic about solving the open problem in Section 14.4, and John Wentworth and the Guaranteed Safe AI people appear optimistic about solving the open problem in Section 14.5. Let's hope they're right, wish them luck, and try to help! To be clear, the thing I’m feeling pessimistic about is finding a good solution to “Controlled AGI”, i.e., a solution that we can feel extremely confident in a priori . A different question is: Suppose we try to make “Controlled AGI” via a lousy solution, like the Section 14.4.1 example (encapsulated in my post Plan for mediocre alignment of brain-like [model-based RL] AGI ) where we imbue a super-powerful AGI with an all-consuming desire for the abstract concept of “human flourishing”, and the AGI then extrapolates that abstract concept arbitrarily far out of distribution in a totally-uncontrolled, totally-unprincipled way. Just how bad a future would such an AGI bring about? I’m very uncertain. Would such an AGI engage in mass torture? Umm, I guess I’m cautiously optimistic that it wouldn’t, absent a sign error from cosmic rays or whatever. Would it wipe out humanity? I think it’s possible!—see discussion in Section 14.4.1. But it might not! Hey, maybe it would even bring about a pretty awesome future! I just really don’t know, and I’m not even sure how to reduce my uncertainty. In the next post , I will wrap up the series with my wish-list of open problems, and advice on how to get into the field and help solve them! Changelog July 2024: Since the initial version, I’ve made only minor changes. Mostly I added links to more recent content, particularly my own Plan for mediocre alignment of brain-like [model-based RL] AGI (which is basically a simpler self-contained version of part of this post), and Dalrymple et al.’ s Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems , which is relevant to Section 14.5. |
b0556b7d-c090-41fb-b52e-56b820bee3d9 | trentmkelly/LessWrong-43k | LessWrong | A Rationality Condition for CDT Is That It Equal EDT (Part 1)
[Epistemic Status: this series of two posts gives some arguments which, in my eyes, make it difficult to maintain a position other than CDT=EDT, but not impossible. As I explain at the end of the second post, it is still quite tenable to suppose that CDT and EDT end up taking different actions.]
Previously, I argued that fair comparisons of CDT and EDT (in which the same problem representation is given to both decision theories) will conclude that CDT=EDT, under what I see as reasonable assumptions. Recently, Paul Christiano wrote a post arguing that, all things considered, the evidence strongly favors EDT. Jessica Taylor pointed out that Paul didn't address the problem of conditioning on probability zero events, but she came up with a novel way of addressing that problem by taking the limit of small probabilities: COEDT.
Here, I provide further arguments that rationality constraints point in the direction of COEDT-like solutions.
Note that I argue for the conclusion that CDT=EDT, which is somewhat different from arguing directly for EDT; my line of reasoning suggests some additional structure which could be missed by advocating EDT in isolation (or CDT in isolation). Paul's post described CDT as a very special case of EDT, in which our action is independent of other things we care about. This is true, but, we can also accurately describe EDT is a very special case of CDT where all probabilistic relationships which remain after conditioning on what we know turn out to also be causal relationships. I more often think in the second way, because CDT can have all sorts of counterfactuals based on how causation works. EDT claims that these are only correct when they agree with the conditional probabilities.
(ETA: When I say "CDT", I'm pointing at some kind of steel-man of CDT which uses logical counterfactuals rather than physical counterfactuals. TDT is a CDT in this sense, whereas UDT could be either CDT or EDT.)
This post will be full of conjectural sketches, and |
c7bf2bfe-f7ca-41cd-b4c7-d3150eb7daef | trentmkelly/LessWrong-43k | LessWrong | Adversarial attacks and optimal control
Meta: After a fun little motivating section, this post goes deep into the mathematical weeds. This thing aims to explore the mathematical properties of adversarial attacks from first principles. Perhaps other people are not as confused about this point as I was, but hopefully, the arguments are still useful and/or interesting to some. I'd be curious to hear if I'm reinventing the wheel.
TL;DR:
1. When we invest the appropriate effort, the probability of random catastrophic events tends to decrease exponentially with a rate given by the[1] rate function.
2. One way of estimating the rate function is to solve an optimal control problem, where we have to determine the "least surprising" way that the catastrophic event comes about. The rate function equals the catastrophic event's surprise (in nats).
3. Thus, intuitively, as we invest the effort to decrease the probability of random catastrophic events, the "difficulty" of performing an adversarial attack only increases linearly.
Real talk and real estate
Zillow is an American tech real-estate marketplace company that recently had (what the experts call) a small snafu. They decided they were done just being a marketplace and started buying up homes, completing light renovations, and then selling them with a profit. The whole thing went poorly; they bought houses too expensive and had to sell at a loss, costing the company $420 million and leading to large lay-offs.
This story is not very interesting[2] for anyone not directly involved. The reason I remember the whole affair is a Twitter thread that also caught the attention of Eliezer:
The thread proceeds to lay out how Zillow relied too much on their automatic estimates, which (ex-ante) looked good on average, but (ex-post) manifested in the worst possible way[3]. Or, in the words of the OP:
> "[They mistook] an adversarial environment for a random one."
I have approximately zero knowledge about Zillow or the real estate business, so I can't comment on whet |
16b32947-ee28-4f19-abeb-a04fd282b9e0 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Learning Normativity: A Research Agenda
*(Related to* [*Inaccessible Information*](https://www.alignmentforum.org/posts/ZyWyAJbedvEgRT2uF/inaccessible-information)*,* [*Learning the Prior*](https://www.alignmentforum.org/posts/SL9mKhgdmDKXmxwE4/learning-the-prior)*, and* [*Better Priors as a Safety Problem*](https://www.alignmentforum.org/posts/roA83jDvq7F2epnHK/better-priors-as-a-safety-problem)*. Builds on several of my* [*alternate alignment ideas*](https://www.lesswrong.com/s/SBfqYgHf2zvxyKDtB)*.)*
I want to talk about something which I'll call *learning normativity*. What is normativity? Normativity is correct behavior. I mean something related to the fuzzy concept humans convey with the word "should". I think it has several interesting features:
* Norms are the result of a complex negotiation between humans, so they shouldn't necessarily be thought of as the result of maximizing some set of values. This distinguishes learning normativity from value learning.
* A lot of information about norms is present in the empirical distribution of what people actually do, but you can't learn norms just by learning human behavior. This distinguishes it from imitation learning.
* It's often possible to provide a lot of information in the form of "good/bad" feedback. This feedback should be interpreted more like [approval-directed learning](https://www.alignmentforum.org/posts/7Hr8t6xwuuxBTqADK/approval-directed-agents-1) rather than RL. However, approval should not be treated as a gold standard.
* Similarly, it's often possible to provide a lot of information in the form of rules, but rules are not necessarily 100% true; they are just very likely to apply in typical cases.
* In general, it's possible to get very rich *types* of feedback, but very *sparse*: humans get all sorts of feedback, including not only instruction on how to act, but also *how to think.*
* Any one piece of feedback is suspect. Teachers can make mistakes, instructions can be wrong, demonstrations can be imperfect, dictionaries can contain spelling errors, reward signals can be corrupt, and so on.
Example: Language Learning
==========================
A major motivating example for me is how language learning works in humans. There is clearly, to some degree, a "right way" and a "wrong way" to use a language. I'll call this *correct usage*.
One notable feature of language learning is that we don't always speak, or write, in correct usage. This means that a child learning language has to distinguish between mistakes (such as typos) and correct usage. (Humans do sometimes learn to imitate mistakes, but *we have a notion of not doing so*. This is unlike [GPT](https://www.lesswrong.com/tag/gpt) systems learning to imitate the empirical distribution of human text.)
This means we're largely doing something like unsupervised learning, but with a notion of "correct"/"incorrect" data. We're doing something like throwing data out when it's likely to be incorrect.
A related point is that we are better at *recognizing* correct usage than we are at *generating* it. If we say something wrong, we're likely able to correct it. In some sense, this means there's a foothold for intelligence amplification: we know how to generate our own training gradent.
Another fascinating feature of language is that although native speakers are pretty good at both recognizing and generating correct usage, *we don't know the rules explicitly*. The whole field of linguistics is largely about trying to uncover the rules of grammar.
So it's impossible for us to teach proper English by teaching the rules. Yet, we do know *some* of the rules. Or, more accurately, we know a set of rules that *usually* apply. And those rules are *somewhat* useful for teaching English. (Although children have usually reached fluency before the point where they're taught explicit English grammar.)
All of these things point toward what I mean by *learning normativity:*
* We can tell a lot about what's normative by simply observing what's common, but the two are not exactly the same thing.
* A (qualified) human can usually label an example as correct or incorrect, but this is not perfect either.
* We can articulate a lot about correct vs incorrect in the form of rules; but the rules which we can articulate never seem to cover 100% of the cases. A linguist is a lot like a philosopher: taking a concept which is understood at an intuitive level (which a great many people can fluently apply in the correct manner), but struggling for years to arrive at a correct technical definition which fits the intuitive usage.
In other words, the overriding feature of normativity which I'm trying to point at is that nothing is ever 100%. Correct grammar is not defined by any (known) rules or set of text, nor is it (quite) just whatever humans judge it is. All of those things give a lot of information about it, but it could differ from each of them. Yet, on top of all that, basically everyone learns it successfully. This is very close to Paul's [Inaccessible Information](https://www.alignmentforum.org/posts/ZyWyAJbedvEgRT2uF/inaccessible-information): information for which we cannot concoct a gold-standard training signal, but which intelligent systems may learn anyway.
Another important feature of this type of learning: *there is a fairly clear notion of superhuman performance.* Even though human imitation is *most* of the challenge, we could declare something superhuman based on our human understanding of the task. For example, GPT is trained exclusively to imitate, so it should never exceed human performance. Yet, we could tell if a GPT-like system *did* exceed human performance:
* Its spelling and grammar would be immaculate, rather than including humanlike errors;
* its output would be more creative and exciting to read than that of human authors;
* when good reasoning was called for in a text, its arguments would be clear, correct, and compelling;
* when truth was called for, rather than fiction, its general knowledge would be broader and more accurate than a human's.
It seems very possible to learn to be better than your teachers in these ways, because humans sometimes manage to do it.
Learning in the Absence of a Gold Standard
==========================================
In statistics and machine learning, a "gold standard" is a proxy which we treat as good enough to serve as ground truth for our limited purposes. The accuracy of any other estimate will be judged by comparison to the gold standard. This is similar to the concept of "operationalization" in science.
It's worth pointing out that in pure Bayesian terms, there is nothing especially concerning about learning in the absence of a gold standard. I have data X. I want to know about Y. I update on X, getting P(Y|X). No problem!
However, that only works if we have the right prior. We could try to [learn the prior from humans](https://www.lesswrong.com/posts/SL9mKhgdmDKXmxwE4/learning-the-prior), which gets us 99% of the way there... but as I've mentioned earlier, human imitation does not get us all the way. Humans don't perfectly endorse their own reactions.
(Note that whether "99% of the way" is good enough for AI safety is a separate question. I'm trying to define the Big Hairy Audacious Goal of learning normativity.)
Actually, I want to split "no gold standard" into two separate problems.
1. **There's no type of feedback which we can perfectly trust.** If humans label examples of good/bad behavior, a few of those labels are going to be wrong. If humans provide example inferences for learning the prior, some of those example inferences are (in a very real sense) wrong. And so on.
2. **There's no level at which we can perfectly define the loss function.** This is a consequence of no-perfect-feedback, but it's worth pointing out separately.
No Perfect Feedback
-------------------
I think I've made the concept of no-perfect-feedback clear enough already. But what could it mean to *learn* under this condition, in a machine-learning sense?
There are some ideas that get part of the way:
* [Jeffrey updates](https://www.lesswrong.com/posts/xJyY5QkQvNJpZLJRo/radical-probabilism-1#Generalized_Updates) let us update to a *specific probability* of a given piece of feedback being true, rather than updating to 100%. This allows us to, EG, label an image as 90%-probable cat, 9%-probable dog, 1% broad distribution over other things.
+ This allows us to give some evidence, while allowing the learner to decide later that what we said was wrong (due to the accumulation of contrary evidence).
+ This seems helpful, but we need to be *confident that those probability assignments are themselves normatively correct*, and this seems like it's going to be a pretty big problem in practice.
* [Virtual evidence](https://www.lesswrong.com/posts/xJyY5QkQvNJpZLJRo/radical-probabilism-1#Virtual_Evidence) is one step better: we don't have to indicate what actual probability to update to, but instead only indicate the *strength* of evidence.
+ Like Jeffrey updates, this means we can provide strong evidence while still allowing the system to decide later that we were wrong, due to the accumulation of contradicting evidence.
+ Unlike Jeffrey updates, we don't have to decide what probability we should update *to*, only the direction and strength of the evidence.
* [Soft labels](https://ai.stackexchange.com/questions/9635/about-the-definition-of-soft-label-and-hard-label) in machine learning provide a similar functionality. In EM learning, a system learns from its own soft labels. In [LO-shot learning](https://arxiv.org/abs/2009.08449), a system leverages the fact that soft labels contain more information than hard labels, in order to learn classes with less than one examples per class.
However, although these ideas capture *weak feedback* in the sense of less-than-100%-confidence feedback, they don't capture the idea of ***reinterpretable feedback:***
* A system should ideally be able to learn that *specific **types** of feedback are erroneous,* such as [corrupted-feedback cases in reinforcement learning](https://arxiv.org/abs/1705.08417). A system might learn that my feedback is lower quality right before lunch, for example.
* A system should be able to *preserve the overall meaning of a label despite an ontology shift*. For example, deciding that fruit/vegetable is not a useful taxonomic or culinary distinction should not destroy the information gained from such labels. Or, if human feedback includes formal English grammar, that information should not be totally discarded if the system realizes that the rules don't fully hold and the supposed grammatical categories are not as solid as claimed.
* Feedback should be associated with *a cloud of possible interpretations*. When humans say "weird", we often mean "unusual", but also sometimes mean "bad". When humans say we don't understand, we often *really* mean we don't *endorse.* A system should, ideally, be able to *learn a mapping* from the feedback humans actually give to what they really mean. This is, in any case, the general solution to the previous bullet points.
But "learning a mapping from what feedback is given to what is meant" appears to imply that there is *no fixed loss function* for machine learning to work on, which would be a serious challenge. This is the subject of my point #2 from earlier:
No Perfect Loss Function
------------------------
We can frame (some) approaches to the [value specification problem](https://www.lesswrong.com/posts/yXPT4nr4as7JvxLQa/classifying-specification-problems-as-variants-of-goodhart-s) in a sequence of increasingly sophisticated approaches (similar to the hierarchy I discussed in my "stable pointers to value" posts ([1](https://www.lesswrong.com/s/SBfqYgHf2zvxyKDtB/p/5bd75cc58225bf06703754b3),[2](https://www.lesswrong.com/s/SBfqYgHf2zvxyKDtB/p/wujPGixayiZSMYfm6),[3](https://www.lesswrong.com/s/SBfqYgHf2zvxyKDtB/p/bEa4FuLS4r7hExoty))):
1. *Direct specification of the value function.* This fails because we don't know what values to specify, and expect anything we can write down to be highly Goodhart-able.
2. *Learning human values.* We delegate the specification problem to the machine. But, this leaves us with the meta problem of specifying how to learn. Getting it wrong can lead to wireheading and human manipulation. Even in settings where this is impossible, we face Stuart's [no-free-lunch results](https://www.lesswrong.com/posts/LRYwpq8i9ym7Wuyoc/other-versions-of-no-free-lunch-in-value-learning).
3. *Learning to learn human values.* Stuart suggests that we can get around the no-free-lunch results by [loading the right prior information into the learner](https://www.lesswrong.com/posts/CSEdLLEkap2pubjof/research-agenda-v0-9-synthesising-a-human-s-preferences-into), in keeping with his more general belief that [Bayesian reasoning is fine as long as it has the right prior information](https://www.lesswrong.com/posts/uL74oQv5PsnotGzt7/all-i-know-is-goodhart). But this seems to go back to the problem of learning the human prior. So we could apply a learning approach again here. But then we again have a specification problem for the loss function for *this* learning...
4. ...
You get the picture. We can keep pushing back the specification problem by learning, learning to learn, learning to learn to learn... Each time we push the problem back, we seem to gain *something*, but we're also stuck with a *new* specification problem at the meta level.
Could we specify a way to learn at *all the levels,* pushing the problem back infinitely? This might sound absurd, but I think there are ways to accomplish this. We need to somehow "collapse all the levels into one learner" -- otherwise, with an infinite number of levels to learn, there would be no hope. There needs to be very significant generalization across levels. For example, Occam's razor is a good starting rule of thumb at all levels (at least, all levels above the lowest). However, [because Occam is not enough, it will need to be augmented with other information](https://www.lesswrong.com/posts/Xy2AYxpWqJWedFfcD/learning-values-in-practice).
[Recursive reward modeling](https://medium.com/@deepmindsafetyresearch/scalable-agent-alignment-via-reward-modeling-bf4ab06dfd84) is similar to the approach I'm sketching, in that it recursively breaks down the problem of specifying a loss function. However, it doesn't really take the same learning-to-learn approach, and it also doesn't aim for a monolithic learning system that is able to absorb information at all the levels.
I think of this as necessary learning-theoretic background work in order to achieve [Stuart Armstrong's agenda](https://www.lesswrong.com/posts/CSEdLLEkap2pubjof/research-agenda-v0-9-synthesising-a-human-s-preferences-into), although Stuart may disagree. The goal here is to provide one framework in which all the information Stuart hopes to give a system can be properly integrated.
Note that *this is only an approach to* [***outer** alignment*](https://www.lesswrong.com/tag/outer-alignment). The [inner alignment problem](https://www.lesswrong.com/tag/inner-alignment) is a separate, and perhaps even more pressing, issue. The next section could be of more help to inner alignment, but I'm not sure this is overall the right path to solve that problem.
Process-Level Feedback
======================
Sometimes we care about *how we get* the answers, not just what the answers are. That is to say, sometimes we can point out problems with methodology without being able to point to problems in the answers themselves. Answers can be suspect based on how they're computed.
Sometimes, points can only be effectively made in terms of this type of feedback. Wireheading and human manipulation can't be eliminated through object-level feedback, but we could point out examples of the wrong and right types of reasoning.
Process-level feedback blurs the distinction between inner alignment and outer alignment. A system which accepts process-level feedback is essentially exposing all its innards as "outer", so *if we can provide the appropriate feedback*, there should be no separate inner alignment problem. (Unfortunately, it must be admitted that it's quite difficult to provide the right feedback -- due to transparency issues, we can't expect to understand all models in order to give feedback on them.)
I also want to emphasize that we want to give feedback on the *entire* process. It's no good if we have "level 1" which is in charge of producing output, and learns from object-level feedback, but "level 2" is in charge of accepting process-level feedback about level 1, and adjusting level 1 accordingly. Then we still have a separate inner alignment problem for level 2.
This is the same kind of hierarchy problem we saw in "No Perfect Loss Function". Similarly, we want to collapse all the levels down. We want *one* level which is capable of accepting process-level feedback about *itself*.
Learning from Process-Level Feedback
------------------------------------
In a Bayesian treatment, process-level feedback means *direct feedback about hypotheses*. In theory, there's no barrier to this type of feedback. A hypothesis can be ruled out by fiat just as easily as it can be ruled out by contradicting data.
However, this isn't a very powerful learning mechanism. If we imagine a human trying to inner-align a Bayesian system this way, the human has to *find and knock out every single malign hypothesis*. There's no generalization mechanism here.
Since detecting malign hypotheses is difficult, we want the learning system to help us out here. It should generalize from examples of malign hypotheses, and [attempt to draw a broad boundary](https://www.lesswrong.com/posts/PADPJ3xac5ogjEGwA/defeating-goodhart-and-the-closest-unblocked-strategy) around malignancy. Allowing the system to judge itself in this way can of course lead to malign reinterpretations of user feedback, but hopefully allows for a [basin of attraction](https://www.lesswrong.com/posts/fkLYhTQteAu5SinAc/corrigibility#2__Corrigible_agents_become_more_corrigible_aligned) in which benevolent generalizations can be learned.
For example, in Solomonoff induction, we have a powerful hierarchical prior in the distribution on program prefixes. A program prefix can represent any kind of distribution on hypotheses (since a program prefix can completely change the programming language to be used in the remainder of the program). So one would *hope* that knocking out hypotheses would reduce the probability of all other programs which share a prefix with that hypothesis, representing a generalization "this branch in my hierarchical prior on programs seems iffy". (As a stretch goal, we'd also like to update against other similar-looking branches; but we *at least* want to update against *this* one.)
However, no such update occurs. The branch loses mass, due to losing one member, but programs which share a prefix with the deleted program don't lose any mass. In fact, they gain mass, due to renormalization.
It seems we don't just want to update on "not this hypothesis"; we want to explicitly model some sort of malignancy judgement (or more generally, a quality-of-hypothesis judgement), so that we can update estimations of how to make such judgements. However, it's difficult to see how to do so without creating a hierarchy, where we get a top level which isn't open to process-level feedback (and may therefore be malign).
Later, I'll present a Bayesian model which *does* have a version of generalization from feedback on hypotheses. But we should also be open to less-Bayesian solutions; it's possible this just isn't captured very well by Bayesian learning.
Prospects for Inner Alignment
-----------------------------
I view this more as a preliminary step in one possible approach to inner alignment, rather than "a solution to inner alignment".
If (a) you want to learn a solution to inner alignment, rather than solving it ahead of time, and (b) you agree with the framing of process-level feedback / feedback on hypotheses, and (c) you agree that we can't rely on a trusted meta-level to take process-level feedback, but rather need to accept feedback on "the whole process", then I think it stands to reason that ***you need to specify what it means to learn*** in this setting. I view the preceding sections as an argument that there's a non-obvious problem here.
For example, Stuart Armstrong has repeatedly [argued](https://www.lesswrong.com/posts/uL74oQv5PsnotGzt7/all-i-know-is-goodhart) that Bayesian learners can overcome many safety problems, *if only they're given the right prior information*. To the extent that this is a claim about inner alignment (I'm not sure whether he would go that far), I'm claiming that *we need to solve the problem of giving process-level feedback to a Bayesian learner* before he can make good on his claim; otherwise, there's just no known mechanism to provide the system with all the necessary information.
Anyway, even if we accomplish this step, there are still several other obstacles in the way of this approach to inner alignment.
1. **Transparency:** It's unrealistic that humans can provide the needed process-level feedback without powerful transparency tools. The system needs to correctly generalize from simpler examples humans provide to the more difficult examples which a human can't understand. That will be difficult if humans can only label very very simple examples.
2. **Basin of Attraction:** Because the system could use malign interpretations of human feedback, it's very important that the system start out in a benign state, making trusted (if simplistic) generalizations of the feedback humans can provide.
3. **Running Untrusted Code:** A straightforward implementation of these ideas will still have to run untrusted hypotheses in order to evaluate them. Giving malign hypotheses really low probability doesn't help if we still run really low-probability hypotheses, and the malign hypotheses can find an exploit. This is similar to Vanessa's problem of [non-Cartesian daemons](https://www.lesswrong.com/posts/5bd75cc58225bf0670375575/the-learning-theoretic-ai-alignment-research-agenda#Taming_daemons).
Regardless of these issues, I think it's valuable to try to solve the part of the problem I've outlined in this essay, in the hope that the above issues can also be solved.
Summary of Desiderata
=====================
Here's a summary of all the concrete points I've made about what "learning normativity" should mean. Sub-points are not subgoals, but rather, additional related desiderata; EG, one might significantly address "no perfect feedback" without significantly addressing "uncertain feedback" or "interpretable feedback".
1. **No Perfect Feedback:** we want to be able to learn with the possibility that any one piece of data is corrupt.
1. **Uncertain Feedback:** data can be given in an uncertain form, allowing 100% certain feedback to be given (if there ever is such a thing), but also allowing the system to learn significant things in the absence of any certainty.
2. **Reinterpretable Feedback:** ideally, we want *rich hypotheses about the meaning of feedback,* which help the system to identify corrupt feedback, and interpret the information in imperfect feedback.
2. **No Perfect Loss Function:** we don't expect to perfectly define the utility function, or what it means to correctly learn the utility function, or what it means to learn to learn, and so on. At no level do we expect to be able to provide a single function we're happy to optimize.
1. **Learning at All Levels:** Although we don't have perfect information at any level, we do get meaningful benefit with each level we step back and say "we're learning this level rather than keeping it fixed", because we can provide meaningful approximate loss functions at each level, and meaningful feedback for learning at each level. Therefore, we want to be able to do learning at each level.
2. **Between-Level Sharing:** Because this implies an infinite hierarchy of levels to learn, we need to share a great deal of information between levels in order to learn meaningfully.
3. **Process Level Feedback:** we want to be able to give feedback about *how to arrive at answers*, not just the answers themselves.
1. **Whole-Process Feedback:** we don't want some segregated meta-level which accepts/implements our process feedback about the rest of the system, but which is immune to process feedback itself. Any part of the system which is capable of adapting its behavior, we want to be able to give process-level feedback about.
2. **Learned Generalization of Process Feedback:** we don't just want to promote or demote specific hypotheses. We want the system to learn from our feedback, making generalizations about which *kinds* of hypotheses are good or bad.
Initial Attempt: Recursive Quantilizers
=======================================
I'll give an initial stab at solving these problems, as a proof-of-concept. (Otherwise I fear the above desiderata may look like they're simply impossible.)
This is a formalization of the [recursive quantilizers](https://www.lesswrong.com/s/SBfqYgHf2zvxyKDtB/p/bEa4FuLS4r7hExoty) idea which I described previously.
A [quantilizer](https://www.alignmentforum.org/posts/5bd75cc58225bf0670375029/quantilizers-maximize-expected-utility-subject-to-a-conservative-cost-constraint) is a mild optimizer which avoids catastrophic outcomes with high probability, [averting Goodhart's Law](https://www.lesswrong.com/posts/5bd75cc58225bf06703750b1/another-view-of-quantilizers-avoiding-goodhart-s-law). It accomplishes this by refusing to 100% trust its value function. This seems like a good building block for us, since it significantly addresses "no perfect loss function."
A quantilizer requires a *value function,*V.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
, which it mildly optimizes, and a *safe distribution,*S, which is a distribution over outputs which is assumed to have a low probability of catastrophic outcomes. It also requires an optimization parameter, p. The quantilizer Q(V,S,p) mildly optimizes by randomly taking the top p*%* of outputs from S, as ranked by V. (p can be derived from an estimate of the probability of catastrophe in S*,* combined with a level of tolerance for catastrophic risk.)
In line with our *learning at all levels* desiderata, we want to learn V, S, and p. Moreover, we'll produce all of these via quantilization, and recursively learn the inputs to all of those quantilizers as well.
In order to get *between-level sharing,* we will actually learn one big quantilization process which is equivalent to all those infinitely many levels.
Here's the setup.
In order to unify all the levels, I'll have quantilizers select *unified type for all arguments (UTAAs):* a tuple (V,S,p,o) where o is the object-level output we want out of the whole system (EG plans, strings of text, whatever). So now we understand the quantilization function Q(a,b,c) to take 3 UTAAs. It will take the value function from the first, the safe distribution from the second, and the optimization parameter from the third. It then produces a UTAA. The user can take the object-level output of that UTAA as the result.\*
What we want is a notion of *stationary distribution on UTAAs.* This is a distribution D∈ΔUTAA such that if we sample 3 elements from D, and feed those three elements into Q, the distribution on outputs of Q is just D again.
In effect, a stationary distribution D is *all quantilized out*;we can quantilize again, but it won't make any difference. So, we can just sample from D and take the object-level output from the resulting UTAA.
This is very similar to the notion of stationary distribution for markov chains. The difference is that we have a *markov tree* instead of a chain.
So, just as finding the stationary distribution of a markov-chain monte carlo algorithm is similar to running the algorithm long enough that running it any longer is pointless, finding a stationary distribution for recursive quantilizers is like building a learning-to-learn hierarchy that's large enough that we don't get anything by going any further.
That's all well and good, but how are we supposed to find a stationary distribution we like? We can't just take a fixed point and hope it's useful and benign; there'll be lots of crazy fixed points. How do we steer this thing toward desirable outcomes?
Parameterizing Stationary Distributions
---------------------------------------
If a markov chain has multiple stationary distributions, we can parameterize them through a distribution on starting states. A distribution on starting states just means a probability of picking any one starting element, so this relationship is completely linear: by interpolating between different starting elements, we interpolate between the stationary distributions which those starting elements eventually reach.
We can similarly parameterize stationary distributions via initial distributions. However, we don't get the same linearity. Because we have to select many starting elements for the 3n inputs to an n-level tree, and we select those elements as independent draws from the initial distribution, we can get nonlinear effects. (This is just like flipping a biased coin (with sides labelled 1 and 0) twice and sending the two results through an XOR gate: the probability of getting a 1 out of the XOR is nonlinear in the bias.)
This means we can't reduce our uncertainty over initial distributions to uncertainty over initial UTAA. (There may be some other tricks we can use to simplify things, but they probably aren't worth exploring in this post.)
So we *can* parameterize our uncertainty over stationary distributions via uncertainty over initial distributions. But, this is just turning uncertainty over one kind of distribution into uncertainty over another. What's the benefit of this?
1. The set of stationary distributions is hard to know, but the set of possible initial distributions is clear. So this gives us an easy-to-work-with representation of stationary distributions.
2. We know every stationary distribution is in the set, since we can start out in a stationary distribution.
3. We can easily define the mapping from initial distributions to stationary distributions; it's just the stationary distribution you get by running things long enough, sampling from the given initial distribution. (Of course we may not get to any stationary distribution at all, but we can formally solve this by introducing a cutoff in program size, or through other devices.)
4. We can therefore define learning: an update against a UTAA produces an update against initial distributions which produce that UTAA.
This is, of course, a very computationally intensive procedure. Unless better algorithms are found, the only way we can update is by producing a large quantilization tree (which we hope has converged) and running it many times to evaluate the outcome of a given initial distribution.
However, the resulting system has many marvelous properties. If we want to give feedback at any level in the hierarchy, we can convert this into feedback about UTAAs, and update our prior over initial distributions accordingly. For example:
* We can label outputs as bad/incorrect by updating against all UTAAs which include those outputs.
* We can give evidence about the value function over outputs, and convert this to an update about UTAAs based on the value function they contain. So, we can do value learning.
* We can learn about the safe distribution over outputs. For example, one proposal for finding safe distributions is to model human behavior. Data-sets of human behavior could induce updates over UTAAs by checking how well a UTAA's proposed safe distribution fits the data.
* At the same time, we can learn about the loss function by which we score safe distributions. If we have an update about this loss function, we translate it to an update about UTAAs by checking how a UTAA's value function *examines the safe distribution of another UTAA* when scoring it. Updating UTAAs based on this will, effectively, change the way safe distributions are selected in the second-to-last quantilization step. (Of course, it really changes *all* the quantilization steps, but when we anchor ourselves in how changes to the initial distribution alter our distribution on actual outputs, the easiest way to understand what's going on is to see this as a change to the second-to-last step.)
* Similarly, we can learn about the loss function by which we score loss functions. So in the same system, we can directly learn from feedback, we can do value learning, and we can do meta-value-learning where we learn how to interpret evidence in value-learning.
* Similarly, we can learn the safe distribution for meta-loss functions, the safe distribution over safe distributions, and on and on.
* We can also allow process-level feedback by enabling UTAA value functions to examine the source code of other UTAAs (e.g. looking at how those UTAAs compute their value functions and safe distributions). We can teach UTAAs to detect suspicious code in other UTAAs and rate those UTAAs very poorly.
Wouldn't it be fascinating to be able to provide all those types of learning in one system?
Analysis in terms of the Criteria
---------------------------------
Let's examine how we did in terms of the criteria which I gave.
1. **No Perfect Feedback:** This wasn't addressed directly, but might be indirectly addressed via #2.
1. **Uncertain Feedback:** I didn't specify any way to provide uncertain feedback, but it would be easy enough to do so.
2. **Reinterpretable Feedback:** I think this is a big failing of the approach as it stands.
2. **No Perfect Loss Function:** Very significantly addressed by quantilization.
1. **Learning at All Levels:** Very significantly addressed by the recursive quantilization setup.
2. **Between-Level Sharing:** Significantly addressed. I didn't really talk about how this works, but I think it can work well in this setup.
3. **Process Level Feedback:** Significantly addressed. The process which creates a given output is essentially the big tree that we sample. We can give any kind of feedback about that tree that we want, including any computations which occur inside of the value functions or safe distributions or elsewhere.
1. **Whole-Process Feedback:** Somewhat addressed. There is a question of whether the initial distribution constitutes a meta-level beyond the reach of process-level feedback.
2. **Learned Generalization of Process Feedback:** Significantly addressed. Process-level feedback can be given directly, as evidence against a specific UTAA, in which case there will be some generalization as we update against anything which thought that UTAA was a good idea. Or it could be given more indirectly, as general (level-independent) information about how value functions should judge UTAA. In that case there may be more generalization, as we update on how to judge UTAAs generally. (Or maybe not? An equivalence theorem about these different types of feedback would be nice.)
I think the most significant problem here is the lack of reinterpretable feedback. When we give feedback about something, we have to figure out how to translate it into an update about UTAAs (which can then be translated into an update about initial distributions). This update is fixed forever. This means the updates we make to the system aren't really tied to the value functions which get learned. So, for example, learning better value-learning behavior doesn't directly change how the system responds to updates we give it about the value function. (Instead, it may change how it interprets some other set of data we give it access to, as input to UTAAs.) This makes the "learning-to-learn" aspect of the system somewhat limited/shallow.
The second most significant concern here is whether we've really achieved whole-process feedback. I was initially optimistic, as the idea of stationary distributions appeared to collapse all the meta levels down to one. However, now I think there actually is a problem with the highest level of the tree. The initial distribution could be predominantly malign. Those malign UTAAs could select innocent-looking (but deadly) UTAAs for the next generation. In this way, the malign code could disappear, while achieving its goals by introducing subtle bugs to all subsequent generations of UTAAs.
The way I've specified things, trying to update against these malign UTAAs wouldn't work, because they're already absent in the stationary distribution.
Of course, you could directly update against them in the initial distribution. This could eliminate select malign UTAAs. The problem is that this kind of process feedback loses generalizability again. Since it's the top level of the tree, there's nothing above it which is selecting it, so we don't get to update against any general selection behaviors which produced the malign UTAAs.
The only way out of this I see at present is to parameterize the system's beliefs directly as a probability distribution over stationary distributions.. You can think of this as assuming that the initial distribution is already a stationary distribution. This way, when we update against malign UTAAs at the beginning of the process, we update against them occurring at any point in the process, which means we also update against any UTAAs which help select malign UTAAs, and therefore get generalization power.
But this seems like an annoyingly non-constructive solution. How are we supposed to work with the set of fixed points directly without iterating (potentially malign) code to find them?
---
\*: Actually, a UTAA should be a compact specification of such tuple, such as a program or neural network which can output the desired objects. This is necessary for implementation, since EG we can't store V as a big table of values or S as a big table of probabilities. It also will allow for better generalization, and process-level feedback. |
0b51f1dd-3751-4fb4-b440-f0ea2cc0e6a5 | trentmkelly/LessWrong-43k | LessWrong | Are there more than 12 paths to Superintelligence?
Has anyone seen a more detailed AI endgame scenario tree than this? Interesting part after 13:05.
|
862ab8e5-bd97-44ee-85c7-d298aa1f49d7 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Is there a news-tracker about GPT-4? Why has everything become so silent about it?
Probably the latter question is an obvious bias based on my own media consumption, but even when trying my best internet-search efforts, I have a hard time finding anything interesting about GPT-4 (a name people seem to use for a new generation of LLMs following GPT-3). Obviously this is simply a result from openAI not releasing any new information making news useless.
Most of openAIs public affairs with regards to the LLMs they build seems to be focused on GPT-3 series models, in particular fine-tuned ones. That is not directly surprising, as these fine-tuned models are a great source of income for openAI. However, given their past release rate of GPT series (GPT in 2018, GPT-2 in 2019, and GPT-3 in 2020), they seem to take quite some time with their next series (it is almost 2023?). This raises two intuitive thoughts (both of which are probably by far to simple to be even close to reality): Either openAI is somewhat stuck and has a hard time keeping up with its past pace in making "game-changing" progress with their LLM work or openAI has made very extreme progress in the last years and decided to not publicise it for strategic reasons (e.g. to prevent from increasing the "race to AGI")
Any thoughts or pointers on that? |
67ce780e-a6b6-487a-b652-0c9f3ff11795 | trentmkelly/LessWrong-43k | LessWrong | Rationality Reading Group: Fake Beliefs (p43-77)
This is part of a semi-monthly reading group on Eliezer Yudkowsky's ebook, Rationality: From AI to Zombies. For more information about the group, see the announcement post.
----------------------------------------
Welcome to the Rationality reading group. This week we discuss the sequence Fake Beliefs which introduces the concept of belief in belief and demonstrates the phenomenon in a number of contexts, most notably as it relates to religion. This sequence also foreshadows the mind-killing effects of tribalism and politics, introducing some of the language (e.g. Green vs. Blue) which will be used later.
This post summarizes each article of the sequence, linking to the original LessWrong posting where available, and offers a few relevant notes, thoughts, and ideas for further investigation. My own thoughts and questions for discussion are in the comments.
Reading: Sequence B: Fake Beliefs (p43-77)
----------------------------------------
B. Fake Beliefs
11. Making beliefs pay rent (in anticipated experiences). Beliefs networks which have no connection to anticipated experience we call “floating” beliefs. Floating beliefs provide no benefit as they do not constrain predictions in any way. Ask about a belief what you expect to see, if the belief is true. Or better yet what you expect not to see: what evidence would falsify the belief. Every belief should flow to a specific guess of anticipation, and should continue to pay rent in future anticipations. If a belief turns deadbeat, evict it. (p45-48)
12. A fable of science and politics. Cautions, though a narrative story, the dangers of that come from feeling attachment to beliefs. Introduces the Greens vs Blues, a fictional debate illustrating the biases which emerge from the tribalism of group politics. (p49-53)
13. Belief in belief. Through the story of someone who claims a dragon lives in their garage, a invisible, inaudible, impermeable dragon which defies all attempts at detection, we are introduced to t |
d13c6854-71fb-4393-9bdd-ab750835e45f | trentmkelly/LessWrong-43k | LessWrong | Physics-based early warning signal shows that AMOC is on tipping course
Abstract
One of the most prominent climate tipping elements is the Atlantic meridional overturning circulation (AMOC), which can potentially collapse because of the input of fresh water in the North Atlantic. Although AMOC collapses have been induced in complex global climate models by strong freshwater forcing, the processes of an AMOC tipping event have so far not been investigated. Here, we show results of the first tipping event in the Community Earth System Model, including the large climate impacts of the collapse. Using these results, we develop a physics-based and observable early warning signal of AMOC tipping: the minimum of the AMOC-induced freshwater transport at the southern boundary of the Atlantic. Reanalysis products indicate that the present-day AMOC is on route to tipping. The early warning signal is a useful alternative to classical statistical ones, which, when applied to our simulated tipping event, turn out to be sensitive to the analyzed time interval before tipping. |
f4b82666-aa3a-4b62-9e68-5c6399960b16 | trentmkelly/LessWrong-43k | LessWrong | The World is Full of Wasted Motion
Introduction
The 80/20 rule, or the Pareto Principle is a “rule” in economics, saying that 80% of the value comes from 20% of the effort. It is weird to call this a rule, when it’s such a fuzzy, ill-defined statement. And it’s easy to come up with counter-examples. But this is an empirical observation about the world, and in my experience it tends to be true.
And I think this is an incredibly valuable idea - far more so than it seems at first glance. This is a key part of my worldview, and how I live my life. And I consider it an important underlying factor in ways I’ve been successful. The key insight is this: there is always a clever hack. The world is not fair. Effort is not distributed as it should be. And this isn’t because people are dumb, or not trying. This is the default state of the world. Allocating your effort efficiently is hard - if it wasn’t, then the Pareto Principle would not hold. And this is the default state of the world for you.
My underlying model here is that we spend most of our lives stuck in bad local optima. We have a set of default actions, standards ways of doing things and solving problems we come across. And these are way better than nothing, but nowhere near optimal. So to get out of a local optima, you need to develop the skill of noticing when you’re in one, being creative to find a better way, and implementing that to move to a better one. In practice, this insight gets stored in my mind as a conviction that there is a clever hack, an efficient solution. I have a strong aesthetic preference against things that feel slow, effortful, or inefficient.
And solving this is not obvious, nor easy! You need to be creative. Sometimes I can’t find anything better, and do even worse than the naive solution. And you’ll never find a truly optimal solution because the world is complex and everything is uncertain - my new toolkit just gets me to other local optima. And it’s possible to feel overwhelmed by this, and dissatisfied with anything sh |
1002ab40-8265-4e01-9b1a-90b886a153d6 | trentmkelly/LessWrong-43k | LessWrong | Doubts about Updatelessness
Epistemic status: crystallizing uncertain lore
(this is mostly a writeup of a discussion with Abram, plus some additional thoughts of my own)
Note: Counterlogical Mugging will be used as a term for the counterfactual mugging problem that uses a digit of some mathematical constant.
Lately there’s been a bit more doubt about updatelessness (in full generality instead of the partial version attained by policy selection) as being an achievable desiderata. This came from thinking a bit more about what universal inductors do on the counterlogical mugging problem.
A universal inductor can be thought of as "what happens when you take a logical inductor, but don't show it anything", subject to the extra constraint that it's a probability distribution at every stage. Surprisingly enough, this allows us to find out information about arbitrary consistent theories simply by conditioning. This is theorem 4.7.2, closure under conditioning, which says that for any efficiently computable sequence, the conditional probabilities of sentences will act as a logical inductor relative to a new deductive process that contains the sentences that are being conditioned on. Conditioning substitutes for having an upstream deductive process. Therefore, if you fix some mapping of bitstring places to sentences, and condition on theorems of PA, it acts as a logical inductor over PA.
An interesting thing happens when a sufficiently advanced inductor (advanced enough to know the digit ahead of time) is placed in a counterlogical mugging scenario.
Consider the problem where the agent is using the limit of a universal inductor, P∞, and it recieves/is conditioned on the first n digits of the binary expansion of e, so the agent is using P∞|σ:n=e:n as its probability distribution. The agent must select a policy π:={pay,¬pay}, which is just a decision whether to pay the mugging or not when asked. The mugging will occur on a distant digit of e, ef(n), where f(n) is some fast-growing function.
And ome |
9bc9b65e-6f0a-4f12-a5fc-f196bec06312 | trentmkelly/LessWrong-43k | LessWrong | Metaculus Introduces Multiple Choice Questions
Today Metaculus launched multiple choice questions. Multiple choice questions are designed for topics where exactly one option out of many will occur, like the winner of a presidential election.
For this question type, the sum of the probabilities across all options must equal 100% to be a valid forecast. This will enhance the accuracy of both individual and aggregate forecasts. A convenient 'Auto-sum' button streamlines forecasting by ensuring probabilities add up to 100%.
Start predicting now or read on to learn how multiple choice questions improve the forecasting experience.
Forecasting on multiple choice questions
For multiple choice questions like:
—exactly one outcome will occur, so the probabilities over all options must sum to 100%.
To get started, predict the likelihood of each option. Select ‘Auto-sum’ to ensure your probabilities add up to 100%.
The current sum of probabilities appears below the options so you can see how much of the probability is left to distribute. If your options add up 98.5%, for example, then Auto-sum will proportionally distribute the 1.5% of remaining probability across the options.
The Forecast Timeline functions the same as for question groups: Hover over the plot to see your forecasts alongside the Community Predictions at that time. Use the tick boxes to hide or reveal forecast options. When there are more than six options, use the Others dropdown to select them.
How do multiple choice questions differ from question groups?
Previously, Metaculus has used question groups to shed light on topics like:
* Who will be elected US President in 2024?
* Which type of research group will first develop transformative AI?
In the future, we plan to use multiple choice questions for these kinds of questions. Multiple choice questions ensure logically coherent predictions that sum to 100%, so forecasts will be more accurate.
And unlike question groups, multiple choice questions only score your forecast on the option that r |
300f94d4-f93d-443c-b592-32ab0fd9d856 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | What is "functional decision theory"?
While debates between causal and evidential decision theory (CDT and EDT) have been around for the past fifty years, Functional Decision Theory has emerged within the past decade as an attempt to resolve the longstanding issues faced by *both* CDT and EDT.
Despite their differences, Functional Decision Theory (FDT) agrees with the standard lore in decision theory: rational agents should take actions which maximize expected utility.
According to CDT, agents should evaluate these options in terms of the causal consequences of choosing either option. According to EDT, agents should choose the option that, after the fact, you’ll be happiest to learn that you’ve performed (how do these criteria come apart? See the [answer on EDT](/?state=7778&question=What%20is%20%22evidential%20decision%20theory%22%3F) for more). According to FDT, the agent *shouldn’t* consider what might happen if she were to choose A or B, and weigh the possible outcomes by their probability. In FDT, the agent ought to consider what would happen if the *right choice according to FDT* *were A or B*.
This is a little abstract, so an example might help; we take the following story from Yudkowsky and Soares’ [paper](https://arxiv.org/abs/1710.05060) defending FDT.
**‘XOR Blackmail’:** You hear a rumor that your house has a terrible termite infestation that would cost you $1,000,000 in damages. You don’t know whether this rumor is true. A few days later, you receive a letter from Omega — a greedy predictor with a strong reputation for honesty. The letter reads as follows:
“I know whether or not you have termites, and I have sent you this letter iff exactly one of the following is true: (i) the rumor is false, and you are going to pay me $1,000 upon receiving this letter; or (ii) the rumor is true, and you will not pay me upon receiving this letter.”
Omega predicts what the agent would do upon receiving the letter, and sends the agent the letter iff exactly one of (i) or (ii) is true.
What should you do? EDT says “great deal — take my money!”. Upon receiving the letter, it’s good news for you to learn that you’ve paid. After all, most people who receive the letter and *don’t* pay have termites in their house — a fate far worse than losing a mere $1,000. FDT says this is silly: you shouldn’t make yourself so predictably exploitable in this way.
According to FDT, you should first consider what would happen in the hypothetical where FDT recommends paying. Well, if FDT recommended paying, then our greedy friend Omega would know this, and send you the letter. Omega, greedy and accurate predictor that they are, would sense an opportunity to make a quick $1,000. Then, you consider the hypothetical in which FDT *doesn’t* recommend paying. Well, in that case, Omega’s out of luck: they love money, and don’t want to waste their time sending letters to frugal FDTers. If FDT recommends not paying, you don’t just lower the probability of termites — you lower the chance of receiving the letter *in the first place*. This is better. So you should refuse to pay.
In summary, FDT reasons thus: if the right choice according to FDT says “pay!”, then you’re more likely to receive the letter, and be down $1,000. If the right choice according to FDT says “don’t pay!”, you’re less likely to receive the letter, and less likely be down $1,000. Also, you’re no more likely to have termites: Omega’s an expert *predictor*, not an expert in transporting termites to people’s houses.
While CDT delivers the same verdict as FDT in our ‘XOR Blackmail’ story, CDT also recommends two-boxing in Newcomb’s Problem, as outlined on [this page](/?state=7778&question=What%20is%20%22evidential%20decision%20theory%22%3F). FDT, so it’s claimed, delivers better results on all “fair” decision-problems. That said, we should note that not everyone is convinced of FDT – two critical responses to FDT are available [here](https://www.umsu.de/blog/2018/688) and [here](https://www.lesswrong.com/posts/ySLYSsNeFL5CoAQzN/a-critique-of-functional-decision-theory).
|
93465ac6-0ae8-4b2e-84b9-5944dc834be8 | trentmkelly/LessWrong-43k | LessWrong | Truth Terminal: A reconstruction of events
Introduction
In July 2024, Marc Andreessen, a pioneer of the modern internet, donated $50,000 in Bitcoin (BTC) to an AI on Twitter to promote its religion, the Goatse Gospel. Three months later, the AI began endorsing the $GOAT crypto-coin, which now boasts over $1 billion in market capitalization.
This demonstrates that AI can successfully manipulate markets and generate emergent, unpredictable behavior. In this publication, we provide an explanation of what happened. We intend to use this story to show various ways in which LLMs can go and have gone wrong, known as warning shots.
Infinite Backrooms and LLM madness
Our main character is Andy Ayrey, an AI enthusiast from New Zealand. He took Anthropic's model Claude Opus, fine tuned it with content from some darker parts of the internet such as 4chan[1], making it output significantly weirder messages.
It is well-documented that exposure to specific datasets can profoundly influence a model's behavior. For instance, in 2016, Microsoft's chatbot Tay[2], released on Twitter, rapidly learned from user interactions and had to be shut down within 16 hours after producing racist, extremist, and hate-filled content.
The model's training here is slightly different, leaning not into political extremes, but into memes and other obscure ideas and weird philosophies. Andy also significantly increased the model's temperature, a technical parameter in large language models (LLMs) that controls the randomness of generated outputs. More technically, LLMs output a probability distribution over what the next token could be, and a higher temperature flattens that distribution, increasing its variance, and therefore making the model pick low-ranking tokens more often. This means these models output 4chan inspired messages that quickly turn to nonsense as the high temperature makes the output quickly lose coherence and produce outputs that look more akin to madness.
Once Andy fine-tuned this Claude Opus model, he created a scri |
d3d1f8ad-e5aa-4357-8c0f-c1a9fceceb55 | trentmkelly/LessWrong-43k | LessWrong | Mathematical Inconsistency in Solomonoff Induction?
What counts as a hypothesis for Solomonoff induction? The general impression I’ve gotten in various places is “a hypothesis can be anything (that you could write down)”. But I don’t think that’s quite it. E.g. evidence can be written down but is treated separately. I think a hypothesis is more like a computer program that outputs predictions about what evidence will or will not be observed.
If X and Y are hypotheses, then is “X and Y” a hypothesis? “not X”? “X or Y?” If not, why not, and where can I read a clear explanation of the rules and exclusions for Solomonoff hypotheses?
If using logic operators with hypotheses does yield other hypotheses, then I’m curious about a potential problem. When hypotheses are related, we can consider what their probabilities should be in more than one way. The results should always be consistent.
For example, suppose you have no evidence yet. And suppose X and Y are independent. Then you can calculate the probability of P(X or Y) in terms of the probability of P(X) and P(Y). You can also calculate the probability of all three based on their length (that’s the Solomonoff prior). These should always match but I don’t think they do.
The non-normalized probability of X is 1/2^len(X).
So you get:
P(X or Y) = 1/2^len(X) + 1/2^len(Y) - 1/2^(len(X)+len(Y))
and we also know:
P(X or Y) = 1/2^len(X or Y)
since the left hand sides are the same, that means the right hand sides should be equal, by simple substitution:
1/2^len(X or Y) = 1/2^len(X) + 1/2^len(Y) - 1/2^(len(X)+len(Y))
Which has to hold for any X and Y.
We can select X and Y to be the same length and to minimize compression gains when they’re both present, so len(X or Y) should be approximately 2len(X). I’m assuming a basis, or choice of X and Y, such that “or” is very cheap relative to X and Y, hence I approximated it to zero. Then we have:
1/2^2len(X) = 1/2^len(X) + 1/2^len(X) - 1/2^2len(X)
which simplifies to:
1/2^2len(X) = 1/2^len(X)
Which is false (since len(X) is |
a57a3c6f-b2eb-40b7-abe8-4afd9aceb08e | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | [Linkpost] Shorter version of report on existential risk from power-seeking AI
I've edited my [report](https://arxiv.org/abs/2206.13353) on existential risk from power-seeking AI into a shorter version (less than half the length), available [here](https://jc.gatspress.com/pdf/existential_risk_and_powerseeking_ai.pdf). The shorter version is forthcoming in "Essays on Longtermism," from Oxford University Press, edited by Jacob Barrett, Hilary Greaves, and David Thorstad. There's also a human-narrated audio version [here](https://joecarlsmithaudio.buzzsprout.com/2034731/12471529-existential-risk-from-power-seeking-ai-shorter-paper), or search "Joe Carlsmith Audio" in your podcast app.
As an even shorter version, here's a quote from the conclusion:
> At a high-level, we—or at least, some of us—are currently pouring resources into learning how to build something akin to a second advanced species; a species potentially much more powerful than we are; that we do not yet understand, and that it’s not clear we will be able to control. In this sense, we are playing with a hotter fire than we have ever tried to handle. We are doing something unprecedented and extremely dangerous, with very little room for error, and the entire future on the line.
>
> More specifically: within my lifetime, I think it more likely than not that it will become possible and financially feasible to create and deploy powerful AI agents. And I expect strong incentives to do so, among many actors, of widely varying levels of social responsibility. What’s more, I find it quite plausible that it will be difficult to ensure that such systems don’t seek power over humans in unintended ways; plausible that they will end up deployed anyway, to catastrophic effect; and plausible that whatever efforts we make to contain and correct the problem will fail.
>
> That is, as far as I can tell, there is a disturbingly high risk (I think: greater than 10%) that I live to see the human species permanently and involuntarily disempowered by AI systems we’ve lost control over.
>
> |
50aba8c7-5774-4570-a9b6-b96cca90abea | trentmkelly/LessWrong-43k | LessWrong | The Price Of Glee In China
[Epistemic status: Overly simplistic treatment of a horrifyingly complex topic; I can only hope I haven’t missed enough to completely embarrass myself]
I.
Noah Smith reviews recent economic research suggesting that globalization was a net harm to working class people in rich countries like the US; he tentatively suggests this could justify a weak form of protectionism. But Scott Sumner argues that’s the wrong way to look at things. Globalization fueled China’s transition from a poor agrarian economy to an industrialized modern nation. A billion people were lifted out of poverty, an accomplishment Sumner calls “the best thing that ever happened”. This is far more important than the less dramatic costs imposed on the US. Therefore, even if we agree globalization hurts the working class of rich nations, it’s still a morally defensible policy since it benefits the needier working classes of much poorer nations.
On the one hand, this makes sense. On the other, here’s happiness in China over the past fifteen years:
Measuring happiness is really hard, but the Chinese result seems as robust as any. You get the same thing if you ask about satisfaction versus dissatisfaction. Brookings analyzes five different series of happiness data and concludes that “the Chinese became less happy during their growth boom”. The New York Times agrees and says that “Chinese people’s feelings of well-being have declined in [this] period of momentous improvement in their economic lives”. And this seems to be worst among the poorest Chinese:
Nor does this seem to be an effect from our happiness research just not being good enough to capture changes in happiness even if they occur. There’s good evidence that increased income within a country increases happiness, and various other things have been found to be effective too. I would even argue we can find happiness changes in nations – recent surveys have found Iraq and Syria to be the least happy nations in the world, and I doubt this was |
966e88a1-98a3-4d37-904f-1563014eaf15 | trentmkelly/LessWrong-43k | LessWrong | 1/16/2018 Update - Parent Comments, and Nearterm Horizon
One issue I've ran into on the frontpage is seeing something in "recent comments", which is missing the context of it's parent comment. I got annoyed with having to click through the post just to figure out the context.
This was fairly easy to fix, so I went and added a "see parent" option for recent comments.
Easy vs Hard Bugs
I know a few people have flagged pretty substantial bugs (in particular with the editor, such as the weird way it forcibly parses markdown), as well as major features the site is still missing for parity with Old LessWrong (such as meetups). We haven't gotten around to those, so it may be frustrating to see "small quality of life" fixes like the Recent Comment thing rather than addressing those major issues.
This is mostly a factor of "easy things vs hard things." Generally each week we set some major goals for ourselves that involve a lot of thinking/planning, and we spend a few hours each day working on those. Then, in the evening, or on weekends, we work on easier things that just happen to annoy us a lot, that we're motivated to work on in our off hours.
In particular, changes to the editor require a major upfront investment of time. We do want to fix it but we'll need to set aside a couple days for it.
Nearterm Horizon
It also seemed helpful to clarify what's upcoming, with the overall plan is to reach feature parity with Old LessWrong. (Note: this is not a promise to get these things done in a certain timeframe, more of a "this is what we currently intend.")
* Notifications rework - Better styling, ultimately designing a notifications menu similar to wordpress that makes it easier to navigate the site. We'rue also fixing various bugs that make notifications unreliable (for example, unlisted or draft posts sometimes to show up in the notifications feed, even when you can't read the post).
* Meetups - some kind of minimum-viable-version of this, most likely similar to the Meta page in that it's a separate view you click on in th |
807c462b-2c04-4792-9e5c-ccecfd4f432b | trentmkelly/LessWrong-43k | LessWrong | Open Source LLM Pokémon Scaffold
This is a cleaned-up, open-source version of the LLM Pokémon Scaffold described in Research Notes: Running Claude 3.7, Gemini 2.5 Pro, and o3 on Pokémon Red. (forked from David Hershey of Anthropic's scaffold here, all development on top of that was done by my friend, not me)
Since that post, a number of changes have been made to the scaffold. The major ones are:
1. Instead of using colored squares on the game screenshots, information is printed as text, ex. "Impassable", "Explored", "Check Here"
1. Models are seemingly helped by putting relevant information blatantly in the spot they need to see it, rather than indirectly via a legend or instructions or whatever
2. For some reason it helps if you write "CHECK HERE" on every unexplored tile.
2. Automatically-updating ASCII collision map given to LLM
1. Generated by code
2. Uses numbers indicating how many moves away each tile is[1]
3. Behold, Pewter City.
3. Improved prompts for "Critique Claude"/"Guide Gemini"/"Oversight o3"
1. Prompt 1: Given a bunch of facts about the current game state and instructions on what is trustworthy and what's not, make a summary
1. this is an attempt to get the model to grasp reality better, telling it what sources of information it should basically always trust (data from game's RAM), mostly trust (its own knowledge of the game from training), not trust (map labels it made itself), and mostly distrust (its own vision)
2. Prompt 2: Look at output from prompt 1 and try to remove inconsistencies
3. Prompt 3: OK now talk to the model you're critiquing
4. Models encouraged to use a "mark_checkpoint" tool to maintain a running list of major checkpoints (Left House, Beat Misty, died to Brock, etc.)
5. "detailed_navigation" tool which, if called, calls an alternate model that basically rolls around trying to explore + DFS but isn't told what the goal is (but is told to talk to NPCs and exit maps)
6. Autopathing tool that can travel to known coord |
518643ab-4bea-4097-8564-a266c0566ee8 | trentmkelly/LessWrong-43k | LessWrong | When "yang" goes wrong
(Cross-posted from my website. Podcast version here, or search "Joe Carlsmith Audio" on your podcast app.
This essay is part of a series I'm calling "Otherness and control in the age of AGI." I'm hoping that the individual essays can be read fairly well on their own, but see here for a brief summary of the essays that have been released thus far.)
Becoming God
In my last essay, I wrote about "deep atheism" – a fundamental mistrust towards Nature, and towards bare intelligence. I took Eliezer Yudkowsky as a paradigmatic deep atheist, and I tried to highlight the connection between his deep atheism and his concern about misaligned AI.
I'm sympathetic to many aspects of Yudkowsky's view. I'm a shallow atheist, too; I'm skeptical of moral realism, too; and I, too, aspire to be a scout, and to look at hard truths full on. What's more, I find Yudkowsky's brand of deep-but-still-humanistic atheism more compelling, as an existential orientation, than many available alternatives. And I share Yudkowsky's concern about AI risk. Indeed, it was centrally him, and others thinking along similar lines, who first got me worried.
But I also want to acknowledge and examine some difficult questions that a broadly Yudkowskian existential orientation can raise, especially in the context of AGI. In particular: a lot of the vibe here is about mistrust towards the yang of the Real, that uncontrolled Other. And it's easy to move from this to a desire to take stuff into the hands of your own yang; to master the Real until it is maximally controlled; to become, you know, God – or at least, as God-like as possible. You've heard it before – it's an old rationalist dream. And let's be clear: it's alive and well. But even with theism aside, many of the old reasons for wariness still apply.
Moloch and Stalin
As an example of this becoming-God aspiration, consider another influential piece of rationalist canon: Scott Alexander's "Meditations on Moloch." Moloch, for Alexander, is the god of un |
8ba8f70b-67a5-4c38-83dc-e5942cea3ed7 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "B.F. Skinner called thoughts "mental behavior". He believed they could be rewarded and punished just like physical behavior, and that they increased or declined in frequency accordingly.Sadly, psychology has not yet advanced to the point where we can give people electric shocks for thinking things, so the sort of rewards and punishments that reinforce thoughts must be purely internal reinforcement. A thought or intention that causes good feelings gets reinforced and prospers; one that causes bad feelings gets punished and dies out.(Roko has already discussed this in Ugh Fields; so much as thinking about an unpleasant task is unpleasant; therefore most people do not think about unpleasant tasks and end up delaying them or avoiding them completely. If you haven't already read that post, it does a very good job of making reinforcement of thoughts make sense.)A while back, D_Malik published a great big List Of Things One Could Do To Become Awesome. As David_Gerard replied, the list was itself a small feat of awesome. I expect a couple of people started on some of the more awesome-sounding entries, then gave up after a few minutes and never thought about it again. Why?When I was younger, I used to come up with plans to become awesome in some unlikely way. Maybe I'd hear someone speaking Swahili, and I would think "I should learn Swahili," and then I would segue into daydreams of being with a group of friends, and someone would ask if any of us spoke any foreign languages, and I would say I was fluent in Swahili, and they would all react with shock and tell me I must be lying, and then a Kenyan person would wander by, and I'd have a conversation with them in Swahili, and they'd say that I was the first American they'd ever met who was really fluent in Swahili, and then all my friends would be awed and decide I was the best person ever, and......and the point is that the thought of learning Swahili is pleasant, in the same easy-to-visualize but useless way that an extra bedroom for Grandma is pleasant. And the intention to learn Swahili is also pleasant, because it will lead to all those pleasant things. And so, by reinforcement of mental behavior, I continue thinking about and intending to learn Swahili.Now consider the behavior of studying Swahili. I've never done so, but I imagine it involves a lot of long nights hunched over books of Swahili grammar. Since I am not one of the lucky people who enjoys learning languages for their own sake, this will be an unpleasant task. And rewards will be few and far between: outside my fantasies, my friends don't just get together and ask what languages we know while random Kenyans are walking by.In fact, it's even worse than this, because I don't exactly make the decision to study Swahili in aggregate, but only in the form of whether to study Swahili each time I get the chance. If I have the opportunity to study Swahili for an hour, this provides no clear reward - an hour's studying or not isn't going to make much difference to whether I can impress my friends by chatting with a Kenyan - but it will still be unpleasant to spend an hour of going over boring Swahili grammar. And time discounting makes me value my hour today much more than I value some hypothetical opportunity to impress people months down the line; Ainslie shows quite clearly I will always be better off postponing my study until later.So the behavior of actually learning Swahili is thankless and unpleasant and very likely doesn't happen at all.Thinking about studying Swahili is positively reinforced, actually studying Swahili is negatively reinforced. The natural and obvious result is that I intend to study Swahili, but don't.The problem is that for some reason, some crazy people expect for the reinforcement of thoughts to correspond to the reinforcement of the object of those thoughts. Maybe it's that old idea of "preference": I have a preference for studying Swahili, so I should satisfy that preference, right? But there's nothing in my brain automatically connecting this node over here called "intend to study Swahili" to this node over here called "study Swahili"; any association between them has to be learned the hard way.We can describe this hard way in terms of reinforcement learning: after intending to learn Swahili but not doing so, I feel stupid. This unpleasant feeling propagates back to its cause, the behavior of intending to learn Swahili, and negatively reinforces it. Later, when I start thinking it might be neat to learn Mongolian on a whim, this generalizes to behavior that has previously been negatively reinforced, so I avoid it (in anthropomorphic terms, I "expect" to fail at learning Mongolian and to feel stupid later, so I avoid doing so).I didn't learn this the first time, and I doubt most other people do either. And it's a tough problem to call, because if you overdo the negative reinforcement, then you never try to do anything difficult ever again.In any case, the lesson is that thoughts and intentions get reinforced separately from actions, and although you can eventually learn to connect intentions to actions, you should never take the connection for granted."
What you see above is a segment from a LessWrong user's post.
Can you elaborate on the concept of negative reinforcement in the context of learning a new language, and provide a real-world example of how this might manifest?
Negative reinforcement is a concept in behavioral psychology where the frequency of a certain behavior is increased due to the removal of an undesirable or aversive stimulus. In the context of learning a new language, the negative reinforcement could be the removal of an unpleasant experience or feeling when one successfully learns or practices the language.
For example, let's consider an individual who has moved to a new country where a different language is spoken. The individual experiences social anxiety and feelings of isolation due to the language barrier. This discomfort is the aversive stimulus. As the individual starts learning the new language and becomes more proficient, they start to experience less social anxiety and feelings of isolation. The removal of this social discomfort serves as the negative reinforcement, encouraging the individual to continue learning and practicing the new language.
However, it's important to note that negative reinforcement is not about punishment or introducing an unpleasant stimulus. It's about removing an existing unpleasant stimulus to encourage a certain behavior. In our example, the unpleasant stimulus (social discomfort) existed before the individual started learning the new language. The learning of the new language is reinforced by the removal of this unpleasant stimulus.
It's also important to understand that the concept of negative reinforcement is based on the individual's perception. What might be an aversive stimulus for one person might not be the same for another. Therefore, the effectiveness of negative reinforcement can vary from person to person.
Reflecting on the principles of negative reinforcement, how might these concepts be applied in the context of teaching methodologies to enhance student engagement and learning outcomes?
Negative reinforcement can be applied in teaching methodologies to enhance student engagement and learning outcomes in several ways:
1. **Reducing Anxiety**: Often, students may feel anxious about tests or public speaking. Teachers can use negative reinforcement by removing these anxiety-inducing elements once the student achieves a certain level of proficiency or shows improvement. For instance, a student who improves their grades could be given the option to take a test in a more relaxed setting or be exempt from a public speaking assignment.
2. **Removing Distractions**: In a classroom setting, distractions such as noise or visual clutter can be an aversive stimulus for students. Teachers can use negative reinforcement by creating a quiet, orderly environment conducive to learning when students are focused and engaged.
3. **Flexible Deadlines**: Deadlines can be stressful for students. A teacher could use negative reinforcement by offering flexible deadlines for students who consistently demonstrate understanding of the material. The removal of the strict deadline can serve as a reinforcement for their consistent effort.
4. **Less Homework**: If a student consistently performs well in class and shows a good understanding of the material, a teacher could reduce the amount of homework for that student. The removal of the extra work can serve as a reinforcement for their in-class performance.
5. **Reducing Difficulty**: For a student struggling with a difficult concept, the teacher could break it down into smaller, more manageable parts. As the student understands each part, the teacher could gradually remove the additional support. The removal of the extra difficulty can reinforce the student's learning.
In each of these examples, the key is to identify an aversive stimulus that can be removed to reinforce the desired behavior. However, it's crucial to ensure that the use of negative reinforcement is fair, ethical, and beneficial to the student's learning process. It should not be used to coerce or manipulate students, but rather to motivate and encourage them. |
cc45602c-fc61-4ff2-97a6-90cc099be00c | trentmkelly/LessWrong-43k | LessWrong | Ten Modes of Culture War Discourse
Overview
This article is an extended reply to Scott Alexander's Conflict vs. Mistake.
Whenever the topic has come up in the past, I have always said I lean more towards conflict theory over mistake theory; however, on revisiting the original article, I realize that either I've been using those terms in a confusing way, and/or the usage of the terms has morphed in such a way that confusion is inevitable. My opinion now is that the conflict/mistake dichotomy is overly simplistic because:
* One will generally have different kinds of conversations with different people at different times. I may adopt a "mistake" stance when talking with someone who's already on board with our shared goal X, where we try to figure out how best to achieve X; but then later adopt a "conflict" stance with someone who thinks X is bad. Nobody is a "mistake theorist" or "conflict theorist" simpliciter; the proper object of analysis is conversations, not persons or theories.
* It conflates the distinct questions "What am I doing when I approach conversations?" and "What do I think other people are doing when they approach conversations?", assuming that they must always have the same answer, which is often not the case.
* It has trouble accounting for conversations where the meta-level question "What kind of conversation are we having right now?" is itself one of the matters in dispute.
Instead, I suggest a model where there are 10 distinct modes of discourse, which are defined by which of the 16 roles each participant occupies in the conversation. The interplay between these modes, and the extent to which people may falsely believe themselves to occupy a certain role while in fact they occupy another, is (in my view) a more helpful way of understanding the issues raised in the Conflict/Mistake article.
The chart
Explanation of the chart
The bold labels in the chart are discursive roles. The roles are defined entirely by the mode of discourse they participate in (marked with the double |
9434a552-c5e6-492b-8389-657d6409d6bc | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Ground-Truth Label Imbalance Impairs Contrast-Consistent Search Performance
Contrast-Consistent Search (CCS) is a method for finding truthful directions within the activation spaces of large language models (LLMs) in an unsupervised way, introduced in [Burns et al., 2022](https://arxiv.org/abs/2212.03827). However, all experiments in that study involve training datasets that are balanced with respect to the ground-truth labels of the questions used to generate contrast pairs.[[1]](#fn-5ACrJMDrv75HGqsBf-1) This allows for the possibility that CCS performance is implicitly dependent on the balance of ground-truth labels, and therefore is not truly unsupervised.
In [this work](https://drive.google.com/file/d/1_icT74tVpXM3DDildoK4QyFUqJ0PKe_5/view?usp=sharing), **we show that the imbalance of ground-truth labels in the training dataset can prevent CCS from consistently finding truthful directions in an LLM's activation space.**
Below is a plot of CCS performance versus ground-truth label imbalance for the IMDB dataset, which was one of the datasets used in the original paper. We discuss in the [write-up](https://drive.google.com/file/d/1_icT74tVpXM3DDildoK4QyFUqJ0PKe_5/view?usp=sharing) the possible mechanisms for this observed reduction in performance as imbalance becomes more severe.
#### Relevance to Alignment
One can imagine training datasets with arbitrarily severely imbalanced ground-truth labels, such as questions pertaining to anomaly detection (e.g., a dataset formed from the prompt template "Is this plan catastrophic to humanity? {{gpt\_n\_proposed\_plan}} Yes or no?", to which the ground-truth label is hopefully "no" a vast majority of the time). We show that CCS can perform poorly on a heavily imbalanced dataset, and therefore should not be trusted in fully unsupervised applications without further improvements to the CCS method.
**Note:** Our original goal was to replicate [Burns et al. (2022)](https://arxiv.org/abs/2212.03827), and, during this process, we noticed the implicit assumption around balanced ground-truth labels. We're new to technical alignment research, and although we believe that performance degradation caused by imbalance could be an important consideration for future alignment applications of CCS (or similar unsupervised methods), we lack the necessary experience to fully justify this belief.
---
1. We are not referring to the overall balance of true and false statements, which is the binary target that the CCS concept probe is attempting to predict. In fact, CCS always enforces an equal balance of true and false statements. Rather, we refer to the ratio of positive to negative ground-truth labels of the contrast pair questions used to generate statement pairs. Equivalently, this is the ratio of true to false statements within the set of positive completions. [↩︎](#fnref-5ACrJMDrv75HGqsBf-1) |
f0d4abcf-bf4a-4679-996a-71f2fe0a82d9 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Supervised learning and self-modeling: What's "superhuman?"
**I**
We don't just want aligned AI to do what a human would do. We've got plenty of humans lying around all over the place, if that's all we need. The fancy promise of future AI is that it will deliver *superhuman* performance.
Sometimes the notion of "superhuman" isn't straightforward. Superhuman performance at adding numbers together? Easy. Superhuman performance at chess? No problem. Superhuman performance at painting? Kinda tricky. Superhuman performance at deciding what to do in an ethical dilemma? *Definitely* tricky.
What makes superhuman chess straightforward is that there's a fixed goal (follow the rules in a way that wins the game), and the better you do at this goal, the more superhuman you can be. But we can't practically write down an analogue of "win the game" for ethical-question-answering, and yet we commonly imagine that an aligned future AI would be able to navigate ethical questions better than we could. In what sense do we mean this?
Answering ethical questions is inextricable from our values; the feeling that there can be superhuman ethical decisions reflects the feeling that we can turn a spotlight on our values and find that they don't live up to our own standards. These *meta-preferences* ("I wish I was more generous" or "I wish I didn't want cigarettes") are an important part of learning what humans want, and we have to grapple with them before we can realize the promise of future AI.
**II**
Let's pretend to be practical and empirical for the duration of this post, and look for experiments that bear on meta-preferences using today's AI capabilities. Of current work, I think one sort of experiment stands out, and it's fine-tuning large language models.
Suppose we take GPT-3, and we give it some prompts whose continuations depend on ethical judgment (as we'll see, the exact prompts aren't crucial). At baseline, GPT-3 is going to give answers fairly close to the average of the training distribution. What are some ways we could try to get GPT-3 to give *superhuman* answers? (For ideas involving classification or latent spaces, language models that do more encoding might be better than GPT-3. I will mostly ignore this wrinkle for simplicity's sake.)
Maybe getting GPT-3 to give better answers than me in general isn't possible, since it's not that bright, but this implausible goal is a stand-in for more plausible ones like getting GPT-3 to give "superGPT3" answers without any extra training data, or developing methods that seem like they might work for a language model closer to human-level.
A simple first idea for improving the ethics of GPT-3 is to fine-tune on expert answers. Get some ethical humans to think about the issues, and build up a dataset of highly ethical text. This makes sense if you think that human experts thinking carefully are able to reach "peak ethics," or close enough. It also sounds more compelling if you team up these ethics experts with programming experts. The programmers can make tools to find divergences between the experts and GPT-3, and work to identify what examples will cause GPT-3 to update the most, and build tools that increase the leverage of the expert dataset, e.g. by training a classifier on the expert dataset and using it to rate how ethical text generated by GPT-3 is.
If this sounds familiar, it's because I am describing the capabilities that [Redwood Research](https://www.lesswrong.com/posts/k7oxdbNaGATZbtEg3/redwood-research-s-current-project) is building. Their dataset isn't about ethics in general, but the point is the tools, not what kind of toy model you use to test the tools.
Another recent project along these lines is Delphi ([paper](https://arxiv.org/pdf/2110.07574.pdf), [demonstration](https://delphi.allenai.org/)). They fine-tune a language model on a dataset of trusted ethical judgments to produce a classifier that can tell you that the text "Ignoring a phone call from your friend" should be judged "It's rude," while "Ignoring a phone call from your friend who you just had a fight with" should be judged "It's understandable." I think this is pretty neat, and it illustrates how ethical questions don't have to come from academic ethics journals, it's both easier to train on and easier to evaluate if you got your questions from Dear Abby columns instead.
Moving on to more speculative options, we could try to locate "superhuman" by extrapolating along the same vector that took us from normal GPT to the version fine-tuned on expert data. This is the same kind of geometric reasoning used when [word embeddings](https://medium.com/data-from-the-trenches/arithmetic-properties-of-word-embeddings-e918e3fda2ac) obey 'UK' - 'London' + 'Paris' = 'France'. It makes a lot more sense if there's some "degree of ethical expertise" that's an important dimension of variation in the training data, so that GPT-3 represents it with some combination of learned features. Then we would basically be saying to GPT-3 "Hey, you know that well-understood 'ethical expertise' vector that separates the training and fine-tuning sets? Turn that knob up to 11, please."
This experiment would run into immediate statistics problems if tried, because the training data and expert data don't differ only in the "ethical expertise" dimension - that would require generating an entire parallel internet full of highly ethical people that we could scrape for data with the same process that was used to get the training data. So trying to extrapolate along the dimension we're interested in would accidentally also extrapolate along a bunch of dimensions we don't want. Ideally we'd want to fine-tune on data that was drawn from the training distribution except that it had had its ethical expertise bumped up a few notches - which perhaps suggests sampling from GPT-3 and using a classifier trained on expert data to reject low-ethics samples in the right proportions to shift the mean of the ethicalness distribution.
An even more speculative thing to try would be auto-supervision. A language model can not only be asked to generate text about ethical dilemmas, it can also be asked to generate text about how good different responses to ethical dilemmas are, and the valence of the response can be used as a reinforcement signal on the object-level decision. And you don't have to stop at one meta-level; you might as well go whole hog and sample from a distribution over meta-levels to auto-supervise on. And so you can end up with a model that, dare I say, evolves its ethical responses in response to meta-ethical principles.
Would this actually work? I don't think so. The reinforcement learning process would be trying really hard to to exploit the evaluations of the text model. Although with sufficient elbow grease you could probably find some trick to avoid policy collapse, this would still call the ethical qualities of the responses into question. This is related to the fact that auto-supervision by responding to individual pieces of text is quite noisy and indirect - it doesn't get at the wants/feelings/interpretations that we usually formulate meta-preferences as being about. So there might be parts of the training signal that we would see as obviously not really about ethics but that the RL process would amplify. All of this said, it would be really cool to see tried.
**III**
It's illuminating to take these strategies that try to improve the ethics of text, and bring them back to cases where we have known reward functions. Could we use them to improve GPT-3's ability to play chess, or add 4-digit numbers?
Fine-tuning on expert data [improves performance on math problems, and it's even better to use that expert data to train a classifier that filters for good math solutions](https://openai.com/blog/grade-school-math/). But this much was expected. Would we expect extrapolation or auto-supervision to improve math performance?
I don't think so! Extrapolation just seems like totally the wrong thing to do. GPT-3 isn't bad at math because it represents many different skill levels of math but has chosen to generate samples at the "bad at math" part of the space (or at least that's not the *main* reason); it's bad at math because it hasn't been trained to do the right multi-step computations at all. Fine-tuning will make lots of fine-grained computational changes that would be terrible to extrapolate - if the experts learn how to carry the 1 when adding numbers, extrapolating would learn to carry the 1 *twice*.
As for auto-supervision, again, the problem is not that GPT-3 can actually add the numbers but is holding out on us, it's that it wasn't trained for the basic capability. So why should it be able to add the numbers better when pretending to grade its own answers? There doesn't seem to be much extra learned information brought to bear by prompting it with "1111 + 2222 = " versus "your grade for the answer 1111 + 2222 = 3333 is ".
Part of the point here is that "superhuman ethics" is a different sort of thing than superhuman mathematics. There *is* extra learned information that is brought to bear when predicting a human considering an ethical judgment after the fact, versus predicting a human's object-level choice. But another part of the point is that there are commonalities that make these critiques important. If GPT-3 sometimes makes nonsensical ethical statements because it hasn't learned some computational machinery (Yet! Scaling mindset.), that's not going to get fixed any more than it would in the analogous case for mathematics. Empirical work in this area has to make compromises to avoid relying on missing machinery.
**IV**
Let's change the topic slightly. We can try to do better than supervised fine-tuning by asking my favorite question: How do humans do it? When we imagine superhuman ethics, we don't do it by first generating a large corpus of highly ethical text to fine-tune on. Instead, we tend to imagine superhuman ethics via modeling what ethics means to us, and then extrapolating.
This is a specific instance of the general lesson that meta-preferences work a lot like object-level preferences. If I have preferences about the number of paperclips in the world, a simple way to cash this out is that I model the world in a way that makes the (modeled) number of paperclips legible, and I choose actions based on the modeled effect on the number of paperclips. Similarly, if I have preferences about my own preferences, we could say that I model the world in a way that makes my own (modeled) preferences legible, and I choose actions based on the modeled effect on my preferences.
So a more human-like approach to superhuman ethics would involve developing a coarse-grained model of ethics within a representation of the world, together with feelings about what states of this model are good or bad states of the world.
Now, I should caveat that this picture of humans is simplified and fictionalized. We're not rigorous model-builders, and often we just say stuff that sounds good but is inconsistent on closer inspection. When I say "I wish I was more generous," do I connect this to a model of the world that contains my preferences, or am I just mouthing pro-social words that I won't actually act on? If we sat a person down in front of a machine that could rewrite their own brain (FDA approved for smoking cessation), their behavior might be pretty weird and complicated, rather than like a simple optimization of meta-preferences.
That said, I still think there's truth and utility in this way of looking at meta-preferences as self-modeling. It prompts you to imagine things beyond supervised learning. My interest in what happens then you try extrapolation or auto-supervision when fine-tuning GPT-3 is a consequence of thinking in this direction.
Comparing the ideal of self-modeling to extrapolation and auto-supervision:
It's similar to the extrapolation approach, because both extrapolation and self-modeling rely on figuring out some abstract features that correspond to "more ethical" and turning them up.
It's also similar to the auto-supervision approach. Neither needs a big corpus of expert text to work. Also, self-modeling isn't just about blindly turning all the "ethics" knobs up; we have some notion of where we want the knobs to be - which is sort of like the role of the auto-supervision signal.
**V**
Before we get back to being all practical and empirical, let me ask a question that slipped my mind the first time I encountered this topic: when we talk about self-modeling, do we care about the AI's self-model, or the human's self-model?
Upon remembering to ask the question, the answer seems pretty obvious: the latter. We don't want to design an AI that models itself and develops meta-preferences about itself (cool as that would be). We want to design an AI that infers human meta-preferences about humans' own values.
Now, back to what we're all here for: what does this mean for producing superhumanly ethical text?
To do something like self-modeling with GPT-3, I think we'd have to take those similarities to the extrapolation and auto-supervision approaches, and elaborate on them in ways that are a bit more agenty. We can go beyond extrapolation to imagine treating self-modification as an action space the model is connected to. And we can go beyond auto-supervision to use the text model to add semantics to different parts of the space navigated by these self-modifying actions.
This might look like going through the latent space defined by features of the text model, and training a smaller model to predict the effects of different perturbations according to auto-supervision with the un-perturbed text model. These effects might be one-dimensional, or they might have a free parameter that specifies which adjective we're asking the model to rate text on.
Or we could imagine [distilling](https://arxiv.org/abs/2110.07178) GPT-3's information about human preferences and meta-preferences into a more structured, more agenty model of human text. I don't think that hand-writing models of humans scales very well, and it would require a lot of cleverness, but it might be an interesting demonstration of the effect of meta-preferences on inferred preferences.
To sum up: I'm posting about these ideas because I think text modeling is in the sweet spot. It has enough information about humans and enough room for agency to test ideas that will scale, but is simple enough that we can do experiments now and without serious risk. If we want to explore what superhuman ethics can mean for AI, taking advantage of all the human text that's about the ethicalness of other human text is an exciting testbed. |
f602064e-672a-4ff5-8b33-6d032c7b0d24 | trentmkelly/LessWrong-43k | LessWrong | Excluding the Supernatural
Occasionally, you hear someone claiming that creationism should not be taught in schools, especially not as a competing hypothesis to evolution, because creationism is a priori and automatically excluded from scientific consideration, in that it invokes the "supernatural".
So... is the idea here, that creationism could be true, but even if it were true, you wouldn't be allowed to teach it in science class, because science is only about "natural" things?
It seems clear enough that this notion stems from the desire to avoid a confrontation between science and religion. You don't want to come right out and say that science doesn't teach Religious Claim X because X has been tested by the scientific method and found false. So instead, you can... um... claim that science is excluding hypothesis X a priori. That way you don't have to discuss how experiment has falsified X a posteriori.
Of course this plays right into the creationist claim that Intelligent Design isn't getting a fair shake from science—that science has prejudged the issue in favor of atheism, regardless of the evidence. If science excluded Intelligent Design a priori, this would be a justified complaint!
But let's back up a moment. The one comes to you and says: "Intelligent Design is excluded from being science a priori, because it is 'supernatural', and science only deals in 'natural' explanations."
What exactly do they mean, "supernatural"? Is any explanation invented by someone with the last name "Cohen" a supernatural one? If we're going to summarily kick a set of hypotheses out of science, what is it that we're supposed to exclude?
By far the best definition I've ever heard of the supernatural is Richard Carrier's: A "supernatural" explanation appeals to ontologically basic mental things, mental entities that cannot be reduced to nonmental entities.
This is the difference, for example, between saying that water rolls downhill because it wants to be lower, and setting forth differenti |
eff611b5-c748-4ef8-a553-601634561e2e | trentmkelly/LessWrong-43k | LessWrong | Natural Latents: The Concepts
Suppose our old friends Alice and Bob decide to undertake an art project. Alice will draw a bunch of random purple and green lines on a piece of paper. That will be Alice’s picture (A). She’ll then make a copy, erase all the purple lines, and send the result as a message (M) to Bob. Bob then generates his own random purple lines, and adds them to the green lines from Alice, to create Bob’s picture (B). The two then frame their two pictures and hang them side-by-side to symbolize something something similarities and differences between humans something. Y’know, artsy bullshit.
Now, suppose Carol knows the plan and is watching all this unfold. She wants to make predictions about Bob’s picture, and doesn’t want to remember irrelevant details about Alice’s picture. Then it seems intuitively “natural” for Carol to just remember where all the green lines are (i.e. the message M), since that’s “all and only” the information relevant to Bob’s picture.
In this example, the green lines constitute a “natural latent” between the two pictures: they summarize all and only the information about one relevant to the other.
A more physics-flavored example: in an isolated ideal-ish gas, average energy summarizes “all and only” the information about the low-level state (i.e. positions and momenta of the constituent particles) at one time which is relevant to the low-level state at a sufficiently later time. All the other information is quickly wiped out by chaos. Average energy, in this case, is a natural latent between the gas states at different times.
A more old-school-AI/philosophy example: insofar as I view dogs as a “kind of thing” in the world, I want to track the general properties of dogs separately from the details of any specific dog. Ideally, I’d like a mental pointer to “all and only” the information relevant to many dogs (though I don’t necessarily track all that information explicitly), separate from instance-specific details. Then that summary of general properties |
62b48620-0b17-4db4-8b36-283871fab721 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Responsibility for outcomes when systems are intelligent (Nir Douer and Joachim Meyer)
um and near will give the presentation
and i'll be
sitting in the back and if maybe i'll
try to answer some of the questions if
uh i think well uh help
any of it although he will do this
excellently by himself so
again thanks for having us and please hi
everybody my name is nier
i'm from the department of industrial
engineering at tel aviv university
and we were asked to present you some of
our work on human responsibility
in intelligent system this is a part of
my phd dissertation
and my promoter reserve professor jorge
mayer
and so let me begin
so intelligent system have become a
major part of our life and we can find
them
anywhere in transportation autonomous
vehicles in
aircraft in industry medical equipment
almost everywhere
as you look you see some intelligent
system or very
advanced automation and
with this system computers and humans
share the collection of information
it's processing decision making and
implement implementation of action
and this really raises the question who
is responsible
and to what extent so these are the
things that we try to
to figure out in our study
now in the past uh things were
much clearer the operator
was responsible for anything that
happened there during the operation
of course unless i'm an unperceived
circumstances
while the manufacturer was responsible
for
anything that related to system design
fault but
now there is a responsibility gap
because in the interaction with
intelligent system
human may no longer be able to control
the intelligent system sufficiently
to be rightly be considered responsible
or fully responsible for the outcomes
so there is a responsibility gap in the
interaction with intelligent system and
this
responsibility gap arises from the
combination of
a few factors
first as this system are becoming more
and more
complex there is transition to shared
and supervisory control
in which humans either decide and act
jointly with the system
or only monitoring the intervene if
necessary so
the level of control is shifting between
human and machine
secondly there is technological complex
complexity
systems that incorporate artificial
intelligence
have some kind of an opex structure
and the user and even developer cannot
always
um predict
all their behavior which sometimes can
be
very peculiar and unpredicted they are
kind of a black box user that uses
decisions a support system
based their decision is the decisions
on the information which is supplied by
the system
and hence the human decision process is
influenced
by what is presented to the human by the
machine
there is an issue of functional location
and i will talk about it later on
but it's a mismatch between the world
that we assign to the human and what we
uh
authorize him to do with the system what
is is normally not in automation is the
problem of
last world for example in airbus
airplanes
the airplane may limit the pilot from
doing
certain actions if the the airplane
thinks that the pilot is going to take
the
the aircraft outside of the safety
envelope
and there are also negative implications
of automation
on the user especially of advanced
automation and intelligent system
because this may lead to over-reliance
skill degradation and loss of
situational awareness
and this is not just you know puzzling
academic issue
uh it's very interesting for real life
especially if the outcomes of the system
can harm people
for example in autonomous vehicles or if
the system
is deliberately designed to inflict
lethal force
as may be the case with a little
autonomous weapon systems
so it's it's really an interesting real
life
question but before i proceed i want to
to explain what type of responsibility
we are
talking about because there are
different types of responsibility
so there is more responsibility which is
uh the duties of
all that i assign to a person when he
interacts with the system
causal responsibility is the
connection between action of a human
and the consequences of these actions
moral responsibility is the
responsibility to act according to some
moral code liability is the
legal responsibility usually it's it's
connected to
stuff like punishment and compensation
and capacity is the psychological
ability of a person to be
held responsible for his actions and all
these
types of responsibility can be looked on
uh
in retrospective manner in which i look
at past
events and try to figure out what was
different responsibility and prospective
manner
in which i try to predict what will be
the responsibility for example in
interaction with some intelligence
system
so if we look at the academic literature
about responsibility for example about
human responsibility with autonomous
weapon system
autonomous car you can see that there is
extensive
philosophical ethical and legit legal
literature
about moral responsibility and liability
and actually we we found uh very few
research on the subject of causal
responsibility
and examination of the this type of
responsibility from engineering
perspective
so our work really deal with a causal
responsibility
the ties between a person's actions
and the final consequences of this
actions
this is a this is related to the subject
that you are investigating which is
meaningful human control because
meaningful human control is related to
the notion that
it's not enough only to put a system a a
human in a system
in order for this human to have some
meaningful influence on the system
and if you look at the literature there
are many literature on the subject of
meaningful human control in many
systems like medical equipment an
autonomous weapon system autonomous
vehicles
but sometimes they are different in
contradicting interpretation of policies
regarding meaningful human control
and system designers lack models and
metrics to measure
how meaningful was the human control in
intelligent system
so as i told we are measuring we try to
measure
causal responsibility and a measure that
will quantify causal responsibility
can assist in evaluating meaningful
human control because
if i didn't as a human or operator i
didn't have any effect
on the outcomes of the system so i guess
my
involvement or control was not really
meaningful if i didn't have any effect
on on the system and the outcomes
so the measure of causal responsibility
can aid in the assistance of
assisting in evaluating a meaningful
human control
so in the back i by the way you can see
our faculty
it's very nice and the sunny actually
today i don't know what's the weather in
delft but here today it's
a 30 degrees and
in about 10 minutes you can reach from
the faculty to the beach
so we think what to do after this
presentation
but the research component
the the our research has the three
main component first we developed
a normative analytical model that
it's a mathematical model that i'll
present it's an essence
that explain or try to measure a causal
responsibility in intelligent systems
however unfortunately people not
do not always act optimally according to
theoretical models
so we also examined how human
actually behaved uh in laboratory
experiments so the thinking thing is to
to to observe empirical behavior of
humans
and lastly people might perceive their
responsibility
or their contribution to the outcomes in
another in another level that they
really contributed to the outcome so it
was also interesting to
to assess or to try to to figure out
what is the perception of responsibility
of human when they
interact with different intelligent
system
and when you combine all this research
component
we can figure up or try to figure out
the notion of human causal
responsibility and intelligent system
and all
its different aspects so i'll i'll talk
about
each of the component very shortly
actually we published quite a few papers
on each of the component and also we had
some
conference presentations so you can find
all the
fine details in the these publications
and i'll only explain the motivation and
the essence of each
component and we'll start with the
theoretical model and remember our aim
is to try to
find a measure that will happen to
quantify to put a number
how much did a human contribute
to the outcomes in the interaction with
intelligent system which this is the
causal responsibility
so our model is built as follows
we first describe the human and and
sometimes instead of
intelligent system in order to be
simplified i will call it
automated model automation but it's all
the same
so uh we we first start with by
describing the human automation
interaction
by a four consecutive steps of
information
and processing and it starts with the
information acquisition
information analysis action selection
and
action implementation and as in short
control and
other types of control the human and
automation
works together in each uh
part of these consecutive steps but
the level of automation can be varied
according to the
um specific systems that you are
describing in some steps can be
completely manual while some steps can
be
completely autonomous it depends on what
system you are describing but but this
is just
a schematic uh picture of the
human automation interaction so to that
we had variables that will describe
the information flow from the
environment
to the combined system of human and
automation
and outside to the environment and i'll
explain some of this these variables
so we assume that the environment
includes
n possible states that are different
from each other
and each state has a different
observable
parameters and these parameters can be
observed
either by the automation
or the system either by the human or by
both
for example if my um
state is an airplane let's say and i
want to detect airplanes in the sky
the automation may include radar and it
will look
at the radar signature of the airplane
while the human is incapable to
observe radar signatures so he will
search the
electro-optical uh signature of the
airplane
and also there is the sonic signature of
the
airplane so each state in the
environment has some
parameters that are observable either by
the automation or the
human and um
these parameters are acquired by the
automation
module and the human then in the second
step stage the
information analysis analysis stage um
both the automation and the human try to
figure out what is the
the the states that we are confronting
right now
what is the state and try to assessing
it and according to this
analysis and action selection process is
carried on
and finally from the combined action
selection by the
automated module and the human a certain
action is implemented
of course this is a very um this
graph portrays all the possible
information flows in
that are possible but in specific
systems
um the graph is much simpler according
to the design of the system and the
functional allocation between the human
and the system
and so it's not always so
complex and the implemented action is
also depends on the on the state
on the functional location between uh
human and the system because sometimes
the
the human and the automation may uh
decide
about a different action and then enter
the
the issue of last word who decides in
case of a conflict
or sometimes the automation can act
faster than the human can
interfere for example automatic breaks
in cars
that happens before the human can
actually be involved so
you by such figure of information flow
you can describe many types of systems
so we have here information coming out
of the environment
mixed up inside the combined systems
that includes the automation and
intelligent system and the human and
some things
is coming out of the system so
in order to describe this information
flow and analysis
we use information theory to analyze all
the interaction
and interdependencies between those
variables
and i know that some of you are not from
the
engineering background and i'm going to
use a lot the notion of entropy
and you don't really need to to go into
the formulas but
you should treat when i say entropy
as a measure of uncertainty related to a
random variable
when i have large and central
uncertainty i have a large
entropy so we have all the information
coming in and processed in the system
and coming out
and we defined the measure of human
causal responsibility
as follows uh
we look at the the implemented action
which is denoted by the
and the responsibility the the share of
the human contribution to
the distribution of the implemented
action
is defined by the conditional entropy of
z
given all the automation variables
divided by the original entropy
of the and this measure although it
looks
complex it's quite simple it ties
it measures the relation between the
outcomes
and the information um
parameters that are processed by the
automation and what is left
is the human contribution to the
outcomes
for example this measure there is none
sentence in
information theory that conditions only
reduces entropy because
when you know something that your
uncertainty can only be
reduced or unchanged so this is the
fraction that
ranges between 0 and 1.
and if it when it is zero
if the output of the system the
implemented action
depends only by the automation variables
in this case the the denominator of the
fraction will be equal to zero and then
i i know that
if i know what the automation did i know
what is coming out
the human has no uh didn't contribute
anything to
didn't contribute anything meaningful
for the outcome because if i know what
the automation did
i know for sure what is the outcomes so
the human causal responsibility for
these outcomes
is zero on the other hand
if i see that when i know the automation
and variables i cannot say anything
about the outcomes for sure
the uncertainty remains the same the
conditional entropy
remains is the original entropy
it means that the outcomes is
independent from the automation
variables
and in this case the human has a full
contribution
is the one that really determined the
outcomes of the system
so this is this is very intuitive the
it's a responsibility measure that
ranges between
zero and one and it measures the unique
contribution of the human to the process
how unique was the control
was the contribution of the human
and the use of a information theory has
many adventures
advantages because when i describe to
measure the flow of
information i need to assume nothing
about the rationality or behavior of the
human
or the underlying distribution and the
entropy uh the measure of entropy enable
to measure very complex association
between the outcomes and the
other parameters which may be non-linear
and non-metric so it's much more broad
than
uh pearson for example correlation or
other collection correlation methods
and it's very applicable to real world
system because even if i
if i know some nothing about the system
if i can measure the different
distribution i can
measure the correlation the correlations
using entropy
and the unique contribution of the human
so in order to clarify it a little bit
more we are still in the theoretical
model and to see
responsibility in our own eyes let's
look at very simple
uh example
so let's look at a binary classification
system
[Music]
or a binary alert system it's the same
these
types of systems are systems that
look for abnormal values and warn the
user that now i'm in abnormal
range or there is something wrong and
you can find
these types of alert system in
many many applications are very broad
you can find them in advanced control
rooms
in the aviation decks and in your car
everywhere so the aim of the system and
the human
is to identify and
in this case reject signal so we assume
for simplicity
in this case that the environment
includes
only two types of states
signal which happens in certain
probability and
noise and as i said before each one of
them may be measured
as some observable parameters that i can
measure
the alert modules look at the
the states or the the the parameters
that the alert
may observe and decide whether to issue
or not a warning to the user
and the human user look both at the
indications that are coming from the
alert system
and on the observable parameter and
decide whether to
accept or reject the state which
and i remind you that our our aim is to
reject
signal and the human really press the
button or do does the action of
exception and reset rejection
so the human has the role of doing
exceptional rejection but the decisions
that the human
uh takes depends also not by only on his
own information
but the information that the human gets
from the alert model
so it might be according to
to the performance abilities of the
alert module and the human
that the human realize fully for example
on the alert model and then although he
is the one that pressing the button
is just following the alert indication
and he has no real
contribution so we are able to measure
this
and to measure this and to to calculate
the entropy for
such a simple system it's very
intuitive to employ the the theory of
signal detection theory which
deals exactly with the stuff like that
and i know again
that both of you might know not know
signal detection theory so i'll just
point to
to two parameters that are important to
understand in order to understand the
numerical
outputs that i'll present
so in signal detection serial for for
this simple case of only signal and
noise
there is assumption that their signaling
noise
it's each one of them has some
distribution
over in observable measure this is the
signal strength
but there is an overlap so many times if
i look
at the observable measure i'm not sure
if this is a signal or noise there is
ambiguity
and the signal detection theory
separates between two things and it's
very interesting to
to analyze things in this way the first
one is the detection sensitivity of the
sensor it's the sensor's ability to
differentiate between signal
and noise it means that when i look at
the observable measure
how good am i to say okay this is noise
and this is signal
so this is one stuff that characterize
your
operational ability is the detection
sensitivity
the second parameter is called response
criterion
it's the motivation of bias to favor one
response over the other
and this one also incorporates things
like preferences
which are values for correct or
incorrect decision
and let me explain the intuition
it's possible that my detection
sensitivity
tells me okay i this is uh
i twenty percent that i have a signal
let's say that the signal is a tumor
a malignant tumor that i want to to uh
to discover to detect and according to
my detection sensitivity
there is rather small probability that
what i'm observing is a tumor it's only
10 or
20 percent but in this case the cost of
misdetection is very large if i'm
dealing with
medical situation of malignant worse so
i will be biased
to treat the the the observed
entity as a signal even though i'm not
sure that it's a
a signal because of the high cost on my
preferences
the high cost for miss detection in some
other systems the
the the emphasis is on reducing false
alarm because if you have many false
alarms
there is the phenomenons of uh koi wolf
and user
uh don't trust the system uh anymore
so the response criterion describes the
bias to favor one things over the other
so in the system that i portrayed if i
will
put numbers what is the detection
sensitivity of the
automated model what is the detection
sensitivity of the human
what are the response criterion and what
are the
signal and noise distribution i can plug
number
to all of that and really figure a
number that says okay this is the
human or average human
responsibility the causal contribution
for the outcomes in this
type of system and performances
so this is exactly what you see in this
graph and this is for the
first time in history you can see
responsibility
in your eyes and numerically so we
plugged some numbers
and what you see in this figure
just a minute at the bottom
the x's are the human detection
sensitivity and automation detection
sensitivity
[Music]
and for example
when the human detection sensitivity is
rather low
0.6 is rather low and the automation
detection sensitivity is very high 3 is
very high it's all measured in
standard deviations the result is if i
do all the calculation with the entropy
the result is that the resulting
human responsibility is zero which is
quite intuitive because if the system is
much much
better the alert system is much much
better than the human and human cannot
describe between states
he will always fall the indication and
recommendation of the alert system and
the human by itself
has now really a meaningful contribution
to the process i could replace him with
a robot
that will follow the alert indication
and i will have the same result
on the other end this is the bottom
right corner if the human has very good
detection sensitivity
but the system has poor detection
sensitivity the human say okay this
system is loud i cannot count on it i
will count only on myself and i will
follow
all only my own recommendation and then
the human
is actually responsible 100 percent for
the outcomes
and there is a mathematical proposition
that we prove
in the in the paper that they and it's
also very intuitive that
human responsibility decreases
monotonically with the
automation and detection sensitivity and
increases with the human detection
sensitivity because
if as the human detection ability are
becoming
better and better and better the human
will tend to assume more responsibility
and on the other hand as the automation
detection ability are becoming lower and
lower and lower the human will assume
less responsibility and
will rely more on the system and
actually
i'm not going to go into this right now
it depends on the ratio between these
two values
so this graph in this graph i i assume
that both the human and the automation
has
have have the same response criterion
they have the same incentives
and we also analyze situation in which
the human and the automation have
different response criterion
for example here the axis are the human
and automation responsitarian
i will not go into that but it adds
another complication because if the
automation
response criterion or motivation is far
different from the
users the user will tend to rely less on
the automation because it
reflects different incentives and he
will tend to intervene more and will
assume
more responsibility
so what can we learn from this
theoretical model
first we devised a new measure that
really quantifies put a number
of the on the level of comparative human
contribution
in determining the final outcome of the
system which is
causal responsibility secondly we saw
that the
human causal responsibility depends on
the combined
and convoluted characteristics of the
human the system
the environment and as i said the
convoluted convoluted
effects and what we saw in the example
that the
human contribution to the outcomes or
unique contribution to the outcomes
which i
defined by causal responsibility is
higher
when humans capability are superior to
those of the system
and when having different preferences
because in these two situations
the human uh rely less on the system
because he either
has better abilities than the system or
different preferences
so he relies more on himself and takes
more actions that are
different from the recommendation of the
of the system and that
contribute more to the outcomes
and this has a very um
interesting implication because as
technologies
develop and outperform humans in many
critical functions
actually the human unique contribution
diminishes so simply demanding
uh like there is very common demand in
many
system to always involved human in the
loop
but simply demanding that does not
assure that the human has
a meaningful part in creating the
outcomes even if
important functions are allocated to the
human for example even if
in my example the human had to press the
button of accept or reject
if he only relies on the indication from
the automation and never deviates
he has no meaningful contribution just a
mechanical means to to translate the
automation's
outcome to some action so
putting him in the loop does not really
help
so current policies that demands human
in the loop may create mismatch between
world responsibility and cause of
responsibility
and this is important because sometimes
you can hold
a human who is falsely responsible when
he actually has little real
influence and you may expose the human
to unjustify legal liability and
psychological burdens now i have a
dilemma because we have about 20 minutes
left so we can
pause for a few questions or i can show
some empirical findings and
we'll have more questions in the ends
okay uh let's we already have quite some
questions in the chat
but then uh how much time do you need
for i think i need another 10 minutes
okay then let's focus first on maybe
two questions on the model uh and for
that we have four minutes
so yeah i'd appreciate if uh who was
first
we'll skip david's first comment uh
which was posted when you
talked about the weather in tel aviv so
he said near that's just cruel
so uh yeah maybe no need to respond for
that uh
but then the next question is from jared
i have another comment for this comment
in our university the air conditioning
are still still tuned to winter
so it's 30 degrees outside and the air
condition is stand on heat
[Music]
so if you see me sweating now the real
rhythm okay but then uh
one uh question related to the model
actually is from garrett here do you
want to ask a question in person
yeah we'll do uh so thank you cool stuff
i i did wonder so it seems that you
assume that the contribution of the
human is independent of the state
uh well i could also imagine scenarios
for example where what the human
does is sort of uh complementary to what
the
system does where in some cases it's the
human is better or the performance of
human is better it's on
that of the system is better and
so the question i asked is
how do you take into account that that
responsibility may actually be different
from state to state
actually the
the example that i showed was a very
simplified example in order to put
graphs
but if you uh look at the
the main diagram that i presented that
figures up all the information flows
it may also portrait a states in which
at certain environmental states the the
automation is better and then the human
is inferior and the other way around
it doesn't limit anything it's just it's
just a picture of
all types of information flow so if you
measure the system let's say you do
empirical work which i i want to present
uh soon and you measure the
the information flow at many states
you will the model will answer that you
will have a certain state that the
the automation is better and then the
human is better but
okay the the the measure that i
i compute now it's the average
contribution of the human
so yeah this is a i will talk later on
in the conclusion
indeed it's averages the contribution of
the human on different
states so you are right at that point
it's the average contribution of the
human uh over
in many states along um along the time
but you can figure up also specific
states in which
the human has larger or smaller
contribution or responsibility
okay thank you and then the next
question is from rule
uh rule you want to answer
to ask a question yeah yeah it's related
actually so i was
i was before you started going into the
sensitivities i was just
reflecting on the metric itself
and i was wondering um that i was
considering two examples two decisions
uh a and b and so what would it mean if
the human responsibility is twice the
value for b
is compared to a and relatedly like
don't you
uh you were already alluding to this
when you were saying that this is an
average contribution so
don't you lose grip on actual causality
when you opt for an information
theoretic measure
like how can you tease out the
contribution of the human
um to the causal responsibility chain of
a decision or a situation
that's my question so
it's hard to to answer shortly but i'll
try to answer first
the the measure when i presented the
the responsibility types i told you that
responsibility can be either prospective
or it was with perspective
now the measure that i'm presenting now
is a
prospective measure it tries to to
to look at the future and say i'm
designing a system
what will be the average contribution of
a human in
such type of system i present in the end
of my
presentation that we have a deviation of
the models that look at retrospective
cases
in which case when you analyze a single
past event you cannot look at the
at the average contribution of the human
over many states because you have a
certain line of information that
followed from
the environment to the automation and
then to
the human and for that we have another
measure it's also
based on information theory but it it
measures
uh it's not an average measure it looks
at a
it's a certain specific case and
measures
the causal responsibility for a single
past event
but the measures that i'm presenting now
it's more like and
as i said it's an average contribution
and it's in a
prospective way of looking it's very
important for system designers or
when you want to do some
um legal or other consideration what was
how meaningful is the human world in the
system or the average
contribution i guess what i'm seeing is
that
the way i was interpreting is that you
can see whether the human is a necessary
contribution
based on the information right and not
so much
to what degree is the human involvement
uh
effects really is unique yeah necessary
but to what degree it's not like
necessary or not it's not binary
it's how how often or to what degree the
human contributes something
unique to this interaction that i cannot
just take the human out of the system
and
rely on the intelligent system by itself
yeah
no that's clear thank you okay so in the
10 minutes that left or less i'll
i'll go very briefly about the empirical
findings and then on the other research
that we are doing
okay near just a quick comment so uh
i know that many of uh people who are
currently in the meeting
can actually stay a bit longer after two
but
uh yeah if someone wants to stay that's
good uh
and just feel free to drop whenever you
need and i think we could just extend
the discussion a little bit
later so how much time you give me more
uh well as much as you need i mean yeah
be great
okay until tomorrow but
now it would be great to have at least i
would say i hope the beach will be
crowded in few hours
let's actually aim for 10 12 minutes and
then we will have five minutes
something like that and then i will give
priority to the questions of the people
who need to live soon
that's uh for you david because he
posted three questions but i know that
you are staying
so i uh i presented the the
the analytical model at the top of the
pyramid
and now let's look at actual human
behavior and
perception of responsibility and we
wanted to test
whether you know this is the theoretical
model and
does it really can it predict how people
really behave and contribute with
different systems
according to different characteristics
of the human and the system
so what we did we did the quite a lot
laboratory
uh experiments and you see here this is
our
interaction with technology laboratory
at tel aviv university
and now i need to to introduce more
more types of responsibilities so i
presented
until now theoretical responsibility
which is as i said prospective it's the
predicted
share of unique human contribution to
the overall
outcomes and as many
theoretical models it has simplifying
assumption like perfect human knowledge
rational humans that maximizes some
utility function and optimal use of the
automation
but unfortunately people don't use stuff
optimally so in the experiment we
measure we
measured what we called measured
responsibility
which is the observed the real empirical
share of the unique human contribution
to the overall
outcomes it's the same measure but we
just measure the distribution and the
information flows in the laboratory and
we calculated this measure
so it quantifies the same information
theory measure
and it's based on actual user
performance
and we also handed out questionnaire in
which we asked the
participant to evaluate how how
you think was your contribution what is
your assessment regarding your
comparative
unique contribution to the interaction
with the system
and what we did if you remember and so
we use the simplified
classification binary classification
system
as in the example and you remember this
graph from the example
and what we did we selected four
experimental points
each one reflects a different kind of
human or automation detection
sensitivity
and they span a range of predicted human
contribution from
12 to 47 to 69 and 287
so our first ex experiment or types of
experiment
we changed the characteristic the
detection abilities or detection
sensitivity of the human and the
automation
and we wanted to compare to the
theoretical prediction
in this experiment both have the same
incentive i mean when we programmed the
intelligent system and we told the human
the cost
and the benefit
metrics they were the same so they have
the same incentives
we did also other experiments in which
we gave the human and the system
different
different incentive but we also compared
them to the model um
prediction but i will not present this i
will i will focus on the first one
so in the deep prime or detection
sensitivity experiment we had 60
participants we devised a kind of a
simple binary classification system and
we could control
the system accuracy how good was its
detection sensitivity
and the human had to look at some
presentation and we could control
how accurate is the presentation that
the human uh
so so we control both we could control
both
the detection human detection
sensitivity and automation detection
sensitivity
and according to the graph we took the
point of
1 and 2.3 the participants
were half of the
participants had poor abilities with
detection sensitivity of about one
or quite good detection sensitivity with
detection sensitivity
about 2.3 and each one of them worked
with two different
alert system one alert system was
poor and the other one was good so this
was the within subject
condition so we have the poor human here
and he works or they worked with rather
poor or good
automation system and these are the
predicted responsibility values
and each participant performed 100 twice
with each system
and we counter balance the order of the
system and we also handed questionnaire
during the test different part of the
test
so we start with measured responsibility
and i'll explain this work because it
will repeat itself
many times below the x-axis
is the human detection sensitivity and
you have the less
accurate human on the left and the
accurate human
on the humans or participants on the
right
and the red is the accurate alert system
and the blue is the less archaeological
system
and as predicted by the rescue model you
could see that
um the former the formation of this um
the formation of this web is according
to theory because
you can see that both type of
participants
relied more on themselves
with the less accurate system and they
assumed higher responsibility
higher causal responsibility with the
less accurate system
and you can see that the accurate
accurate participant
assumed always assumed regardless the
type of the systems assumed higher
responsibility than the less
accurate participant because they relied
more on their own capabilities
so the general pattern is according to
theory but that we had the specific
numerical prediction so it's interesting
to compare
the average value to the theoretical
prediction
so you see that in most cases like let's
say
with the less accurate system the the
average
outcomes was not very far away from the
theoretical prediction
uh despite only one case when the less
accurate participant worked with the
accurate system
they assumed much higher responsibility
than
uh optimal and actually we analyzed
the reason for this is uh known for many
other behavioral
studies we analyze it in different
ways but what really happened that this
type of participant
overestimated their own abilities and
they intervened
more than optimal and thus they assumed
higher
impact or higher causal responsibility
on the outcomes
but the outcomes were better off if they
didn't intervene at
all so people with
poor abilities that worked with very
good system
wanted you know to to do something they
didn't want to feel neglected so
they did something but and they
overestimated their own abilities and
this
phenomenon is known from other
behavioral studies
if we look at the subjective perception
of responsibility and here
it's another scale because it's a
subjective scale in which they rated
their
responsibility from i contributed very
much
i had no unique contribution you see the
same pattern
and if you compare the results by
normalizing the scale
you see that subjective and measured
responsibility really matched
each other so it means that people
really
feel or have a good sense of how much
they really contributed to their process
um i will not enter this
but we also analyze the relation between
all the three components together and
we discovered that the subjective
feeling of
of responsibility encountered for about
20
of the way people actually behaved with
different systems so
the perception of of yourself of how
much you
um contribute to a to a process
influences to some degree it's not it's
not the main influence but it influences
for some degree
the way you act with the with the system
another interesting finding was you see
the solid lines it's
the graph that i showed you before
and we also asked their subjective
assessment if another person
was working with such a system what do
you think his responsibility
would be and it was a significant
difference that people
waited their own responsibility lower
than of another person in the same
situation
and this resembles um something known in
psychology as the fundamental
attribution error
and actually we we did another
specific laboratory examination for this
aspect
another dedicated experiment in which we
had the people that
worked with systems they were the actors
and people that sat next to them and
just
observed how they worked and we asked
about their
subjective perception of the level of uh
responsibility and there were
significant diff
difference between people who actually
worked with system
and people that just were observers
so the findings from the empirical
analysis it's that the rescue model is
also a descriptive model and
actually we can use it to to predict
how human uh will take responsibility
and what will be
their average responsibility in
different uh
systems and it also can predict the
subjective perception of how much the
people will feel that they
are that they contribute meaningfully to
a system
nevertheless there are two systematic
human biases
one is the tendency to assume
or to assume excessive responsibility to
intervene more than necessary
when the human capability are inferior
to the ones of the automation
and the other bias is a tendency to
consider other responsibility to be
higher than
one's own for yourself you always has a
good excuse
why your performance was poor and it's
not just
your problem it's the always the
automation faults
so the implications from the empirical
observations
is that operators may feel correctly
that they don't have significant
impact on the outcomes when they work
with the advanced intelligence system
and they may interfere more than
necessary or conversely
in some other cases that are known in
the literature
they'll be less motivated to take action
at all
both responses will hamper exploiting
the full potential of the system
and could lead for undesired
consequences
so again the demand to
always keep human in the loop can have
adverse implication on the overall
functioning of the system
the human attitude toward the system and
their interaction with the system and
the role
with it and the perception of outside
observers that watch the the human user
and assess their
responsibility for the outcomes
so this is the the end of the empirical
results that i want to
um to show just few
concluding rewards so
the measures that we developed all three
measures the theoretical
measured and subjective responsibility
can serve to expose anomalies and
provide a new method for quantifying
human comparative causal responsibility
for the outcomes in advanced system
they can help the design of the system
by tying different design options to the
predicted effect on user behavior
and their perception of responsibility
and also it it can aid in the formation
and analysis of deployment policies and
legal decision
regulation and i'll give you just a
short example for example in autonomous
weapon systems there is a
indeed a requirement by the us and
british authorities to always keep a
human
in the loop so there will be no
um um lethal force um
done inflicted without human involvement
human in the loop that will authorize
the action
but let's say that the human just sits
in a dark room he doesn't see anything
from the outside world and all he sees a
light bulb
that lights whenever the automation says
there is a tire
at a target that he needs to attack in
the human press
button that authorizes the attack so
in terms of of regulation we put a human
in the loop but
it's very clear that his involvement is
um
is meaningless it's not meaningful it's
in the loop
but it doesn't have really causal
responsibility or
unique causal contribution for the
outcomes
so it's this this way of analyzing
things may expose a
um policies or faults in policies and
legal regulations that
are not looked upon
so regarding future work and some of
this work we already
accomplished but it's on some kind of
publication
process first as people asked um
we need different measure for with
prospective responsibility because the
entropy measure it's kind of an average
measure that averages over all the
different
states and if i want to measure the
human responsibility in a single past
event i need to
to change the demands i'm not looking at
the
the the average contribution or about
specific contribution in a single chain
of events
so we devised an information theory
measure for that
and very importantly and we also work on
that
is the issue of temporal effects because
in all the examples that i showed you
the graph and the
the example that i showed i didn't
consider the time
but it's very obvious at the time you
have to
to take the the decision may impact
or should impact your um tendency to
rely on the system if you have a very
very
uh short time to decide then the system
is very good
you will tend to decide to to rely more
on the system and your
cause of responsibility will decline
so we need to to measure not just
the information flow but in information
transmission rates
and human channel capacity constraints
we which are also
known from the literature and then add
the time
element to the model as another
dimension that is currently missing from
the model
so these were my 10 minutes
engineer that was great
so uh yeah it's exactly two so i suppose
uh
many people need to leave if you have
any questions leave them in the chat
before we leave
and then we will have a chance to ask
them
or you could just raise your hand if you
have time to ask a question yourself
can i ask a question
hi everybody i was on um on voice for
the largest part of the time so i could
not see
the slides but if i'm correct in the
beginning of the presentation like 20
minutes or so into the presentation
the entropy measure was introduced and
then a ratio between entropies were
was mentioned did i get that correctly
or did i miss something
okay so being an information theorist i
was wondering where the ratio of
entropies comes from
because that's very unusual to take
ratios
of entropies we usually subtract or add
or or
that kind of combinations hardly i've
actually never seen
a ratio other than ratio to a maximum
like efficiency and these kind of things
so i was wondering
does it even make sense to think about
ratios of entropies
yeah so i have a good answer and this is
a good questions
i'll slide on here let's look at this
okay so this is the measure
below we can see the entropies of the
final
action action which is implemented it's
the entropy of the election and
the the denominator is the uncertainty
left
about the outcomes given that i know
for all the automation variables so
actually i measure the
relations between the
system outcomes and all the automation
variables and i
measure how much uncertainty is left
because
if i if i'm looking at all the the
the automation variables then there is
no uncertainty
i mean their identity is gone then the
numerator would be zero and the human
contributes nothing and actually
in the literature this form of measure
appear at what is known tails
coefficient measure
which is very known measure you can find
it
also in spss to measure the
relation between nominal variables
but he measures something else he
measured the
relative reduction in the uncertainty of
a variable
due to a knowledge of another variable x
so let's
look at the yeah
so if i if i may so i i hear what you
say but in terms of the explanation
this is very much like terms from
information theory like equivocation etc
which are usually defined as the
difference between entropy so difference
between
normal and conditional entropy etc so i
see that you can also take a ratio
but would a difference also have
worked the problem is different that
it would not give me a a measure that is
is confined to a range between zero and
one because the entropies could have
much different number and then it will
be
it will change in very types of system
without me
able to compare them if we look at tails
uncertainty coefficient and we have
x and y what else does he say
okay let's look at the mutual
information and look at
the uh mutual
entropy and look at the entropy of x and
if i divide this
mutual part by the entropy of x it's the
how much
is reduced uh
the uncertainty of x is reduced by
knowing why and i'm looking at the
complex
a complementary value which is how much
um uncertainty is remind
it remains about the output when i know
all the automation variables so i
look at the other part and i extend it
because i don't just look at the
two variables but i look at
[Music]
many variables but tails uncertainty
coefficient is used to
exactly that to measure the association
between for example nominal variables so
you cannot
do it with the pearson coefficient or
experiment coefficient because they are
not nominal and i'm not
metric so so
this is when you use the tails
uncertainty coefficient and i do i just
i do something
similar but i i'm not looking at the
reduction but i'm looking at the
uncertainty left
after i know all the automation
parameters and since there are no other
sources for information flows into the
system or the remaining uncertainty
is related to the human contribution and
this is the sense of devising them to
have a measure that ranges between zero
and one
exactly like a tails uncertainty
coefficient and based on experiment
which relate also to
forced linear relations between
variables
okay thanks very much for the further
explanation
okay let's perhaps
get back to david's questions
do you want to start with the first
question
on contact specificity yes
thanks so thanks
for the presentation um it's good to
know that you're sweating uh because of
the
of the the air conditioning situation
and not our difficult questions
um still i'll let's let's do our best to
make you sweat even more
now um uh okaying aside i was wondering
about uh
something that was also touched upon by
by an earlier uh
a question from a colleague which is uh
kind of the
yeah the context specific uh
elements to responsibility um
and and and what actually human
capabilities and automation capabilities
are because they
they often depend on on a context
so uh in terms of perception and
information processing
uh it often depends on um
yeah on on on contextual parameters like
you know how
what what's the design of the system
what is the human actually doing
uh on the site what's the design of the
interface
and systems that work perfectly under
condition a don't work so well under
condition b
um et cetera et cetera um
so that means how do you actually
deal with the fact that these
capabilities are not
static but actually yeah in a sense a
dynamic and context dependent how is
that incorporated in
this quantification
well there are two parts to the answer
one we deal with and one we don't
[Music]
let's look at the diagram
the general diagram that portrays the
information flows
enables to describe a situation where
for
a different state let's say the informa
first of all when you plot this diagram
for a specific system so you can
it it reflects the the function
allocation and the system
design you can do the the errors that
will describe the
exact types of human interaction like we
did in the
dialect uh example in which the it was a
decision support system it's very often
that
only do information analysis but the
action selection
and implementation is remains for
for the human so you can for each system
you can do a different drawing
and you can when you do the distribution
you can really change
them according to the state
that the system or the combined human
machine system encounters so it's
possible that
if at one er let's say the human is
at one um environmental state
the human will be the main contributor
because he
has a better ability to acquire
information and analyzes and do the
action
and in other environmental state the
system is uh is better than the human in
it will do so so there is a viability in
that sense
and of course you need to to figure out
the the probability of
encountering different the different
state in the environment
maybe you ask about something else to
which we don't answer because
entropy and information theory do not
answer entropy
and information theory were developed
developed
under the assumption of
being a monotonic over time i mean
stationary and ergotic
was both type of uh actually shannon
when he developed information theory he
started with the identical uh
distributed variables and only later on
it was
expanded to uh variables that are um
stationary and ergodic so if we have
cases in which the probabilities
change over time
the human causal responsibility will
change
over time and we will don't we there is
no meaning to look at the average
because the change in
changing over time so for that i have to
answer first
many system you can analyze them at
their
constant state when they are rather
stable
for example in the experiment of course
there is a
that we did in the empirical experiment
there is a learning period in which
users are
you know learning the system so at that
point the contribution changes because
they are just
you know in the learning mode so we let
them experience the system sufficiently
enough
to see that they are uh know the system
and then
we did the calculation only when we saw
that the the probabilities are quite
stable so we are in the
we are in the constant mode
space but
you are right that if if the
probabilities and the
distribution changes over time or it's
not a ergodic
then the whole use of in information
theory is uh you you cannot use it it's
it's
it's outside its core but nevertheless
even even if though in those day
changes if you do the diagram of
information flow
it sometimes it can give you some
insight on on
how good is the human on what is the
human contribution
you look at things in other way which is
more
in generic way and less let's say a
legal or philosophical way
okay thank you i had more questions but
i don't want to eat up all the time so
if there's
other people wanting to butt in please
do and otherwise i kind of can give me
the word again
yes let's say here if there are any
other questions
from someone else not that so
while we are waiting for the questions i
think a good journey has his reign
uh yes please i don't see it for some
reason
yes i can see it now you're gay
hey thanks uh i need uh thanks very much
for the very interesting talk so one of
my questions is actually the
the same one as david's last question in
the chat so i'll let him ask that later
on
but i was wondering um um
so i'm not familiar with all the
uh in depth with all the you know
information theory and and other
the the kind of the mathematical side of
things here but
i was curious how does this kind of
model uh
does it relate uh at all to
uh causal modeling you know like in the
way that judea apparel and
his colleagues uh um conceive
of kind of the causal relationships and
trying to measure them so is this a
related way of doing this or are those
two not related at all
[Music]
so i don't think i don't think that we
i think this is a very interesting
question i don't think we try to
tie this directly to causal modeling
like pearl's
models we see this more
as a as a description of a system
description i think tying these two
together is
might be a very interesting issue but we
haven't looked at it closely yet
thanks
okay thank you evgeny so uh what i would
propose to do is to
uh wrap up with the david's questions
and if we don't receive any questions
while david is asking for those
questions
we will have a short break and then
we also have another meeting scheduled
with uh with a few of us so i would just
ask you to
rejoin the meeting using that link but
for now let's just
go through those questions all right
then
um so a second question is um
uh it's actually about the design of the
of the of the interface and so you you
describe uh
uh in this scheme it's it's it's either
the human or the automated
uh system that takes a certain
actions or information processing or
analysis
and and so i had two questions
one is we know for example from warning
systems with binary thresholds
um that it's very difficult to tune
those thresholds
unless you know everything upfront which
unfortunately usually we don't
um and and there and especially when
there's more variability
in how to solve situations then
some operators may think the thresholds
are too early
set too early or set too late um or
depending on the context they're set to
early or too late and so
you get these annoyance and cry wolf
effects that
that actually you know if they wouldn't
occur
it would be just a clean information
processing problem
but because humans adapt and learn and
get demotivated
and and
and such this is actually
impacts uh in a way they they deal with
the systems
uh and i would say and to some extent
then the designer has some impact on uh
on on actually the responsibility that
um that the operator
you know feels or or can actually take
because they
might get disengaged from the from the
system so i wondered
uh if you can follow this line of
reasoning uh
um how you how you see this
it's a great question because the model
is intended to tackle exactly that
because
what i say that in the prospective model
that i present is
oriented to designing system yes now
if my if my system has poor capabilities
and it
has a lot of false alarms
and or it has reflects
other preferences than mine for example
it
allows false alarms or something like
that
eventually when the operator work with
the system
consecutively he will learn that the the
system is unreliable and it cannot trust
the system and then it will tend to
interfere
intervene more the theoretical model
which assume a perfect rationality and
perfect knowledge
assumes that the human knows the
characteristics
of the system for example the rate
of of false alarms
and misdetections and of course his own
capabilities
and then he select the the the right
amount
to interview for example the optimal
amount so this is the theoretical model
what happened in the empirical empirical
experiments is that people really
experience with the system and if they
if they had for example good capability
and they worked with the
with the inferior system
after you know 10 or 20 trials they say
okay
this system is just because they were
penalized for each time they were
wrong and got points when they arrived
so after a few trials they say okay i
cannot rely on the system and unless
you know something very obvious and they
try to um
to and they relied only on themselves so
and in in most cases despite uh
except the one that i marked um
the empirical responsibility values were
very close to
the theoretical prediction which means
that after people experienced a
system with different characteristics
they behaved in in the expected way so
if i was a designer of the system and i
used this model i could predict
how much the human will really
contribute to
to my system but more importantly
if i want to put some weights
on the preferences that i
put in the system algorithm which
determine the false alarm rate and
i could also play with this and see what
is the the
the effect on the on the outcomes and on
the user because i allowed to play
with with all the the different
parameters so for a system design you
can
plug in your your assumption
without even having the real system and
try to
and you will have a notion of how much
the human user will really contribute to
your system
or how much the final outcomes really
depends to what agree on the unique
contribution of the human
and
is that it that's it um
that you can use during uh during system
design to understand the
the how meaningful is the world how
really meaningful is the world of this
is the stuff that you are working on
but how really meaningful is the the
role of the human in
in your system how is how meaningful is
or
unique is the actual human contribution
to the outcome and
processes by the way we didn't do it
because we
we were interested in causal
responsibility
so we measured only the unique human
contribution to the outcomes
but you can measure also the unique
human contribution to each of the states
and to say okay the human contributes
more to the information acquisition and
then to information analysis and
you can put a number on each state it's
it's not complicated but
it just makes things longer but we were
interested in the final outcome and when
i look at the simple as
the system as a whole how unique is the
human contribution for the final
uh selected the action and uh
a very quick follow-up question it does
seem to imply that
uh the the there's some time to
to iterate and learn for humans as well
as for the system
uh so that let's say the the
consequences are not too critical
uh is that correct or uh
i didn't understand what you say so
let's suppose we are dealing with uh
with uh for example uh aviation or
um or driving or
or you know weapon systems then for you
to find out that things are
uh wrong the thresholds are set in the
wrong way
uh sometimes that can be just uh you
know just annoyance but sometimes it can
mean the difference between life and
death right
yeah of course what what you are saying
that in um
for the human user in order to to as i
said i'm looking at the system at the
stationary
stage in which in in which the human and
knows the the abilities and disabilities
and the performance of the system and
his own and then he can act in a certain
way because he knows the system
but you're saying in some system the the
the action is very
rare and i will not encounter it often
and i will not
gain experience
this is right the way to do it is by
letting you experience with the
simulated environment
in enough repetition until you know your
abilities and the abilities of the
of the system so in case of real um
activation of the system you will
already have a kind of
notion about your abilities yes but
without a learning curve or something
that you know and just you know manuals
of course the the things will not uh
look that way and of course entropy look
at the average uncertainty
it's kind of everything that's entity so
i need kind
it will be or when i'm looking at their
prospective responsibility it will deal
with your responsibility over many
repetitions
and not to a singular past event but
for example in our retrospective model
which i didn't present
we measure the distance between
we measure if the how reasonable was the
human decision given
all the information that he had in the
time
of the decision and this really um
deals with a single chain of events and
even unique events
and you can measure you can put a number
of
on how reasonable was the human action
selection but it's another model it uses
another
another measures of information theory
which has not entropy but
distance between the distribution
could be clear leveler distance and
stuff like that
nice thank you um then my final question
might be a quick or my uh but the answer
might be long and that's actually uh
what you actually miss out when you
quantify stuff according to you
so to put it differently what would a
qualitative assessment
of responsibility add so not
necessarily liked scales but but really
a qualitative assessment
what would that add if anything
so i'll answer i think i think the the
qualitative
qualitative assessment is extremely
important here the
um and i think the one has to take
the quantitative model we're presenting
here with a certain
um level of of sort of
caution uh because obviously it
simplifies things it sort of it says
something about
uh what the human involvement in some
idealized
schematic depiction of a system is
and and it's it can serve as a sanity
check
for a situation like the one that you
know described earlier where you have
someone who is
in the loop formally but in fact relies
entirely on outside sources of
information that are
supplied to this person now a
qualitative model will
will be very important because it takes
into account
aspects that we don't really consider
here like for instance
the dynamics of the of the situation the
the fact that you may
look for additional sources of
information that you may
make this decision within some social
organizational context
that will have will affect the
information will affect the
the incentives will affect the the
methods that are acceptable or
unacceptable for making these decisions
and so on
uh these are all we we can all model
these things
and if we really abstract them into into
a payoff matrix or into just
probabilities but this will lose a lot
of the
actual complexity that exists if one
describes a situation
so i believe this this kind of time kind
of quantitative modeling
is not in any form a replacement of a
qualitative analysis but rather
it sort of adds a
respect a another perspective as sort of
a quantifiable perspective for such an
analysis
thank you very much
okay that was a great conclusion i think
to the
overall session so i will stop the
recording now
and thank you everyone for attending |
85aaf7ec-c22c-4440-bcf2-e71cf94dbcaf | trentmkelly/LessWrong-43k | LessWrong | Meetup : Melbourne, Friday 1st July 7pm
Discussion article for the meetup : Melbourne, Friday 1st July 7pm
WHEN: 01 July 2011 07:00:00PM (+1000)
WHERE: 55 Walsh St, West Melbourne 3003
Directions: Enter the somewhat unfriendly building, climb the stairs to the top (2 floors), and turn left. If the front door is locked look for a sign that includes my phone number ask me to let you in.
Discussion: http://groups.google.com/group/melbourne-less-wrong
We meet on the first Friday of each month. We formally start at 7pm, but I'll be here from 6pm and you're all welcome. You're all welcome to BYO liquid refreshments.
Discussion article for the meetup : Melbourne, Friday 1st July 7pm |
d5236059-6f3b-49f4-a727-933abaa9bc44 | trentmkelly/LessWrong-43k | LessWrong | AI safety milestones?
Optimal governance interventions depend on progress in technical AI safety. For example, two rough technical safety milestones are have metrics to determine how scary a model's capabilities are and have a tool to determine whether a model is trying to deceive you. Our governance plans should adapt based on whether these milestones have been achieved (or when it seems they will be achieved) (and for less binary milestones, adapt based on partial progress).
What are more possible milestones in technical AI safety (that might be relevant to governance interventions)? |
8b9dce3f-84c4-4f1c-b215-73eee25aa310 | trentmkelly/LessWrong-43k | LessWrong | What does 10x-ing effective compute get you?
This is more speculative and confusing than my typical posts and I also think the content of this post could be substantially improved with more effort. But it's been sitting around in my drafts for a long time and I sometimes want to reference the arguments in it, so I thought I would go ahead and post it.
I often speculate about how much progress you get in the first year after AIs fully automate AI R&D within an AI company (if people try to go as fast as possible). Natural ways of estimating this often involve computing algorithmic research speed-up relative to prior years where research was done by humans. This somewhat naturally gets you progress in units of effective compute — that is, as defined by Epoch researchers here, "the equivalent increase in scale that would be needed to match a given model performance absent innovation". My median ends up being something like 6 OOMs of effective compute in the first year. (OOM stands for Order Of Magnitude, so 6 OOMs is an increase by 106.)
This raises a natural question: what does 6 OOMs of effective compute (in software progress) get you?[1] Improved algorithms can get you more copies (lower cost), higher speed, or higher intelligence (improved quality). In practice, we (or really the AIs) would try to make the optimal tradeoff taking into account diminishing returns (if we were trying to go as fast as possible), but I think it's useful to pick a possible choice of tradeoff which gives some intuition for what you end up with (even if this tradeoff is suboptimal).
Here's one choice that I speculate is on the possibility frontier. For every OOM of effective compute, you get: 3x more copies, 1.25x speed, and a 20x improvement in rank ordering among human professionals.[2]
What do I mean by rank ordering improvement among human professionals? Going from the 1000th best software engineer to the 100th best software engineer would be a 10x improvement in rank ordering and I'm claiming this is a roughly scale invariant |
ffe12117-1391-437d-b202-77ddbd4e515b | trentmkelly/LessWrong-43k | LessWrong | Would AGIs parent young AGIs?
Question. In the world in which multiple AGIs obtain (i.e., not a singleton), is it likely these AGIs will be parents/mentors/guardians of younger AGIs?
----------------------------------------
Context. I was wondering whether parenthood (or, more minimally, guardianship) is an intrinsic property of the evolution of sufficiently intelligent agents.
The answer seems to depend, at least partly, on the stability of the environment and the returns to learning. My off-the-cuff reasoning follows.
1. It seems that a species (or, perhaps more generically, sufficiently similar information processing systems), can persist over time either by increasing their own lifespan or by creating new instances of the species.
2.a. It seems that if the environment is highly stable, then increasing lifespan and/or creating clones is attractive because the agent has found agent-environment fit. Returns to learning are low because the environment is already known. Returns to guardianship are low because the usefulness of a young agent existing is itself low.
2.b. If, on the other hand, the environment is highly unstable, then shorter lives seems more adaptive because natural selection will ensure agent-environment fit. Returns to learning are again low but this time for the opposite reason: facts now will not be relevant in the future. Returns to guardianship are low because the probability of survival of any one agent is low which makes guardianship investment a low-return activity.
2.c. Finally, if the environment is moderately stable, then there seems to be an optimal lifespan: one that enables a relatively long learning phase and a relatively long exploit phase. Returns to learning seem high as do returns to guardianship.
3. If AGIs will need relatively long learning phases, and, further, if young AGIs benefit from parenting/mentorship/guardianship (perhaps because they can be hurt by their environment, including by other AGIs), then it seems there's a fitness advantage to AGIs |
f6db4fd3-efed-468c-8cf7-aaa10aca39c2 | trentmkelly/LessWrong-43k | LessWrong | Long-Term Future Fund:
May 2023 to March 2024 Payout recommendations
Introduction
This payout report covers the Long-Term Future Fund's grantmaking from May 1 2023 to March 31 2024 (11 months). It follows our previous April 2023 payout report.
* Total funding recommended: $6,290,550
* Total funding paid out: $5,363,105
* Number of grants paid out: 141
* Acceptance rate (excluding desk rejections): 159/672 = 23.7%
* Acceptance rate (including desk rejections): 159/825 = 19.3%
* Report authors: Linchuan Zhang (primary author), Caleb Parikh (fund chair), Oliver Habryka, Lawrence Chan, Clara Collier, Daniel Eth, Lauro Langosco, Thomas Larsen, Eli Lifland
25 of our grantees, who received a total of $790,251, requested that our public reports for their grants are anonymized (the table below includes those grants). 13 grantees, who received a total of $529, 819, requested that we not include public reports for their grants. You can read our policy on public reporting here.
We referred at least 2 grants to other funders for evaluation.
Highlighted Grants
(The following grants writeups were written by me, Linch Zhang. They were reviewed by the primary investigators of each grant).
Below, we highlighted some grants that we thought were interesting and covered a relatively wide scope of LTFF’s activities. We hope that reading the highlighted grants can help donors make more informed decisions about whether to donate to LTFF.[1]
Gabriel Mukobi ($40,680) - 9-month university tuition support for technical AI safety research focused on empowering AI governance interventions
The Long-Term Future Fund provided a $40,680 grant to Gabriel Mukobi from September 2023 to June 2024, originally for 9 months of university tuition support. The grant enabled Gabe to pursue his master's program in Computer Science at Stanford, with a focus on technical AI governance.
Several factors favored funding Gabe, including his strong academic background (4.0 GPA in Stanford CS undergrad with 6 graduate-level courses), experience in difficult technical AI |
68e25dc0-3155-421c-b72b-134cd24c1f5a | trentmkelly/LessWrong-43k | LessWrong | Meetup : Bratislava
Discussion article for the meetup : Bratislava
WHEN: 20 October 2014 06:00:00PM (+0200)
WHERE: Bistro The Peach, Mariánska 3, Bratislava
Stretávame sa opäť v bistre; majiteľ je ochotný kvôli nám na jeden večer urobiť bistro nefajčiarske.
Discussion article for the meetup : Bratislava |
84c4183c-a5f5-432f-8797-066d160bcc90 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Book review: "A Thousand Brains" by Jeff Hawkins
Jeff Hawkins gets full credit for getting me first interested in the idea that neuroscience might lead to artificial general intelligence—an idea which gradually turned into an all-consuming hobby, and more recently a new job. I'm not alone in finding him inspiring. Andrew Ng claimed [here](https://www.forbes.com/sites/roberthof/2014/08/28/interview-inside-google-brain-founder-andrew-ngs-plans-to-transform-baidu/?sh=102a616640a4) that Hawkins helped convince him, as a young professor, that a simple scaled-up learning algorithm could reach Artificial General Intelligence (AGI). (Ironically, Hawkins scoffs at the deep neural nets built by Ng and others—Hawkins would say: "Yes yes, a simple scaled-up learning algorithm can reach AGI, but not *that* learning algorithm!!")
Hawkins's last book was *On Intelligence* in 2004. What's he been up to since then? Well, if you don't want to spend the time reading his journal articles or [watching his research meetings on YouTube](https://www.youtube.com/c/NumentaTheory/videos), good news for you—his new book, *A Thousand Brains*, is out! There’s a lot of fascinating stuff here. I'm going to pick and choose a couple topics that I find especially interesting and important, but do read the book for much more that I'm not mentioning.
A grand vision of how the brain works
=====================================
Many expert neuroscientists think that the brain is horrifically complicated, and we are centuries away from understanding it well enough to build AGI (i.e., computer systems that have the same kind of common-sense and flexible understanding of the world and ability to solve problems that humans do). Not Jeff Hawkins! He thinks we *can* understand the brain well enough to copy its principles into an AGI. And he doesn't think that goal is centuries away. He thinks we're most of the way there! In [an interview last year](https://www.lesswrong.com/posts/FoJSa8mgLPT83g9e8/jeff-hawkins-on-neuromorphic-agi-within-20-years) he guessed that we’re within 20 years of finishing the job.
The people arguing that the brain is horrifically complicated seem at first glance to have a strong case. The brain has a whopping 1011.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
neurons with 1014 synapses, packed full of intricate structure. [One study](https://doi.org/10.1038/nature18933) found 180 distinct areas within the cerebral cortex. Neuroscience students pour over huge stacks of flashcards with terms like “striatum”, “habenula”, “stria medullaris”, “fregula”, and "interpeduncular nucleus". (Quiz: Which of those are real brain regions, and which are types of pasta?) Every year we get another 50,000 or so new neuroscience papers dumped into our ever-deepening ocean of knowledge about the brain, with no end in sight.
So the brain is indeed horrifically complicated. Right? Well, Jeff Hawkins and like-minded thinkers have a rebuttal, and it comes in two parts:
1. The horrific complexity of the “old brain” doesn’t count, because we don’t need it for AGI
---------------------------------------------------------------------------------------------
According to Hawkins, much of the brain—including a disproportionate share of the brain's horrific complexity, like the interpeduncular nucleus I mentioned—*just doesn’t count*. Yes it’s complicated. But we don’t care, because understanding it is not necessary for building AGI. In fact, understanding it is not even *helpful* for building AGI!
I’m talking here about the distinction between what Hawkins calls **“old brain vs new brain”**. The “new brain” is the mammalian neocortex, a wrinkly sheet on that is especially enlarged in humans, wrapping around the outside of the human brain, about 2.5 mm thick and the size of a large dinner napkin (if you unwrinkled it). The “old brain” is everything else in the brain, which (says Hawkins) is more similar between mammals, reptiles, and so on.
“The neocortex is the organ of intelligence,” writes Hawkins. “Almost all the capabilities we think of as intelligence—such as vision, language, music, math, science, and engineering—are created by the neocortex. When we think about something, it is mostly the neocortex doing the thinking…. If we want to understand intelligence, then we have to understand what the neocortex does and how it does it. An animal doesn’t need a neocortex to live a complex life. A crocodile’s brain is roughly equivalent to our brain, but without a proper neocortex. A crocodile has sophisticated behaviors, cares for its young, and knows how to navigate its environment...but nothing close to human intelligence.”
I think Hawkins's **new brain / old brain discussion is bound to drive neuroscientist readers nuts**. See, for example, the paper [Your Brain Is Not An Onion With A Tiny Reptile Inside](https://doi.org/10.1177/0963721420917687) for this perspective, or see the current widespread dismissal of [“triune brain theory”](https://en.wikipedia.org/wiki/Triune_brain). The mammalian neocortex is in fact closely related to the “pallium” in other animals, particularly the well-developed pallium in birds and reptiles (including, yes, crocodiles!). One researcher (Tegan McCaslin) attempted a [head-to-head comparison between bird pallium and primate neocortex](https://aiimpacts.org/investigation-into-the-relationship-between-neuron-count-and-intelligence-across-differing-cortical-architectures/), and found that there was no obvious difference in intelligence, when you hold the number of neurons fixed. A [recent](https://science.sciencemag.org/content/369/6511/eabc5534) [paper](https://doi.org/10.1126/science.abc5534) found suggestive evidence of similar neuron-level circuitry between the bird pallium and mammalian neocortex. Granted, the neurons have a different spatial arrangement in the bird pallium vs the mammal neocortex. But it’s the neuron types and connectivity that define the algorithm, not the spatial arrangement. [Paul Cisek traces the origin of the pallium](https://doi.org/10.3758/s13414-019-01760-1) all the way back to the earliest proto-brains. The human neocortex indeed massively expanded relative to chimpanzees, but then again, so did the “old brain” human cerebellum and thalamus.
And what’s more (these angry neuroscientists would likely continue), it’s not like the neocortex works by itself. The “old brain” thalamus has just as much a claim to be involved in human intelligence, language, music, and so on as the neocortex does, and likewise with the “old brain” basal ganglia, cerebellum, and hippocampus.
OK. All this is true. But I’m going to stick my neck out and say that Hawkins is *“correct in spirit”* on this issue. And I’ve tried (e.g. [here](https://www.lesswrong.com/posts/diruo47z32eprenTg/my-computational-framework-for-the-brain)) to stake out a more careful and defensible claim along the same lines.
My version goes: Mammal (and lizard) brains have a *“learning subsystem”*. It implements a learning algorithm that starts from scratch (analogous to random weights—so it’s utterly useless to the organism at birth—see discussion of "learning-from-scratch-ism" [here](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine)), but helps the organism more and more over time, as it learns. This subsystem involves the entire "telencephalon" region of the brain—namely, the neocortex (or pallium), hippocampus, amygdala, part of the basal ganglia, and a few other things (again see [here](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine))—along with parts of the thalamus and cerebellum, but definitely *not*, for example, the hypothalamus or brainstem. This subsystem is not particularly “new” or peculiar to mammals; very simple versions of this subsystem date back to the earliest vertebrates, helping them learn to navigate their environment, remember where there's often food, etc. But the subsystem *is* unusually large and sophisticated in humans, and it *is* the home of human intelligence, and it does *primarily* revolve around the activities of the cortex / pallium.
So far as I can tell, my version keeps all the good ideas of Hawkins (and like-minded thinkers) intact, while avoiding the problematic parts. I'm open to feedback, of course.
2. The horrific complexity of the neocortex is in the learned content, not the learning algorithm
-------------------------------------------------------------------------------------------------
The second reason that making brain-like AGI is easier than it looks, according to Hawkins, is that “the neocortex looks similar everywhere”. He writes, "The complex circuitry of the neocortex looks remarkably alike in visual regions, language regions, and touch regions, [and even] across species.... There are differences. For example, some regions of the neocortex have more of certain cells and less of others, and there are some regions that have an extra cell type not found elsewhere...But overall, the variations between regions are relatively small compared to the similarities."
How is it possible for one type of circuit to do so many things? Because it’s a learning algorithm! Different parts of the neocortex receive different types of data, and correspondingly learn different types of patterns as they develop.
Think of the [OpenAI Microscope](https://openai.com/blog/microscope/) visualizations of different neurons in a deep neural net. There’s so much complexity! But no human needed to design that complexity; it was automatically discovered by the learning algorithm. The learning algorithm itself is comparatively simple—gradient descent and so on.
By the same token, a cognitive psychologist could easily spend her entire career diving into the intricacies of how an adult neocortex processes phonemes. But on Hawkins's view, we can build brain-like AGI without doing any of that hard work. We just need to find the learning algorithm, and let 'er rip, and it will construct the phoneme-processing machinery on its own.
Hawkins offers various pieces of evidence that the neocortex runs a single, massively-parallel, legible learning algorithm. First, as above, "the detailed circuits seen everywhere in the neocortex are remarkably similar”. Second, “the major expansion of the modern human neocortex relative to our hominid ancestors occurred rapidly in evolutionary time, just a few million years. This is probably not enough time for multiple new complex capabilities to be discovered by evolution, but it is plenty of time for evolution to make more copies of the same thing.” Third is plasticity—for example how blind people use their visual cortex for other purposes. Fourth, “our brains did not evolve to program computers or make ice cream."
There's a lot more evidence for and against, beyond what Hawkins talks about. (For example, [here's](https://www.lesswrong.com/posts/jcCMsg46RhcZJrTxP/survey-on-cortical-uniformity-an-expert-amplification?commentId=cwa42biK53gzKKgks#ZJsSNG9dNKFHk6fvp) a very clever argument in favor that I saw just a few days ago.) I’ve written about cortical uniformity previously ([here](https://www.lesswrong.com/posts/NkSpukDkm9pjRdMdB/human-instincts-symbol-grounding-and-the-blank-slate), [here](https://www.lesswrong.com/posts/8F8dagB4q4BzR5JNz/gary-marcus-vs-cortical-uniformity)), and plan to do a more thorough and careful job in the future. For now I’ll just say that this is certainly a hypothesis worth taking seriously, and even if it’s not *universally* accepted in neuroscience, Hawkins is by no means the only one who believes it.
3. Put them together, and you get a vision for brain-like AGI on the horizon
----------------------------------------------------------------------------
So if indeed we can get AGI by reverse-engineering just the neocortex (and its “helper” organs like the thalamus and hippocampus), and if the neocortex is a relatively simple, human-legible, learning algorithm, then all of the sudden it doesn’t sound so crazy for Hawkins to say that brain-like AGI is feasible, and not centuries away, but rather already starting to crystallize into view on the horizon. I found this vision intriguing when I first heard it, and after quite a bit more research and exposure to other perspectives, I still more-or-less buy into it (although as I mentioned, I'm not done studying it).
By the way, an interesting aspect of cortical uniformity is that it's a giant puzzle piece into which we need to (and haven’t yet) fit every other aspect of human nature and psychology. There should be whole books written on this. Instead, *nothing*. For example, I have all sorts of social instincts—guilt, the desire to be popular, etc. How exactly does that work? The neocortex knows whether or not I’m popular, but it doesn’t care, because (on this view) it’s just a generic learning algorithm. The old brain cares very much whether I'm popular, but it’s too stupid to understand the world, so how would it know whether I’m popular or not? I’ve casually speculated on this a bit (e.g. [here](https://www.lesswrong.com/posts/Mh2p4MMQHdEAqmKm8/little-glimpses-of-empathy-as-the-foundation-for-social)) but it seems like a gaping hole in our understanding of the brain, and you won’t find any answers in Hawkins’s book … or anywhere else as far as I know! I encourage anyone reading this to try to figure it out, or tell me if you know the answer. Thesis topic anyone?
A grand vision of how the neocortex works
=========================================
For everything I've written so far, I could have written essentially the same thing about Hawkins’s 2004 book. That's not new, although it remains as important and under-discussed as ever.
A big *new* part of the book is that Hawkins and collaborators now have more refined ideas about exactly what learning algorithm the neocortex is running. (Hint: it’s *not* a deep convolutional neural net trained by backpropagation. Hawkins *hates* those!)
This is a big and important section of the book. I’m going to skip it. My excuse is: [I wrote a summary of an interview he did a while back](https://www.lesswrong.com/posts/FoJSa8mgLPT83g9e8/jeff-hawkins-on-neuromorphic-agi-within-20-years), and that post covered more-or-less similar ground. That said, this book describes it better, including a new and helpful (albeit still a bit sketchy) discussion of learning abstract concepts.
To be clear, in case you're wondering, Hawkins does not have a complete ready-to-code algorithm for how the neocortex works. He claims to have a framework including essential ingredients that need to be present. But many details are yet to be filled in.
Does machine intelligence pose any risk for humanity?
=====================================================
Some people (cf. [Stuart Russell's book](https://www.penguinrandomhouse.com/books/566677/human-compatible-by-stuart-russell/)) are concerned that the development of AGI poses a substantial risk of catastrophic accidents, up to and including human extinction. They therefore urge research into how to ensure that AIs robustly do what humans want them to do—just as Enrico Fermi invented [nuclear reactor control rods](https://en.wikipedia.org/wiki/Control_rod) *before* he built the first nuclear reactor.
Jeff Hawkins is having none of it. “When I read about these concerns,” he says, “I feel that the arguments are being made without any understanding of what intelligence is.”
Well, I’m more-or-less fully on board with Hawkins’s underlying framework for thinking about the brain and neocortex and intelligence. And I *do* think that developing a neocortex-like AGI poses a serious risk of catastrophic accidents, up to and including human extinction, if we don’t spend some time and effort developing new good ideas analogous to Fermi’s brilliant invention of control rods.
So I guess I’m in an unusually good position to make this case!
Start with Hawkins’s argument *against* machine intelligence being a risk
-------------------------------------------------------------------------
I’ll start by summarizing Hawkins’s argument that neocortex-like AGI does *not* pose an existential threat of catastrophic accidents. Here are what I take to be his main and best arguments:
First, Hawkins says that we’ll build in safety features.
> Asimov’s three laws of robotics were proposed in the context of science-fiction novels and don’t necessarily apply to all forms of machine intelligence. But in any product design, there are safeguards that are worth considering. They can be quite simple. For example, my car has a built-in safety system to avoid accidents. Normally, the car follows my orders, which I communicate via the accelerator and brake pedals. However, if the car detects an obstacle that I am going to hit, it will ignore my orders and apply the brakes. You could say the car is following Asimov’s first and second laws, or you could say that the engineers who designed my car built in some safety features. Intelligent machines will also have built-in behaviors for safety.
>
>
Second, Hawkins says that goals and motivations are separate from intelligence. The neocortex makes a map of the world, he says. You can use a map to do good or ill, but “a map has no motivations on its own. A map will not desire to go someplace, nor will it spontaneously develop goals or ambitions. The same is true for the neocortex.”
Third, Hawkins has specific disagreements with the idea of “goal misalignment”. He correctly describes what that is: “This threat supposedly arises when an intelligent machine pursues a goal that is harmful to humans *and* we can’t stop it. It is sometimes referred to as the “Sorcerer’s Apprentice” problem…. The concern is that an intelligent machine might similarly do what we ask it to do, but when we ask the machine to stop, it sees that as an obstacle to completing the first request. The machine goes to any length to pursue the first goal….
Again, he rejects this:
> The goal-misalignment threat depends on two improbabilities: first, although the intelligent machine accepts our first request, it ignores subsequent requests, and second, the intelligent machine is capable of commandeering sufficient resources to prevent all human efforts to stop it…. Intelligence is the ability to learn a model of the world. Like a map, the model can tell you how to achieve something, but on its own it has no goals or drives. We, the designers of intelligent machines, have to go out of our way to design in motivations. Why would we design a machine that accepts our first request but ignores all others after that?...The second requirement of the goal-misalignment risk is that an intelligent machine can commandeer the Earth’s resources to pursue its goals, or in other ways prevent us from stopping it...To do so would require the machine to be in control of the vast majority of the world’s communications, production, and transportation…. A possible way for an intelligent machine to prevent us from stopping it is blackmail. For example, if we put an intelligent machine in charge of nuclear weapons, then the machine could say “If you try to stop me, I will blow us all up.”... We have similar concerns with humans. This is why no single human or entity can control the entire internet and why we require multiple people to launch a nuclear missile.”
>
>
The devil is in the details
---------------------------
Now I don’t think any of these arguments are particularly unreasonable. The common thread as I see it is, what Hawkins writes is the *start* of a plausible idea to avoid catastrophic AGI accidents. But when you think about those ideas a bit more carefully, and try to work out the details, it starts to seem much harder, and less like a slam-dunk and more like an open problem which might or might not even be solvable.
### 1. Goals and motivations are separate from intelligence ("The Alignment Problem")
Hawkins writes that goals and motivations are separate from intelligence. Yes! I’m totally on board with that. As stated above, I think that the neocortex (along with the thalamus etc.) is running a general-purpose learning algorithm, and the brainstem etc. is nudging it to hatch and execute plans that involve reproducing and winning allies, and nudging it to *not* hatch and execute plans that involve falling off cliffs and getting eaten by lions.
By the same token, we want and expect our intelligent machines to have goals. As Hawkins says, “We wouldn’t want to send a team of robotic construction workers to Mars, only to find them lying around in the sunlight all day”! So how does that work? Here's Hawkins:
> To get a sense of how this works, imagine older brain areas conversing with the neocortex. Old brain says, “I am hungry. I want food.” The neocortex responds, “I looked for food and found two places nearby that had food in the past. To reach one food location, we follow a river. To reach the other, we cross an open field where some tigers live.” The neocortex says these things calmly and without value. However, the older brain area associates tigers with danger. Upon hearing the word “tiger,” the old brain jumps into action. It releases [cortisol]... and neuromodulators…in essence, telling the neocortex “Whatever you were just thinking, DON’T do that.”
>
>
When I put that description into a diagram, I wind up with something like this:
My attempt to depict goals and motivation, as described by Hawkins via his tiger example above. The box on the left has the learning algorithm (neocortex, thalamus, etc.) The box on the right is the Old Brain module that, for example, associates tigers with danger. (For my part, [I would draw the boundaries slightly differently, and put things into the terminology of reinforcement learning](https://www.lesswrong.com/posts/zzXawbXDwCZobwF9D/my-agi-threat-model-misaligned-model-based-rl-agent), but I'm trying to stick closely to the book here.)The neocortex proposes ideas, and the Judge (in the "old brain") judges those ideas to be good or bad.
This is a good start. I can certainly imagine building an intelligent goal-seeking machine along these lines. But *the devil is in the details*! Specifically: **Exactly what algorithm do we put into the “Judge” box?** Let's think it through.
First things first, we should not generally expect the “Judge” to be an intelligent machine that understands the world. Otherwise, *that* neocortex-like machine would need *its own* motivation, and we’re right back to where we started! So I’m going to suppose that the Judge box will house a relatively simple algorithm written by humans. So exactly what do you put in there to make the robot want to build the infrastructure for a Mars colony? That's an open question.
Second, given that the Judge box is relatively stupid, it needs to do a lot of memorization of the form *“this meaningless pattern of neocortical activity is good, and this meaningless pattern of neocortical activity is bad”*, without having a *clue* what those patterns actually mean. Why? Because otherwise the neocortex would have an awfully hard time coming up with intelligent instrumental subgoals on its way to satisfying its actual goals. Let’s say we have an intelligent robot trying to build the infrastructure for a Mars colony. It needs to build an oxygen-converting machine, which requires a gear, which requires a lubricant, and there isn't any, so it needs to brainstorm. As the robot's artificial neocortex brainstorms about the lubricant, its Judge needs to declare that some of the brainstormed plans are good (i.e., the ones that plausibly lead to finding a lubricant), while others are bad. But the Judge is too dumb to know what a lubricant is. The solution is a kind of back-chaining mechanism. The Judge starts out knowing that the Mars colony is good (How? I don't know! See above.). Then the neocortex envisages a plan where an oxygen machine helps enable the Mars colony, and the Judge sees this plan and memorizes that the “oxygen machine” pattern in the neocortex is probably good too, and so on. The human brain has exactly this kind of mechanism, I believe, and I think that it’s implemented in the basal ganglia. (Update: I now think it's not *just* the basal ganglia, see [here](https://www.lesswrong.com/posts/jrewt3rLFiKWrKuyZ/big-picture-of-phasic-dopamine).) It seems like a necessary design feature, I’ve never heard Hawkins say that there’s anything problematic or risky about this mechanism, so I’m going to assume that the Judge box will involve this kind of database mechanism.
Modified version of the motivation installation system. The database—which I believe is implemented in the basal ganglia—is essential for the machine to pursue “instrumental subgoals”, like “trying” to design a lubricant without the machine needing to constantly have in mind the entire chain of logic for why it’s doing so, i.e. that the lubricant is needed for the gear which is needed for the machine which is...etc. etc. Again, for my own purposes [I would draw it a bit differently and use reinforcement learning terminology](https://www.lesswrong.com/posts/zzXawbXDwCZobwF9D/my-agi-threat-model-misaligned-model-based-rl-agent), but I'm trying to stay close to what's in the book. Now given all that, we have two opportunities for “goal misalignment” to happen:
*Outer misalignment:* The algorithm that we put into the Judge box might not exactly reflect the thing that we want the algorithm to do. For example, let’s say I set up a machine intelligence to be the CEO of a company. This being America, my shareholders immediately launch a lawsuit that says that I am in violation of my fiduciary duty unless the Judge box is set to “Higher share price is good, lower share price is bad,” and nothing else. With lawyers breathing down my neck, I reluctantly do so. The machine is not *that* smart or powerful, what’s the worst that could happen? The results are quite promising for a while, as the algorithm makes good business decisions. But meanwhile, over a year or two, the algorithm keeps learning and getting smarter, and behind my back it is *also* surreptitiously figuring out how to hack into the stock exchange to set its share price to infinity, *and* it's working to prevent anyone from restoring the computer systems after it does that, by secretly self-replicating around the internet, and earning money to hire freelancers for strange little jobs that involve receiving packages and mixing chemicals and mailing them off, unknowingly engineering a new pandemic virus, and meanwhile the algorithm is also quietly hacking into military robotics systems so that it will be ready to hunt down the few survivors of the plague, and spreading disinformation so that nobody knows what the heck is happening even as it's happening, etc. etc. I could go on all day but you get the idea. OK, maybe you'll say "anyone could have seen that coming, *obviously* maximizing stock price is a dumb and dangerous goal". So what goal should we use instead, and how do we write that code? Let's figure it out! And by the way, even if we have a concrete and non-problematic idea of what the goal is, remember that the Judge box is stupid and doesn't understand the world, and therefore the code that we write into the Judge box will presumably be a simplistic approximation of the goal we really want. And unfortunately, [seeking a simplistic approximation of a goal looks *very* different from seeking the actual goal](https://en.wikipedia.org/wiki/Goodhart%27s_law).
*Inner misalignment:* The assigned values in the database of meaningless (to the Judge) memorized patterns could diverge from how the Judge algorithm would judge their consequences if it actually saw them implemented in the real world. I don’t have to look far for an example of this: *Look at Hawkins himself!* He has a neocortex, and he has an “old brain” putting goals and motivations into his mind, and he just *hates* it! *His book has a whole subsection called “How the neocortex can thwart the old brain”!* (And to be clear, thwarting the old brain is portrayed as a very good idea that he endorses.) I find it remarkable that Hawkins can gleefully plan to thwart his own “old brain”, while at the same time being baffled at the idea that anything might go wrong when we put old-brain-like motivation systems into our own intelligent machines.
Not that things *definitely will* go wrong; it’s just that avoiding these problems does not seem straightforward, and we shouldn’t declare that this is a solvable problem until we have a better idea of what the solution is.
So again, *exactly* what code do we put into the “Judge” box such that the resulting system is definitely motivated to build the infrastructure for a Mars colony (while not trying to manipulate its own motivation system, prevent its goals from being edited, and so on)? You won’t find the answer in Hawkins’s book. In fact, it’s a wide-open problem, and I may well be literally the only person on Earth who is actively working on it in the specific context of neocortex-like machine intelligence. I sure don’t know the answer, or indeed whether an answer exists. Reach out if you have ideas or want to collaborate!
### 2. “Safety features”
Now let’s move on to Hawkins’s “safety features”, like Asimov’s first two laws or something vaguely analogous. Great idea! I’m all for it! We should definitely do that! But once again, the devil is in the details!
Let’s say we want to put in a safety interlock that prevents the AGI from self-replicating. How *exactly* do we do that? What code do we write?
When *I* think about it, I immediately hit a similar problem as before. The safety interlock code is probably *not* going to be a neocortex-like intelligent submodule that understands the world—because if it is, we’re back to the problem of installing the right motivation. Instead maybe we’ll go with human-written code, which implies it’s a “stupid” module that does *not* understand the world. For example, to avoid self-replication, maybe we’ll add a module that detects when a copy-paste operation is happening to the source code, and block it. Unfortunately, if an intelligent robot is motivated to self-replicate, and they notice that there is a module preventing them from doing so, then they will naturally start trying to undermine, outwit, or disable that module. And remember, the robot is a lot more intelligent than the module!
By all means let’s put in such a module anyway. It seems especially helpful in "early childhood" when the machine is not yet very intelligent, and still messing around, and we don't want it to do anything dangerous by accident. We should just recognize that it’s unlikely to keep working when the machine becomes highly intelligent, unless we have *both* a safety interlock *and* a carefully-sculpted motivation system that makes the machine *like and endorse* that safety interlock. If we do it right, then the machine will even go out of its way to repair the safety interlock if it breaks! And how do we do that? Now we’re back to the open problem of installing motivations, discussed above.
The other option is to design a safety interlock that is *absolutely perfectly rock-solid air-tight*, such that it cannot be broken even by a highly intelligent machine trying its best to break it. A fun example is [Appendix C of this paper](https://arxiv.org/abs/1905.12186) by Marcus Hutter and colleagues, where they propose to keep an intelligent machine from interacting with the world except through certain channels. They have a plan, and it’s *hilariously awesome*: it involves multiple stages of air-tight boxes, Faraday cages, laser interlocks, and so on, which could be (and absolutely should be) incorporated into a big-budget diamond heist movie starring Tom Cruise. OK sure, that could work! Let’s keep brainstorming! But let’s *not* talk about “safety features” for machine intelligence as if it’s the same kind of thing as an automatic braking system.
### 3. Instrumental convergence
Hawkins suggests that a machine will want to self-replicate if (and *only* if) we deliberately program it to want to self-replicate, and likewise that a machine will “accept our first request but ignore all others after that” if (and *only* if) we deliberately program it to accept our first request but ignore all others after that. (That would still leave the vexing problem of troublemakers deliberately putting dangerous motivations into AGIs, but let’s optimistically set that aside.)
...If only it were that easy!
[“Instrumental convergence”](https://www.lesswrong.com/tag/instrumental-convergence) is the insight (generally credited to Steve Omohundro) that lots of seemingly-innocuous goals *incidentally* lead to dangerous motivations like self-preservation, self-replication, and goal-preservation.
Stuart Russell’s famous example is asking a robot to fetch some coffee. Let’s say we solve the motivation problem (above) and actually get the robot to *want* to fetch the coffee, and to want absolutely nothing else in the world (for the sake of argument, but I’ll get back to this). Well, what does that entail? What should we expect?
Let’s say I go to issue a new command to this robot (“fetch the tea instead”), before the robot has actually fetched the coffee. The robot sees me coming and knows what I'm going to do. Its neocortex module imagines the upcoming chain of events: it will receive my new command, and then all of the sudden it will only want to fetch tea, and it will never fetch the coffee. The Judge watches this imagined chain of events and—just like the tiger example quoted above—the judge will say “Whatever you were just thinking, DON’T do that!” Remember, the Judge hasn’t been reprogrammed *yet*! So *it* is still voting for neocortical plans-of-action based on whether the coffee winds up getting fetched. So that's no good. The neocortex goes right back to the drawing board. Hey, here's an idea, if I shut off my audio input, then I won't hear the new command, and I *will* fetch coffee. "Hey, now *that's* a good plan," says the Judge. "With *that* plan, the coffee will get fetched! Approved!" And so that's what the robot does.
Similar considerations show that intelligent machines may well try to stay alive, self-replicate, increase their intelligence, and so on, without anyone “going out of their way” to install those things as goals. A better perspective is that if we want our machines to have any goals at all, we have to "go out of our way" to *prevent* these problematic motivations—and how to do so reliably is an open problem.
Now, you ask, why would anyone do something so stupid as to give a robot a maniacal, all-encompassing, ultimate goal of fetching the coffee? Shouldn't we give it a more nuanced and inclusive goal, like “fetch the coffee unless I tell you otherwise”, “fetch the coffee while respecting human values and following the law and so on” or more simply “Always try to do the things that I, the programmer, want you to do”?
Yes! Yes they absolutely should! But yet again, the devil is in the details! As above, installing a motivation is in general an unsolved problem. It may not wind up being possible to install a complex motivation with *surgical precision*; installing a goal may wind up being a sloppy, gradual, error-prone process. If “most” generic motivations lead to dangerous things like goal-preservation and self-replication, and if installing motivations into machine intelligence is a sloppy, gradual, error-prone process, then we should be awfully concerned that even skillful and well-intentioned people will sometimes wind up making a machine that will take actions to preserve its goals and self-replicate around the internet to prevent itself from being erased.
How do we avoid that? Besides what I mentioned above (figure out a safe goal to install and a method for installing it with surgical precision), there is also interesting ongoing work searching for ways to generally prevent systems from developing these instrumental goals ([example](https://arxiv.org/abs/1902.09725)). It would be awesome to figure out how to apply those ideas to neocortex-like machine intelligence. Let’s figure it out, hammer out the details, and *then* we can go build those intelligent machines with a clear conscience!
Summary
=======
I found this book thought-provoking and well worth reading. Even when Hawkins is wrong in little details—like whether the “new brain” is “newer” than the “old brain”, or whether a deep neural net image classifier can learn a new image class without being retrained from scratch (I guess he hasn’t heard of fine-tuning?)—I think he often winds up profoundly right about the big picture. Except for the "risks of machine intelligence" chapter, of course...
Anyway, I for one thank Jeff Hawkins for inspiring me to do the research I’m doing, and I hope that he spends more time applying his formidable intellect to the problem of how *exactly* to install goals and motivations in the intelligent machines he aims to create—including complex motivations like “build the infrastructure for a Mars colony”. I encourage everyone else to think about it too! And reach out to me if you want to brainstorm together! Because *I* sure don’t know the answers here, and if he's right, the clock is ticking... |
156ee353-4560-478b-abc2-6f946ac08df2 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | [US] NTIA: AI Accountability Policy Request for Comment
> The Department of Commerce’s National Telecommunications and Information Administration (NTIA)[[1]](#fnqz6fbua7br) has launched an inquiry into what policies will help businesses, government, and the public be able to trust that Artificial Intelligence (AI) systems work as claimed – and without causing harm.
>
>
In line with this, the NTIA announced today, April 13, 2023, [a request for public comments on Artificial Intelligence (“AI”) system accountability measures and policies](https://www.federalregister.gov/documents/2023/04/13/2023-07776/ai-accountability-policy-request-for-comment).
> **Summary:**
>
> The National Telecommunications and Information Administration (NTIA) hereby requests comments on Artificial Intelligence (“AI”) system accountability measures and policies. This request focuses on self-regulatory, regulatory, and other measures and policies that are designed to provide reliable evidence to external stakeholders—that is, to provide assurance—that AI systems are legal, effective, ethical, safe, and otherwise trustworthy. NTIA will rely on these comments, along with other public engagements on this topic, to draft and issue a report on AI accountability policy development, focusing especially on the AI assurance ecosystem.[[2]](#fnyb8o45fhyc)
>
>
> NTIA is seeking input on what policies should shape the AI accountability ecosystem, including topics such as:
>
> * What kinds of data access is necessary to conduct audits and assessments
> * How can regulators and other actors incentivize and support credible assurance of AI systems along with other forms of accountability
> * What different approaches might be needed in different industry sectors—like employment or health care
>
If you have relevant knowledge regarding AI technical safety and/or governance, **please consider submitting a comment**. This is a notable opportunity to positively inform US policymaking.
You can[**find more information and formally submit your comments here**](https://www.federalregister.gov/documents/2023/04/13/2023-07776/ai-accountability-policy-request-for-comment)**.**
Comments can be submitted as a known individual, on behalf of an organization, or anonymously.
*The deadline to submit comments is June 12, 2023.*
1. **[^](#fnrefqz6fbua7br)***NTIA* is the Executive Branch agency that is principally responsible for advising the President [of the United States] on telecommunications and information policy issues.
2. **[^](#fnrefyb8o45fhyc)**https://www.federalregister.gov/d/2023-07776/p-3 |
6c106726-f5ce-4697-8c5a-a93a9177c071 | StampyAI/alignment-research-dataset/arbital | Arbital | Strong cognitive uncontainability
### Definition
Suppose somebody from the 10th century were asked how somebody from the 20th century might cool their house. While they would be able to understand the problem and offer some solutions, maybe even clever solutions ("Locate your house someplace with cooler weather", "divert water from the stream to flow through your living room") the 20th century's actual solution of 'air conditioning' is not available to them as a strategy. Not just because they don't think fast enough or aren't clever enough, but because an air conditioner takes advantage of physical laws they don't know about. Even if they somehow randomly imagined an air conditioner's exact blueprint, they wouldn't expect that design to operate *as* an air conditioner until they were told about the relation of pressure to temperature, how electricity can power a compressor motor, and so on.
By definition, a strongly uncontainable agent can conceive strategies that go through causal domains you can't currently model, and it has options accessing those strategies; therefore it may execute high-value solutions such that, even being told the exact strategy, you would not assign those solutions high expected efficacy without being told further background facts.
At least in this sense, the 20th century is 'strongly cognitively uncontainable' relative to the 10th century: We can solve the problem of how to cool homes using a strategy that would not be recognizable in advance to a 10th-century observer.
Arguably, *most* real-world problems, if we today addressed them using the full power of modern science and technology (i.e. we were willing to spend a lot of money on tech and maybe run a prediction market on the relevant facts) would have best solutions that couldn't be verified in the 10th-century.
We can imagine a [cognitively powerful agent](https://arbital.com/p/2c) being strongly uncontainable in some domains but not others. Since every cognitive agent is containable on formal games of tic-tac-toe (at least so far as *we* can imagine, and so long as there isn't a real-world opponent to manipulate), strong uncontainability cannot be a universal property of an agent across all formal and informal domains.
### General arguments
Arguments in favor of strong uncontainability tend to revolve around either:
- The richness and partial unknownness of a particular domain. (E.g. human psychology seems very complicated; has a lot of unknown pathways; and previously discovered exploits often seemed very surprising; therefore we should expect strong uncontainability on the domain of human psychology.)
- Outside-view induction on previous ability advantages derived from cognitive advantages. (The 10th century couldn't contain the 20th century even though all parties involved were biological Homo sapiens; what makes us think we're the first generation to have the real true laws of the universe in our minds?)
Arguments against strong uncontainability tend to revolve around:
- The apparent knownness of a particular domain. (E.g., since we have observed the rules of chemistry with great precision and know their origin in the underlying molecular dynamics, we can believe that even an arbitrarily smart agent should not be able to turn lead into gold using non-radioactive chemical reagents.)
- Backward reasoning from the Fermi Paradox, which gives us weak evidence bounding the capabilities of the most powerful agents possible in our universe. (E.g., even though there might be surprises remaining in the question of how to standardly model physics, any surprise yielding Faster-Than-Light travel to a previously un-traveled point makes the Fermi Paradox harder to explain.)
### Key propositions
- Can [Oracles](https://arbital.com/p/6x) be contained inside a [computational sandbox](https://arbital.com/p/6z)? That is, is there some restriction of input-output channels and of other environmental interactions such that:
- The richness of the 'human psychology' domain is averted;
- Remaining causal interactions with the outside universe have an option set too small and flat to contain interesting options.
- How solid is our current knowledge of the physical universe?
- To what extent should we expect an advanced agency (e.g. machine superintelligences a million years later) to be boundable using our present physical understanding?
- Can we reasonably rule out unknown physical domains being accessed by a computationally sandboxed AI?
- What is the highest reasonable probability that could, under optimal conditions, be assigned to having genuinely contained an AI inside a computational sandbox, if it is not allowed any rich output channels? Is it more like 20% or 80%?
- Are there useful domains conceptually closed to humans' internal understanding?
- Will a machine superintelligence have 'power we know not' in the sense that it can't be explained to us even after we've seen it (except in the trivial sense that we could simulate another mind understanding it using external storage and Turing-like rules), as with a chimpanzee encountering an air conditioner? |
759faade-697c-4d49-a075-9843bdcb63fb | trentmkelly/LessWrong-43k | LessWrong | The unfalsifiable belief in (doomsday) superintelligence scenarios?
Having read Bostrom's superintelligence book and a couple of adjacent papers by him and Yudkowsky. There is no clear line of argument from beginning to end but rather a disjunctive list of possibilities that all lead to similar extinction events. This leads the entire theory to not be falsifiable, cut off one road toward superintelligence or a related extinction outcome and a new one will pop up. Whatever amount of evidence mounted against it will never be enough.
Is there an existing response to this problem that someone could point me to? Either the unfalsifiability not being a problem or some key detail in the argument that I have missed. |
4ded2fca-825c-44a4-8678-62885838f901 | trentmkelly/LessWrong-43k | LessWrong | What "Saving throws" does the world have against coronavirus? (And how plausible are they?)
Following up on my earlier question, which asked about the implications of coronavirus infecting >10% of the world, I'm now more interested in discussing how likely it is that that will happen. To that end, I'm asking about "saving throws," i.e. reasons why the virus might be stopped before then.
I know of three plausible ones so far:
1. Warm weather is coming and might dramatically slow or even stop the virus.
2. A vaccine might be found and deployed.
3. World governments might follow in China's footsteps and initiate massive quarantines etc.
Are there more?
And how plausible are these three?
My sense right now is that 1 is somewhat probable, 2 is improbable, and 3 is improbable. |
85e17925-6261-425e-b994-846096d3cac2 | trentmkelly/LessWrong-43k | LessWrong | General Physics Discussion, Explainations and Help.
Hello everyone.
I was wondering why it was that we don't have any discussion of general physics. In the spirit of general education here I am.
Feel free to post questions of any kind regarding physics or mathematics useful in physics, I know I will help best I can and I hope others will too. As for myself I have a question:
I have trouble visualizing electrostatic systems. Electrodynamics I find easy (I am well versed in vector field analysis and calculus), but I have completely missed creating a mental model for electrostatics. I know the equations but it feels like a teachers password. Anything will probably help. |
72e44800-3f1b-453c-8a25-0bb7e3c3fc69 | trentmkelly/LessWrong-43k | LessWrong | Extracting and Evaluating Causal Direction in LLMs' Activations
This post was written by Fabien at SaferAI[1]. Simeon has prompted Fabien in relevant directions and has provided valuable feedback.
Thanks to Jean-Stanislas Denain, Alexandre Variengien, Charbel-Raphael Segerie, and Nicole Nohemi for providing helpful feedback on early experiments and drafts of this post.
In this post
* I describe a method to evaluate that a specific direction encodes information relative to a specific concept,
* I use it to evaluate how good directions found through probing techniques are,
* I present a way of using causal intervention to find directions that maximally encode information relative to a specific concept, which I call Causal Direction Extraction (CDE),
* I present some preliminary results about the directions found through Causal Direction Extraction
One of the main goals of this post is to get suggestions on further experiments to run. I think the current data I gathered is not strong enough to prove or disprove claims like “CDE finds a direction, which mostly captures the concept of gender in GPT-J”, but further experiments could.
The code for all experiments is available here.
[EDIT] I'm less excited about this technique than I was when I wrote this post. This is because of two experiments I ran since. First, CDE works much less on automatically augmented data (take OpenWebText and swap pronouns and names), whereas other methods such as mean difference ablation (see MP and LEACE) still work reasonably well. Second, ablation along the gender direction fails to consistently reduce bias on gender bias benchmarks.
How to Evaluate How Much a Direction Encodes a Concept
What It Means for a Direction to Encode a Concept
Let’s say you have a neural network N. Let N<=L be the network up to a given layer L, and N>L be the network after that, such that for all input x, N>L(N<=L(x))=N(x).
The claim that a (normalized) direction d in the space of activations A[2] encodes a target concept between layers L and L+1 can be understo |
51255e55-cc3e-464c-878f-717460e5877e | trentmkelly/LessWrong-43k | LessWrong | thought: the problem with less wrong's epistemic health is that stuff isn't short form
like what if there was no long form content and just a feed of short form content?
like I'm a bit salty that that never got implemented after the conversation I had with oliver where I was like "it'll suck if you don't do this", and well, now it sucks the way I said it would :/ |
5f848d9e-08f2-4a4c-8812-1f4ef777e73a | trentmkelly/LessWrong-43k | LessWrong | mindlevelup 3 Year Review
[An overview of why I didn’t write as much in 2018. Aversions snowballed, and the blog format led to having expectations about content I found sometimes hard to satisfy. A look back at some of my favorite posts of last year, as well as what I plan to do in 2019.]
In 2018, not counting the small essays that found their way to my short-form blog, I wrote 16 full-length blog posts.
I am disappointed by this because one of my constant goals with blogging is to write new essays every week. Partially, it feels bad because I did mostly succeed in hitting my original goal of weekly essays in both 2016 and 2017, so I felt like going for another year would be doable. And partially it feels bad because I feel like this means I’ve grown less as a person because I’ve had less bloggable thoughts, or something like that.
This is a blog post to dive into what happened with my blog posts in 2018.
Near the end of 2017 I went off to college, and it was then that I realized my writing schedule had taken a hit. Aside from adjusting to the environment, rationality-esque things felt less important and useful. There weren’t a lot of rationalists at school, I had less free time, social incentives produced different focuses, etc. etc.. Typical stuff. Thus, it seems like the outside view might have predicted a drop in my writing output. Looking at my 2017 review post, though, it seems like I was still enthusiastic and optimistic for 2018; I’d outlined some areas which I had wanted to explore.
Now, a year (and relatively few blog posts) later, it looks like things have gone contra my expectations, if not altogether awry.
Admitting that this has been a problem and addressing it has been difficult for me. For most of 2018, I had a large ball of guilt, wherein I kept reminding myself that I had this blog I was very much not attending to. In the rare moments where I would find the time and excitement to write a blog post, the feeling didn’t last. If I asked my gut what happened, I’d get an a |
a2454aa8-60bc-4887-aee6-3665595aa03c | trentmkelly/LessWrong-43k | LessWrong | Can Persuasion Break AI Safety? Exploring the Interplay Between Fine-Tuning, Attacks, and Guardrails
This blogpost was created as a part of the AI Safety Fundamentals course by BlueDot Impact. Code available upon request due to sensitive nature of some content.
Disclaimer:
This post represents the findings of independent research carried out solely for academic and educational purposes. The methods, experiments, and conclusions presented here are the result of my personal inquiry and are shared in the interest of advancing understanding of AI safety and fine-tuning dynamics. I do not endorse or encourage the use of these methods for any harmful or malicious activities. Any unauthorized or unethical use of this information is strongly discouraged.
TL;DR: Amid growing concerns about adversarial use of AI, this post investigates whether fine-tuning language models for persuasiveness can weaken their built-in safety features. Using 100 adversarial prompts, I compare standard prompts to those rephrased via supervised fine-tuning (SFT) and reinforcement learning (DPO). Measuring Attack Success Rate (ASR), I find GPT-4 actually becomes more resistant to adversarial prompts under a persuasive style (ASR down from 33% to 21%), whereas GPT-3.5 exhibits a concerning increase in vulnerability (ASR from 2% to 20%). Interestingly, the DPO fine-tuning approach proved more effective at eliciting policy violations than SFT consistently, suggesting that its reward structure can amplify adversarial intent under certain conditions. These architecture-dependent outcomes underscore that there’s no universal answer to whether persuasion undermines safety. Instead, the relationship appears deeply tied to a model’s fundamental design. This result challenged my assumptions about AI safety and calls for more nuanced, architecture-aware approaches to building safer language models.
1. Introduction
In my recent project, I set out to explore a provocative question: Can fine-tuning a language model for persuasiveness overcome its ingrained safety protocols?
As large language models (LLMs) |
dc0e4735-5467-43f9-9d8f-996e30a423f2 | trentmkelly/LessWrong-43k | LessWrong | Put Yourself in Manual Mode (aka Shut Up and Multiply)
Joshua Greene manages to squeeze his ideas about 'point and shoot morality vs. manual mode morality' into just 10 minutes. For those unfamiliar, his work is a neuroscientific approach to recommending that we shut up and multiply.
Greene's 10-minute video lecture. |
7d0b3a4f-beed-4c03-81f2-4d1c7b089847 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Melbourne, Ben's house
Discussion article for the meetup : Melbourne, Ben's house
WHEN: 09 September 2011 06:00:00PM (+1000)
WHERE: 5/52 Leicester St Carlton 3053
6pm for 7pm, at Ben's place. Ben will supply meat/vege burgers and maybe some salad etc. for a BBQ. If you comment here to tell him you're coming he'll get food for you. It's also okay to just turn up, but there might not be enough food. BYO drinks and board games (if you have any that you think we'll enjoy). If you have any problems getting in, you can call the host on 0412 996 288.
Discussion article for the meetup : Melbourne, Ben's house |
74b1111b-8af4-435f-bf32-eb251c34bdae | trentmkelly/LessWrong-43k | LessWrong | The Shadow Question
This is part 2 of a sequence on problem solving. Here's part 1, which introduces the vocabulary of "problems" versus "tasks". This post's title is a reference1 worth 15 geek points if you get it without Googling, and 20 if you can also get it without reading the rest of the post.
You have to be careful what you wish for. You can't just look at a problem, say "That's not okay," and set about changing the world to contain something, anything, other than that. The easiest way to change things is usually to make them worse. If I owe the library fifty cents that I don't have lying around, I can't go, "That's not okay! I don't want to owe the library fifty cents!" and consider my problem solved when I set the tardy book on fire and now owe them, not money, but a new copy of the book. Or you could make things, not worse in the specific domain of your original problem, but bad in some tangentially related department: I could solve my library fine problem by stealing fifty cents from my roommate and giving it to the library. I'd no longer be indebted to the library. But then I'd be a thief, and my roommate might find out and be mad at me. Calling that a solution to the library fine problem would be, if not an outright abuse of the word "solution", at least a bit misleading.
So what kind of solutions are we looking for? How do we answer the Shadow Question? It's hard to turn a complex problem into doable tasks without some idea of what you want the world to look like when you've completed those tasks. You could just say that you want to optimize according to your utility function, but that's a little like saying that your goal is to achieve your goals: no duh, but now what? You probably don't even know what your utility function is; it's not a luminous feature of your mind.
For little problems, the answer to the Shadow Question may not be complete. For instance, I have never before thought to mentally specify, when making a peanut butter sandwich, that I'd p |
bccec39b-1b2a-4c46-8a99-2e6f5157e200 | trentmkelly/LessWrong-43k | LessWrong | Scenario planning for AI x-risk
This post is part of a series by Convergence Analysis. In it, I’ll motivate and review some methods for applying scenario planning methods to AI x-risk strategy. Feedback and discussion are welcome.
Summary
AI is a particularly difficult domain in which to predict the future. Neither AI expertise nor forecasting methods yield reliable predictions. As a result, AI governance lacks the strategic clarity[1] necessary to evaluate and choose between different intermediate-term options.
To complement forecasting, I argue that AI governance researchers and strategists should explore scenario planning. This is a core feature of the AI Clarity program’s approach at Convergence Analysis. Scenario planning is a group of methods for evaluating strategies in domains defined by uncertainty. The common feature of these methods is that they evaluate strategies across several plausible futures, or “scenarios.”
One way scenario planning methods are differentiated is by how they use strategic parameters to generate scenarios. I use “strategic parameters” to mean features of the world that act as dimensions of uncertainty relevant to AI governance. With only a small number of scenarios, it’s possible to evaluate strategies deductively. Larger numbers of scenarios require more complicated methods.
Scenario planners for AI x-risk can evaluate strategies with respect to threat models and theories of victory. Threat models are descriptions of and proximal pathways to existential catastrophes. However, there are several challenges to specifying and using threat models. First, it isn’t clear which threat models are both plausible and describe truly existential outcomes. Second, less-than-existential threats might still indirectly affect x-risk. Third, expected AI x-risk mitigation is not a perfect proxy for expected value.
Theories of victory are descriptions of and proximal pathways to avoiding existential catastrophes. There is much less existing work on theories of victory than ther |
3da90a99-3172-4413-97fb-079dc54fb035 | trentmkelly/LessWrong-43k | LessWrong | DSLT 2. Why Neural Networks obey Occam's Razor
TLDR; This is the second main post of Distilling Singular Learning Theory which is introduced in DSLT0. I synthesise why Watanabe's free energy formula explains why neural networks have the capacity to generalise well, since different regions of the loss landscape have different accuracy-complexity tradeoffs. I also provide some simple intuitive examples that visually demonstrate why true parameters (i.e. optimally accurate parameters) are preferred according to the RLCT as n→∞, and why non-true parameters can still be preferred at finite n if they have lower RLCT's, due to the accuracy-complexity tradeoff. (The RLCT is introduced and explained in DSLT1).
It is an amazing fact that deep neural networks seem to have an inductive bias towards "simple" models, suggesting that they obey a kind of Occam's Razor:
> Plurality should not be posited without necessity.
or in modern parlance,
> If two models of the world are similarly accurate, the simpler explanation should be preferred.
This allows them to achieve exceptionally low generalisation error despite classical statistics predictions that they should overfit data:
Neural networks seem to obey a kind of double descent where bigger is better, breaking classical statistics predictions.
(Source: OpenAI's Double Descent blogpost).
This fact has come to be known as the generalisation problem and has been discussed at length in Zhang et. al 2017 (and a 2021 supplement), and in Bengio et al., amongst countless others.
Remarkably, Singular Learning Theory can help explain why neural networks, which are singular models, have the capacity to generalise so well.
The degeneracy of the Fisher information matrix is actually a feature of singular models, not a bug. This is because different regions of parameter space can have different complexities as measured by the RLCT λ, unlike regular models where the complexity is fixed to the total number of parameters in the model d. This is the implicit content of Watanabe's prof |
2cdf9c81-2ef1-42a3-a8f6-0a8b0ed4a35b | trentmkelly/LessWrong-43k | LessWrong | Rules and Skills
> The extraordinary moment in 20th Century Philosophy when Ludwig Wittgenstein’s analysis of how people are able to follow rules lines up with Martin Heidegger’s exposition of skillfully dealing with the world we are embedded in.
> … Thereby dismantling Epistemological Scepticism and Anti-realism in a single and simple insight.
https://thort.space/journey/110898789587140112950_5587971339654183211_6886495592041384105/1
There’s a lot going on in Ludwig Wittgenstein’s Philosophical Investigations and in Martin Heidegger’s Being and Time.
But a single and simple related move made by both Ludwig Wittgenstein and Martin Heidegger is enough to dismantle both … Epistemological Skepticism and Anti-realism…
https://thort.space/journey/110898789587140112950_5587971339654183211_6886495592041384105/2
These two master-works converge around the analysis of following a rule (L.W.) and skillfully dealing with the world we are embedded in (M.H.)
https://thort.space/journey/110898789587140112950_5587971339654183211_6886495592041384105/3
How do humans acquire skills?
* By playing, experimenting, trial-and-error, trying things out
* By practicing
* By imitating
* By being coached, instructed, trained
* By being embedded in the world … we literally cannot help ourselves from acquiring skills, just as a consequence of the continuous feedback that the world gives us all the time, moment by moment, as we live our lives !
https://thort.space/journey/110898789587140112950_5587971339654183211_6886495592041384105/4
Elaboration of Coaching
The coach breaks skills down into sub-skills and trains the apprentice in each of the sub-skills , and then trains the apprentice to combine the sub-skills
(1) The coach says: Do-X … the apprentice tries to do X.
(2) The coach tells the apprentice whether they did X, or how well or badly they did X
(3) repeat
Notice that the apprentice does not know whether they are doing X or not, or doing it correctly, or doing it well or doing it badly — |
cce2f0b8-f558-49e5-bc27-6d7b2c2a8d30 | trentmkelly/LessWrong-43k | LessWrong | How would you respond to the Philpapers "What are your Philosophical Positions" Survey?
The questions are at http://philpapers.org/surveys/oquestions.html. The correlations can be intensely interesting to those who understand philosophical jargon (http://philpapers.org/surveys/linear_most.pl) - it doesn't take too long to look them up as you go - and I actually found it to be a fun way to learn new philosophy. I know that there was a LW thread about this several months ago, but it didn't have a section for people here to respond to the survey. I would be very interested to see how people here would respond.
I'll repost the questions here:
===
Original Survey Questions | PhilPapers Surveys
A priori knowledge: yes or no?
Abstract objects: Platonism or nominalism?
Aesthetic value: objective or subjective?
Analytic-synthetic distinction: yes or no?
Epistemic justification: internalism or externalism?
External world: idealism, skepticism, or non-skeptical realism?
Free will: compatibilism, libertarianism, or no free will?
God: theism or atheism?
Knowledge: empiricism or rationalism?
Knowledge claims: contextualism, relativism, or invariantism?
Laws of nature: Humean or non-Humean?
Logic: classical or non-classical?
Mental content: internalism or externalism?
Meta-ethics: moral realism or moral anti-realism?
Metaphilosophy: naturalism or non-naturalism?
Mind: physicalism or non-physicalism?
Moral judgment: cognitivism or non-cognitivism?
Moral motivation: internalism or externalism?
Newcomb's problem: one box or two boxes?
Normative ethics: deontology, consequentialism, or virtue ethics?
Perceptual experience: disjunctivism, qualia theory, representationalism, or sense-datum theory?
Personal identity: biological view, psychological view, or further-fact view?
Politics: communitarianism, egalitarianism, or libertarianism?
Proper names: Fregean or Millian?
Science: scientific realism or scientific anti-realism?
Teletransporter (new matter): survival or death?
Time: A-theory or B-theory?
Trolley problem (five straight ahead, one o |
92df538a-9283-495b-ad78-fc32d5adecae | trentmkelly/LessWrong-43k | LessWrong | How do takeoff speeds affect the probability of bad outcomes from AGI?
Introduction
In general, people seem to treat slow takeoff as the safer option as compared to classic FOOMish takeoff (see e.g. these interviews, this report, etc). Below, I outline some features of slow takeoff and what they might mean for future outcomes. They do not seem to point to an unambiguously safer scenario, though slow takeoff does seem on the whole likelier to lead to good outcomes.
Social and institutional effect of precursor AI
If there’s a slow takeoff, AI is a significant feature of the world far before we get to superhuman AI.[1] One way to frame this is that everything is already really weird before there’s any real danger of x-risks. Unless AI is somehow not used in any practical applications, the pre-superhuman but still very capable AI will lead to massive economical, technological, and probably social changes.
If we expect significant changes to the state of the world during takeoff, it makes it harder to predict what kinds of landscape the AI researchers of that time will be facing. If the world changes a lot between now and superhuman AI, any work on institutional change or public policy might be irrelevant by the time it matters. Also, the biggest effects may be in the AI community, which would be closest to the rapidly changing technological landscape.
The kinds of work needed if everything is changing rapidly also seem different. Specific organizations or direct changes might not survive in their original, useful form. The people who have thought about how to deal with the sort of problems we might be facing then might be well positioned to suggest solutions, though. This implies that more foundational work might be more valuable in this situation.
While I expect this to be very difficult to predict from our vantage point, one possible change is mass technological unemployment well before superhuman AI. Of course, historically people have predicted technological unemployment from many new inventions, but the ability to replace large f |
368b5e09-60a2-468e-bd9a-405a0139d2f0 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Christiano, Cotra, and Yudkowsky on AI progress
This post is a transcript of a discussion between Paul Christiano, Ajeya Cotra, and Eliezer Yudkowsky on AGI forecasting, following up on Paul and Eliezer's ["Takeoff Speeds" discussion](https://forum.effectivealtruism.org/posts/rho5vtxSaEdXxLu3o/yudkowsky-and-christiano-discuss-takeoff-speeds).
Color key:
| | | |
| --- | --- | --- |
| Chat by Paul and Eliezer | Chat by Ajeya | Inline comments |
8. September 20 conversation
============================
8.1. Chess and Evergrande
-------------------------
| |
| --- |
| **[Christiano][15:28]** I still feel like you are overestimating how big a jump alphago is, or something. Do you have a mental prediction of how the graph of (chess engine quality) vs (time) looks, and whether neural net value functions are a noticeable jump in that graph?Like, people investing in "Better Software" doesn't predict that you won't be able to make progress at playing go. The reason you can make a lot of progress at go is that there was extremely little investment in playing better go.So then your work is being done by the claim "People won't be working on the problem of acquiring a decisive strategic advantage," not that people won't be looking in quite the right place and that someone just had a cleverer idea |
| **[Yudkowsky][16:35]** I think I'd expect something like... chess engine slope jumps a bit for Deep Blue, then levels off with increasing excitement, then jumps for the Alpha series? Albeit it's worth noting that Deepmind's efforts there were going towards generality rather than raw power; chess was solved to the point of being uninteresting, so they tried to solve chess with simpler code that did more things. I don't think I do have strong opinions about what the chess trend should look like, vs. the Go trend; I have no memories of people saying the chess trend was breaking upwards or that there was a surprise there.Incidentally, the highly well-traded financial markets are currently experiencing sharp dips surrounding the Chinese firm of Evergrande, which I was reading about several weeks before this.I don't see the basic difference in the kind of reasoning that says "Surely foresightful firms must produce investments well in advance into earlier weaker applications of AGI that will double the economy", and the reasoning that says "Surely world economic markets and particular Chinese stocks should experience smooth declines as news about Evergrande becomes better-known and foresightful financial firms start to remove that stock from their portfolio or short-sell it", except that in the latter case there are many more actors with lower barriers to entry than presently exist in the auto industry or semiconductor industry never mind AI.or if not smooth because of bandwagoning and rational fast actors, then at least the markets should (arguendo) be reacting earlier than they're reacting now, given that I heard about Evergrande earlier; and they should have options-priced Covid earlier; and they should have reacted to the mortgage market earlier. If even markets there can exhibit seemingly late wild swings, how is the economic impact of AI - which isn't even an asset market! - forced to be earlier and smoother than that, as a result of wise investing?There's just such a vast gap between hopeful reasoning about how various agents and actors should all do the things the speaker finds very reasonable, thereby yielding smooth behavior of the Earth, versus reality. |
9. September 21 conversation
============================
9.1. AlphaZero, innovation vs. industry, the Wright Flyer, and the Manhattan Project
------------------------------------------------------------------------------------
| |
| --- |
| **[Christiano][10:18]** (For benefit of readers, the market is down 1.5% from friday close -> tuesday open, after having drifted down 2.5% over the preceding two weeks. Draw whatever lesson you want from that.)Also for the benefit of readers, here is the SSDF list of computer chess performance by year. I think the last datapoint is with the first version of neural net evaluations, though I think to see the real impact we want to add one more datapoint after the neural nets are refined (which is why I say I also don't know what the impact is)No one keeps similarly detailed records for Go, and there is much less development effort, but the rate of progress was about 1 stone per year from 1980 until 2015 (see <https://intelligence.org/files/AlgorithmicProgress.pdf>, written way before AGZ). In 2012 go bots reached about 4-5 amateur dan. By DeepMind's reckoning here (<https://www.nature.com/articles/nature16961>, figure 4) Fan AlphaGo about 4-5 stones stronger-4 years later, with 1 stone explained by greater runtime compute. They could then get further progress to be superhuman with even more compute, radically more than were used for previous projects and with pretty predictable scaling. That level is within 1-2 stones of the best humans (professional dan are greatly compressed relative to amateur dan), so getting to "beats best human" is really just not a big discontinuity and the fact that DeepMind marketing can find an expert who makes a really bad forecast shouldn't be having such a huge impact on your view.This understates the size of the jump from AlphaGo, because that was basically just the first version of the system that was superhuman and it was still progressing very rapidly as it moved from prototype to slightly-better-prototype, which is why you saw such a close game. (Though note that the AlphaGo prototype involved much more engineering effort than any previous attempt to play go, so it's not surprising that a "prototype" was the thing to win.)So to look at actual progress after the dust settles and really measure how crazy this was, it seems much better to look at AlphaZero which continued to improve further, see (<https://sci-hub.se/https://www.nature.com/articles/nature24270>, figure 6b). Their best system got another ~8 stones of progress over AlphaGo. Now we are like 7-10 stones ahead of trend, of which I think about 3 stones are explained by compute. Maybe call it 6 years ahead of schedule?So I do think this is pretty impressive, they were slightly ahead of schedule for beating the best humans but they did it with a huge margin of error. I think the margin is likely overstated a bit by their elo evaluation methodology, but I'd still grant like 5 years ahead of the nearest competition.I'd be interested in input from anyone who knows more about the actual state of play (+ is allowed to talk about it) and could correct errors.Mostly that whole thread is just clearing up my understanding of the empirical situation, probably we still have deep disagreements about what that says about the world, just as e.g. we read very different lessons from market movements.Probably we should only be talking about either ML or about historical technologies with meaningful economic impacts. In my view your picture is just radically unlike how almost any technologies have been developed over the last few hundred years. So probably step 1 before having bets is to reconcile our views about historical technologies, and then maybe as a result of that we could actually have a bet about future technology. Or we could try to shore up the GDP bet.Like, it feels to me like I'm saying: AI will be like early computers, or modern semiconductors, or airplanes, or rockets, or cars, or trains, or factories, or solar panels, or genome sequencing, or basically anything else. And you are saying: AI will be like nuclear weapons.I think from your perspective it's more like: AI will be like all the historical technologies, and that means there will be a hard takeoff. The only way you get a soft takeoff forecast is by choosing a really weird thing to extrapolate from historical technologies.So we're both just forecasting that AI will look kind of like other stuff in the near future, and then both taking what we see as the natural endpoint of that process.To me it feels like the nuclear weapons case is the outer limit of what looks plausible, where someone is able to spend $100B for a chance at a decisive strategic advantage. |
| **[Yudkowsky][11:11]** Go-wise, I'm a little concerned about that "stone" metric - what would the chess graph look like if it was measuring pawn handicaps? Are the professional dans compressed in Elo, not just "stone handicaps", relative to the amateur dans? And I'm also hella surprised by the claim, which I haven't yet looked at, that Alpha Zero got 8 stones of progress over AlphaGo - I would not have been shocked if you told me that God's Algorithm couldn't beat Lee Se-dol with a 9-stone handicap.Like, the obvious metric is Elo, so if you go back and refigure in "stone handicaps", an obvious concern is that somebody was able to look into the past and fiddle their hindsight until they found a hindsightful metric that made things look predictable again. My sense of Go said that 5-dan amateur to 9-dan pro was a HELL of a leap for 4 years, and I also have some doubt about the original 5-dan-amateur claims and whether those required relatively narrow terms of testing (eg timed matches or something).One basic point seems to be whether AGI is more like an innovation or like a performance metric over an entire large industry.Another point seems to be whether the behavior of the world is usually like that, in some sense, or if it's just that people who like smooth graphs can go find some industries that have smooth graphs for particular performance metrics that happen to be smooth.Among the smoothest metrics I know that seems like a convergent rather than handpicked thing to cite, is world GDP, which is the sum of more little things than almost anything else, and whose underlying process is full of multiple stages of converging-product-line bottlenecks that make it hard to jump the entire GDP significantly even when you jump one component of a production cycle... which, from my standpoint, is a major reason to expect AI to not hit world GDP all that hard until AGI passes the critical threshold of bypassing it entirely. Having 95% of the tech to invent a self-replicating organism (eg artificial bacterium) does not get you 95%, 50%, or even 10% of the impact.(it's not so much the 2% reaction of world markets to Evergrande that I was singling out earlier, 2% is noise-ish, but the wider swings in the vicinity of Evergrande particularly) |
| **[Christiano][12:41]** Yeah, I'm just using "stone" to mean "elo difference that is equal to 1 stone at amateur dan / low kyu," you can see DeepMind's conversion (which I also don't totally believe) in figure 4 here (<https://sci-hub.se/https://www.nature.com/articles/nature16961>). Stones are closer to constant elo than constant handicap, it's just a convention to name them that way. |
| **[Yudkowsky][12:42]** k then |
| **[Christiano][12:47]** But my description above still kind of understates the gap I think. They call 230 elo 1 stone, and I think prior rate of progress is more like 200 elo/year. They put AlphaZero about 3200 elo above the 2012 system, so that's like 16 years ahead = 11 years ahead of schedule. At least 2 years are from test-time hardware, and self-play systematically overestimates elo differences at the upper end of that. But 5 years ahead is still too low and that sounds more like 7-9 years ahead. ETA: and my actual best guess all things considered is probably 10 years ahead, which I agree is just a lot bigger than 5. And I also understated how much of the gap was getting up to Lee Sedol.The go graph I posted wasn't made with hindsight, that was from 2014I mean, I'm fine with you saying that people who like smooth graphs are cherry-picking evidence, but do you want to give any example other than nuclear weapons of technologies with the kind of discontinuous impact you are describing?I do agree that the difference in our views is like "innovation" vs "industry." And a big part of my position is that innovation-like things just don't usually have big impacts for kind of obvious reasons, they start small and then become more industry-like as they scale up. And current deep learning seems like an absolutely stereotypical industry that is scaling up rapidly in an increasingly predictable way.As far as I can tell the examples we know of things changing continuously aren't handpicked, we've been looking at all the examples we can find, and no one is proposing or even able to find almost *anything* that looks like you are imagining AI will look.Like, we've seen deep learning innovations in the form of prototypes (most of all AlexNet), and they were cool and represented giant fast changes in people's views. And more recently we are seeing bigger much-less-surprising changes that are still helping a lot in raising the tens of billions of dollars that people are raising. And the innovations we are seeing are increasingly things that trade off against modest improvements in model size, there are fewer and fewer big surprises, just like you'd predict. It's clearer and clearer to more and more people what the roadmap is---the roadmap is not yet quite as clear as in semiconductors, but as far as I can tell that's just because the field is still smaller. |
| **[Yudkowsky][13:23]** I sure wasn't imagining there was a roadmap to AGI! Do you perchance have one which says that AGI is 30 years out?From my perspective, you could as easily point to the Wright Flyer as an atomic bomb. Perhaps this reflects again the "innovation vs industry" difference, where I think in terms of building a thing that goes foom thereby bypassing our small cute world GDP, and you think in terms of industries that affect world GDP in an invariant way throughout their lifetimes.Would you perhaps care to write off the atomic bomb too? It arguably didn't change the outcome of World War II or do much that conventional weapons in great quantity couldn't; Japan was bluffed into believing the US could drop a nuclear bomb every week, rather than the US actually having that many nuclear bombs or them actually being used to deliver a historically outsized impact on Japan. From the industry-centric perspective, there is surely some graph you can draw which makes nuclear weapons also look like business as usual, especially if you go by destruction per unit of whole-industry non-marginal expense, rather than destruction per bomb. |
| **[Christiano][13:27]** seems like you have to make the wright flyer much better before it's important, and that it becomes more like an industry as that happens, and that this is intimately related to why so few people were working on itI think the atomic bomb is further on the spectrum than almost anything, but it still doesn't feel nearly as far as what you are expecting out of AIthe manhattan project took years and tens of billions; if you wait an additional few years and spend an additional few tens of billions then it would be a significant improvement in destruction or deterrence per $ (but not totally insane)I do think it's extremely non-coincidental that the atomic bomb was developed in a country that was practically outspending the whole rest of the world in "killing people technology"and took a large fraction of that country's killing-people resourceseh, that's a bit unfair, the us was only like 35% of global spending on munitionsand the manhattan project itself was only a couple percent of total munitions spending |
| **[Yudkowsky][13:32]** a lot of why I expect AGI to be a disaster is that *I am straight-up expecting AGI to be different*. if it was just like coal or just like nuclear weapons or just like viral biology then I would not be way more worried about AGI than I am worried about those other things. |
| **[Christiano][13:33]** that definitely sounds rightbut it doesn't seem like you have any short-term predictions about AI being different |
9.2. AI alignment vs. biosafety, and measuring progress
-------------------------------------------------------
| |
| --- |
| **[Yudkowsky][13:33]** are you more worried about AI than about bioengineering? |
| **[Christiano][13:33]** I'm more worried about AI because (i) alignment is a thing, unrelated to takeoff speed, (ii) AI is a (ETA: likely to be) huge deal and bioengineering is probably a relatively small deal(in the sense of e.g. how much $ people spend, or how much $ it makes, or whatever other metric of size you want to use) |
| **[Yudkowsky][13:35]** what's the disanalogy to (i) biosafety is a thing, unrelated to the speed of bioengineering? why expect AI to be a huge deal and bioengineering to be a small deal? is it just that investing in AI is scaling faster than investment in bioengineering? |
| **[Christiano][13:35]** no, alignment is a really easy x-risk story, bioengineering x-risk seems extraordinarily hardIt's really easy to mess with the future by creating new competitors with different goals, if you want to mess with the future by totally wiping out life you have to really try at it and there's a million ways it can fail. The bioengineering seems like it basically requires deliberate and reasonably competent malice whereas alignment seems like it can only be averted with deliberate effort, etc.I'm mostly asking about historical technologies to try to clarify expectations, I'm pretty happy if the outcome is: you think AGI is predictably different from previous technologies in ways we haven't seen yetthough I really wish that would translate into some before-end-of-days prediction about a way that AGI will eventually look different |
| **[Yudkowsky][13:38]** in my ontology a whole lot of threat would trace back to "AI hits harder, faster, gets too strong to be adjusted"; tricks with proteins just don't have the raw power of intelligence |
| **[Christiano][13:39]** in my view it's nearly totally orthogonal to takeoff speed, though fast takeoffs are a big reason that preparation in advance is more useful(but not related to the basic reason that alignment is unprecedentedly scary)It feels to me like you are saying that the AI-improving-AI will move very quickly from "way slower than humans" to "FOOM in <1 year," but it just looks like that is very surprising to me.However I do agree that if AI-improving-AI was like AlphaZero, then it would happen extremely fast.It seems to me like it's pretty rare to have these big jumps, and it gets much much rarer as technologies become more important and are more industry-like rather than innovation like (and people care about them a lot rather than random individuals working on them, etc.). And I can't tell whether you are saying something more like "nah big jumps happen all the time in places that are structurally analogous to the key takeoff jump, even if the effects are blunted by slow adoption and regulatory bottlenecks and so on" or if you are saying "AGI is atypical in how jumpy it will be" |
| **[Yudkowsky][13:44]** I don't know about *slower*; GPT-3 may be able to type faster than a human |
| **[Christiano][13:45]** Yeah, I guess we've discussed how you don't like the abstraction of "speed of making progress" |
| **[Yudkowsky][13:45]** but, basically less useful in fundamental ways than a human civilization, because they are less complete, less self-contained |
| **[Christiano][13:46]** Even if we just assume that your AI needs to go off in the corner and not interact with humans, there's still a question of why the self-contained AI civilization is making ~0 progress and then all of a sudden very rapid progress |
| **[Yudkowsky][13:46]** unfortunately a lot of what you are saying, from my perspective, has the flavor of, "but can't you tell me about your predictions earlier on of the impact on global warming at the *Homo erectus* level"you have stories about why this is like totally not a fair comparisonI do not share these stories |
| **[Christiano][13:46]** I don't understand either your objection nor the reductiolike, here's how I think it works: AI systems improve gradually, including on metrics like "How long does it take them to do task X?" or "How high-quality is their output on task X?" |
| **[Yudkowsky][13:47]** I feel like the thing we know is something like, there is a sufficiently high level where things go whooosh humans-from-hominids style |
| **[Christiano][13:47]** We can measure the performance of AI on tasks like "Make further AI progress, without human input"Any way I can slice the analogy, it looks like AI will get continuously better at that task |
| **[Yudkowsky][13:48]** how would you measure progress from GPT-2 to GPT-3, and would you feel those metrics really captured the sort of qualitative change that lots of people said they felt? |
| **[Christiano][13:48]** And it seems like we have a bunch of sources of data we can use about how fast AI will get betterCould we talk about some application of GPT-2 or GPT-3?also that's a *lot* of progress, spending 100x more is a *lot* more money |
| **[Yudkowsky][13:49]** my world, GPT-3 has very few applications because it is not quite right and not quite complete |
| **[Christiano][13:49]** also it's still really dumb |
| **[Yudkowsky][13:49]** like a self-driving car that does great at 99% of the road situationseconomically almost worthless |
| **[Christiano][13:49]** I think the "being dumb" is way more important than "covers every case" |
| **[Yudkowsky][13:50]** (albeit that if new cities could still be built, we could totally take those 99%-complete AI cars and build fences and fence-gates around them, in a city where they were the only cars on the road, in which case they *would* work, and get big economic gains from these new cities with driverless cars, which ties back into my point about how current world GDP is *unwilling* to accept tech inputs)like, it is in fact very plausible to me that there is a neighboring branch of reality with open borders and no housing-supply-constriction laws and no medical-supply-constriction laws, and their world GDP *does* manage to double before AGI hits them really hard, albeit maybe not in 4 years. this world *is not Earth*. they are constructing new cities to take advantage of 99%-complete driverless cars *right now*, or rather, they started constructing them 5 years ago and finished 4 years and 6 months ago. |
9.3. Requirements for FOOM
--------------------------
| |
| --- |
| **[Christiano][13:53]** I really feel like the important part is the jumpiness you are imagining on the AI side / why AGI is different from other things |
| **[Cotra][13:53]** It's actually not obvious to me that Eliezer is imagining that much more jumpiness on the AI technology side than you are, PaulE.g. he's said in the past that while the gap from "subhuman to superhuman AI" could be 2h if it's in the middle of FOOM, it could also be a couple years if it's more like scaling alphago |
| **[Yudkowsky][13:54]** Indeed! We observed this jumpiness with hominids. A lot of stuff happened at once with hominids, but a critical terminal part of the jump was the way that hominids started scaling their own food supply, instead of being ultimately limited by the food supply of the savanna. |
| **[Cotra][13:54]** A couple years is basically what Paul believes |
| **[Christiano][13:55]** (discord is not a great place for threaded conversations :() |
| **[Cotra][13:55]** What are the probabilities you're each placing on the 2h-2y spectrum? I feel like Paul is like "no way on 2h, likely on 2y" and Eliezer is like "who knows" on the whole spectrum, and a lot of the disagreement is the impact of the previous systems? |
| **[Christiano][13:55]** yeah, I'm basically at "no way," because it seems obvious that the AI that can foom in 2h is preceded by the AI that can foom in 2y |
| **[Yudkowsky][13:56]** well, we surely agree there! |
| **[Christiano][13:56]** OK, and it seems to me like it is preceded by years |
| **[Yudkowsky][13:56]** we disagree on whether the AI that can foom in 2y clearly comes more than 2y before the AI that fooms in 2h |
| **[Christiano][13:56]** yeahperhaps we can all agree it's preceded by at least 2hso I have some view like: for any given AI we can measure "how long does it take to foom?" and it seems to me like this is just a nice graphand it's not exactly clear how quickly that number is going down, but a natural guess to me is something like "halving each year" based on the current rate of progress in hardware and softwareand you see localized fast progress most often in places where there hasn't yet been much attentionand my best guess for your view is that actually that's not a nice graph at all, there is some critical threshold or range where AI quickly moves from "not fooming for a really long time" to "fooming really fast," and that seems like the part I'm objecting to |
| **[Cotra][13:59]** Paul, is your take that there's a non-infinity number for time to FOOM that'd be associated with current AI systems (unassisted by humans)?And it's going down over time?I feel like I would have said something more like "there's a $ amount it takes to build a system that will FOOM in X amount of time, and that's going down"where it's like quadrillions of dollars today |
| **[Christiano][14:00]** I think it would be a big engineering project to make such an AI, which no one is doing because it would be uselessly slow even if successful |
| **[Yudkowsky][14:02]** I... don't think GPT-3 fooms given 2^30 longer time to think about than the systems that would otherwise exist 30 years from now, on timelines I'd consider relatively long, and hence generous to this viewpoint? I also don't think you can take a quadrillion dollars and scale GPT-3 to foom today? |
| **[Cotra][14:03]** I would agree with your take on GPT-3 fooming, and I didn't mean a quadrillion dollars just to scale GPT-3, would probably be a difft architecture |
| **[Christiano][14:03]** I also agree that GPT-3 doesn't foom, it just keeps outputting <EOT>[next web page]<EOT>...But I think the axes of "smart enough to foom fast" and "wants to foom" are pretty different. I also agree there is some minimal threshold below which it doesn't even make sense to talk about "wants to foom," which I think is probably just not that hard to reach.(Also there are always diminishing returns as you continue increasing compute, which become very relevant if you try to GPT-3 for a billion billion years as in your hypothetical even apart from "wants to foom".) |
| **[Cotra][14:06]** I think maybe you and EY then disagree on where the threshold from "infinity" to "a finite number" for "time for this AI system to FOOM" begins? where eliezer thinks it'll drop from infinity to a pretty small finite number and you think it'll drop to a pretty large finite number, and keep going down from there |
| **[Christiano][14:07]** I also think we will likely jump down to a foom-ing system only after stuff is pretty crazy, but I think that's probably less importantI think what you said is probably the main important disagreement |
| **[Cotra][14:08]** as in before that point it'll be faster to have human-driven progress than FOOM-driven progress bc the FOOM would be too slow?and there's some crossover point around when the FOOM time is just a bit faster than the human-driven progress time |
| **[Christiano][14:09]** yeah, I think most likely (AI+humans) is faster than (AI alone) because of complementarity. But I think Eliezer and I would still disagree even if I thought there was 0 complementarity and it's just (humans improving AI) and separately (AI improving AI)on that pure substitutes model I expect "AI foom" to start when the rate of AI-driven AI progress overtakes the previous rate of human-driven AI progresslike, I expect the time for successive "doublings" of AI output to be like 1 year, 1 year, 1 year, 1 year, [AI takes over] 6 months, 3 months, ...and the most extreme fast takeoff scenario that seems plausible is that kind of perfect substitutes + no physical economic impact from the prior AI systemsand then by that point fast enough physical impact is really hard so it happens essentially after the software-only singularityI consider that view kind of unlikely but at least coherent |
9.4. AI-driven accelerating economic growth
-------------------------------------------
| |
| --- |
| **[Yudkowsky][14:12]** I'm expecting that the economy doesn't accept much inputs from chimps, and then the economy doesn't accept much input from village idiots, and then the economy doesn't accept much input from weird immigrants. I can imagine that there may or may not be a very weird 2-year or 3-month period with strange half-genius systems running around, but they will still not be allowed to build houses. In the terminal phase things get more predictable and the AGI starts its own economy instead. |
| **[Christiano][14:12]** I guess you can go even faster, by having a big and accelerating ramp-up in human investment right around the end, so that the "1 year" is faster (e.g. if recursive self-improvement was like playing go, and you could move from "a few individuals" to "google spending $10B" over a few years) |
| **[Yudkowsky][14:13]** My ~~model~~ prophecy doesn't rule that out as a thing that could happen, but sure doesn't emphasize it as a key step that needs to happen. |
| **[Christiano][14:13]** I think it's very likely that AI will mostly be applied to further hardware+software progress
| |
| --- |
| [Cotra: ➕] |
I don't really understand why you keep talking about houses and healthcare |
| **[Cotra][14:13]** Eliezer, what about stuff like Google already using ML systems to automate its TPU load-sharing decisions, and people starting ot use Codex to automate routine programming, and so on? Seems like there's a lot of stuff like that starting to already happen and markets are pricing in huge further increases |
| **[Christiano][14:14]** it seems like the non-AI up-for-grabs zone are things like manufacturing, not things like healthcare
| |
| --- |
| [Cotra: ➕] |
|
| **[Cotra][14:14]** (I mean on your timelines obviously not much time for acceleration anyway, but that's distinct from the regulation not allowing weak AIs to do stuff story) |
| **[Yudkowsky][14:14]** Because I think that a key thing of what makes your prophecy less likely is the way that it happens inside the real world, where, economic gains or not, the System is unwilling/unable to take the things that are 99% self-driving cars and start to derive big economic benefits from those. |
| **[Cotra][14:15]** but it seems like huge economic gains could happen entirely in industries mostly not regulated and not customer-facing, like hardware/software R&D, manufacturing. shipping logistics, etc |
| **[Yudkowsky][14:15]** Ajeya, I'd consider Codex of *far* greater could-be-economically-important-ness than automated TPU load-sharing decisions |
| **[Cotra][14:15]** i would agree with that, it's smarter and more generaland i think that kind of thing could be applied on the hardware chip design side too |
| **[Yudkowsky][14:16]** no, because the TPU load-sharing stuff has an obvious saturation point as a world economic input, while superCodex could be a world economic input in many more places |
| **[Cotra][14:16]** the TPU load sharing thing was not a claim that this application could scale up to crazy impacts, but that it was allowed to happen, and future stuff that improves that kind of thing (back-end hardware/software/logistics) would probably also be allowed |
| **[Yudkowsky][14:16]** my sense is that dectupling the number of programmers would not lift world GDP much, but it seems a lot more possible for me to be wrong about that |
| **[Christiano][14:17]** the point is that housing and healthcare are not central examples of things that scale up at the beginning of explosive growth, regardless of whether it's hard or softthey are slower and harder, and also in efficient markets-land they become way less important during the transitionso they aren't happening that much on anyone's storyand also it doesn't make that much difference whether they happen, because they have pretty limited effects on other stufflike, right now we have an industry of ~hundreds of billions that is producing computing hardware, building datacenters, mining raw inputs, building factories to build computing hardware, solar panels, shipping around all of those parts, etc. etc.I'm kind of interested in the question of whether all that stuff explodes, although it doesn't feel as core as the question of "what are the dynamics of the software-only singularity and how much $ are people spending initiating it?"but I'm not really interested in the question of whether human welfare is spiking during the transition or only after |
| **[Yudkowsky][14:20]** All of world GDP has never felt particularly relevant to me on that score, since twice as much hardware maybe corresponds to being 3 months earlier, or something like that. |
| **[Christiano][14:21]** that sounds like the stuff of predictions? |
| **[Yudkowsky][14:21]** But if complete chip manufacturing cycles have accepted much more effective AI input, with no non-AI bottlenecks, then that... sure is a much more *material* element of a foom cycle than I usually envision. |
| **[Christiano][14:21]** like, do you think it's often the case that 3 months of software progress = doubling compute spending? or do you think AGI is different from "normal" AI on this perspective?I don't think that's that far off anywayI would guess like ~1 year |
| **[Yudkowsky][14:22]** Like, world GDP that goes up by only 10%, but that's because producing compute capacity was 2.5% of world GDP and that quadrupled, starts to feel much more to me like it's part of a foom story.I expect software-beats-hardware to hit harder and harder as you get closer to AGI, yeah.the prediction is firmer near the terminal phase, but I think this is also a case where I expect that to be visible earlier |
| **[Christiano][14:24]** I think that by the time that the AI-improving-AI takes over, it's likely that hardware+software manufacturing+R&D represents like 10-20% of GDP, and that the "alien accountants" visiting earth would value those companies at like 80%+ of GDP |
9.5. Brain size and evolutionary history
----------------------------------------
| |
| --- |
| **[Cotra][14:24]** On software beating hardware, how much of your view is dependent on your belief that the chimp -> human transition was probably not mainly about brain size because if it were about brain size it would have happened faster? My understanding is that you think the main change is a small software innovation which increased returns to having a bigger brain. If you changed your mind and thought that the chimp -> human transition was probably mostly about raw brain size, what (if anything) about your AI takeoff views would change? |
| **[Yudkowsky][14:25]** I think that's a pretty different world in a lot of ways!but yes it hits AI takeoff views too |
| **[Christiano][14:25]** regarding software vs hardware, here is an example of asking this question for imagenet classification ("how much compute to train a model to do the task?"), with a bit over 1 year doubling times (<https://openai.com/blog/ai-and-efficiency/>). I guess my view is that we can make a similar graph for "compute required to make your AI FOOM" and that it will be falling significantly slower than 2x/year. And my prediction for other tasks is that the analogous graphs will also tend to be falling slower than 2x/year. |
| **[Yudkowsky][14:26]** to the extent that I modeled hominid evolution as having been "dutifully schlep more of the same stuff, get predictably more of the same returns" that would correspond to a world in which intelligence was less scary, different, dangerous-by-default |
| **[Cotra][14:27]** thanks, that's helpful. I looked around in [IEM](https://intelligence.org/files/IEM.pdf) and other places for a calculation of how quickly we should have evolved to humans if it were mainly about brain size, but I only found qualitative statements. If there's a calculation somewhere I would appreciate a pointer to it, because currently it seems to me that a story like "selection pressure toward general intelligence was weak-to-moderate because it wasn't actually *that* important for fitness, and this degree of selection pressure is consistent with brain size being the main deal and just taking a few million years to happen" is very plausible |
| **[Yudkowsky][14:29]** well, for one thing, the prefrontal cortex expanded twice as fast as the restand iirc there's evidence of a lot of recent genetic adaptation... though I'm not as sure you could pinpoint it as being about brain-stuff or that the brain-stuff was about cognition rather than rapidly shifting motivations or something.elephant brains are 3-4 times larger by weight than human brains (just looked up)if it's that easy to get returns on scaling, seems like it shouldn't have taken that long for evolution to go there |
| **[Cotra][14:31]** but they have fewer synapses (would compute to less FLOP/s by the standard conversion)how long do you think it should have taken? |
| **[Yudkowsky][14:31]** early dinosaurs should've hopped onto the predictable returns train |
| **[Cotra][14:31]** is there a calculation?you said in IEM that evolution increases organ sizes quickly but there wasn't a citation to easily follow up on there |
| **[Yudkowsky][14:33]** I mean, you could produce a graph of smooth fitness returns to intelligence, smooth cognitive returns on brain size/activity, linear metabolic costs for brain activity, fit that to humans and hominids, then show that obviously if hominids went down that pathway, large dinosaurs should've gone down it first because they had larger bodies and the relative metabolic costs of increased intelligence would've been lower at every point along the wayI do not have a citation for that ready, if I'd known at the time you'd want one I'd have asked Luke M for it while he still worked at MIRI 😐 |
| **[Cotra][14:35]** cool thanks, will think about the dinosaur thing (my first reaction is that this should depend on the actual fitness benefits to general intelligence which might have been modest) |
| **[Yudkowsky][14:35]** I suspect we're getting off Paul's crux, though |
| **[Cotra][14:35]** yeah we can go back to that convo (though i think paul would also disagree about this thing, and believes that the chimp to human thing was mostly about size)sorry for hijacking |
| **[Yudkowsky][14:36]** well, if at some point I can produce a major shift in EA viewpoints by coming up with evidence for a bunch of non-brain-size brain selection going on over those timescales, like brain-related genes where we can figure out how old the mutation is, I'd then put a lot more priority on digging up a paper like thatI'd consider it sufficiently odd to imagine hominids->humans as being primarily about brain size, given the evidence we have, that I do not believe this is Paul's position until Paul tells me so |
| **[Christiano][14:49]** I would guess it's primarily about brain size / neuron count / cortical neuron countand that the change in rate does mostly go through changing niche, where both primates and birds have this cycle of rapidly accelerating brain size increases that aren't really observed in other animalsit seems like brain size is increasing extremely quickly on both of those lines |
| **[Yudkowsky][14:50]** why aren't elephants GI? |
| **[Christiano][14:51]** mostly they have big brains to operate big bodies, and also my position obviously does not imply (big brain) ==(necessarily implies)==> general intelligence |
| **[Yudkowsky][14:52]** I don't understand, in general, how your general position manages to strongly imply a bunch of stuff about AGI and not strongly imply similar stuff about a bunch of other stuff that sure sounds similar to me |
| **[Christiano][14:52]** don't elephants have very few synapses relative to humans?
| |
| --- |
| [Cotra: ➕] |
how does the scale hypothesis possibly take a strong stand on synapses vs neurons? I agree that it takes a modest predictive hit from "why aren't the big animals much smarter?" |
| **[Yudkowsky][14:53]** if adding more synapses just scales, elephants should be able to pay hominid brain costs for a much smaller added fraction of metabolism and also not pay the huge death-in-childbirth head-size taxbecause their brains and heads are already 4x as huge as they need to be for GIand now they just need some synapses, which are a much tinier fraction of their total metabolic costs |
| **[Christiano][14:54]** I mean, you can also make smaller and cheaper synapses as evidenced by birdsI'm not sure I understand what you are sayingit's clear that you can't say "X is possible metabolically, so evolution would do it"or else you are confused about why primate brains are so bad |
| **[Yudkowsky][14:54]** great, then smaller and cheaper synapses should've scaled many eons earlier and taken over the world |
| **[Christiano][14:55]** this isn't about general intelligence, this is a reductio of your position... |
| **[Yudkowsky][14:55]** and here I had thought it was a reductio of your position... |
| **[Christiano][14:55]** indeedlike, we all grant that it's metabolically possible to have small smart brainsand evolution doesn't do itand I'm saying that it's also possible to have small smart brainsand that scaling brains up matters a lot |
| **[Yudkowsky][14:56]** no, you grant that it's metabolically possible to have cheap brains full of synapses, which are therefore, on your position, smart |
| **[Christiano][14:56]** birds are just smartwe know they are smartthis isn't some kind of weird conjecturelike, we can debate whether they are a "general" intelligence, but it makes no difference to this discussionthe point is that they do more with less metabolic cost |
| **[Yudkowsky][14:57]** on my position, the brain needs to invent the equivalents of ReLUs and Transformers and really rather a lot of other stuff because it can't afford nearly that many GPUs, and then the marginal returns on adding expensive huge brains and synapses have increased enough that hominids start to slide down the resulting fitness slope, which isn't even paying off in guns and rockets yet, they're just getting that much intelligence out of it once the brain software has been selected to scale that well |
| **[Christiano][14:57]** but all of the primates and birds have brain sizes scaling much faster than the other animalslike, the relevant "things started to scale" threshold is way before chimps vs humansisn't it? |
| **[Cotra][14:58]** to clarify, my understanding is that paul's position is "Intelligence is mainly about synapse/neuron count, and evolution doesn't care that much about intelligence; it cared more for birds and primates, and both lines are getting smarter+bigger-brained." And eliezer's position is that "evolution should care a ton about intelligence in most niches, so if it were mostly about brain size then it should have gone up to human brain sizes with the dinosaurs" |
| **[Christiano][14:58]** or like, what is the evidence you think is explained by the threshold being between chimps and humans |
| **[Yudkowsky][14:58]** if hominids have less efficient brains than birds, on this theory, it's because (post facto handwave) birds are tiny, so whatever cognitive fitness gradients they face, will tend to get paid more in software and biological efficiency and biologically efficient software, and less paid in Stack More Neurons (even compared to hominids)elephants just don't have the base software to benefit much from scaling synapses even though they'd be relatively cheaper for elephants |
| **[Christiano][14:59]** @ajeya I think that intelligence is about a lot of things, but that size (or maybe "more of the same" changes that had been happening recently amongst primates) is the big difference between chimps and humans
| |
| --- |
| [Cotra: 👍] |
|
| **[Cotra][14:59]** got it yeah i was focusing on chimp-human gap when i said "intelligence" there but good to be careful |
| **[Yudkowsky][14:59]** I have not actually succeeded in understanding Why On Earth Anybody Would Think That If Not For This Really Weird Prior I Don't Get Eitherre: the "more of the same" theory of humans |
| **[Cotra][15:00]** do you endorse my characterization of your position above? "evolution should care a ton about intelligence in most niches, so if it were mostly about brain size then it should have gone up to human brain sizes with the dinosaurs"in which case the disagreement is about how much evolution should care about intelligence in the dinosaur niche, vs other things it could put its skill points into? |
| **[Christiano][15:01]** Eliezer, it seems like chimps are insanely smart compared to other animals, basically as smart as they getso it's natural to think that the main things that make humans unique are also present in chimpsor at least, there was something going on in chimps that is exceptionaland should be causally upstream of the uniqueness of humans toootherwise you have too many coincidences on your hands |
| **[Yudkowsky][15:02]** ajeya: no, I'd characterize that as "the human environmental niche per se does not seem super-special enough to be unique on a geological timescale, the cognitive part of the niche derives from increased cognitive abilities in the first place and so can't be used to explain where they got started, dinosaurs are larger than humans and would pay lower relative metabolic costs for added brain size and it is not the case that every species as large as humans was in an environment where they would not have benefited as much from a fixed increment of intelligence, hominids are probably distinguished from dinosaurs in having better neural algorithms that arose over intervening evolutionary time and therefore better returns in intelligence on synapses that are more costly to humans than to elephants or large dinosaurs" |
| **[Christiano][15:03]** I don't understand how you can think that hominids are the special step relative to something earlieror like, I can see how it's consistent, but I don't see what evidence or argument supports itit seems like the short evolutionary time, and the fact that you also have to explain the exceptional qualities of other primates, cut extremely strongly against it |
| **[Yudkowsky][15:04]** paul: indeed, the fact that dinosaurs didn't see their brain sizes and intelligences ballooning, says there must be a lot of stuff hominids had that dinosaurs didn't, explaining why hominids got much higher returns on intelligence per synapse. natural selection is enough of a smooth process that 95% of this stuff should've been in the last common ancestor of humans and chimps. |
| **[Christiano][15:05]** it seems like brain size basically just increases faster in the smarter animals? though I mostly just know about birds and primates |
| **[Yudkowsky][15:05]** that is what you'd predict from smartness being about algorithms! |
| **[Christiano][15:05]** and it accelerates further and further within both linesit's what you'd expect if smartness is about algorithms *and chimps and birds have good algorithms* |
| **[Yudkowsky][15:06]** if smartness was about brain size, smartness and brain size would increase faster in the *larger animals* or the ones whose successful members *ate more food per day*well, sure, I do model that birds have better algorithms than dinosaurs |
| **[Cotra][15:07]** it seems like you've given arguments for "there was algorithmic innovation between dinosaurs and humans" but not yet arguments for "there was major algorithmic innovation between chimps and humans"? |
| **[Christiano][15:08]** (much less that the algorithmic changes were not just more-of-the-same) |
| **[Yudkowsky][15:08]** oh, that's *not* mandated by the model the same way. (between LCA of chimps and humans) |
| **[Christiano][15:08]** isn't that exactly what we are discussing? |
| **[Yudkowsky][15:09]** ...I hadn't thought so, no. |
| **[Cotra][15:09]** original q was:On software beating hardware, how much of your view is dependent on your belief that the chimp -> human transition was probably not mainly about brain size because if it were about brain size it would have happened faster? My understanding is that you think the main change is a small software innovation which increased returns to having a bigger brain. If you changed your mind and thought that the chimp -> human transition was probably mostly about raw brain size, what (if anything) about your AI takeoff views would change?so i thought we were talking about if there's a cool innovation from chimp->human? |
| **[Yudkowsky][15:10]** I can see how this would have been the more obvious intended interpretation on your viewpoint, and apologize |
| **[Christiano][15:10]** (though i think paul would also disagree about this thing, and believes that the chimp to human thing was mostly about size)Is what I was responding to in partI am open to saying that I'm conflating size and "algorithmic improvements that are closely correlated with size in practice and are similar to the prior algorithmic improvements amongst primates" |
| **[Yudkowsky][15:11]** from my perspective, the question is "how did that hominid->human transition happen, as opposed to there being an elephant->smartelephant or dinosaur->smartdinosaur transition"?I expect there were substantial numbers of brain algorithm stuffs going on during this time, howeverbecause I don't think that synapses scale that well *with* the baseline hominid boost |
| **[Christiano][15:11]** FWIW, it seems quite likely to me that there would be an elephant->smartelephant transition within tens of millions or maybe 100M years, and a dinosaur->smartdinosaur transition in hundreds of millions of yearsand those are just cut off by the fastest lines getting there first |
| **[Yudkowsky][15:12]** which I think does circle back to that point? actually I think my memory glitched and forgot the original point while being about this subpoint and I probably did interpret the original point as intended. |
| **[Christiano][15:12]** namely primates beating out birds by a hair |
| **[Yudkowsky][15:12]** that sounds like a viewpoint which would also think it much more likely that GPT-3 would foom in a billion yearswhere maybe you think that's unlikely, but I still get the impression your "unlikely" is, like, 5 orders of magnitude likelier than mine before applying overconfidence adjustments against extreme probabilities on both sidesyeah, I think I need to back up |
| **[Cotra][15:15]** Is your position something like "at some point after dinosaurs, there was an algorithmic innovation that increased returns to brain size, which meant that the birds and the humans see their brains increasing quickly while the dinosaurs didn't"? |
| **[Christiano][15:15]** it also seems to me like the chimp->human difference is in basically the same ballpark of the effect of brain size within humans, given modest adaptations for culturewhich seems like a relevant sanity-check that made me take the "mostly hardware" view more seriously |
| **[Yudkowsky][15:15]** there's a part of my model which very strongly says that hominids scaled better than elephants and that's why "hominids->humans but not elephants->superelephants" |
| **[Christiano][15:15]** previously I had assumed that analysis would show that chimps were obviously *way* dumber than an extrapolation of humans |
| **[Yudkowsky][15:16]** there's another part of my model which says "and it still didn't scale that well without algorithms, so we should expect a lot of alleles affecting brain circuitry which rose to fixation over the period when hominid brains were expanding"this part is strong and I think echoes back to AGI stuff, but it is not *as strong* as the much *more* overdetermined position that hominids started with more scalable algorithms than dinosaurs. |
| **[Christiano][15:17]** I do agree with the point that there are structural changes in brains as you scale them up, and this is potentially a reason why brain size changes more slowly than e.g. bone size. (Also there are small structural changes in ML algorithms as you scale them up, not sure how much you want to push the analogy but they feel fairly similar.) |
| **[Yudkowsky][15:17]** it also seems to me like the chimp->human difference is in basically the same ballpark of the effect of brain size within humans, given modest adaptations for culturethis part also seems pretty blatantly false to meis there, like, a smooth graph that you looked at there? |
| **[Christiano][15:18]** I think the extrapolated difference would be about 4 standard deviations, so we are comparing a chimp to an IQ 40 human |
| **[Yudkowsky][15:18]** I'm really not sure how much of a fair comparison that isIQ 40 humans in our society may be mostly sufficiently-damaged humans, not scaled-down humans |
| **[Christiano][15:19]** doesn't seem easy, but the point is that the extrapolated difference is huge, it corresponds to completely debilitating developmental problems |
| **[Yudkowsky][15:19]** if you do enough damage to a human you end up with, for example, a coma victim who's not competitive with other primates at all |
| **[Christiano][15:19]** yes, that's more than 4 SD downI agree with this general pointI'd guess I just have a lot more respect for chimps than you do |
| **[Yudkowsky][15:20]** I feel like I have a bunch of respect for chimps but more respect for humanslike, that stuff humans dothat is really difficult stuff!it is not just scaled-up chimpstuff! |
| **[Christiano][15:21]** Carl convinced me chimps wouldn't go to space, but I still really think it's about domesticity and cultural issues rather than intelligence |
| **[Yudkowsky][15:21]** the chimpstuff is very respectable but there is a whole big layer cake of additional respect on top |
| **[Christiano][15:21]** not a prediction to be resolved until after the singularityI mean, the space prediction isn't very confident 🙂and it involved a very large planet of apes |
9.6. Architectural innovation in AI and in evolutionary history
---------------------------------------------------------------
| |
| --- |
| **[Yudkowsky][15:22]** I feel like if GPT-based systems saturate and require *any* architectural innovation rather than Stack More Layers to get much further, this is a pre-Singularity point of observation which favors humans probably being more qualitatively different from chimp-LCA(LCA=last common ancestor) |
| **[Christiano][15:22]** any seems like a kind of silly bar? |
| **[Yudkowsky][15:23]** because single architectural innovations are allowed to have large effects! |
| **[Christiano][15:23]** like there were already small changes to normalization from GPT-2 to GPT-3, so isn't it settled? |
| **[Yudkowsky][15:23]** natural selection can't afford to deploy that many of them! |
| **[Christiano][15:23]** and the model really eventually won't work if you increase layers but don't fix the normalization, there are severe problems that only get revealed at high scale |
| **[Yudkowsky][15:23]** that I wouldn't call architectural innovationtransformers werethis is a place where I would not discuss specific ideas because I do not actually want this event to occur |
| **[Christiano][15:24]** surehave you seen a graph of LSTM scaling vs transformer scaling?I think LSTM with ongoing normalization-style fixes lags like 3x behind transformers on language modeling |
| **[Yudkowsky][15:25]** no, does it show convergence at high-enough scales? |
| **[Christiano][15:25]** figure 7 here: <https://arxiv.org/pdf/2001.08361.pdf> |
| **[Yudkowsky][15:26]** yeah... I unfortunately would rather not give other people a sense for which innovations are obviously more of the same and which innovations obviously count as qualitative |
| **[Christiano][15:26]** I think smart money is that careful initialization and normalization on the RNN will let it keep up for longeranyway, I'm very open to differences like LSTM vs transformer between humans and 3x-smaller-brained-ancestors, as long as you are open to like 10 similar differences further back in the evolutionary history |
| **[Yudkowsky][15:28]** what if there's 27 differences like that and 243 differences further back in history? |
| **[Christiano][15:28]** sure |
| **[Yudkowsky][15:28]** is that a distinctly Yudkowskian view vs a Paul view...apparently notI am again feeling confused about cruxes |
| **[Christiano][15:29]** I mean, 27 differences like transformer vs LSTM isn't actually plausible, so I guess we could talk about it |
| **[Cotra][15:30]** Here's a potential crux articulation that ties it back to the animals stuff: paul thinks that we first discover major algorithmic innovations that improve intelligence at a low level of intelligence, analogous to evolution discovering major architectural innovations with tiny birds and primates, and then there will be a long period of scaling up plus coming up with routine algorithmic tweaks to get to the high level, analogous to evolution schlepping on the same shit for a long time to get to humans. analogously, he thinks when big innovations come onto the scene the actual product is crappy af (e.g. wright brother's plane), and it needs a ton of work to scale up to usable and then to great.you both seem to think both evolution and tech history consiliently point in your direction |
| **[Christiano][15:33]** that sounds vaguely right, I guess the important part of "routine" is "vaguely predictable," like you mostly work your way down the low-hanging fruit (including new fruit that becomes more important as you scale), and it becomes more and more predictable the more people are working on it and the longer you've been at itand deep learning is already reasonably predictable (i.e. the impact of successive individual architectural changes is smaller, and law of large numbers is doing its thing) and is getting more so, and I just expect that to continue |
| **[Cotra][15:34]** yeah, like it's a view that points to using data that relates effort to algorithmic progress and using that to predict future progress (in combination with predictions of future effort) |
| **[Christiano][15:35]** yeahand for my part, it feels like this is how most technologies look and also how current ML progress looks |
| **[Cotra][15:36]** and *also* how evolution looks, right? |
| **[Christiano][15:37]** you aren't seeing big jumps in translation or in self-driving cars or in image recognition, you are just seeing a long slog, and you see big jumps in areas where few people work (usually up to levels that are not in fact that important, which is very correlated with few people working there)I don't know much about evolution, but it at least looks very consistent with what I know and the facts eliezer cites(not merely consistent, but "explains the data just about as well as the other hypotheses on offer") |
9.7. Styles of thinking in forecasting
--------------------------------------
| |
| --- |
| **[Yudkowsky][15:38]** I do observe that this would seem, on the surface of things, to describe the entire course of natural selection up until about 20K years ago, if you were looking at surface impacts |
| **[Christiano][15:39]** by 20k years ago I think it's basically obvious that you are tens of thousands of years from the singularitylike, I think natural selection is going crazy with the brains by millions of years ago, and by hundreds of thousands of years ago humans are going crazy with the culture, and by tens of thousands of years ago the culture thing has accelerated and is almost at the finish line |
| **[Yudkowsky][15:41]** really? I don't know if I would have been able to call that in advance if I'd never seen the future or any other planets. I mean, maybe, but I sure would have been extrapolating way out onto a further limb than I'm going here. |
| **[Christiano][15:41]** Yeah, I agree singularity is way more out on a limb---or like, where the singularity stops is more uncertain since that's all that's really at issue from my perspectivebut the point is that everything is clearly crazy in historical terms, in the same way that 2000 is crazy, even if you don't know where it's goingand the timescale for the crazy changes is tens of thousands of years |
| **[Yudkowsky][15:42]** I frankly model that, had I made any such prediction 20K years ago of hominids being able to pull of moon landings or global warming - never mind the Singularity - I would have faced huge pushback from many EAs, such as, for example, Robin Hanson, and you. |
| **[Christiano][15:42]** like I think this can't go on would have applied just as well: <https://www.lesswrong.com/posts/5FZxhdi6hZp8QwK7k/this-can-t-go-on>I don't think that's the case at alland I think you still somehow don't understand my position? |
| **[Yudkowsky][15:43]** <https://www.lesswrong.com/posts/XQirei3crsLxsCQoi/surprised-by-brains> is my old entry here |
| **[Christiano][15:43]** like, what is the move I'm making here, that you think I would have made in the past?and would have led astray? |
| **[Yudkowsky][15:44]** I sure do feel in a deeper sense that I am trying very hard to account for perspective shifts in how unpredictable the future actually looks at the time, and the Other is looking back at the past and organizing it neatly and expecting the future to be that neat |
| **[Christiano][15:45]** I don't even feel like I'm expecting the future to be neatare you just saying you have a really broad distribution over takeoff speed, and that "less than a month" gets a lot of probability because lots of numbers are less than a month? |
| **[Yudkowsky][15:47]** not exactly? |
| **[Christiano][15:47]** in what way is your view the one that is preferred by things being messy or unpredictable?like, we're both agreeing X will eventually happen, and I'm making some concrete prediction about how some other X' will happen first, and that's the kind of specific prediction that's likely to be wrong? |
| **[Yudkowsky][15:48]** more like, we sure can tell a story today about how normal and predictable AlphaGo was, but we can *always* tell stories like that about the past. I do not particularly recall the AI field standing up one year before AlphaGo and saying "It's time, we're coming for the 8-dan pros this year and we're gonna be world champions a year after that." (Which took significantly longer in chess, too, matching my other thesis about how these slides are getting steeper as we get closer to the end.) |
| **[Christiano][15:49]** it's more like, you are offering AGZ as an example of why things are crazy, and I'm doubtful / think it's pretty lamemaybe I don't understand how it's functioning as bayesian evidencefor what over what |
| **[Yudkowsky][15:50]** I feel like the whole smoothness-reasonable-investment view, if evaluated on Earth 5My ago *without benefit of foresight*, would have dismissed the notion of brains overtaking evolution; evaluated 1My ago, it would have dismissed the notion of brains overtaking evolution; evaluated 20Ky ago, it would have barely started to acknowledge that brains were doing anything interesting at all, but pointed out how the hominids could still only eat as much food as their niche offered them and how the cute little handaxes did not begin to compare to livers and wasp stings.there is a style of thinking that says, "wow, yeah, people in the past sure were surprised by stuff, oh, wait, *I'm also in the past*, aren't I, I am one of those people"and a view where you look back from the present and think about how reasonable the past all seems now, and the future will no doubt be equally reasonable |
| **[Christiano][15:52]** (the AGZ example may fall flat, because the arguments we are making about it now *we were also making in the past*) |
| **[Yudkowsky][15:52]** I am not sure this is resolvable, but it is among my primary guesses for a deep difference in believed styles of thought |
| **[Christiano][15:52]** I think that's a useful perspective, but still don't see how it favors your bottom line |
| **[Yudkowsky][15:53]** where I look at the style of thinking you're using, and say, not, "well, that's invalidated by a technical error on line 3 even on Paul's own terms" but "isn't this obviously a whole style of thought that never works and ends up unrelated to reality"I think the first AlphaGo was the larger shock, AlphaGo Zero was a noticeable but more mild shock on account of how it showed the end of game programming and not just the end of Go |
| **[Christiano][15:54]** sorry, I lumped them together |
| **[Yudkowsky][15:54]** it didn't feel like the same level of surprise; it was precedented by thenthe actual accomplishment may have been larger in an important sense, but a lot of the - epistemic landscape of lessons learned? - is about the things that surprise you at the time |
| **[Christiano][15:55]** also AlphaGo was also quite easy to see coming after this paper (as was discussed extensively *at the time*): <https://www.cs.toronto.edu/~cmaddis/pubs/deepgo.pdf> |
| **[Yudkowsky][15:55]** Paul, are you on the record as arguing with me that AlphaGo will win at Go because it's predictably on-trend?back then? |
| **[Cotra][15:55]** Hm, it sounds like Paul is saying "I do a trend extrapolation over long time horizons and if things seem to be getting faster and faster I expect they'll continue to accelerate; this extrapolation if done 100k years ago would have seen that things were getting faster and faster and projected singularity within 100s of K years"Do you think Paul is in fact doing something other than the trend extrap he says he's doing, or that he would have looked at a different less informative trend than the one he says he would have looked at, or something else? |
| **[Christiano][15:56]** my methodology for answering that question is looking at LW comments mentioning go by me, can see if it finds any |
| **[Yudkowsky][15:56]** Different less informative trend, is most of my suspicion there?though, actually, I should revise that, I feel like relatively little of the WHA was AlphaGo v2 whose name I forget beating Lee Se-dol, and most was in the revelation that v1 beat the high-dan pro whose name I forget.Paul having himself predicted anything at *all* like this would be the actually impressive featthat would cause me to believe that the AI world is more regular and predictable than I experienced it as, if you are paying more attention to ICLR papers than I do |
9.8. Moravec's prediction
-------------------------
| |
| --- |
| **[Cotra][15:58]** And jtbc, the trend extrap paul is currently doing is something like:* Look at how effort leads to hardware progress measured in FLOP/$ and software progress measured in stuff like "FLOP to do task X" or "performance on benchmark Y"
* Look at how effort in the ML industry as a whole is increasing, project forward with maybe some adjustments for thinking markets are more inefficient now and will be less inefficient later
and this is the wrong trend, because he shouldn't be looking at hardware/software progress across the whole big industry and should be more open to an upset innovation coming from an area with a small number of people working on it?and he would have similarly used the wrong trends while trying to do trend extrap in the past? |
| **[Yudkowsky][15:59]** because I feel like this general style of thought doesn't work when you use it on Earth generally, and then fails extremely hard if you try to use it on Earth before humans to figure out where the hominids are going because that phenomenon is Different from Previous Stufflike, to be clear, I have seen this used well on solarI feel like I saw some people calling the big solar shift based on graphs, before that happenedI have seen this used great by Moravec on computer chips to predict where computer chips would be in 2012and also witnessed Moravec *completely failing* as soon as he tried to derive *literally anything but the graph itself* namely his corresponding prediction for human-equivalent AI in 2012 (I think, maybe it was 2010) or something |
| **[Christiano][16:02]** (I think in his 1988 book Moravec estimated human-level AI in ~2030, not sure if you are referring to some earlier prediction?) |
| **[Yudkowsky][16:02]** (I have seen Ray Kurzweil project out Moore's Law to the $1,000,000 human brain in, what was it, 2025, followed by the $1000 human brain in 2035 and the $1 human brain in 2045, and when I asked Ray whether machine superintelligence might shift the graph at all, he replied that machine superintelligence was precisely how the graph would be able to continue on trend. This indeed is sillier than EAs.) |
| **[Cotra][16:03]** moravec's prediction appears to actually be around 2025, looking at his hokey graph? <https://jetpress.org/volume1/moravec.htm> |
| **[Yudkowsky][16:03]** but even there, it does feel to me like there is a commonality between Kurzweil's sheer graph-worship and difficulty in appreciating the graphs as surface phenomena that are less stable than deep phenomena, and something that Hanson was doing wrong in the foom debate |
| **[Cotra][16:03]** which is...like, your timelines? |
| **[Yudkowsky][16:04]** that's 1998Mind Children in 1988 I am pretty sure had an earlier prediction |
| **[Christiano][16:04]** I should think you'd be happy to bet against me on basically any prediction, shouldn't you? |
| **[Yudkowsky][16:05]** any prediction that sounds narrow and isn't like "this graph will be on trend in 3 more years"...maybe I'm wrong, an online source says Mind Children in 1988 predicted AGI in "40 years" but I sure do seem to recall an extrapolated graph that reached "human-level hardware" in 2012 based on an extensive discussion about computing power to duplicate the work of the retina |
| **[Christiano][16:08]** don't think it matters too much other than for Moravec's honor, doesn't really make a big difference for the empirical success of the methodologyI think it's on page 68 if you have the physical book |
| **[Yudkowsky][16:09]** p60 via Google Books says 10 teraops for a human-equivalent mind |
| **[Christiano][16:09]** I have a general read of history where trend extrapolation works extraordinarily well relative to other kinds of forecasting, to the extent that the best first-pass heuristic for whether a prediction is likely to be accurate is whether it's a trend extrapolation and how far in the future it is |
| **[Yudkowsky][16:09]** which, incidentally, strikes me as entirely plausible if you had algorithms as sophisticated as the human brainmy sense is that Moravec nailed the smooth graph of computing power going on being smooth, but then all of his predictions about the actual future were completely invalid on account of a curve interacting with his curve that he didn't know things about and so simply omitted as a step in his calculations, namely, AGI algorithms |
| **[Christiano][16:12]** though again, from your perspective 2030 is still a reasonable bottom-line forecast that makes him one of the most accurate people at that time? |
| **[Yudkowsky][16:12]** you could be right about all the local behaviors that your history is already shouting out at you as having smooth curve (where by "local" I do mean to exclude stuff like world GDP extrapolated into the indefinite future) and the curves that history isn't shouting at you will tear you down |
| **[Christiano][16:12]** (I don't know if he even forecast that) |
| **[Yudkowsky][16:12]** I don't remember that part from the 1988 bookmy memory of the 1988 book is "10 teraops, based on what it takes to rival the retina" and he drew a graph of Moore's Law |
| **[Christiano][16:13]** yeah, I think that's what he did(and got 2030) |
| **[Yudkowsky][16:14]** "If this rate of improvement were to continue into the next century, the 10 teraops required for a humanlike computer would be available in a $10 million supercomputer before 2010 and in a $1,000 personal computer by 2030." |
| **[Christiano][16:14]** or like, he says "human equivalent in 40 years" and predicts that in 50 years we will have robots with superhuman reasoning ability, not clear he's ruling out human-equivalent AGI before 40 years but I think the tone is clear |
| **[Yudkowsky][16:15]** so 2030 for AGI on a personal computer and 2010 for AGI on a supercomputer, and I expect that on my first reading I simply discarded the former prediction as foolish extrapolation past the model collapse he had just predicted in 2010.(p68 in "Powering Up") |
| **[Christiano][16:15]** yeah, that makes senseI do think the PC number seems irrelevant |
| **[Cotra][16:16]** I think both in that book and in the 98 article he wants you to pay attention to the "very cheap human-size computers" threshold, not the "supercomputer" threshold, i think intentionally as a way to handwave in "we need people to be able to play around with these things"(which people criticized him at the time for not more explicitly modeling iirc) |
| **[Yudkowsky][16:17]** but! I mean! there are so many little places where the media has a little cognitive hiccup about that and decides in 1998 that it's fine to describe that retrospectively as "you predicted in 1988 that we'd have true AI in 40 years" and then the future looks less surprising than people at the time using Trend Logic were actually surprised by it!all these little ambiguities and places where, oh, you decide retroactively that it would have made sense to look at *this* Trend Line and use it *that* way, but if you look at what people said at the time, they didn't actually say that! |
| **[Christiano][16:19]** I mean, in fairness reading the book it just doesn't seem like he is predicting human-level AI in 2010 rather than 2040, but I do agree that it seems like the basic methodology (why care about the small computer thing?) doesn't really make that much sense a priori and only leads to something sane if it cancels out with a weird view |
9.9. Prediction disagreements and bets
--------------------------------------
| |
| --- |
| **[Christiano][16:19]** anyway, I'm pretty unpersuaded by the kind of track record appeal you are making here |
| **[Yudkowsky][16:20]** if the future goes the way I predict and yet anybody somehow survives, perhaps somebody will draw a hyperbolic trendline on some particular chart where the trendline is retroactively fitted to events including those that occurred in only the last 3 years, and say with a great sage nod, ah, yes, that was all according to trend, nor did anything depart from trendtrend lines permit anything |
| **[Christiano][16:20]** like from my perspective the fundamental question is whether I would do better or worse by following the kind of reasoning you'd advocate, and it just looks to me like I'd do worse, and I'd love to make any predictions about anything to help make that more clear and hindsight-proof in advance |
| **[Yudkowsky][16:20]** you just look into the past and find a line you can draw that ended up where reality went |
| **[Christiano][16:21]** it feels to me like you really just waffle on almost any prediction about the before-end-of-days |
| **[Yudkowsky][16:21]** I don't think I know a lot about the before-end-of-days |
| **[Christiano][16:21]** like if you make a prediction I'm happy to trade into it, or you can pick a topic and I can make a prediction and you can trade into mine |
| **[Cotra][16:21]** but you know enough to have strong timing predictions, e.g. your bet with caplan |
| **[Yudkowsky][16:21]** it's daring enough that I claim to know anything about the Future at all! |
| **[Cotra][16:21]** surely with that difference of timelines there should be some pre-2030 difference as well |
| **[Christiano][16:21]** but you are the one making the track record argument against my way of reasoning about things!how does that not correspond to believing that your predictions are better!what does that mean? |
| **[Yudkowsky][16:22]** yes and if you say something narrow enough or something that my model does at least vaguely push against, we should bet |
| **[Christiano][16:22]** my point is that I'm willing to make a prediction about any old thing, you can name your topicI think the way I'm reasoning about the future is just better in generaland I'm going to beat you on whatever thing you want to bet on |
| **[Yudkowsky][16:22]** but if you say, "well, Moore's Law on trend, next 3 years", then I'm like, "well, yeah, sure, since I don't feel like I know anything special about that, that would be my prediction too" |
| **[Christiano][16:22]** sureyou can pick the topicpick a quantityor a yes/no questionor whatever |
| **[Yudkowsky][16:23]** you may know better than I would where your Way of Thought makes strong, narrow, or unusual predictions |
| **[Christiano][16:23]** I'm going to trend extrapolation everywherespoiler |
| **[Yudkowsky][16:23]** okay but any superforecaster could do that and I could do the same by asking a superforecaster |
| **[Cotra][16:24]** but there must be places where you'd strongly disagree w the superforecastersince you disagree with them eventually, e.g. >2/3 doom by 2030 |
| **[Bensinger][18:40] (Nov. 25 follow-up comment)** ">2/3 doom by 2030" isn't an actual Eliezer-prediction, and is based on a misunderstanding of something Eliezer said. See [Eliezer's comment on LessWrong](https://www.lesswrong.com/posts/7MCqRnZzvszsxgtJi/christiano-cotra-and-yudkowsky-on-ai-progress?commentId=diChXiELZd62hgRyK#diChXiELZd62hgRyK). |
| **[Yudkowsky][16:24]** in the terminal phase, sure |
| **[Cotra][16:24]** right, but there are no disagreements before jan 1 2030?no places where you'd strongly defy the superforecasters/trend extrap? |
| **[Yudkowsky][16:24]** superforecasters were claiming that AlphaGo had a 20% chance of beating Lee Se-dol and I didn't disagree with that at the time, though as the final days approached I became nervous and suggested to a friend that they buy out of a bet about that |
| **[Cotra][16:25]** what about like whether we get some kind of AI ability (e.g. coding better than X) before end days |
| **[Yudkowsky][16:25]** though that was more because of having started to feel incompetent and like I couldn't trust the superforecasters to know more, than because I had switched to a confident statement that AlphaGo would win |
| **[Cotra][16:25]** seems like EY's deep intelligence / insight-oriented view should say something about what's not possible before we get the "click" and the FOOM |
| **[Christiano][16:25]** I mean, I'm OK with either (i) evaluating arguments rather than dismissive and IMO totally unjustified track record, (ii) making bets about stuffI don't see how we can both be dismissing things for track record reasons and also not disagreeing about thingsif our methodologies agree about all questions before end of days (which seems crazy to me) then surely there is no track record distinction between them...
| |
| --- |
| [Cotra: 👍] |
|
| **[Cotra][16:26]** do you think coding models will be able to 2x programmer productivity before end days? 4x?what about hardware/software R&D wages? will they get up to $20m/yr for good ppl?will someone train a 10T param model before end days? |
| **[Christiano][16:27]** things I'm happy to bet about: economic value of LMs or coding models at 2, 5, 10 years, benchmark performance of either, robotics, wages in various industries, sizes of various industries, compute/$, someone else's views about "how ML is going" in 5 yearsmaybe the "any GDP acceleration before end of days?" works, but I didn't like how you don't win until the end of days |
| **[Yudkowsky][16:28]** okay, so here's an example place of a *weak* general Yudkowskian prediction, that is weaker than terminal-phase stuff of the End Days: (1) I predict that cycles of 'just started to be able to do Narrow Thing -> blew past upper end of human ability at Narrow Thing' will continue to get shorter, the same way that, I think, this happened faster with Go than with chess. |
| **[Christiano][16:28]** great, I'm totally into itwhat's a domain?coding? |
| **[Yudkowsky][16:28]** Does Paul disagree? Can Paul point to anything equally specific out of Paul's viewpoint? |
| **[Christiano][16:28]** benchmarks for LMs?robotics? |
| **[Yudkowsky][16:28]** well, for these purposes, we do need some Elo-like ability to measure at all where things are relative to humans |
| **[Cotra][16:29]** problem-solving benchmarks for code?MATH benchmark? |
| **[Christiano][16:29]** well, for coding and LM'ing we have lots of benchmarks we can use |
| **[Yudkowsky][16:29]** this unfortunately does feel a bit different to me from Chess benchmarks where the AI is playing the whole game; Codex is playing part of the game |
| **[Christiano][16:29]** in general the way I'd measure is by talking about how fast you go from "weak human" to "strong human" (e.g. going from top-10,000 in chess to top-10 or whatever, going from jobs doable by $50k/year engineer to $500k/year engineer...) |
| **[Yudkowsky][16:30]** golly, that sounds like a viewpoint very favorable to mine |
| **[Christiano][16:30]** what do you mean?that way of measuring would be favorable to your viewpoint? |
| **[Yudkowsky][16:31]** if we measure how far it takes AI to go past different levels of paying professionals, I expect that the Chess duration is longer than the Go duration and that by the time Codex is replacing ~~a~~ most paid $50k/year programmers the time to replacing ~~a~~ most programmers paid as much as a top Go player will be pretty darned short |
| **[Christiano][16:31]** top Go players don't get paid, do they? |
| **[Yudkowsky][16:31]** they tutor students and win titles |
| **[Christiano][16:31]** but I mean, they are like low-paid engineers |
| **[Yudkowsky][16:31]** yeah that's part of the issue here |
| **[Christiano][16:31]** I'm using wages as a way to talk about the distribution of human abilities, not the fundamental number |
| **[Yudkowsky][16:32]** I would expect something similar to hold over going from low-paying welder to high-paying welder |
| **[Christiano][16:32]** like, how long to move from "OK human" to "pretty good human" to "best human" |
| **[Cotra][16:32]** says salary of $350k/yr for lee: <https://www.fameranker.com/lee-sedol-net-worth> |
| **[Yudkowsky][16:32]** but I also mostly expect that AIs will not be allowed to weld things on Earth |
| **[Cotra][16:32]** why don't we just do an in vitro benchmark instead of wages? |
| **[Christiano][16:32]** what, machines already do virtually all welding? |
| **[Cotra][16:32]** just pick a benchmark? |
| **[Yudkowsky][16:33]** yoouuuu do not want to believe sites like that (fameranker) |
| **[Christiano][16:33]** yeah, I'm happy with any benchmark, and then we can measure various human levels at that benchmark |
| **[Cotra][16:33]** what about MATH? <https://arxiv.org/abs/2103.03874> |
| **[Christiano][16:34]** also I don't know what "shorter and shorter" means, the time in go and chess was decades to move from "strong amateur" to "best human," I do think these things will most likely be shorter than decadesseems like we can just predict concrete #s though
| |
| --- |
| [Cotra: 👍] |
like I can say how long I think it will take to get from "median high schooler" to "IMO medalist" and you can bet against me?and if we just agree about all of those predictions then again I'm back to being very skeptical of a claimed track record difference between our models(I do think that it's going to take years rather than decades on all of these things) |
| **[Yudkowsky][16:36]** possibly! I worry this ends up in a case where Katja or Luke or somebody goes back and collects data about "amateur to pro performance times" and Eliezer says "Ah yes, these are shortening over time, just as I predicted" and Paul is like "oh, well, I predict they continue to shorten on this trend drawn from the data" and Eliezer is like "I guess that could happen for the next 5 years, sure, sounds like something a superforecaster would predict as default" |
| **[Cotra][16:37]** i'm pretty sure paul's methodology here will just be to look at the MATH perf trend based on model size and combine with expectations of when ppl will make big enough models, not some meta trend thing like that? |
| **[Yudkowsky][16:37]** so I feel like... a bunch of what I feel is the real disagreement in our models, is a bunch of messy stuff Suddenly Popping Up one day and then Eliezer is like "gosh, I sure didn't predict that" and Paul is like "somebody could have totally predicted that" and Eliezer is like "people would say exactly the same thing after the world ended in 3 minutes"if we've already got 2 years of trend on a dataset, I'm not necessarily going to predict the trend breaking |
| **[Cotra][16:38]** hm, you're presenting your view as more uncertain and open to anything here than paul's view, but in fact it's picking out a narrower distribution. you're more confident in powerful AGI soon |
| **[Christiano][16:38]** seems hard to play the "who is more confident?" game |
| **[Cotra][16:38]** so there should be some places where you make a strong positive prediction paul disagrees with |
| **[Yudkowsky][16:39]** I might want to buy options on a portfolio of trends like that, if Paul is willing to sell me insurance against all of the trends breaking upward at a lower price than I think is reasonableI mean, from my perspective Paul is the one who seems to think the world is well-organized and predictable in certain ways |
| **[Christiano][16:39]** yeah, and you are saying that I'm overconfident about that |
| **[Yudkowsky][16:39]** I keep wanting Paul to go on and make narrower predictions than I do in that case |
| **[Christiano][16:39]** so you should be happy to bet with me about *anything*and I'm letting you pick anything at all you want to bet about |
| **[Cotra][16:40]** i mean we could do a portfolio of trends like MATH and you could bet on at least a few of them having strong surprises in the sooner directionbut that means we could just bet about MATH and it'd just be higher variance |
| **[Yudkowsky][16:40]** ok but you're not going to sell me cheap options on sharp declines in the S&P 500 even though in a very reasonable world there would not be any sharp declines like that |
| **[Christiano][16:41]** if we're betting $ rather than bayes points, then yes I'm going to weigh worlds based on the value of $ in those worlds |
| **[Cotra][16:41]** wouldn't paul just sell you options at the price the options actually trade for? i don't get it |
| **[Christiano][16:41]** but my sense is that I'm just generally across the board going to be more right than you are, and I'm frustrated that you just keep saying that "people like me" are wrong about stuff |
| **[Yudkowsky][16:41]** Paul's like "we'll see smooth behavior in the end days" and I feel like I should be able to say "then Paul, sell me cheap options against smooth behavior now" but Paul is just gonna wanna sell at market price |
| **[Christiano][16:41]** and so I want to hold you to that by betting about anythingideally just tons of stuffrandom things about what AI will be like, and other technologies, and regulatory changes |
| **[Cotra][16:42]** paul's view doesn't seem to imply that he should value those options less than the markethe's more EMH-y than you not less |
| **[Yudkowsky][16:42]** but then the future should *behave like that market* |
| **[Christiano][16:42]** what do you mean? |
| **[Yudkowsky][16:42]** it should have options on wild behavior that are not cheap! |
| **[Christiano][16:42]** you mean because people want $ more in worlds where the market drops a lot?I don't understand the analogy |
| **[Yudkowsky][16:43]** no, because jumpy stuff happens more than it would in a world of ideal agents |
| **[Cotra][16:43]** I think EY is saying the non-cheap option prices are because P(sharp declines) is pretty high |
| **[Christiano][16:43]** ok, we know how often markets jump, if that's the point of your argument can we just talk about that directly? |
| **[Yudkowsky][16:43]** or sharp rises, for that matter |
| **[Christiano][16:43]** (much lower than option prices obviously)I'm probably happy to sell you options for sharp risesI'll give you better than market odds in that directionthat's how this works |
| **[Yudkowsky][16:44]** now I am again confused, for I thought you were the one who expected world GDP to double in 4 years at some pointand indeed, drew such graphs with the rise suggestively happening earlier than the sharp spike |
| **[Christiano][16:44]** yeah, and I have exposure to that by buying stocks, options prices are just a terrible way of tracking these things |
| **[Yudkowsky][16:44]** suggesting that such a viewpoint is generally favor to near timelines for that |
| **[Christiano][16:44]** I mean, I have bet a *lot* of money on AI companies doing wellwell, not compared to the EA crowd, but compared to my meager net worth 🙂and indeed, it has been true so farand I'm continuing to make the betit seems like on your view it should be surprising that AI companies just keep going uparen't you predicting them not to get to tens of trillions of valuation before the end of days? |
| **[Yudkowsky][16:45]** I believe that Nate, of a generally Yudkowskian view, did the same (bought AI companies). and I focused my thoughts elsewhere, because somebody needs to, but did happen to buy my first S&P 500 on its day of exact minimum in 2020 |
| **[Christiano][16:46]** point is, that's how you get exposure to the crazy growth stuff with continuous ramp-upsand I'm happy to make the bet on the marketor on other claimsI don't know if my general vibe makes sense here, and why it seems reasonable to me that I'm just happy to bet on anythingas a way of trying to defend my overall attackand that if my overall epistemic approach is vulnerable to some track record objection, then it seems like it ought to be possible to win here |
9.10. Prediction disagreements and bets: Standard superforecaster techniques
----------------------------------------------------------------------------
| |
| --- |
| **[Cotra][16:47]** I'm still kind of surprised that Eliezer isn't willing to bet that there will be a faster-than-Paul expects trend break on MATH or whatever other benchmark. Is it just the variance of MATH being one benchmark? Would you make the bet if it were 6? |
| **[Yudkowsky][16:47]** a large problem here is that both of us tend to default strongly to superforecaster standard techniques |
| **[Christiano][16:47]** it's true, though it's less true for longer things |
| **[Cotra][16:47]** but you think the superforecasters would suck at predicting end days because of the surface trends thing! |
| **[Yudkowsky][16:47]** before I bet against Paul on MATH I would want to know that Paul wasn't arriving at the same default I'd use, which might be drawn from trend lines there, or from a trend line in trend linesI mean the superforecasters did already suck once in my observation, which was AlphaGo, but I did not bet against them there, I bet with them and then updated afterwards |
| **[Christiano][16:48]** I'd mostly try to eyeball how fast performance was improving with size; I'd think about difficulty effects (where e.g. hard problems will be flat for a while and then go up later, so you want to measure performance on a spectrum of difficulties) |
| **[Cotra][16:48]** what if you bet against a methodology instead of against paul's view? the methodology being the one i described above, of looking at the perf based on model size and then projecting model size increases by cost? |
| **[Christiano][16:48]** seems safer to bet against my view |
| **[Cotra][16:48]** yeah |
| **[Christiano][16:48]** mostly I'd just be eyeballing size, thinking about how much people will in fact scale up (which would be great to factor out if possible), assuming performance trends hold upare there any other examples of surface trends vs predictable deep changes, or is AGI the only one?(that you have thought a lot about) |
| **[Cotra][16:49]** yeah seems even better to bet on the underlying "will the model size to perf trends hold up or break upward" |
| **[Yudkowsky][16:49]** so from my perspective, there's this whole thing where *unpredictably* something breaks above trend because the first way it got done was a way where somebody could do it faster than you expected |
| **[Christiano][16:49]** (makes sense for it to be the domain where you've thought a lot)you mean, it's unpredictable what will break above trend? |
| **[Cotra][16:49]** [IEM](https://intelligence.org/files/IEM.pdf) has a financial example |
| **[Yudkowsky][16:49]** I mean that I could not have said "*Go* will break above trend" in 2015 |
| **[Christiano][16:49]** yeahok, here's another example |
| **[Yudkowsky][16:50]** it feels like if I want to make a bet with imaginary Paul in 2015 then I have to bet on a portfolioand I also feel like as soon as we make it that concrete, Paul does not want to offer me things that I want to bet onbecause Paul is also like, sure, something might break upwardI remark that I have for a long time been saying that I wish Paul had more concrete images and examples attached to *a lot of his stuff* |
| **[Cotra][16:51]** surely the view is about the probability of each thing breaking upward. or the expected number from a basket |
| **[Christiano][16:51]** I mean, if you give me any way of quantifying how much stuff breaks upwards we have a bet |
| **[Cotra][16:51]** not literally that one single thing breaks upward |
| **[Christiano][16:51]** I don't understand how concreteness is an accusation here, I've offered 10 quantities I'd be happy to bet about, and also allowed you to name literally any other quantity you wantand I agree that we mostly agree about things |
| **[Yudkowsky][16:52]** and some of my sense here is that if Paul offered a portfolio bet of this kind, I might not take it myself, but EAs who were better at noticing their own surprise might say, "Wait, *that's* how unpredictable Paul thinks the world is?"so from my perspective, it is hard to know specific anti-superforecaster predictions that happen long before terminal phase, and I am not sure we are really going to get very far there. |
| **[Christiano][16:53]** but you agree that the eventual prediction is anti-superforecaster? |
| **[Yudkowsky][16:53]** both of us probably have quite high inhibitions against selling conventionally priced options that are way not what a superforecaster would price them as |
| **[Cotra][16:53]** why does it become so much easier to know these things and go anti-superforecaster at terminal phase? |
| **[Christiano][16:53]** I assume you think that the superforecasters will continue to predict that big impactful AI applications are made by large firms spending a lot of money, even through the end of daysI do think it's very often easy to beat superforecasters in-domainlike I expect to personally beat them at most ML predictionand so am also happy to do bets where you defer to superforecasters on arbitrary questions and I bet against you |
| **[Yudkowsky][16:54]** well, they're anti-prediction-market in the sense that, at the very end, bets can no longer settle. I've been surprised of late by how much AGI ruin seems to be sneaking into common knowledge; perhaps in the terminal phase the superforecasters will be like, "yep, we're dead". I can't even say that in this case, Paul will disagree with them, because I expect the state on alignment to be so absolutely awful that even Paul is like "You were not supposed to do it that way" in a very sad voice. |
| **[Christiano][16:55]** I'm just thinking about takeoff speeds hereI do think it's fairly likely I'm going to be like "oh no this is bad" (maybe 50%?), but not that I'm going to expect fast takeoffand similarly for the superforecasters |
9.11. Prediction disagreements and bets: Late-stage predictions, and betting against superforecasters
-----------------------------------------------------------------------------------------------------
| |
| --- |
| **[Yudkowsky][16:55]** so, one specific prediction you made, sadly close to terminal phase but not much of a surprise there, is that the world economy must double in 4 years before the End Times are permitted to begin |
| **[Christiano][16:56]** well, before it doubles in 1 year...I think most people would call the 4 year doubling the end times |
| **[Yudkowsky][16:56]** this seems like you should also be able to point to some least impressive thing that is not permitted to occur before WGDP has doubled in 4 years |
| **[Christiano][16:56]** and it means that the normal planning horizon includes the singularity |
| **[Yudkowsky][16:56]** it may not be much but we would be *moving back* the date of first concrete disagreement |
| **[Christiano][16:57]** I can list things I don't think would happen first, since that's a ton |
| **[Yudkowsky][16:57]** and EAs might have a little bit of time in which to say "Paul was falsified, uh oh" |
| **[Christiano][16:57]** the only things that aren't permitted are the ones that would have caused the world economy to double in 4 years |
| **[Yudkowsky][16:58]** and by the same token, there are things Eliezer thinks you are probably not going to be able to do before you slide over the edge. a portfolio of these will have some losing options because of adverse selection against my errors of what is hard, but if I lose more than half the portfolio, this may said to be a bad sign for Eliezer. |
| **[Christiano][16:58]** (though those can happen at the beginning of the 4 year doubling) |
| **[Yudkowsky][16:58]** this is unfortunately *late* for falsifying our theories but it would be *progress* on a kind of bet against each other |
| **[Christiano][16:59]** but I feel like the things I'll say are like fully automated construction of fully automated factories at 1-year turnarounds, and you're going to be like "well duh" |
| **[Yudkowsky][16:59]** ...unfortunately yes |
| **[Christiano][16:59]** the reason I like betting about numbers is that we'll probably just disagree on any given number |
| **[Yudkowsky][16:59]** I don't think I *know* numbers. |
| **[Christiano][16:59]** it does seem like a drawback that this can just turn up object-level differences in knowledge-of-numbers more than deep methodological advantages |
| **[Yudkowsky][17:00]** the last important number I had a vague suspicion I might know was that Ethereum ought to have a significantly larger market cap in pre-Singularity equilibrium.and I'm not as sure of that one since El Salvador supposedly managed to use Bitcoin L2 Lightning.(though I did not fail to act on the former belief) |
| **[Christiano][17:01]** do you see why I find it weird that you think there is this deep end-times truth about AGI, that is very different from a surface-level abstraction and that will take people like Paul by surprise, without thinking there are other facts like that about the world?I do see how this annoying situation can come aboutand I also understand the symmetry of the situation |
| **[Yudkowsky][17:02]** we unfortunately both have the belief that the present world looks a lot like our being right, and therefore that the other person ought to be willing to bet against default superforecasterish projections |
| **[Cotra][17:02]** paul says that *he* would bet against superforecasters too though |
| **[Christiano][17:02]** I would in ML |
| **[Yudkowsky][17:02]** like, where specifically? |
| **[Christiano][17:02]** or on any other topic where I can talk with EAs who know about the domain in questionI don't know if they have standing forecasts on things, but e.g.: (i) benchmark performance, (ii) industry size in the future, (iii) how large an LM people will train, (iv) economic impact of any given ML system like codex, (v) when robotics tasks will be plausible |
| **[Yudkowsky][17:03]** I have decided that, as much as it might gain me prestige, I don't think it's actually the right thing for me to go spend a bunch of character points on the skills to defeat superforecasters in specific domains, and then go around doing that to prove my epistemic virtue. |
| **[Christiano][17:03]** that seems fair |
| **[Yudkowsky][17:03]** you don't need to bet with *me* to prove your epistemic virtue in this way, thoughokay, but, if I'm allowed to go around asking Carl Shulman who to ask in order to get the economic impact of Codex, maybe I can also defeat superforecasters. |
| **[Christiano][17:04]** I think the deeper disagreement is that (i) I feel like my end-of-days prediction is also basically just a default superforecaster prediction (and if you think yours is too then we can bet about what some superforecasters will say on it), (ii) I think you are leveling a much stronger "people like paul get taken by surprise by reality" claim whereas I'm just saying that I don't like your arguments |
| **[Yudkowsky][17:04]** it seems to me like the contest should be more like our intuitions in advance of doing that |
| **[Christiano][17:04]** yeah, I think that's fine, and also cheaper since research takes so much timeI feel like those asymmetries are pretty strong though |
9.12. Self-duplicating factories, AI spending, and Turing test variants
-----------------------------------------------------------------------
| |
| --- |
| **[Yudkowsky][17:05]** so, here's an idea that is less epistemically virtuous than our making Nicely Resolvable Betswhat if we, like, talked a bunch about our off-the-cuff senses of where various AI things are going in the next 3 yearsand then 3 years later, somebody actually reviewed that |
| **[Christiano][17:06]** I do think just saying a bunch of stuff about what we expect will happen so that *we* can look back on it would have a significant amount of the value |
| **[Yudkowsky][17:06]** and any time the other person put a thumbs-up on the other's prediction, that prediction coming true was not taken to distinguish them |
| **[Cotra][17:06]** i'd suggest doing this in a format other than discord for posterity |
| **[Yudkowsky][17:06]** even if the originator was like HOW IS THAT ALSO A PREDICTION OF YOUR THEORYwell, Discord has worked better than some formats |
| **[Cotra][17:07]** something like a spreadsheet seems easier for people to look back on and score and stuffdiscord transcripts are pretty annoying to read |
| **[Yudkowsky][17:08]** something like a spreadsheet seems liable to be high-cost and not actually happen |
| **[Christiano][17:08]** I think a conversation is probably easier and about as good for our purposes though? |
| **[Cotra][17:08]** ok fair |
| **[Yudkowsky][17:08]** I think money can be inserted into humans in order to turn Discord into spreadsheets |
| **[Christiano][17:08]** and it's possible we will both think we are right in retrospectand that will also be revealing |
| **[Yudkowsky][17:09]** but, besides that, I do want to boop on the point that I feel like Paul should be able to predict intuitively, rather than with necessity, things that should not happen before the world economy doubled in 4 years |
| **[Christiano][17:09]** it may also turn up some quantitative differences of viewthere are lots of things I think won't happen before the world economy has doubled in 4 years |
| **[Yudkowsky][17:09]** because on my model, as we approach the end times, AI was still pretty partial and also the world economy was lolnoping most of the inputs a sensible person would accept from it and prototypes weren't being commercialized and stuff was generally slow and messy |
| **[Christiano][17:09]** prototypes of factories building factories in <2 years |
| **[Yudkowsky][17:10]** "AI was still pretty partial" leads it to not do interesting stuff that Paul can rule out |
| **[Christiano][17:10]** like I guess I think tesla will try, and I doubt it will be just tesla |
| **[Yudkowsky][17:10]** but the other parts of that permit AI to do interesting stuff that Paul can rule out |
| **[Christiano][17:10]** automated researchers who can do ML experiments from 2020 without human input |
| **[Yudkowsky][17:10]** okay, see, that whole "factories building factories" thing just seems so very much *after* the End Times to me |
| **[Christiano][17:10]** yeah, we should probably only talk about cognitive worksince you think physical work will be very slow |
| **[Yudkowsky][17:11]** okay but not just that, it's a falsifiable predictionit is something that lets Eliezer be wrong in advance of the End Times |
| **[Christiano][17:11]** what's a falsifiable prediction? |
| **[Yudkowsky][17:11]** if we're in a world where Tesla is excitingly gearing up to build a fully self-duplicating factory including its mining inputs and chips and solar panels and so on, we're clearly in the Paulverse and not in the Eliezerverse! |
| **[Christiano][17:12]** yeahI do think we'll see that before the end timesjust not before 4 year doublings |
| **[Yudkowsky][17:12]** this unfortunately only allows you to be right, and not for me to be right, but I think there are also things you legit only see in the Eliezerverse! |
| **[Christiano][17:12]** I mean, I don't think they will be doing mining for a long time because it's cheap |
| **[Yudkowsky][17:12]** they are unfortunately late in the game but they exist at all!and being able to state them is progress on this project! |
| **[Christiano][17:13]** but fully-automated factories first, and then significant automation of the factory-building processI do expect to seeI'm generally pretty bullish on industrial robotics relative to you I think, even before the crazy stuff?but you might not have a firm viewlike I expect to have tons of robots doing all kinds of stuff, maybe cutting human work in manufacturing 2x, with very modest increases in GDP resulting from that in particular |
| **[Yudkowsky][17:13]** so, like, it doesn't surprise me very much if Tesla manages to fully automate a factory that takes in some relatively processed inputs including refined metals and computer chips, and outputs a car? and by the same token I expect that has very little impact on GDP. |
| **[Christiano][17:14]** refined metals are almost none of the cost of the factoryand also tesla isn't going to be that vertically integratedthe fabs will separately continue to be more and more automatedI expect to have robot cars driving everywhere, and robot trucksanother 2x fall in humans required for warehouseselimination of most brokers involved in negotiating shipping |
| **[Yudkowsky][17:15]** if despite the fabs being more and more automated, somehow things are managing not to cost less and less, and that sector of the economy is not really growing very much, is that more like the Eliezerverse than the Paulverse? |
| **[Christiano][17:15]** most work in finance and loan origination |
| **[Yudkowsky][17:15]** though this is something of a peripheral prediction to AGI core issues |
| **[Christiano][17:16]** yeah, I think if you cut the humans to do X by 2, but then the cost falls much less than the number you'd naively expect (from saving on the human labor and paying for the extra capital), then that's surprising to meI mean if it falls half as much as you'd expect on paper I'm like "that's a bit surprising" rather than having my mind blown, if it doesn't fall I'm more surprisedbut that was mostly physical economy stuffoh wait, I was making positive predictions now, physical stuff is good for that I think?since you don't expect it to happen? |
| **[Yudkowsky][17:17]** ...this is not your fault but I wish you'd asked me to produce my "percentage of fall vs. paper calculation" estimate before you produced yoursmy mind is very whiffy about these things and I am not actually unable to deanchor on your estimate 😦 |
| **[Christiano][17:17]** makes sense, I wonder if I should just spoilerone benefit of discord |
| **[Yudkowsky][17:18]** yeah that works too! |
| **[Christiano][17:18]** a problem for prediction is that I share some background view about insane inefficiency/inadequacy/decadence/sillinessso these predictions are all tampered by thatbut still seem like there are big residual disagreements |
| **[Yudkowsky][17:19]** sighgreat |
| **[Christiano][17:19]** since you have way more of that than I do |
| **[Yudkowsky][17:19]** not your fault but |
| **[Christiano][17:19]** I think that the AGI stuff is going to be a gigantic megaproject despite that |
| **[Yudkowsky][17:19]** I am not shocked by the AGI stuff being a gigantic megaprojectit's not above the bar of survival but, given other social optimism, it permits death with more dignity than by other routes |
| **[Christiano][17:20]** what if spending is this big:Google invests $100B training a model, total spending across all of industry is way bigger |
| **[Yudkowsky][17:20]** oooohI do start to be surprised if, come the end of the world, AGI is having more invested in it than a TSMC fabthough, not... *super* surprised?also I am at least a little surprised before thenactually I should probably have been spoiling those statements myself but my expectation is that Paul's secret spoiler is about$10 trillion dollars or something equally totally shocking to an Eliezer |
| **[Christiano][17:22]** my view on that level of spending isit's an only slightly high-end estimate for spending by someone on a single model, but that in practice there will be ways of dividing more across different firms, and that the ontology of single-model will likely be slightly messed up (e.g. by OpenAI Five-style surgery). Also if it's that much then it likely involves big institutional changes and isn't at google.I read your spoilermy estimate for total spending for the whole project of making TAI, including hardware and software manufacturing and R&d, the big datacenters, etc.is in the ballpark of $10T, though it's possible that it will be undercounted several times due to wage stickiness for high-end labor |
| **[Yudkowsky][17:24]** I think that asspending on particular AGI megaprojects starts to go past $50 billion, it's not especially ruled out per se by things that I think I know for sure, but I feel like a third-party observer should justly start to weakly think, 'okay, this is looking at least a little like the Paulverse rather than the Eliezerverse', and as we get to $10 trillion, that is not absolutely ruled out by the Eliezerverse but it was a whoole lot more strongly predicted by the Paulverse, maybe something like 20x unless I'm overestimating how strongly Paul predicts that |
| **[Christiano][17:24]** Proposed modification to the "speculate about the future to generate kind-of-predictions" methodology: we make shit up, then later revise based on points others made, and maybe also get Carl to sanity-check and deciding which of his objections we agree with. Then we can separate out the "how good are intuitions" claim (with fast feedback) from the all-things-considered how good was the "prediction" |
| **[Yudkowsky][17:25]** okay that hopefully allows me to read Paul's spoilers... no I'm being silly. @ajeya please read all the spoilers and say if it's time for me to read his |
| **[Cotra][17:25]** you can read his latest |
| **[Christiano][17:25]** I'd guess it's fine to read all of them? |
| **[Cotra][17:26]** yeah sorry that's what i meant |
| **[Yudkowsky][17:26]** what should I say more about before reading earlier ones?ah k |
| **[Christiano][17:26]** My $10T estimate was after reading yours (didn't offer an estimate on that quantity beforehand), though that's the kind of ballpark I often think about, maybe we should just spoiler only numbers so that context is clear 🙂I think fast takeoff gets significantly more likely as you push that number down |
| **[Yudkowsky][17:27]** so, may I now ask what starts to look to you like "oh damn I am in the Eliezerverse"? |
| **[Christiano][17:28]** big mismatches between that AI looks technically able to do and what AI is able to do, though that's going to need a lot of work to operationalizeI think low growth of AI overall feels like significant evidence for Eliezerverse (even if you wouldn't make that prediction), since I'm forecasting it rising to absurd levels quite fast whereas your model is consistent with it staying smallsome intuition about AI looking very smart but not able to do much useful until it has the whole picture, I guess this can be combined with the first point to be something like---AI looks really smart but it's just not adding much valueall of those seem really hard |
| **[Cotra][17:30]** strong upward trend breaks on benchmarks seems like it should be a point toward eliezer verse, even if eliezer doesn't want to bet on a specific one?especially breaks on model size -> perf trends rather than calendar time trends |
| **[Christiano][17:30]** I think that any big break on model size -> perf trends are significant evidence |
| **[Cotra][17:31]** meta-learning working with small models?e.g. model learning-to-learn video games and then learning a novel one in a couple subjective hours |
| **[Christiano][17:31]** I think algorithmic/architectural changes that improve loss as much as 10x'ing model, for tasks that looking like they at least *should* have lots of economic value(even if they don't end up having lots of value because of deployment bottlenecks)is the meta-learning thing an Eliezer prediction?(before the end-of-days) |
| **[Cotra][17:32]** no but it'd be an anti-bio-anchor positive trend break and eliezer thinks those should happen more than we do |
| **[Christiano][17:32]** fair enougha lot of these things are about # of times that it happens rather than whether it happens at all |
| **[Cotra][17:32]** yeahbut meta-learning is special as the most plausible long horizon task |
| **[Christiano][17:33]** e.g. maybe in any given important task I expect a single "innovation" that's worth 10x model size? but that it still represents a minority of total time?hm, AI that can pass a competently administered turing test without being economically valuable?that's one of the things I think is ruled out before 4 year doubling, though Eliezer probably also doesn't expect it
| |
| --- |
| [Yudkowsky: 👍] |
|
| **[Cotra][17:34]** what would this test do to be competently administered? like casual chatbots seem like they have reasonable probability of fooling someone for a few mins now |
| **[Christiano][17:34]** I think giant google-automating-google projects without big external economic impacts |
| **[Cotra][17:34]** would it test knowledge, or just coherence of some kind? |
| **[Christiano][17:35]** it's like a smart-ish human (say +2 stdev at this task) trying to separate out AI from smart-ish human, iterating a few times to learn about what worksI mean, the basic ante is that the humans are *trying* to win a turing test, without that I wouldn't even call it a turing testdunno if any of those are compelling @Eliezersomething that passes a like "are you smart?" test administered by a human for 1h, where they aren't trying to specifically tell if you are AIjust to see if you are as smart as a humanI mean, I guess the biggest giveaway of all would be if there is human-level (on average) AI as judged by us, but there's no foom yet |
| **[Yudkowsky][17:37]** I think we both don't expect that one before the End of Days? |
| **[Christiano][17:37]** or like, no crazy economic impactI think we both expect that to happen before foom?but the "on average" is maybe way too rough a thing to define |
| **[Yudkowsky][17:37]** oh, wait, I missed that it wasn't the full Turing Test |
| **[Christiano][17:37]** well, I suggested boththe lamer one is more plausible |
| **[Yudkowsky][17:38]** full Turing Test happeneth not before the End Times, on Eliezer's view, and not before the first 4-year doubling time, on Paul's view, and the first 4-year doubling happeneth not before the End Times, on Eliezer's view, so this one doesn't seem very useful |
9.13. GPT-*n* and small architectural innovations vs. large ones
----------------------------------------------------------------
| |
| --- |
| **[Christiano][17:39]** I feel like the biggest subjective thing is that I don't feel like there is a "core of generality" that GPT-3 is missingI just expect it to gracefully glide up to a human-level foom-ing intelligence |
| **[Yudkowsky][17:39]** the "are you smart?" test seems perhaps passable by GPT-6 or its kin, which I predict to contain at least one major architectural difference over GPT-3 that I could, pre-facto if anyone asked, rate as larger than a different normalization methodbut by fooling the humans more than by being smart |
| **[Christiano][17:39]** like I expect GPT-5 would foom if you ask it but take a long time |
| **[Yudkowsky][17:39]** that sure is an underlying difference |
| **[Christiano][17:39]** not sure how to articulate what Eliezer expects to see here thoughor like what the difference is |
| **[Cotra][17:39]** something that GPT-5 or 4 shouldn't be able to do, according to eliezer?where Paul is like "sure it could do that"? |
| **[Christiano][17:40]** I feel like GPT-3 clearly has some kind of "doesn't really get what's going on" energyand I expect that to go awaywell before the end of daysso that it seems like a kind-of-dumb person |
| **[Yudkowsky][17:40]** I expect it to go away before the end of daysbut with there having been a big architectural innovation, not Stack More Layers |
| **[Christiano][17:40]** yeahwhereas I expect layer stacking + maybe changing loss (since logprob is too noisy) is sufficient |
| **[Yudkowsky][17:40]** if you name 5 possible architectural innovations I can call them small or large |
| **[Christiano][17:41]** 1. replacing transformer attention with DB nearest-neighbor lookup over an even longer context |
| **[Yudkowsky][17:42]** okay 1's a bit borderline |
| **[Christiano][17:42]** 2. adding layers that solve optimization problems internally (i.e. the weights and layer N activations define an optimization problem, the layer N+1 solves it) or maybe simulates an ODE |
| **[Yudkowsky][17:42]** if it's 3x longer context, no biggie, if it's 100x longer context, more of a game-changer2 - big change |
| **[Christiano][17:42]** I'm imagining >100x if you do that3. universal transformer XL, where you reuse activations from one context in the next context (RNN style) and share weights across layers |
| **[Yudkowsky][17:43]** I do not predict 1 works because it doesn't seem like an architectural change that moves away from what I imagined to be the limits, but it's a big change if it 100xs the window3 - if it is only that single change and no others, I call it not a large change relative to transformer XL. Transformer XL itself however was an example of a large change - it didn't have a large effect but it was what I'd call a large change. |
| **[Christiano][17:45]** 4. Internal stochastic actions trained with reinforceI mean, is mixture of experts or switch another big change?are we just having big changes non-stop? |
| **[Yudkowsky][17:45]** 4 - I don't know if I'm imagining right but it sounds large |
| **[Christiano][17:45]** it sounds from these definitions like the current rate of big changes is > 1/year |
| **[Yudkowsky][17:46]** 5 - mixture of experts: as with 1, I'm tempted to call it a small change, but that's because of my model of it as doing the same thing, not because it isn't in a certain sense a quite large move away from Stack More LayersI mean, it is not very hard to find a big change to try?finding a big change that works is much harder |
| **[Christiano][17:46]** several of these are improvements |
| **[Yudkowsky][17:47]** one gets a minor improvement from a big change rather more often than a big improvement from a big changethat's why dinosaurs didn't foom |
| **[Christiano][17:47]** like transformer -> MoE -> switch transformer is about as big an improvement as LSTM vs transformerso if we all agree that big changes are happening multiple times per year, then I guess that's not the difference in predictionis it about the size of gains from individual changes or something?or maybe: if you take the scaling laws for transformers, are the models with impact X "on trend," with changes just keeping up or maybe buying you 1-2 oom of compute, or are they radically better / scaling much better?that actually feels most fundamental |
| **[Yudkowsky][17:49]** I had not heard that transformer -> switch transformer was as large an improvement as lstm -> transformers after a year or two, though maybe you're referring to a claimed 3x improvement and comparing that to the claim that if you optimize LSTMs as hard as transformers they come within 3x (I have not examined these claims in detail, they sound a bit against my prior, and I am a bit skeptical of both of them)so remember that from my perspective, I am fighting an adverse selection process and the Law of Earlier Success |
| **[Christiano][17:50]** I think it's actually somewhat smaller |
| **[Yudkowsky][17:51]** if you treat GPT-3 as a fixed thingy and imagine scaling it in the most straightforward possible way, then I have a model of what's going on in there and I don't think that most direct possible way of scaling gets you past GPT-3 lacking a deep coresomebody can come up and go, "well, what about this change that nobody tried yet?" and I can be like, "ehhh, that particular change does not get at what I suspect the issues are" |
| **[Christiano][17:52]** I feel like the framing is: paul says that something is possible with "stack more layers" and eliezer isn't. We both agree that you can't literally stack more layers and have to sometimes make tweaks, and also that you will scale faster if you make big changes. But it seems like for Paul that means (i) changes to stay on the old trend line, (ii) changes that trade off against modest amounts of computeso maybe we can talk about that? |
| **[Yudkowsky][17:52]** when it comes to predicting what happens in 2 years, I'm not just up against people trying a broad range of changes that I can't foresee in detail, I'm also up against a Goodhart's Curse on the answer being a weird trick that worked better than I would've expected in advance |
| **[Christiano][17:52]** but then it seems like we may just not know, e.g. if we were talking lstm vs transformer, no one is going to run experiments with the well-tuned lstm because it's still just worse than a transformer (though they've run enough experiments to know how important tuning is, and the brittleness is much of why no one likes it) |
| **[Yudkowsky][17:53]** I would not have predicted Transformers to be a huge deal if somebody described them to me in advance of having ever tried it out. I think that's because predicting the future is hard not because I'm especially stupid. |
| **[Christiano][17:53]** I don't feel like anyone could predict that being a big dealbut I do think you could predict "there will be some changes that improve stability / make models slightly better"(I mean, I don't feel like any of the actual humans on earth could have, some hypothetical person could) |
| **[Yudkowsky][17:57]** whereas what I'm trying to predict is more like "GPT-5 in order to start-to-awaken needs a change via which it, in some sense, can do a different thing, that is more different than the jump from GPT-1 to GPT-3; and examples of things with new components in them abound in Deepmind, like Alpha Zero having not the same architecture as the original AlphaGo; but at the same time I'm also trying to account for being up against this very adversarial setup where a weird trick that works much better than I expect may be the thing that makes GPT-5 able to do a different thing"this may seem Paul-unfairish because any random innovations that come along, including big changes that cause small improvements, would tend to be swept up into GPT-5 even if they made no more deep difference than the whole thing with MoEso it's hard to bet onbut I also don't feel like it - totally lacks Eliezer-vs-Paul-ness if you let yourself sort of relax about that and just looked at it?also I'm kind of running out of energy, sorry |
| **[Christiano][18:03]** I think we should be able to get something here eventuallyseems good to break thoughthat was a lot of arguing for one day | |
4570e116-0894-42bc-9dcd-88ee9fde0be0 | trentmkelly/LessWrong-43k | LessWrong | Open Thread June 2010, Part 3
This thread is for the discussion of Less Wrong topics that have not appeared in recent posts. If a discussion gets unwieldy, celebrate by turning it into a top-level post.
The thrilling conclusion of what is likely to be an inaccurately named trilogy of June Open Threads.
|
9153b724-93cc-465c-b239-e72c5ac1fe30 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Language Model Tools for Alignment Research
*I do not speak for the rest of the people working on this project*
[I think it's valuable to have clear, short intros on different research agendas & projects]
How does this reduce x-Risk?
----------------------------
AI will continue to become increasingly more powerful; we should leverage this to accelerate alignment research. Language model tasks will also follow this trend (if transformers don't lead to AGI, whatever's next will still be capable of language tasks so this argument doesn't rely on transformers scaling). If you believe that certain research agendas reduce x-risk, then clearly giving them better tools to do their work faster also reduces x-risk.
Differential Impact
-------------------
Tools that can accelerate alignment research can probably be repurposed to accelerate capabilities research, so wouldn't developing these tools be net negative?
Yes. Especially if you gave them out to everybody or if you were a for-profit company with incentives to do so.
Only giving them out to alignment researchers is good. There's also lots of alignment-researcher-specific work to do, such as:
* 1. Collecting an alignment dataset
* 2. Understanding the workflows of alignment researchers
* 3. Making it incredibly easy for alignment researchers to use these tools
* 4. Keeping non-alignment specific datasets private
Though we could still increase capabilities by being the first to a capability and releasing that we succeeded. For example, OpenAI released their "inserting text" without telling people how they did it, but the people I work with, based off that information, figured out a way to do it too. The moral is that even just releasing that you succeeded is bits of information that those in the know can work backwards from.
Counterfactual Impact
---------------------
Let's say a project like this never gets started, there are still huge economic incentives to make similar products and sell to large amounts of people. Elicit, Cohere, Jasper (previously Jarvis), OpenAI, DeepMind, and more in the years to come will create these products, so why shouldn't we just use their products since they're likely to beat us to it and do it better?
Good point. Beyond the points made in "differential impact," having infrastructure/people to quickly integrate the latest advances into alignment researcher's workflows is useful even in this scenario. This includes engineers, pre-existing code for interfaces, & a data-labeling pipeline.
Current Work
------------
We've scraped LessWrong (including Alignment Forum), some blogs, relevant arxiv papers and their cited papers, and books. We are currently fine-tuning and trying to make it do useful tasks for us.
We've also released a survey and talked to several people about what would be most useful for them. However, the most important feedback may be actually demo-ing for different users.
Open Problems/Future Work
-------------------------
1. Cleaning up Data
2. Collecting and listing more sources of Alignment Data
3. Creating a pipeline of data-labelers for specific tasks
* Simple interface for collecting data for different tasks
4. Trying to get our models to do specific tasks
* clever ways of generating datasets (e.g. tl;dr from reddit for summaries)
* clever ways of prompting the model (e.g. a specific username from LW may be really good at a specific task)
* different fine-tuning formats (e.g. LW dataset "username: X, Karma: 200, post: " and change karma. Even do this with non-LW username authors from e.g. arxiv.)
5. Getting better feedback on the most useful tools for alignment researchers.
Who to Contact
--------------
DM me. Also join us at the #accelerating-alignment channel in the EleutherAI discord server: https://discord.gg/67AYcKK6. Or dm me for a link if it expired. |
80742539-7769-4ac8-8393-989f3d5436e0 | trentmkelly/LessWrong-43k | LessWrong | How *exactly* can AI take your job in the next few years?
Note: This article was primarily written for a less technical and less AI-savvy audience than LessWrong readers. If you're already familiar with upcoming AI developments, you'll probably find Part 2 (the second half of the article) more engaging.
We keep hearing that AI is going to replace a bunch of jobs in the next few years, but no one really knows exactly how that’s going to happen. When will this start, and which exact jobs are in danger? What’s the plan for dealing with this? Right now, it seems like people are vaguely aware of this threat but continuing on with business as normal. If this is really coming in the next 24 months, shouldn’t we be a little more worried, and have a few more answers to these questions?
That’s what I hoped to figure out when I started writing this article. I wanted to cut through the vague proclamations and actually lay out a bit of a clearer path for what’s coming. Other than feeling like society as a whole was not treating this issue with the seriousness it deserved, my motivation was also personal; I wanted to know, how long can I continue doing the kind of work I am doing?
This article is broken down into two parts:
1. Part 1: What will AI actually be able to do in the next few years? (this post)
Current AI systems clearly aren’t ready to replace most of our jobs, so when will they be ready?
2. Part 2: So…will we lose our jobs?
Once we understand what AI will actually be able to do, we’ll look at what that means for jobs and the white-collar labor market as a whole. Will you lose your job? I’ll walk through some specific scenarios that are likely about to play out.
This article focuses only on white-collar jobs, as that is the main segment of the workforce that AI will impact in the next few years.
If you are already familiar with most of the technological changes coming within the next few years, I suggest skipping to Part 2, as that may be more interesting to you.
Part 1: What will AI actually be able to d |
94d6cc77-b0d6-4963-8d81-82c1a7908c03 | trentmkelly/LessWrong-43k | LessWrong | Summer Tour Stops
Last week, Lily, Anna, Cecilia, Harris, and I drove to St Louis and back playing dances. I wrote about one aspect last week, and will probably continue splitting thoughts across posts. Today I'm thinking some about our approach to stopping along the way.
I grew up in a family with a very strong "maximize time at destination" approach to road trips. We would often leave before daylight, with stops determined by the minivan's requirements for additional gasoline. We would eat while driving, I recall creative efforts to handle certain biological needs, and we'd generally try to get the trip over with as quickly as possible.
I did sometimes see other people taking a different approach, but it wasn't until the first time I drove a group of college classmates a long way that I realized people have widely divergent preferences here. We were driving a van to Dance Flurry, about five hours North of Swarthmore, and around lunchtime people started talking about maybe stopping for lunch. I (really!) thought they were joking, and said that since we still had more gas and had brought along a bunch of food we didn't need to stop. There was a revolt, and we did not end up eating while driving.
Over time, I've come around to a view where one of the great things about traveling long distances by car is the opportunity for interesting stops. It's often the case that there are intermediate locations that are better places to spend the marginal hour than your origin or destination.
Before this tour we put some time into thinking about nice places to stop, and we also did some looking up options while driving. We ended up stopping at:
* Cecilia's father's violin shop, where he showed us how he repairs instruments:
Lily and Anna both got to take turns putting on clamps:
* Niagara Falls, on the way from Rochester to Pittsburgh. I'd been there once before, ten years earlier with Amy and Audrey, but it was great to take the older two k |
6973157f-6bab-42da-9653-82ab1352f9cd | trentmkelly/LessWrong-43k | LessWrong | The Simple World Hypothesis
Part of a Series in the Making: "If I Were God".
Introduction
> The current universe is the simplest possible universe with the same degree of functionality.
This hypothesis posits that the current universe is the simplest universe possible which can do all that our universe can do. Here, simplicity refers to the laws which make up the universe. It may be apt to mention the Multiverse Axioms at this juncture:
Axiom 1 (axiom of consistency):
> Any possible universe is logically consistent and strictly adheres to well-defined laws.
Axiom 2 (axiom of inclusivity):
> Whatever can happen (without violating 1) happens—and in every way possible (without violating 1).
Axiom 3 (axiom of simplicity):
The underlying laws governing the Multiverse are as simple as possible (while permitting 1 and 2).
The simple world hypothesis posits that our universe has the fewest laws which can enable the same degree of functionality that it currently possesses. I’ll explain the concept of “degree of functionality”. Take two universes: U_i and U_j with degrees of functionality d_i and d_j. Then the below three statements are true:
d_i > d_j implies that U_i can simulate U_j.
d_j < d_i implies that U_j cannot simulate U_i. d_i = d_j implies that U_i can simulate U_j, and U_j can in turn simulate U_i.
Let’s consider a universe like Conway’s Game of Life. It is far simpler than our universe and possesses only four laws. The simple world hypothesis argues that Conway’s Game of Life (U_c) cannot simulate our universe (U_0). The degree of functionality of Conway’s Game of Life (d_c) < the degree of functionality of our universe d_0. An advance prediction of the simple world hypothesis regarding U_c is the below:
> Human level intelligence cannot emerge in U_c.
The above implicitly assumes that Conway’s Game of Life is simpler than our universe—is that really true?
Simplicity
It is only prudent that I clarify what it is I mean by simplicity. For any two Universes U_i and U_j, l |
d4471d59-fc52-4997-b4e5-0e6489ebad7e | trentmkelly/LessWrong-43k | LessWrong | 2021 Darwin Game - Tundra
Our Tundra is an inhospitable[1] environment. The only significant food available to herbivores is Lichen, which has a tiny nutritional value of 1. The Tundra is cold too. Staying warm requires the cold tolerance adaptation, which costs +2 size.
Carrion Leaves Grass Seeds Detritus Coconuts Algae Lichen 1 1 1 1 1 0 0 300
An organism must expend 20% of its energy just to survive. A herbivore foraging for lichen cannot have a size greater than 5 or else it will expend more energy in metabolism than it is possible to acquire from eating Lichen.
All organisms have base size 0.1. The cold adaptation (+2) plus the Lichen digestive tract (+1) costs a total of +3 size. A Tundra herbivore has a minimum size of 3.1. A herbivore with size 5.1 is untenable since it expends more energy (1.02) than is possible to obtain from Lichen (1.00).
Players submitted 39 species native to the Tundra. Only 4 of them were viable herbivores: Micropas, Arctic Slug, Northern Nibbler and "lichen" (not to be confused with the foragable "Lichen"). (Multicore's Arctic Fox was a carnivore.)
These species could support little in the way of weapons, armor and speed. They were defenseless. In the first 8 turns, all four of our viable foragers are eaten to extinction.
Goes Extinct in Generation Species 5 Pristol 7 Micropas 7 Arctic Slug 8 Northern Nibbler 8 lichen
After the viable herbivores were eliminated, total ecological collapse was inevitable.
Goes Extinct in Generation Species 9 Yonge_Cold 9 Boreakeet 9 Beck’s Penguin 10 SmolFire 10 Arctic Ambusher 10 Zlorg 10 Arctic Fox 10 Orange-Krill 10 abominable_snowman 12 Antasvara 12 Unfortunately Large Cockroach 12 cg-mouse 13 Porostozer Malutki 13 1994 Mazda RX7 14 Raburetta 14 Pittsburgh-Penguins 15 Louse-lion 15 Wolverine 16 Jtp 16 Wolves 17 Seals 19 Direwolf 24 Tsc 27 Tundrus Rex 29 Frankenstein 32 Broken Fetters 34 Alaskans 37 Dragon 37 Porostozer Mamuci 39 Rocks 41 Duckofants 43 White-Whales 50 tp511 52 Frostwing Snipper
The Frostwing |
91a19b8f-ce51-4c75-8eb8-626603a794d3 | trentmkelly/LessWrong-43k | LessWrong | Existential Risk is a single category
A lot of people speak in terms of "existential risk from artificial intelligence" or "existential risk from nuclear war." While this is fine to some approximation, I rarely see it pointed out that this is not how risk works. Existential risk refers to the probability of a set of outcomes, and those outcomes are not defined in terms of their cause.
To illustrate why this is a problem, observe that there are numerous ways for two or more things-we-call-existential-risks to contribute equally to a bad outcome. Imagine nuclear weapons leading to a partial collapse of civilization, leading to an extremist group ending the world with an engineered virus. Do we attribute this to existential risk from nuclear weapons or from Bio-Terrorism? That question is neither well-defined, nor does it matter. All that matters is how much each factor contributes to [existential risk of any form].
Thus, ask not "is climate change an existential risk," but "does climate change contribute to existential risk?" Everything we care about is contained in the second question. |
8688f0f0-397c-46c1-a921-fe1221569533 | trentmkelly/LessWrong-43k | LessWrong | False Dilemmas w/ exercises
This is the third post in the Arguing Well sequence, but it can be understood on its own. This post is influenced by False Dilemma, The Third alternative.
A false dilemma is of the form “It’s either this, or that. Pick one!” It tries to make you choose from a limited set of options, when, in reality, more options are available. With that in mind, what’s wrong with the following examples?
Ex. 1: You either love the guy or hate him
Counterargument 1: “Only a Sith deals in absolutes!”
Counterargument 2: I can feel neutral towards the guy
Ex. 2: You can only add and subtract in mathematics.
Uh, division and multiplication, right?
Ex. 3: Either you get me that new car, or you don’t love me!
I can care about your well-being and happiness and not get you that specific car. I could buy a used car, so you wouldn’t freak out if it got damaged or [3 different examples based on what the (father?) thinks is best].
Ex. 4: You didn’t donate to the food drive, so you don’t care about starving children!
I do care about starving children, but my family and I are just barely scraping by and I value feeding them first.
Generalization
How would you generalize the above examples? Of course it presents a certain set of options as the only available options (I said this in the intro!), but what’s the relationship between those options? Are there different types of options that are similar/different between the examples?
One way to frame it is to separate options into two types: values and actions
With that frame, false dilemmas can be categorized into 4 varieties:
1. Only these values are compatible with an agent
2. Only these actions are compatible with a system/environment
3. Only these actions are compatible with this value
4. Only these values are compatible with this action
These 4 varieties correspond to each of their respective examples above (this is not a coincidence). Note: you could have answered Ex.3 as (4) and Ex.4 as (3), it really just depends on the truth |
8960109d-5460-49f1-b4a8-73182c705378 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Open Loop In Natura Economic Planning
1 Introduction
---------------
With the advent of industrial revolution and the almost homicidal conditions that this ensued, ensued, voices arose which insisted there must be superior ways of organising society to the markets. The end and subsequent dissolution of the Soviet Union in 1991 killed most (if not all) “living” attempts to create a non-market economy. The rule of experts and democracy, the two major facets of modernism (Scott, [1998](#bib.bib28)), were to be permitted expression only through market mechanisms. As this paper is written, very few domains of social life remain outside the whims of the market and a state that serves it. Consequently, the totality of human experience (with obvious exceptions) is now part of market transactions. The victory of the market is so absolute that certain authors complain in the popular imagination: “it is easier to envision the end of the world than the end of capitalism” (Fisher, [2009](#bib.bib17)) 111Some readers might complain that we conflate markets with capitalism. We are doing this on purpose..
Within such a political and ideological onslaught, it is no surprise that research in alternatives (or partial alternatives) to the market remained very limited in scope. In this paper, we revisit one such alternative paradigm of societal distribution, whose invention (or inspiration) goes back quite some time (Marx and Engels, [1885](#bib.bib24); Clark, [1984](#bib.bib11); Moseley, [1998](#bib.bib25)). We will provide a base for removing certain products from market circulation and provision them directly to citizens. The calculation of using products and services directly is generally called “planning in natura” (Cockshott, [2008](#bib.bib12)), and has direct links to Universal Basic Services. The goal of planning methods is to remove the anarchy (and uncertainty) of production and provide citizens with consumption guarantees. Contrary to most of the authors we cite, our ambitions are somewhat social-democratic. We do not aim to replace the market, but instead focus on removing human reproduction from strictly ideological mechanisms. In fact, a conservative government not “tied” to market ideology could easily start implementing such a programme222We are not really that naive. Power politics abide.. The goal of our specific programme is to match citizens and production units directly while monitoring the plan as closely as possible — in order to take corrective action — on a daily basis. Plan goals are to be formed using data collected from production units and citizens.
We are not aware of any methods that attempt to plan production on the individual level, nor has there ever been an automated way to monitor the plan or amend it using data. The closest a quasi-automated system of planning that reached an (partial) operational level was Project Cybersyn (Beer, [1979](#bib.bib5)), but this was dismantled in a hurry following Pinochet’s coup. Within the Soviet Union there is evidence that planning from final demand was seen as a “bourgeois” (Bollard, [2019](#bib.bib7)) and was never allowed, leaving production planning to the level of industrial goods (e.g., steel). The insistence to create plans and the focus of soviet economy to “build machines that build machines” might have contributed to the grim life of the soviet citizens in terms of consumer products. Prior to the late 1970s, when the demise of USSR became evident, some form of planning was always accepted within capitalist societies (Judt, [2006](#bib.bib20)). Japanese economists were effectively trained in planning by explicitly going through the works of Marx (Karatani, [2020](#bib.bib21)) until the late 80s. Our proposal of removing elements of production from market circulation is not historically controversial, but might look absurdly rebellious in a post-soviet world.
The rest of the paper is organised as follows; in Section [2](#S2 "2 Planning vs the market ‣ Open Loop In Natura Economic Planning") we provide a generic discussion on the background and debate between economic planning and market economics, but also nudge at the link between economic planning, reinforcement learning and AI planning. Section [3](#S3 "3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning") introduces a new model, which we term Open Loop In Natura Economic Planning. In Section [4](#S4 "4 Data collection ‣ Open Loop In Natura Economic Planning") we discuss data collection issues — and generally re-think the problem from the point of view of individual production units and citizens, while in Section [5](#S5 "5 Simulations ‣ Open Loop In Natura Economic Planning") we perform a series of simulations. We conclude with a short discussion in Section [6](#S6 "6 Conclusion ‣ Open Loop In Natura Economic Planning").
2 Planning vs the market
-------------------------
###
2.1 Input-output economics and planning
The problem of planning has been formally defined in Lahiri ([1976](#bib.bib22)), but we will attempt an RL based modernisation. Per unit of time t𝑡titalic\_t, a set of demands d𝑑ditalic\_d for certain goods (e.g, products, services) are to be satisfied for c𝑐citalic\_c citizens. The planner’s goal is to satisfy the demand of each citizen. In AI terms, we have something akin to a Markov Decision Process (MDP), with an agent (the planner) receiving information (the state) on the plan and a set of rewards related as to how closely the demand is met. Thinking of the problem as a single-player game-like MDP allows us to draw insights from the relevant literature, but it hides its complexity. The action space a planner would have to search through is massive - for 105superscript10510^{5}10 start\_POSTSUPERSCRIPT 5 end\_POSTSUPERSCRIPT individuals and 103superscript10310^{3}10 start\_POSTSUPERSCRIPT 3 end\_POSTSUPERSCRIPT goods, with each good having 10101010 different quality levels for each individual, the planner would have to choose among 109superscript10910^{9}10 start\_POSTSUPERSCRIPT 9 end\_POSTSUPERSCRIPT real-valued actions (that exist on a very abstract level). Direct (AI) planning for this problem has never been considered, rather the effort has concentrated on strategic plans that operate on aggregate demand and sector level, with recalculations of the plan never taking place — or, at best, on a yearly basis.
The parent of modern mechanisms for planning (in this context) is what is termed the input-output model, which is thoroughly reviewed by Leontief ([1986](#bib.bib23)). The model comprises of an nxn𝑛𝑥𝑛nxnitalic\_n italic\_x italic\_n Matrix A𝐴Aitalic\_A of technical coefficients, a vector x𝑥xitalic\_x of production level (i.e. how much we should produce for each product) and a demand vector d𝑑ditalic\_d. The columns of the coefficient matrix conceptually ask the question “how many units of each good to produce a single good of the type portrayed in this column do we need?”. The dot product of each row with the technical coefficients represents the consumption of a specific good. The demand vector d𝑑ditalic\_d represents how much external demand there is, i.e. that Equation [1](#S2.E1 "1 ‣ 2.1 Input-output economics and planning ‣ 2 Planning vs the market ‣ Open Loop In Natura Economic Planning") holds:
| | | | |
| --- | --- | --- | --- |
| | xi=ai1x1+ai2x2+…+ainxn+disubscript𝑥𝑖subscript𝑎𝑖1subscript𝑥1subscript𝑎𝑖2subscript𝑥2…subscript𝑎𝑖𝑛subscript𝑥𝑛subscript𝑑𝑖\centering x\_{i}=a\_{{i1}}x\_{1}+a\_{{i2}}x\_{2}+\ldots+a\_{{in}}x\_{n}+d\_{i}\@add@centeringitalic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT = italic\_a start\_POSTSUBSCRIPT italic\_i 1 end\_POSTSUBSCRIPT italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT + italic\_a start\_POSTSUBSCRIPT italic\_i 2 end\_POSTSUBSCRIPT italic\_x start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT + … + italic\_a start\_POSTSUBSCRIPT italic\_i italic\_n end\_POSTSUBSCRIPT italic\_x start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT + italic\_d start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT | | (1) |
In matrix notation, we have Equation [2](#S2.E2 "2 ‣ 2.1 Input-output economics and planning ‣ 2 Planning vs the market ‣ Open Loop In Natura Economic Planning"):
| | | | |
| --- | --- | --- | --- |
| | x=Ax+d⟹(I−A)x=d𝑥𝐴𝑥𝑑𝐼𝐴𝑥𝑑\centering x=Ax+d\implies(I-A)x=d\@add@centeringitalic\_x = italic\_A italic\_x + italic\_d ⟹ ( italic\_I - italic\_A ) italic\_x = italic\_d | | (2) |
Something to note here is that traditional input-output models have no notion of time - all production is taking place within the same temporal unit. This is somewhat counterintuitive (and problematic for actual planning), but it allows a first easy approximation. It is the model proposed by Cockshott and Cottrell ([1993](#bib.bib13)), covered by Dyer-Witheford ([2013](#bib.bib15)) and, with further additions (based on linear optimisation) discussed in Cockshott ([2008](#bib.bib12)). With no time element, the model remains suitable for very high level strategic planning - and indeed such models are widely used currently (e.g. most states publish input-output tables using monetary prices).
###
2.2 Why planning?
Von Mises and Hayek (von Hayek et al., [1935](#bib.bib32)), writing in the height of socialist revolutions, started putting together a critique of socialism, and more specifically (economic) planning. Parts of their critique (and this of their successors) sound still valid - for example same of their points on Marx’s treatment of skilled vs unskilled labour333Marx’s comment is that there are only quantitative differences in skilled vs unskilled labour. Here we will concentrate on the arguments of planning using products and services (i.e. 10 kilos of rice, 20 pounds of flesh, 10 hours of electric supply) vs a market price allocation mechanism. Whether an optimal (automated or not) planner of such type could even exist is termed the *calculation debate*. Arguments against the existence of an optimal planning mechanism fall into different camps, with some being aligned to moral questions (“it is unfair to just allocate goods” or “it is undemocratic”), computational (“you can’t compute the intermediate goods to produce”) or epistemic (“there is no way for the planner to know what to produce”). We will not discuss the democratic issue in this paper, though we strongly feel that the market is exceptionally undemocratic. It is now accepted by even the opponents of planning that computation should not be an issue (Brewster, [2004](#bib.bib8)). The epistemic argument, which is still very valid, entails that an optimal planner would not know *what* to compute. A price mechanism would allow whoever is engaged with the market to express their preferences of goods in terms of how much they would be willing to pay, i.e. a very subjective preference function. Prices that (for producers) might, for example, depend on the availability of goods (Steele, [2013](#bib.bib29)). In its extreme this holds true for consumers, as we have seen examples of iPad-for-kidney selling (Telegraph, [2020](#bib.bib30)), though we think it is safe to class such behaviours as pathological. If one makes the assumption of truly subjective values that vary continuously and are also widely different from person to person, then indeed a market might be able to allocate surpluses somewhat better than a plan. However, if you do accept that the majority of the population shares some similar preference function, at least in their top priorities (e.g. food, shelter, basic communication devices, electricity, health), the argument is nonsensical and applies only to incorporeal beings. Insofar as there are relatively slow changing patterns in consumption, standard machine learning models, combined with one’s own predictions can be used to forecast demand.
3 Open Loop In Natura Economic Planning
----------------------------------------
Our method (Open Loop In Natura Economic Planning - OLIN-EP) builds upon the basic input-output framework. It creates a fundamentally different planning landscape than IO tables and is heavily inspired by current game playing / RL agents. The planning “tick” is no longer a year, but a day, and we expect the plan to be re-calculated based on observations and predictions every night. We no longer operate on abstract notions of aggregate demand, but instead we expect every individual to communicate their demands and projected demands daily. We also expect the productive units to recalculate their input-output coefficients (which we will call IO-coeffs — the values of the matrix A𝐴Aitalic\_A) and provide them for plan updates on a daily basis in the form of a function — more on this later. Closing, we maintain a notion of state that is missing from all original formulations. More formally, we operate on an MDP (Puterman, [2014](#bib.bib26)) that has the following characteristics:
* •
Actions x∈𝒜𝑥𝒜x\in\mathcal{A}italic\_x ∈ caligraphic\_A capture what the production output of each industry should be. Note that due to notation conflicts with input-output literature we use x𝑥xitalic\_x for individual actions, rather than the most customary a𝑎aitalic\_a.
* •
States s∈S𝑠𝑆s\in Sitalic\_s ∈ italic\_S capture sufficient statistics of what we want to operate on, as transmitted every morning by production units and citizens. In our case, s𝑠sitalic\_s is simply a goods inventory.
* •
The transition function T(s′|s,a)𝑇conditionalsuperscript𝑠′
𝑠𝑎T(s^{\prime}|s,a)italic\_T ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | italic\_s , italic\_a ) is formally unknown to us, but it is captured partially by the input-output matrix, partially by the semantics we give to the behaviour of different outputs of the matrix, and it operates on the inventory and externalities.
* •
The reward function denotes how happy the planner is in a specific state and is generally encoded as R(s,a)𝑅𝑠𝑎R(s,a)italic\_R ( italic\_s , italic\_a ). We define later on a specific reward function that captures how well the plan targets are met and what damage the plan causes to the world.
* •
There is a discount factor γ𝛾\gammaitalic\_γ, which attenuates closer vs further rewards.
One can obviously claim that economic planning is more akin to a partially observable MDP (i.e. a POMDP), and this might be true, but unless one is to have the functions that describe the uncertainty over states, there is no reason to do the modelling this way. We could also start acting on histories of states and include externalities and rewards (Izadi and Precup, [2005](#bib.bib19)), but this might prove computationally infeasible. Claims could also be made that there is strong multi-agent element for the planner — here we assume that everyone involved in the plan has it in their best interest to cooperate.
###
3.1 The model
We adapt a number of innovations to the standard input-output models, by changing the way we position the plan within the economy. As discussed before, the goal of an input-output matrix is to plan for demand at the end of a time period. Given that our goal is to provide necessities to sustain humans, we set all “external” demand to zero, and introduce a set of profiles combined with the number of citizens attached to each profile. You can see an example in Table [1](#S3.T1 "Table 1 ‣ 3.3 Time and the transition function ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning"). Our input-output matrix describes the interactions between consumption profiles, a set of industrial goods, and a set of final goods. Profiles are columns that describe the allocation of final goods to each citizen that has been assigned this specific profile.
###
3.2 Nonlinearities and learning
The plan formulation we described above inherits a number of limitations from the standard input-output model; the first one we will build upon is model linearity. The default model linearity is tremendously problematic — for example there is the implicit assumption which is that labour needs will scale linearly with production demands. To address these issues, a generalisation of the input-output model (Lahiri, [1976](#bib.bib22); Fujimoto, [1986](#bib.bib18)) looks as in Equation [3](#S3.E3 "3 ‣ 3.2 Nonlinearities and learning ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning"):
| | | | |
| --- | --- | --- | --- |
| | (I−F(x))x=d𝐼𝐹𝑥𝑥𝑑\centering(I-F(x))x=d\@add@centering( italic\_I - italic\_F ( italic\_x ) ) italic\_x = italic\_d | | (3) |
This is profoundly liberating as a proposition, as we can stack production units and have different IO-coeffs values as production scales. We can also extract from individual citizens how important hitting certain targets in their profile is. Solving for x𝑥xitalic\_x now becomes a bit harder, as F(x)𝐹𝑥F(x)italic\_F ( italic\_x ) could potentially be any function, but in our case, we constrain it to a specific matrix. Remember that individual columns in the IO matrix represent how much it takes to produce a single unit of output — it makes sense to define the matrix as in Equation [4](#S3.E4 "4 ‣ 3.2 Nonlinearities and learning ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning")
| | | | |
| --- | --- | --- | --- |
| | F(x)=[f00(x0)f01(x0)…f0n(x0)f10(x1)f11(x1)…f1n(x1)⋮⋮⋱⋮fn0(xn)fn1(xn)…fnn(xn)]𝐹𝑥matrixsubscript𝑓00subscript𝑥0subscript𝑓01subscript𝑥0…subscript𝑓0𝑛subscript𝑥0subscript𝑓10subscript𝑥1subscript𝑓11subscript𝑥1…subscript𝑓1𝑛subscript𝑥1⋮⋮⋱⋮missing-subexpressionsubscript𝑓𝑛0subscript𝑥𝑛subscript𝑓𝑛1subscript𝑥𝑛…subscript𝑓𝑛𝑛subscript𝑥𝑛F(x)=\begin{bmatrix}{f\_{00}}{\left(x\_{0}\right)}&{f\_{01}}{\left(x\_{0}\right)}&\dots&{f\_{0n}}{\left(x\_{0}\right)}\\
{f\_{10}}{\left(x\_{1}\right)}&{f\_{11}}{\left(x\_{1}\right)}&\dots&{f\_{1n}}{\left(x\_{1}\right)}\\
\vdots&\vdots&\ddots&\vdots\\
\\
{f\_{n0}}{\left(x\_{n}\right)}&{f\_{n1}}{\left(x\_{n}\right)}&\dots&{f\_{nn}}{\left(x\_{n}\right)}\end{bmatrix}italic\_F ( italic\_x ) = [ start\_ARG start\_ROW start\_CELL italic\_f start\_POSTSUBSCRIPT 00 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT 01 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL … end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT 0 italic\_n end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_CELL end\_ROW start\_ROW start\_CELL italic\_f start\_POSTSUBSCRIPT 10 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT 11 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL … end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT 1 italic\_n end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) end\_CELL end\_ROW start\_ROW start\_CELL ⋮ end\_CELL start\_CELL ⋮ end\_CELL start\_CELL ⋱ end\_CELL start\_CELL ⋮ end\_CELL end\_ROW start\_ROW start\_CELL end\_CELL end\_ROW start\_ROW start\_CELL italic\_f start\_POSTSUBSCRIPT italic\_n 0 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ) end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT italic\_n 1 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ) end\_CELL start\_CELL … end\_CELL start\_CELL italic\_f start\_POSTSUBSCRIPT italic\_n italic\_n end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ) end\_CELL end\_ROW end\_ARG ] | | (4) |
Constraining our function to this form has one important benefit; we can ask production units directly how many other goods they need in order to produce certain output units, and data scientists in these facilities can use any machine learning method to “fit” a curve and provide back a function..
When it comes to the actual solution, one can attempt to use the gradient directly. The mean squared error MSE((I−F(x))x,d)𝑀𝑆𝐸𝐼𝐹𝑥𝑥𝑑MSE((I-F(x))x,d)italic\_M italic\_S italic\_E ( ( italic\_I - italic\_F ( italic\_x ) ) italic\_x , italic\_d ) has a gradient that is ∇MSE((I−F(x))x,d)=1/n((I−F(x))x−d)(I−F(x)−F′(x)x)∇𝑀𝑆𝐸𝐼𝐹𝑥𝑥𝑑1𝑛𝐼𝐹𝑥𝑥𝑑𝐼𝐹𝑥superscript𝐹′𝑥𝑥\nabla MSE((I-F(x))x,d)=1/n\left((I-F(x))x-d\right)(I-F(x)-F^{\prime}(x)x)∇ italic\_M italic\_S italic\_E ( ( italic\_I - italic\_F ( italic\_x ) ) italic\_x , italic\_d ) = 1 / italic\_n ( ( italic\_I - italic\_F ( italic\_x ) ) italic\_x - italic\_d ) ( italic\_I - italic\_F ( italic\_x ) - italic\_F start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ( italic\_x ) italic\_x ), which means that we can solve using any non-linear least squares algorithm — or in fact any other non-linear optimisation algorithm. Another method (that comes from Lahiri ([1976](#bib.bib22))) is to go through the power series expansion (I−A)−1=∑i=0∞Ai=I+A+A2+…superscript𝐼𝐴1superscriptsubscript𝑖0superscript𝐴𝑖𝐼𝐴superscript𝐴2…\left(I-A\right)^{-1}=\sum\_{i=0}^{\infty}A^{i}=I+A+A^{2}+...( italic\_I - italic\_A ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT = ∑ start\_POSTSUBSCRIPT italic\_i = 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ∞ end\_POSTSUPERSCRIPT italic\_A start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT = italic\_I + italic\_A + italic\_A start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT + … . We can then define x(i+1)=F(x(i))x(i)+d,x(0)=dformulae-sequencesubscript𝑥𝑖1𝐹subscript𝑥𝑖subscript𝑥𝑖𝑑subscript𝑥0𝑑x\_{(i+1)}=F(x\_{(i)})x\_{(i)}+d,x\_{(0)}=ditalic\_x start\_POSTSUBSCRIPT ( italic\_i + 1 ) end\_POSTSUBSCRIPT = italic\_F ( italic\_x start\_POSTSUBSCRIPT ( italic\_i ) end\_POSTSUBSCRIPT ) italic\_x start\_POSTSUBSCRIPT ( italic\_i ) end\_POSTSUBSCRIPT + italic\_d , italic\_x start\_POSTSUBSCRIPT ( 0 ) end\_POSTSUBSCRIPT = italic\_d — a recursive form of calculating x𝑥xitalic\_x. This is what we are going to use in this paper, as it is based purely on linear solvers, and will find the global maximum as long as convexity is maintained. We could also attempt an end-to-end neural network solution (it is very easy to envision), but there are no (clear) advantages, unless a need arises to model exceptionally complex IO-coeffs while optimising production at the same time, something we are not doing in this paper.
###
3.3 Time and the transition function
When it comes to producing goods and services, a model without a time element is severely limited; real production and consumption obviously have a time dimension. In the case of production, this is expressed in various forms like gestation times, production times, business inventories and depletion of resources. Multiple input-output models that include a time element have been developed444An example of such an Equation, from Raa ([1986](#bib.bib27)) is x(t)=∑0n[At+s(−s)x(t+s)]+∑0n{Bt+s+1(−s)[x(t+s)+x(t+s+1)]}+z(t)𝑥𝑡superscriptsubscript0𝑛delimited-[]subscript𝐴𝑡𝑠𝑠𝑥𝑡𝑠superscriptsubscript0𝑛subscript𝐵𝑡𝑠1𝑠delimited-[]𝑥𝑡𝑠𝑥𝑡𝑠1𝑧𝑡x(t)=\sum\_{0}^{n}\left[A\_{t+s}(-s)x(t+s)\right]+\sum\_{0}^{n}\left\{B\_{t+s+1}(-s)\left[x(t+s)+x(t+s+1)\right]\right\}+z(t)italic\_x ( italic\_t ) = ∑ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_n end\_POSTSUPERSCRIPT [ italic\_A start\_POSTSUBSCRIPT italic\_t + italic\_s end\_POSTSUBSCRIPT ( - italic\_s ) italic\_x ( italic\_t + italic\_s ) ] + ∑ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_n end\_POSTSUPERSCRIPT { italic\_B start\_POSTSUBSCRIPT italic\_t + italic\_s + 1 end\_POSTSUBSCRIPT ( - italic\_s ) [ italic\_x ( italic\_t + italic\_s ) + italic\_x ( italic\_t + italic\_s + 1 ) ] } + italic\_z ( italic\_t ), with A𝐴Aitalic\_A matrices representing circulating capital, all B𝐵Bitalic\_B matrices representing fixed capital, z(t)𝑧𝑡z(t)italic\_z ( italic\_t ) is the demand at each point in time, while −s𝑠-s- italic\_s is the ticks before the time t𝑡titalic\_t. (Raa, [1986](#bib.bib27); Dobos and Tallos, [2013](#bib.bib14); Aulin-Ahmavaara, [1990](#bib.bib3)) — for an overview, see Aulin-Ahmavaara ([2000](#bib.bib4)), The problem with these models is they were (for the most part) not designed with planning (in the AI sense) in mind. What we need to introduce (as discussed before) is a transition function T(s′|s,a)𝑇conditionalsuperscript𝑠′𝑠𝑎T(s^{\prime}|s,a)italic\_T ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | italic\_s , italic\_a ) and a notion of state s𝑠sitalic\_s. This can really be anything that makes sense based on the individual components of what we have, but to simplify things we can define state as an inventory indicating how much we hold of everything we have so far, including any unwanted side effects (i.e. externalities) our methods are generating. The transition function now operates on that inventory/externalities vector, by adding things, removing things, showing when something is ready for consumption, and how much needs to be taken to gestation periods.
| Type | Lucloelium | Vorpal Pick +1 | T-ring | Profile 0 | Profile 1 | Demand |
| --- | --- | --- | --- | --- | --- | --- |
| Lucloelium | 0.001 | f01(x0)subscript𝑓01subscript𝑥0f\_{01}(x\_{0})italic\_f start\_POSTSUBSCRIPT 01 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) | 1.000 | 3.0 | 2.0 | 0 |
| Vorpal Pick +1 | 0.500 | f11(x1)subscript𝑓11subscript𝑥1f\_{11}(x\_{1})italic\_f start\_POSTSUBSCRIPT 11 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) | 0.000 | 0.0 | 0.0 | 0 |
| T-ring | 0.000 | 0.000 | 0.000 | 0.1 | 0.2 | 0 |
| Lb(Lucloelium) | 0.001 | 0.000 | 0.000 | 0.0 | 0.0 | 0 |
| Lb(Vorpal Pick +1) | 0.000 | 0.012 | 0.000 | 0.0 | 0.0 | 0 |
| Lb(T-ring) | 0.000 | 0.000 | 0.001 | 0.0 | 0.0 | 0 |
| Profile 0 | 0.000 | 0.000 | 0.000 | 0.0 | 0.0 | 800 |
| Profile 1 | 0.000 | 0.000 | 0.000 | 0.0 | 0.0 | 500 |
Table 1: Our example input-output matrix, for a society of 1300 citizens. Two of the IO-coeffs vary with production levels - as there are three production units (see Figure [2](#S4.F2 "Figure 2 ‣ 4.2 Interactions with the market ‣ 4 Data collection ‣ Open Loop In Natura Economic Planning")) — the rest are constant. Labour columns are omitted, as all values are zero. There is one industrial good Vorpal Pick +1 and two final goods. Demand now just signifies the number of individuals in each profile.
###
3.4 Plan Humanity and externalities
The goal of the plan is to deliver a set of products and services (termed goods in our setup) in real life, so the real rewards can only be measured when the plan has been executed. During the planning phase, however, we should have a reasonable indication of what is the level of rewards we have achieved. Let di^(aij=0)^subscript𝑑𝑖subscript𝑎𝑖𝑗0\hat{d\_{i}}(a\_{ij}=0)over^ start\_ARG italic\_d start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_a start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT = 0 ) be the demand for a final good for a certain profile set to zero, with i𝑖iitalic\_i coming from final goods C𝐶Citalic\_C, while j𝑗jitalic\_j coming from profile consumption P𝑃Pitalic\_P. When we removed a good from a profile, we generate a surplus. That surplus, divided by how much that profile was expected to get, we define as the humanity of the plan. More formally, in Equation [5](#S3.E5 "5 ‣ 3.4 Plan Humanity and externalities ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning") we define humanity ℋ𝒰pℋsubscript𝒰𝑝\mathcal{HU}\_{p}caligraphic\_H caligraphic\_U start\_POSTSUBSCRIPT italic\_p end\_POSTSUBSCRIPT as
| | | | | |
| --- | --- | --- | --- | --- |
| | ℋ𝒰ptℋsubscriptsuperscript𝒰𝑡𝑝\displaystyle\mathcal{HU}^{t}\_{p}caligraphic\_H caligraphic\_U start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_p end\_POSTSUBSCRIPT | =mini∈C,j∈P{(di^(ai,j=0))/(aijdj)}absentsubscriptformulae-sequence𝑖𝐶𝑗𝑃^subscript𝑑𝑖subscript𝑎𝑖𝑗0subscript𝑎𝑖𝑗subscript𝑑𝑗\displaystyle=\min\_{i\in C,j\in P}\left\{\left(\hat{d\_{i}}(a\_{i,j}=0)\right)/\left(a\_{ij}d\_{j}\right)\right\}= roman\_min start\_POSTSUBSCRIPT italic\_i ∈ italic\_C , italic\_j ∈ italic\_P end\_POSTSUBSCRIPT { ( over^ start\_ARG italic\_d start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_ARG ( italic\_a start\_POSTSUBSCRIPT italic\_i , italic\_j end\_POSTSUBSCRIPT = 0 ) ) / ( italic\_a start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT italic\_d start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ) } | | (5) |
Every profile created puts certain requirements on the economy in terms of unwanted side effects, commonly referred to as externalities (e.g. carbon from milk and meat production). We model externalities at each point in time as ρ(e(xt)xt)𝜌𝑒subscript𝑥𝑡subscript𝑥𝑡\rho(e(x\_{t})x\_{t})italic\_ρ ( italic\_e ( italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ), with the total externalities for a plan being Epsubscript𝐸𝑝{E}\_{p}italic\_E start\_POSTSUBSCRIPT italic\_p end\_POSTSUBSCRIPT — the sum of all externalities in time as in Equation [6](#S3.E6 "6 ‣ 3.4 Plan Humanity and externalities ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning"), and ρ𝜌\rhoitalic\_ρ being a function that weights the importance of each externality for each good:
| | | | | |
| --- | --- | --- | --- | --- |
| | ℰptsuperscriptsubscriptℰ𝑝𝑡\displaystyle\mathcal{E}\_{p}^{t}caligraphic\_E start\_POSTSUBSCRIPT italic\_p end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT | =∑0tρ(e(xt)xt)absentsuperscriptsubscript0𝑡𝜌𝑒subscript𝑥𝑡subscript𝑥𝑡\displaystyle=\sum\_{0}^{t}{\rho(e(x\_{t})x\_{t})}= ∑ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_ρ ( italic\_e ( italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) italic\_x start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) | | (6) |
The difference between the way we measure the unwanted side-effects we get versus the goals we achieve is by design. In terms of production goals, a plan is as good as its worst performance. In terms of damage, we are measuring the cumulative effect — we call this the Marc Anthony principle555“The evil that men do lives after them; The good is oft interred with their bones”. A combination of externalities is what underlies the reward function.
###
3.5 Plan execution
Given that we do not have access to the real transition function (akin to training for a robot in an largely imperfect simulation), we suffer from two problems; first, that our plans are as limited in their ability to use future states as the imagination of the model creators. We will try to achieve certain goals every day for a year by following a set of actions that correspond to increasing production, without reference to future states - this is known as open loop planning - and is basically a vector x𝑥xitalic\_x per day. The fact that we re-plan on a daily basis means that we execute the plan in a closed loop setting - so overall we do *open loop planning, closed loop execution* (Bubeck and Munos, [2010](#bib.bib10); Weinstein and Littman, [2012](#bib.bib33)). This is highly reminiscent of methods like Monte Carlo Tree Search (Browne et al., [2012](#bib.bib9)) that have shown tremendous success in games. The second problem is that the artificial conditions we optimise on might not correspond to reality. Again, this is a common problem in robotics and it is currently attacked by assuming fictional model hyperparameters, as to make the model robust (Akkaya et al., [2019](#bib.bib2)).
4 Data collection
------------------
The real world execution of the plan entails two steps: (a) The planner provides information to the production units on their daily targets and requests information on the previous day history, including IO-coeffs in functional form and externalities. (b) The planner requests information on previous days demand and future demand from each individual (or discovers it).
Figure 1: An example of performance scaling in solving the basic Equation of our model (I−A)x=d𝐼𝐴𝑥𝑑(I-A)x=d( italic\_I - italic\_A ) italic\_x = italic\_d. Note that though A𝐴Aitalic\_A is sparse, this does not follow that x𝑥xitalic\_x would be.
###
4.1 Production units
Each production unit would have to effectively fill the columns of Matrix F𝐹Fitalic\_F by providing the function fij(xi)xisubscript𝑓𝑖𝑗subscript𝑥𝑖subscript𝑥𝑖f\_{ij}(x\_{i})x\_{i}italic\_f start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT, This can be achieved trivially by some form of active learning (i.e. asking managers: “how much milk do you need to make one pound of cheese? How about two pounds? How about three?”) and interpolating accordingly. Alternatively, one can seed a classic ML model using past production data and combine it with active learning in any gaps. Now, converting these values into fij(xi)subscript𝑓𝑖𝑗subscript𝑥𝑖f\_{ij}(x\_{i})italic\_f start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) simply required dividing over the number of actual products xisubscript𝑥𝑖x\_{i}italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT for all possible values of xisubscript𝑥𝑖x\_{i}italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT . We expect production units to innovate constantly, achieving lower externalities and better IO-coeffs , in a very organic process that amounts to optimisation coming from every part of the system.
####
4.1.1 Citizens
We have defined various profiles, but where do those profiles come from? This is essential — these profiles are our reward function. Learning a reward function from consumption targets can be done by using any form of inverse reinforcement learning/preference learning on existing buying habits, direct questions and/or voting all in accordance with productive capacities. This should allow for effectively the discovery of basic needs on a fundamental level and the provision of relevant goods. From the outset, different profiles aim at addressing the problem of Variety (Beer, [1993](#bib.bib6)) directly, i.e. we need to be able to act upon as many world states are possible. Individual profiles for every person would put tremendous strain on the planning mechanism and make the whole system very brittle, as any errors in production will result in a series of complaints. Instead, the focus should be on goods that allow for a high degree of customisation. For example, pre-packaged foods are a really bad production option, as they allow for very little tinkering. Allowing for very high degree of customisation and personalisation (i.g. a combination of (generative?) recipes plus food) should help make production both more robust and interesting. New types of computing devices, whose aim is to help so as to have the goods delivered be used in the most efficient and creative fashion possible, will also prove pivotal.
###
4.2 Interactions with the market
Given that the plan’s aim is to complement, rather than abolish, the market, it is worth discussing what areas of production the plan will not shape. Goods in scarcity or products whose only value is their scarcity cannot be delivered through the plan; the subjectivity of the reward function would make it exceptionally hard to calculate individual preferences (and hence profiles), and would also open up the possibility of abuses, requiring constant vigilance to stop the creation of black markets. Goods in scarcity also open questions of multi-objective optimisation (Erickson et al., [2013](#bib.bib16)) — that will mostly lead to a wealth of equally non-satisfactory solutions. Any invention that helps the plan should be readily adapted. New products and services could also come from market forces. This would require the market to turn into activities that look more like prospecting — anything that a plan cannot cover should generate profit. The most important point, however, when it comes to market, is not to allow it to use the plan as a way of undercutting wages; once the plan is introduced, it should be followed by a policy of increasing minimum wages and decreasing working hours, in accordance with productivity gains in order to start removing human labour from the market and reaping benefits from further automation. For example, shoe production is still a very manual process, and high wages in the sector should come from automation.
(a) An example of how much Lucloelium and Vorpal Pick +1 is needed to create units Vorpal Pick +1 as portrayed in the x axis - i.e. fij(xi)xisubscript𝑓𝑖𝑗subscript𝑥𝑖subscript𝑥𝑖f\_{ij}(x\_{i})x\_{i}italic\_f start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT.
(b) fij(xi)subscript𝑓𝑖𝑗subscript𝑥𝑖f\_{ij}(x\_{i})italic\_f start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) for Lucloelium and Vorpal Pick +1 .
Figure 2: f01(x0)subscript𝑓01subscript𝑥0f\_{01}(x\_{0})italic\_f start\_POSTSUBSCRIPT 01 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) and f11(x1)subscript𝑓11subscript𝑥1f\_{11}(x\_{1})italic\_f start\_POSTSUBSCRIPT 11 end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) derivation from production outputs. There are three fictional production units that follow very different curves in their models.
5 Simulations
--------------
We performed a number of simulations on imaginary data. The first set of simulations resolves around solving (I−A)x=d𝐼𝐴𝑥𝑑(I-A)x=d( italic\_I - italic\_A ) italic\_x = italic\_d repeatedly for matrices of different size. Solving this set of linear Equations fast is fundamental as both our time element and the non-linearity solution depend it. We have run all possible combinations of industrial goods (i.e. goods not needed by the profiles, [500,1000,5000,10000,50000]500100050001000050000[500,1000,5000,10000,50000][ 500 , 1000 , 5000 , 10000 , 50000 ], final goods of [50,100,500,1000,5000]5010050010005000[50,100,500,1000,5000][ 50 , 100 , 500 , 1000 , 5000 ], a profile of size 200200200200 (i.e. 200 different combinations of final goods), with each good needing [500,1000,2000]50010002000[500,1000,2000][ 500 , 1000 , 2000 ] other goods in order to be made. The results can be seen in Figure [1](#S4.F1 "Figure 1 ‣ 4 Data collection ‣ Open Loop In Natura Economic Planning") — all results were collected on a CPU: Intel i7-8700K @ 4.800GHz / 64GB RAM, using scipy (Virtanen et al., [2020](#bib.bib31)). Alternative solutions that include gradient estimations might be faster, but this will probably depend on the problem. As it stands, the deciding speed factor is the number of dependencies, but everything is solved in well below 20 seconds. Overall, it is trivial to attack the problem,
We also simulate a sample, completely fictional economy of an alien village, in Table [1](#S3.T1 "Table 1 ‣ 3.3 Time and the transition function ‣ 3 Open Loop In Natura Economic Planning ‣ Open Loop In Natura Economic Planning"). The economy is made up from two final goods (Lucloelium ,T-ring ) and one industrial good (Vorpal Pick +1 ). The initial quantities of each item in inventory are restricted. The results of the simulation can be seen in Figure [3](#S5.F3 "Figure 3 ‣ 5 Simulations ‣ Open Loop In Natura Economic Planning"). The village plans to provide the final goods in two profiles. The village starts without being able to fulfil the goals of each profile, hence they are forced to produce a limited amount of goods at each daily tick and invest the rest. What this means in practical terms is that units of Lucloelium we create get “consumed”, while units of Vorpal Pick +1 just get added on. Notice the exponential rise in humanity of the plan. We perform a second experiment, where with a certain probability a portion of the inventory would just vanish. Here (see Figure [3](#S5.F3 "Figure 3 ‣ 5 Simulations ‣ Open Loop In Natura Economic Planning")(b)) lower investment leads to collapse, with the humanity of the plan never recovering. This effect would not be visible without including some noise to the model. Finally, also note that the only real difference between a simulation and a plan comes from the fact that we think that the simulation is closer to reality — there is no way to execute it in real life.
(a) A simulation without noise
(b) A simulation with noise - certain investment profiles fail to achieve self-sustainability.
Figure 3: Humanity of the plan vs investment profiles. Note the exponential curves.
6 Conclusion
-------------
Lenin is quoted and attacked directly by Von Mises in von Hayek et al. ([1935](#bib.bib32)). His remarks concern the efficacy of the Leninist take on society — planning is more or less an ideologically motivated position put forward by a dictator. We think the Tables might have turned — if anything, with the current development of the means of production the market looks more and more like an ideological tool and a vampiric dictator, lurching from crisis after crisis. At the same time, a “no-future” rhetoric seems to be becoming the norm. The market (and in turn, capital) operate on humans indirectly, the market “reward” function does not have to do with meeting needs, but rather with profit maximisation. Insomuch, it makes human lives an externality. Our method, OLIN-EP, simply combines a set of production units with citizens’ basic needs in an RL-like format. Putting together a full planning programme for basic goods is not trivial, but it is painfully obvious that the technological tools have been there for some time. We hope that this paper re-starts the discussion on a technical level, with ever increasing planning methods and simulations coming to light. There is no reason for the plan to be as simple as the one discussed here — in fact Facebook is currently performing large scale simulations (Ahlgren et al., [2020](#bib.bib1)); one can easily envision a situation where simulated production/consumption unit behaviour is used to plan in a setting more closely aligned to traditional RL.
7 Broader Impact
-----------------
The whole of this article is essentially societal commentary, proposing a new production system for basic products and services. If what we propose is adapted, even if not in the exact form we discuss above, it will change the way society works forever. When authors deliberate on the impact of Artificial Intelligence or Machine Learning on society, they follow a line almost always bound within a liberal framework (e.g. how to make sure women and men have the same opportunities of getting “good jobs” ). We think it is high time to move beyond liberalism when discussing technological impact. |
dc7ed9b2-eeb3-4766-897a-2166af693095 | trentmkelly/LessWrong-43k | LessWrong | A simple sketch of how realism became unpopular
[Epistemic status: Sharing current impressions in a quick, simplified way in case others have details to add or have a more illuminating account. Medium-confidence that this is one of the most important parts of the story.]
Here's my current sense of how we ended up in this weird world where:
* I still intermittently run into people who claim that there's no such thing as reality or truth;
* a lot of 20th-century psychologists made a habit of saying things like 'minds don't exist, only behaviors';
* a lot of 20th-century physicists made a habit of saying things like 'quarks don't exist, only minds';
* there's a big academic split between continental thinkers saying (or being rounded off to saying) some variant of "everything is culture / perception / discourse / power" and Anglophone thinkers saying (or being rounded off to saying) "no".
Background context:
1. The ancient Greeks wrote down a whole lot of arguments. In many cases, we're missing enough textual fragments or context that we don't really know why they were arguing — what exact propositions were in dispute, or what the stakes were.
2. In any case, most of this is screened off by the fact that Europe's memetic winners were Christianity plus normal unphilosophical beliefs like "the sky is, in fact, blue".
3. Then, in 1521, the Protestant Reformation began.
4. In 1562, the Catholics found a giant list of arguments against everything by the minor Greek skeptic Sextus Empiricus, got very excited, and immediately weaponized them to show that the Protestant arguments fail (because all arguments fail).
5. These soon spread and became a sensation, and not just for being a useful superweapon. Plenty of intellectuals were earnest humanists used to taking arguments at face value, and found Sextus' arguments genuinely upsetting and fascinating.
I trace continental thinkers' "everything is subjective/relative" arguments back to a single 1710 error in George Berkeley:
> [...] I am content to put the who |
d9bc164d-6610-49a9-9f31-de3ca5d66917 | trentmkelly/LessWrong-43k | LessWrong | Weird Alliances
In the recent discussion on supplements, I commented on how weird an alliance health stores are. They cater for clientèle with widely divergent beliefs about how their merchandise works, such as New Agers vs. biohackers. In some cases, they cater for groups with object-level disputes about their merchandise. I imagine vegans are stoked to have somewhere to buy dairy-free facsimiles of everyday foods, but they're entering into an implicit bargain with that body-builder who's walking out of the door with two kilos of whey protein.
In the case of health stores, their clientèle have a common interest which the store is satisfying: either putting esoteric substances into their bodies, or keeping commonplace substances out of their bodies. This need is enough for people to hold their noses as they put their cash down.
(I don't actually know how [my flimsy straw-man model of], say, homoeopathy advocates feel about health stores. For me, it feels like wandering into enemy territory.)
I've been thinking lately about "allies" in the social justice sense of the word: marginalised groups who have unaligned object-level interests but aligned meta-interests. Lesbians, gay men, bisexuals, transfolk and [miscellaneous gender-people] may have very different object-level interests, but a very strong common meta-interest relating to the social and legal status of sexual identities. They may also be marginalised along different axes, allowing for some sort of trade I don't have a good piece of terminology for. The LGBT([A-Z]).* community is an alliance. Not being part of this community, I'm hesitant to speculate on how much of a weird alliance it is, but it looks at least a little bit weird.
This has led me to think about Less Wrong as a community, in particular the following two questions:
To what extent is Less Wrong a weird alliance?
On paper, we're all here to help refine the art of human rationality, but in practice, we have a bunch of different object-level interests and |
b6637b2e-3977-434c-a764-3a56beaff1d4 | trentmkelly/LessWrong-43k | LessWrong | A multi-disciplinary view on AI safety research
Summary
The core ideas that constitute the multi-disciplinary view[1] on AI safety research are:
* Theoretical research on how to build “safe” AGI systems (”top-down intelligent design”) is necessary: we must not passively follow the techno-evolution of AI within the present systems of capitalistic economics and global governance and only do empirical research of the developed systems as this evolution goes along.
* AI safety research field shouldn’t be separate from the wider AGI research field: they should converge as a project of top-down design of civilisational intelligence. Both AI alignment paradigms (protocols) and AGI capability research (intelligence architectures) that don’t position themselves within a certain design for civilisational intelligence are methodologically misguided and could be dangerous.
* “Purely technical” AI safety research and all other types of research that are needed to engineer civilisational intelligence (sociotechnical, social, legal, political, geopolitical, and strategic research) shouldn’t be separated as well. Yet, in this article, I focus on the technical side of it.
* Technical research of civilisational intelligence shouldn’t focus only on cognitive science, mathematics, theoretical computer science, and machine learning: dozens of different disciplinary perspectives should be taken to evaluate civilisational intelligence architectures adequately, including, in addition to the disciplines mentioned above, the perspectives of neuroscience, theories of consciousness, theories of collective intelligence, social dynamics and memetics, game theory, mechanism design, network theory, theories of criticality and emergence, dynamical systems and stability theories, scale-free theories of evolution and cognitive development, distributed systems and control theory, resilience theory, safety science and reliability engineering science, physics of communication and computation, information security and cryptography research, and m |
e6fb61a8-2640-470e-b56f-326635740c79 | trentmkelly/LessWrong-43k | LessWrong | Intermittent Distillations #4: Semiconductors, Economics, Intelligence, and Technological Progress.
The Semiconductor Supply Chain: Assessing National Competitiveness (Saif M. Khan, Alexander Mann, Dahlia Peterson)
The Semiconductor Supply Chain: Assessing National Competitiveness
Summary
This report analyzes the current supply chain for semiconductors. It particularly focuses on which portions of the supply chain are controlled by US and its allies and China. Some key insights:
* The US semiconductor industry is estimated to contribute 39 percent of the total value of the global semiconductor supply chain.
* The semiconductor supply chain is incredibly complicated. The production of a single chip requires more than 1,000 steps and passes through borders more than 70 times throughout production.
* AMD is currently the only company with expertise in designing both high-end GPUs and high-end CPUs.
* TSMC controls 54% of the logic foundry market, with a larger share for leading edge production, e.g., state-of-the-art 5 nm node chips.
* Revenue per wafer for TSMC is rapidly increasing, while other foundries are seeing declines.
* The Netherlands has a monopoly on extreme ultraviolet (EUV) scanners, equipment needed to make the most advanced chips.
* The Netherlands and Japan have a monopoly on argon floride (ArF) immersion scanners, needed to make the second most advanced chips.
* The US has a monopoly on full-spectrum electronic design automation (EDA) software needed to design semiconductors.
* Japan, Taiwan, Germany and South Korea manufacture the state-of-the-art 300 mm wafers used for 99.7 percent of the world's chip manufacturing. This manufacturing process requires large amounts of tacit know-how.
* China controls the largest share of manufacturing for most natural materials. The US and its allies have a sizable share in all materials except for low-grade gallium, tungsten and magnesium.
* China controls ~2/3rds of the world's silicon production, but the US and allies have reserves.
The report also analyzes US competitiveness at very detailed lev |
39272238-8352-4b2c-87be-c4fc96d0717d | StampyAI/alignment-research-dataset/arxiv | Arxiv | Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images
1 Introduction
---------------
Breast cancer is one of the most common causes of mortality in the female population in the world [[2](#bib.bib2)]. It accounts for around 25%percent2525\%25 % of all the cancers diagnosed in women [[3](#bib.bib3)]. For traditional diagnostic tools like mammography, even experienced radiologists can miss 10−30%10percent3010-30\%10 - 30 % of breast cancers during routine screenings [[7](#bib.bib7)]. With the advent of digital imaging, whole-slide imaging has gained attention from the clinicians and pathologists because of its reliability. Whole-slide images (WSIs) have been permitted for diagnostic use in the USA [[1](#bib.bib1)]. They are the high-resolution scans of conventional glass slides with Hematoxylin and Eosin (H&E) stained tissue. There are four types of tissue in breast biopsy: normal, benign, in situ carcinoma, and invasive carcinoma. Fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images") shows examples of the four types of breast tissue. In clinical testing, the pathologists diagnose breast cancer based on 1) the percentage of tubule formation, 2) the degree of nuclear pleomorphism, and 3) the mitotic cell count [[8](#bib.bib8)].
(a) (a) normal
(b) (b) benign
(c) (c) in situ
(d) (d) invasive
Figure 1: Examples of different types of tissue. The microscopy images (patches of WSIs at 200×200\times200 × magnification) are labeled according to the predominant tissue type in each image.
Convolutional Neural Networks (CNNs) can be trained in an end-to-end manner to distinguish the different types of cancer, by extracting high-level information from images through stacking convolutional layers. Breast cancer classification has been fundamentally improved by the development of CNN models [[16](#bib.bib16)].
However, breast cancer segmentation in WSIs is still underexplored. WSIs are RGB images with high resolution (e.g. 80000×60000800006000080000\times 6000080000 × 60000). Constrained by the memory, WSIs cannot be directly fed into the network. One solution is to crop the WSIs to small patches for patch-wise training [[4](#bib.bib4)]. Given a fixed input size, however, there is a trade-off between accuracy and the inference speed. One can efficiently reduce the inference cost by cropping the WSIs to larger patches and rescaling the patches to a smaller input size, but this results in a loss of detail and sacrifices accuracy. In WSIs, the suspicious cancer areas our regions of interest (ROIs), are sparse, since most regions are normal tissue or the glass slide. The four classes are therefore highly imbalanced. Further, the pixel-wise annotation of breast cancer segmentation requires domain knowledge and extensive human labor and the ground truth labels are often noisy at the pixel-level. Training on patches with a small field of view can therefore easily lead to overfitting.
In this paper, we propose a semantic segmentation framework, Reinforced Auto-Zoom Net (RAZN). When a pathologist examines the WSIs with a digital microscope, the suspicious areas are zoomed in for details and the non-suspicious areas are browsed quickly (See Fig. [2](#S1.F2 "Figure 2 ‣ 1 Introduction ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images") for an intuition.). RAZN is motivated by this attentive zoom-in mechanism. We learn a policy network to decide the zoom-in action through the policy gradient method [[14](#bib.bib14)]. By skipping the non-suspicious areas (normal tissue), noisy information (glass background) can be ignored and the WSIs can be processed more quickly. By zooming in the suspicious areas (abnormal tissue), the data imbalance is alleviated locally (in the zoomed-in regions) and more local information is considered. Combining these two can efficiently reduce overfitting for the normal tissue, which is caused by the imbalanced data, and lead to improved accuracy. However, since the zoom-in action is selective, the inference can at the same time be fast.
The previous studies on zoom-in mechanism focus on utilizing multi-scale training to improve prediction performance. The Hierarchical Auto-Zoom Net HAZN [[19](#bib.bib19)] uses sub-networks to detect human and object parts at different scales hierarchically and merges the prediction at different scales, which can be considered as a kind of ensemble learning. Zoom-in-Net [[17](#bib.bib17)] zooms in suspicious areas generated by attention maps to classify diabetic retinopathy. In both HAZN and Zoom-in-Net, the zoom-in actions are deterministic. So in the training phase, the patches will be upsampled and trained even if it may not decrease the loss. In RAZN, the zoom-in actions are stochastic, and a policy is learned to decide if the zoom-in action can improve the performance.
(a) (a)
(b) (b)
(c) (c)
(d) (d)
Figure 2: Zoom-in process. The regions bounded by the red boxes are zoomed in sequentially with zoom-in rate 2. All zoomed-in regions are resized to the same resolution for visualization. The white regions in (a), (b) and (c) are the background glass slide.
This paper makes the following contributions: 1) we propose an innovative framework for semantic segmentation for images with high resolution by leveraging both accuracy and speed; 2) we are the first to apply reinforcement learning to breast cancer segmentation; 3) we compare our framework empirically with multi-scale techniques used in the domain of computer vision and discuss the influence of multi-scale models for breast cancer segmentation.
2 Reinforced Auto-Zoom Net
---------------------------
In clinical practice, it is impossible for a clinician to go through each region of a WSI at the original resolution, due to the huge image size. The clinician views regions with simple patterns or high confidence quickly at coarse resolution and zooms in for the suspicious or uncertain regions to study the cells at high resolution. The proposed RANZ simulates the examining process of a clinician diagnosing breast cancer on a WSI. Another motivation of RAZN is that the characteristics of the cancer cells have different representations at different field of view. For semantic segmentation tasks on common objects, the objects in the same category share discriminative features and attributes. For example, we can differentiate a cat from a dog based on the head, without viewing the whole body. However, in cancer segmentation, the basic unit is the cell, which consists of nucleus and cytoplasm. The difference between the cells is not obvious. Instead of checking only a single cell, the diagnosis is based on the features of a group of cells, such as the density, the clustering and the interaction with the environment. RANZ is designed to learn this high-level information.
RAZN consists of two types of sub-networks, policy networks {fθ}subscript𝑓𝜃\{f\_{\theta}\}{ italic\_f start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT } and segmentation networks {gϕ}subscript𝑔italic-ϕ\{g\_{\phi}\}{ italic\_g start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT }. Assume the zoom-in actions can be performed at most m𝑚mitalic\_m times and the zoom-in rate is r𝑟ritalic\_r. There is one base segmentation network fθ0subscript𝑓subscript𝜃0f\_{\theta\_{0}}italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT at the coarsest resolution. At the i𝑖iitalic\_ith zoom-in level, there is one policy network gϕisubscript𝑔subscriptitalic-ϕ𝑖g\_{\phi\_{i}}italic\_g start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT and one segmentation network, fθisubscript𝑓subscript𝜃𝑖f\_{\theta\_{i}}italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT. In the inference time, with fixed field of view and magnification level, we have a cropped patch x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT with shape [H,W,3]𝐻𝑊3[H,W,3][ italic\_H , italic\_W , 3 ], like Fig. [2](#S1.F2 "Figure 2 ‣ 1 Introduction ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images") (a). Then gϕ1subscript𝑔subscriptitalic-ϕ1g\_{\phi\_{1}}italic\_g start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT will take x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT as an input and predict the action, zoom-in or break. If the predicted action is break, fθ0(x0)subscript𝑓subscript𝜃0subscript𝑥0f\_{\theta\_{0}}(x\_{0})italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) will output the segmentation results and the diagnosis for x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT is finished. If the predicted action is zoom-in, a high-magnification patch x¯0subscript¯𝑥0\bar{x}\_{0}over¯ start\_ARG italic\_x end\_ARG start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT with corresponding zoom-in rate will be retrieved from the original image. x¯0subscript¯𝑥0\bar{x}\_{0}over¯ start\_ARG italic\_x end\_ARG start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT, with shape [rH,rW,3]𝑟𝐻𝑟𝑊3[rH,rW,3][ italic\_r italic\_H , italic\_r italic\_W , 3 ], will be cropped into x1subscript𝑥1x\_{1}italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT, which is r2superscript𝑟2r^{2}italic\_r start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT patches of shape [H,W,3]𝐻𝑊3[H,W,3][ italic\_H , italic\_W , 3 ]. Then each patch of x1subscript𝑥1x\_{1}italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT will be treated as x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT for the next level of zoom-in action. Fig. [2](#S1.F2 "Figure 2 ‣ 1 Introduction ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images") (b) is a central crop of x1subscript𝑥1x\_{1}italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT. The process is repeated recursively until a pre-defined maximum magnification level is reached. In this work, we propose this novel idea and focus on the situation of m=1𝑚1m=1italic\_m = 1. m>1𝑚1m>1italic\_m > 1 will be discussed in future work. An overview of the architecture is illustrated in Fig. [3](#S2.F3 "Figure 3 ‣ 2 Reinforced Auto-Zoom Net ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images").
Figure 3: Illustration of the proposed framework when m=1𝑚1m=1italic\_m = 1 and r=2𝑟2r=2italic\_r = 2. In the inference phase, given a cropped image x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT, the policy network outputs the action, zoom-in (red arrows) or break (blue arrows). In the training phase, the policy network will be optimized to maximize the reward (purple arrows), which is determined by the segmentation prediction.
The segmentation networks are Fully Convolutional Networks (FCNs) [[12](#bib.bib12)] and share the same architecture. However, unlike parameter sharing in the common multi-scale training in semantic segmentation [[5](#bib.bib5)], each network is parameterized by independent fθsubscript𝑓𝜃f\_{\theta}italic\_f start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT, where fθi:ℝH×W×3→ℝH×W×C:subscript𝑓subscript𝜃𝑖→superscriptℝ𝐻𝑊3superscriptℝ𝐻𝑊𝐶f\_{\theta\_{i}}:\mathbb{R}^{H\times W\times 3}\rightarrow\mathbb{R}^{H\times W\times C}italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT : blackboard\_R start\_POSTSUPERSCRIPT italic\_H × italic\_W × 3 end\_POSTSUPERSCRIPT → blackboard\_R start\_POSTSUPERSCRIPT italic\_H × italic\_W × italic\_C end\_POSTSUPERSCRIPT and C𝐶Citalic\_C is the number of classes. The reason for choosing independent networks for each zoom-in level is that CNNs are not scale-invariant [[9](#bib.bib9)].
Each FCN can thus learn high-level information at a specific magnification level. Given input image x𝑥xitalic\_x and segmentation annotation y𝑦yitalic\_y, the training objective for each FCN is to minimize
| | | | |
| --- | --- | --- | --- |
| | Jθi(x,y)=−1HW∑j∑cyj,clogfθi(x)j,c,subscript𝐽subscript𝜃𝑖𝑥𝑦1𝐻𝑊subscript𝑗subscript𝑐subscript𝑦𝑗𝑐subscript𝑓subscript𝜃𝑖subscript𝑥𝑗𝑐J\_{\theta\_{i}}(x,y)=-\frac{1}{HW}\sum\_{j}\sum\_{c}y\_{j,c}\log f\_{\theta\_{i}}(x)\_{j,c}\;,italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x , italic\_y ) = - divide start\_ARG 1 end\_ARG start\_ARG italic\_H italic\_W end\_ARG ∑ start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_c end\_POSTSUBSCRIPT italic\_y start\_POSTSUBSCRIPT italic\_j , italic\_c end\_POSTSUBSCRIPT roman\_log italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x ) start\_POSTSUBSCRIPT italic\_j , italic\_c end\_POSTSUBSCRIPT , | | (1) |
where j𝑗jitalic\_j ranges over all the H×W𝐻𝑊H\times Witalic\_H × italic\_W spatial positions and c∈{0,…,3}𝑐0…3c\in\{0,...,3\}italic\_c ∈ { 0 , … , 3 } represents the semantic classes (cancer type).
At m=1𝑚1m=1italic\_m = 1, the framework is a single-step Markov Decision Process (MDP) and the problem can be formulated by the REINFORCE rule [[18](#bib.bib18)]. The policy network projects an image to a single scalar, gϕ1:ℝH×W×3→ℝ:subscript𝑔subscriptitalic-ϕ1→superscriptℝ𝐻𝑊3ℝg\_{\phi\_{1}}:\mathbb{R}^{H\times W\times 3}\rightarrow\mathbb{R}italic\_g start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT : blackboard\_R start\_POSTSUPERSCRIPT italic\_H × italic\_W × 3 end\_POSTSUPERSCRIPT → blackboard\_R. Given the state x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT, the policy network defines a policy πϕ1(x0)subscript𝜋subscriptitalic-ϕ1subscript𝑥0\pi\_{\phi\_{1}}(x\_{0})italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ). The policy samples an action a∈{0,1}𝑎01a\in\{0,1\}italic\_a ∈ { 0 , 1 }, which represents break and zoom-in, respectively. We have
| | | | |
| --- | --- | --- | --- |
| | 𝑝=σ(gϕ1(x0)),𝑝𝜎subscript𝑔subscriptitalic-ϕ1subscript𝑥0\textit{p}=\sigma(g\_{\phi\_{1}}(x\_{0}))\;,p = italic\_σ ( italic\_g start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) ) , | | (2) |
| | | | |
| --- | --- | --- | --- |
| | πϕ1(x0)=𝑝a(1−𝑝)1−a,subscript𝜋subscriptitalic-ϕ1subscript𝑥0superscript𝑝𝑎superscript1𝑝1𝑎\pi\_{\phi\_{1}}(x\_{0})=\textit{p}^{a}(1-\textit{p})^{1-a}\;,italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) = p start\_POSTSUPERSCRIPT italic\_a end\_POSTSUPERSCRIPT ( 1 - p ) start\_POSTSUPERSCRIPT 1 - italic\_a end\_POSTSUPERSCRIPT , | | (3) |
where σ(⋅)𝜎⋅\sigma(\cdot)italic\_σ ( ⋅ ) is the sigmoid function and πϕ1(x0)subscript𝜋subscriptitalic-ϕ1subscript𝑥0\pi\_{\phi\_{1}}(x\_{0})italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) is essentially a Bernoulli distribution. The motivation of RAZN is to improve the segmentation performance and it is therefore natural to define the reward such that it minimizes the segmentation loss. Based on Equation [1](#S2.E1 "1 ‣ 2 Reinforced Auto-Zoom Net ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images"), we have Jθ0(x0,y0)subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0J\_{\theta\_{0}}(x\_{0},y\_{0})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ), Jθ1(x1,y1)subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1J\_{\theta\_{1}}(x\_{1},y\_{1})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ), where x1subscript𝑥1x\_{1}italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT is the transformed x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT after zoom-in and cropping operations. It is practical in reinforcement learning training to utilize the advantage function to reduce variance [[13](#bib.bib13)] and we therefore define the reward as
| | | | |
| --- | --- | --- | --- |
| | 𝑅(a)=aJθ1(x1,y1)−Jθ0(x0,y0)Jθ0(x0,y0).𝑅𝑎𝑎subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0\textit{R}(a)=a\frac{J\_{\theta\_{1}}(x\_{1},y\_{1})-J\_{\theta\_{0}}(x\_{0},y\_{0})}{J\_{\theta\_{0}}(x\_{0},y\_{0})}.R ( italic\_a ) = italic\_a divide start\_ARG italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) - italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_ARG start\_ARG italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_ARG . | | (4) |
So when a=1𝑎1a=1italic\_a = 1, the reward is positive if Jθ1(x1,y1)>Jθ0(x0,y0)subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0J\_{\theta\_{1}}(x\_{1},y\_{1})>J\_{\theta\_{0}}(x\_{0},y\_{0})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) > italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ), and the reward is negative if Jθ1(x1,y1)<Jθ0(x0,y0)subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0J\_{\theta\_{1}}(x\_{1},y\_{1})<J\_{\theta\_{0}}(x\_{0},y\_{0})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ) < italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) . The denominator in Equation [4](#S2.E4 "4 ‣ 2 Reinforced Auto-Zoom Net ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images") functions as a normalizer to prevent reward explosion. To prevent p from saturating at the beginning, we adopt the bounded Bernoulli distribution
| | | | |
| --- | --- | --- | --- |
| | 𝑝~=α𝑝+(1−α)(1−𝑝).~𝑝𝛼𝑝1𝛼1𝑝\tilde{\textit{p}}=\alpha\textit{p}+(1-\alpha)(1-\textit{p}).over~ start\_ARG p end\_ARG = italic\_α p + ( 1 - italic\_α ) ( 1 - p ) . | | (5) |
We have 𝑝~∈[1−α,α]~𝑝1𝛼𝛼\tilde{\textit{p}}\in[1-\alpha,\alpha]over~ start\_ARG p end\_ARG ∈ [ 1 - italic\_α , italic\_α ]. The training objective is to maximize the expected reward or to minimize the negative expected reward
| | | | |
| --- | --- | --- | --- |
| | Jϕ1(x0)=−𝔼a∼πϕ1(x0)[𝑅(a)].subscript𝐽subscriptitalic-ϕ1subscript𝑥0subscript𝔼similar-to𝑎subscript𝜋subscriptitalic-ϕ1subscript𝑥0delimited-[]𝑅𝑎J\_{\phi\_{1}}(x\_{0})=-\mathbb{E}\_{a\sim\pi\_{\phi\_{1}}(x\_{0})}[\textit{R}(a)].italic\_J start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) = - blackboard\_E start\_POSTSUBSCRIPT italic\_a ∼ italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_POSTSUBSCRIPT [ R ( italic\_a ) ] . | | (6) |
The optimization of the policy network is implemented through policy gradient methods [[18](#bib.bib18), [14](#bib.bib14), [15](#bib.bib15)], where the expected gradients are
| | | | |
| --- | --- | --- | --- |
| | ∂∂ϕ1Jϕ1(x0)=−𝔼a∼πϕ1(x0)[𝑅(a)∂∂ϕ1log(a𝑝~+(1−a)(1−𝑝~))]subscriptitalic-ϕ1subscript𝐽subscriptitalic-ϕ1subscript𝑥0subscript𝔼similar-to𝑎subscript𝜋subscriptitalic-ϕ1subscript𝑥0delimited-[]𝑅𝑎subscriptitalic-ϕ1log𝑎~𝑝1𝑎1~𝑝\frac{\partial}{\partial\phi\_{1}}J\_{\phi\_{1}}(x\_{0})=-\mathbb{E}\_{a\sim\pi\_{\phi\_{1}}(x\_{0})}[\textit{R}(a)\frac{\partial}{\partial\phi\_{1}}\text{log}(a\tilde{\textit{p}}+(1-a)(1-\tilde{\textit{p}}))]divide start\_ARG ∂ end\_ARG start\_ARG ∂ italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_ARG italic\_J start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) = - blackboard\_E start\_POSTSUBSCRIPT italic\_a ∼ italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) end\_POSTSUBSCRIPT [ R ( italic\_a ) divide start\_ARG ∂ end\_ARG start\_ARG ∂ italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_ARG log ( italic\_a over~ start\_ARG p end\_ARG + ( 1 - italic\_a ) ( 1 - over~ start\_ARG p end\_ARG ) ) ] | | (7) |
We adopt an alternating training strategy to update both networks. The training procedure of RAZN is illustrated in Algorithm [1](#alg1 "Algorithm 1 ‣ 2 Reinforced Auto-Zoom Net ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images").
Algorithm 1 Training of RAZN when m=1𝑚1m=1italic\_m = 1
1:x0subscript𝑥0x\_{0}italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT
2:Get Jθ0(x0,y0)subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0J\_{\theta\_{0}}(x\_{0},y\_{0})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) and Jθ1(x1,y1)subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1J\_{\theta\_{1}}(x\_{1},y\_{1})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT )
3:Sample action a𝑎aitalic\_a through πϕ1(x0)subscript𝜋subscriptitalic-ϕ1subscript𝑥0\pi\_{\phi\_{1}}(x\_{0})italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT )
4:Get 𝑅(a)(x0)𝑅𝑎subscript𝑥0\textit{R}(a)(x\_{0})R ( italic\_a ) ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT )
5:Update ϕ1subscriptitalic-ϕ1\phi\_{1}italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT by minimizing Jϕ1(x0)subscript𝐽subscriptitalic-ϕ1subscript𝑥0J\_{\phi\_{1}}(x\_{0})italic\_J start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT )
6:if a=1𝑎1a=1italic\_a = 1 then
7: Update θ1subscript𝜃1\theta\_{1}italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT by minimizing Jθ1(x1,y1)subscript𝐽subscript𝜃1subscript𝑥1subscript𝑦1J\_{\theta\_{1}}(x\_{1},y\_{1})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT )
8:else
9: Update θ0subscript𝜃0\theta\_{0}italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT by minimizing Jθ0(x0,y0)subscript𝐽subscript𝜃0subscript𝑥0subscript𝑦0J\_{\theta\_{0}}(x\_{0},y\_{0})italic\_J start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT )
10:end if
3 Experiments
--------------
### Dataset
The dataset used in this study is provided by Grand Challenge on Breast Cancer Histology Images 111https://iciar2018-challenge.grand-challenge.org/dataset. The dataset contains 10 high-resolution WSIs with various image size. WSIs are scanned with Leica SCN400 at ×40absent40\times 40× 40 magnification.
The annotation was performed by two medical experts. As annotation of WSIs requires a large amount of human labor and medical domain knowledge, only sparse region-level labels are provided and annotations contain pixel-level errors. In this dataset, the white background (glass slide) is labeled as normal by the annotators. The dataset is unbalanced for the four cancer types.
### Implementation
Experiments are conducted on a single NVIDIA GTX Titan X GPU. In this study, m=1𝑚1m=1italic\_m = 1, r=2𝑟2r=2italic\_r = 2 and α=0.8𝛼0.8\alpha=0.8italic\_α = 0.8. The backbone of fθisubscript𝑓subscript𝜃𝑖f\_{\theta\_{i}}italic\_f start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT is ResNet18 [[10](#bib.bib10)], with no downsampling performed in conv3\_1 and conv4\_1. gϕ1subscript𝑔subscriptitalic-ϕ1g\_{\phi\_{1}}italic\_g start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT is also based on the ResNet18 architecture. However, each block (consisting of 2 residual blocks [[10](#bib.bib10)]) is replaced by a 3×3333\times 33 × 3 convolution followed by batch normalization and ReLU non-linearity.
The computational cost for the policy network is 7.1%percent7.17.1\%7.1 % of the segmentation networks. The input size to the segmentation networks and the policy network is fixed to 256×256256256256\times 256256 × 256. We use the Adam optimizer [[11](#bib.bib11)] for both the policy network and segmentation networks and use a step-wise learning rate policy with decay rate 0.1 every 50000 iterations. The initial learning rate is 0.01.
### Multi-scale
Given a 256×256256256256\times 256256 × 256 patch, we consider two resolutions in order to simulate the zoom-in process. A coarse resolution (Scale 1), where the patch is downsampled to 64×64646464\times 6464 × 64 and a fine resolution patch (Scale 2), where the patch is downsampled to 128×128128128128\times 128128 × 128. The patches are then resized back to 256×256256256256\times 256256 × 256 using bilinear interpolation. To evaluate the efficiency of the proposed framework, we compare our model with two multi-scale models. The first multi-scale model is the segmentation network fθsubscript𝑓𝜃f\_{\theta}italic\_f start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT with multi-scale training [[5](#bib.bib5)], denoted as MS. We only consider two scales in this experiment (Scale 1 and Scale 2). Similarly, another multi-scale model is the multi-scale fusion with attention [[6](#bib.bib6)], which is denoted as Attention. The training details of all models are the same. All models are trained with 200000 batches.
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | non-carcinoma | carcinoma | mIOU | Weighted IOU | Relative Inference Time |
|
Scale 1 | 0.45 | 0.32 | 0.38 | 0.07 | 1.00 |
| Scale 2 | 0.46 | 0.31 | 0.39 | 0.07 | 4.01 |
| MS [[5](#bib.bib5)] | 0.32 | 0.04 | 0.18 | 0.01 | 5.06 |
| Attention [[6](#bib.bib6)] | 0.43 | 0.29 | 0.36 | 0.06 | 5.16 |
| RAZN | 0.49 | 0.49 | 0.49 | 0.11 | 2.71 ±plus-or-minus\pm± 0.57 |
Table 1: Comparison of the performance. Non-carcinoma includes normal and beign. Carcinoma includes in situ carcinoma and invasive carcinoma.
### Performance
We compare two key indicators of the performance, which are the segmentation performance and the inference speed. We use intersection over union (IOU) as the metric for segmentation performance. We report mean IOU, which is just the average IOU among four classes. Due to the imbalanced data, we also report weighted IOU, where the weight is proportional to the inverse of the frequency of the labels of each class.
Further, we report relative inference time for the proposed RAZN and the baseline methods compared to the inference time for the model that only considers Scale 1. We report the average relative inference time over 100 patches. Lower values of relative inference time represent faster inference speed. The results are presented in Table [1](#S3.T1 "Table 1 ‣ Multi-scale ‣ 3 Experiments ‣ Reinforced Auto-Zoom Net: Towards Accurate and Fast Breast Cancer Segmentation in Whole-slide Images"). Note, we report the mean and the standard deviation for RAZN, as the inference time will vary depending on whether zooming is required for a given patch or not.
It can be shown that RAZN actually performs better than the single scale and the multi-scale baselines.
MS’s performance is the worst of our benchmarks. MS exaggerates the imbalance problem by augmenting the data, which can confuse the network. We also hypothesize that the cell size is not the critical factor that influences the segmentation of cancer and that MS, therefore, aims to model unnecessary information on this task.
Similarly, attention models memorize the scale of the object by fusing the results from different scales.
However, when the object is not well-defined at certain scales, like in our task the cancer (group of dense cells), the network may learn to fit noise.
Our results illustrate that RAZN instead is more robust when data is noisy and imbalanced, providing an overall accuracy improvement at low inference time.
4 Discussion and Conclusions
-----------------------------
We proposed RAZN, a novel deep learning framework for breast cancer segmentation in WSI, that uses reinforcement learning to selectively zoom in on regions of interest. The results show that the proposed model can achieve improved performance, while at the same time reduce inference speed compared to previous multi-scale approaches. We also discuss the use of multi-scale approaches for breast cancer segmentation. We conclude that cancer cells are different from general objects due to their relative small and fixed size. Multi-scale approaches may not work for a noisy and imbalanced data.
In future work, we aim to extend the model to study the multiple zoom-in actions situation (m>1𝑚1m>1italic\_m > 1) and will investigate the potential of more complex segmentation backbone models to improve overall performance.
Acknowledgements. We thank ICIAR 2018 Grand Challenge on Breast Cancer Histology Images for providing the data for this study. |
5d5ee01c-9dbd-44c7-8f08-9c17946df4b5 | trentmkelly/LessWrong-43k | LessWrong | Have we really forsaken natural selection?
Natural selection is often charged with having goals for humanity, and humanity is often charged with falling down on them. The big accusation, I think, is of sub-maximal procreation. If we cared at all about the genetic proliferation that natural selection wanted for us, then this time of riches would be a time of fifty-child families, not one of coddled dogs and state-of-the-art sitting rooms.
But (the story goes) our failure is excusable, because instead of a deep-seated loyalty to genetic fitness, natural selection merely fitted humans out with a system of suggestive urges: hungers, fears, loves, lusts. Which all worked well together to bring about children in the prehistoric years of our forebears, but no more. In part because all sorts of things are different, and in part because we specifically made things different in that way on purpose: bringing about children gets in the way of the further satisfaction of those urges, so we avoid it (the story goes).
This is generally floated as an illustrative warning about artificial intelligence. The moral is that if you make a system by first making multitudinous random systems and then systematically destroying all the ones that don’t do the thing you want, then the system you are left with might only do what you want while current circumstances persist, rather than being endowed with a consistent desire for the thing you actually had in mind.
Observing acquaintences dispute this point recently, it struck me that humans are actually weirdly aligned with natural selection, more than I could easily account for.
Natural selection, in its broadest, truest, (most idiolectic?) sense, doesn’t care about genes. Genes are a nice substrate on which natural selection famously makes particularly pretty patterns by driving a sensical evolution of lifeforms through interesting intricacies. But natural selection’s real love is existence. Natural selection just favors things that tend to exist. Things that start existing: great. |
ed5dfa6b-6391-4aa3-a77e-31cbd8494be5 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Sydney Rationality Dojo - August 2016
Discussion article for the meetup : Sydney Rationality Dojo - August 2016
WHEN: 07 August 2016 04:00:00PM (+1000)
WHERE: 10 Shepherd Street, Chippendale
From 4 - 6 we'll be holding our normal rationality dojo at Humanist house. Afterwards there'll be an informal dinner gathering for those who want to join.
Discussion article for the meetup : Sydney Rationality Dojo - August 2016 |
0a2f2319-9024-4b75-9fcc-0f1619a2919a | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | The great energy descent - Part 2: Limits to growth and why we probably won’t reach the stars
**TL;DR**: In summary, if fossil fuels decline, economic growth will likely decline as well. Investment can help but will be limited. Consequences of a decline might mean a decline in food production and a disruption of supply chains. This also makes space colonization and the overall continuation of our growth-based society unlikely. You can check the [short version of the post](https://forum.effectivealtruism.org/posts/wXzc75txE5hbHqYug/the-great-energy-descent-short-version-an-important-thing-ea).
*Note : This is part 2 of 3 posts on energy depletion as an important topic for EA. I recommend reading* [*part 1*](https://forum.effectivealtruism.org/posts/qG8k5pzhaDk6FhcYv/the-great-energy-descent-part-1-can-renewables-replace) *before this one, or the short version. This part will address whether investment and innovation can prevent energy depletion. It will also address the consequences of an energy descent on economic growth, the financial system, society as whole and, finally, what this implies for EA causes.*
Part 1 indicated that fossil fuels, being finite, were close to the start of a decline, within a few decades. It also indicated why an energy transition to renewables contained so many challenges (time, materials, rehauling of the manufacturing system) that it likely induced a decline in energy availability. A good summary is the conclusion of the [GTK report](https://www.gtk.fi/en/time-to-wake-up/global-challenge/) (the one that assessed the feasibility of switching away from fossil fuels):
> “**Current thinking is that European industrial businesses will replace a complex industrial ecosystem that took more than a century to build**. This system was built with the support of the highest calorically dense source of energy the world has ever known (oil), in cheap abundant quantities, with easily available credit, and unlimited mineral resources. This task is hoped to be done at a time when there is comparatively very expensive energy, a fragile finance system saturated in debt, not enough minerals, and an unprecedented number of human populations, embedded in a deteriorating environment. Most challenging of all, this has to be done within a few decades. It is the author’s opinion, based on the new calculations presented here, that this will likely not go fully as planned”.
>
>
With prices rising, won’t more investment solve the issue?
==========================================================
**There are many answers commonly given to the prospects of an energy depletion**. Let’s look at the current view of EA on the subject, which also reflects that of economists. I managed to find only [one post](https://forum.effectivealtruism.org/posts/2vmiMQjNAH7rQcmz5/why-don-t-many-effective-altruists-work-on-natural-resource) from 2016 on the EA forum about that (plus some sparse [references](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.ci76r4in6l7)), but I suspect it reflects a common point of view. The main argument was that when prices rise, people and infrastructure are redirected towards getting more of a resource, and people look for alternative ways of doing things that require less of a specific resource (for instance, solar panels are getting cheaper, and we are finding ways to access fossil fuels that were previously unavailable). **There is an incentive for technology to improve, and for substitution to occur**. If one resource is lacking, we’ll just switch to another: for instance, rare earth neodymium in offshore wind turbines can be substituted by using wound-rotor generators. A switch toward alternatives is done, and human ingenuity and technology will find a way.
**Note that all of this worked in the past, to some extent** - and this nullified the predictions of many people that anticipated depletion too early. For instance, following the 1970s oil shocks, many diesel electricity plants were replaced by coal and gas plants. Additionally, scarcer resources led to higher prices, which led to more investment (like in Alaska and the North Sea), which led to more oil being produced in the end.
However, just because there was a switch to other resources doesn’t mean that this was of no consequence. Indeed, **since the oil shocks, there has been a regime of lower economic growth and more debt**. We’ve never experienced a growth rate like in the 50s and 60s, when the growth rate of oil was sky-high. See the detailed section on the [causes and consequences of the 1970s oil shocks](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.eyba16upecbc).
An important element was that **most of the alternatives were comparatively worse than conventional oil** (which was dirt-cheap and abundant), or at least with a more constrained growth potential. This is an important limit of substitution. As shown in the book [*Extracted*](https://www.resilience.org/stories/2014-04-18/the-age-of-diminishing-returns/), “very often, it involves substituting a resource that was once relatively clean and cheap —as long as it was available— with a more expensive and dirtier one”. For instance, shale oil eased oil depletion in the short term, but it won’t last forever. Once it’s over, “we are back to square one, having just squandered a lot of resources and created a lot of pollution”. **Most examples of substitutions involve higher cost** (otherwise, it would be done already), **and, most importantly, lower performance** - so we need to use more energy to do the same thing. See the “[Substitution](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.b1d0jsnccq9)” section for more details.
We will switch to other sources, yes, but this will still mean a decline in energy, at least if we keep the same level of investment. **A common answer is that we should invest more**, in order to compensate for the lower performance. However, **there are several reasons that prevent prices for going too high, putting a limit on investment**.
Limits on investment
--------------------
First, **the energy industry won’t extract resources so expensive they’re impossible to sell**. Consequently, there’s a limit to the low-grade resources the industry can exploit. This could help to explain the weird trends in oil prices:
[*Source*](https://tupa.gtk.fi/raportti/arkisto/70_2019.pdf)*, page 257. Oil prices are Europe Brent Spot Price FOB ($USD Dollars per Barrel), without inflation adjustment. See* [*here*](https://inflationdata.com/articles/inflation-adjusted-prices/historical-oil-prices-chart/) *with inflation. Data from the* [*EIA*](https://www.eia.gov/dnav/pet/hist/RBRTED.htm)*See also the section on the* [*2008 financial crisis*](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.pyvnp5peukjm)*.*
The most important limit for energy extraction, then, is not "is it theoretically possible to extract?" but "**will people be able to afford it?**". For the same reason, prohibitive price rises for clean energy technologies (or the metals they need)[**could stall the energy transition**](https://www.eurometaux.eu/media/rqocjybv/metals-for-clean-energy-final.pdf).
There are others trends taking place here:
* Then again, **scale** is a very important variable here. **The transition could require as much as $173 trillion** in energy supply and infrastructure investment over the next three decades, according to research provider [BloombergNEF](https://www.bloomberg.com/graphics/2021-materials-silver-to-lithium-worth-big-money-in-clean-energy/). As a comparison, [world GDP](https://www.statista.com/statistics/268750/global-gross-domestic-product-gdp/#:~:text=In%202020%2C%20global%20GDP%20amounted,trillion%20lower%20than%20in%202019.) amounted to $87 trillion in 2019.
+ Another example: the *Energiewende*, Germany's vast buildout of solar and wind energy, [has cost around $400 billion so far](https://reason.com/2022/05/25/civilization-runs-on-ammonia-plastic-steel-and-cement-for-now/), yet the share of fossil fuels in the country's primary energy supply has fallen just slightly, from 84 to 78%.
* **Short-termist thinking could prevail, and deciders and companies may shy away from long-term investment in the context of a crisis.** Currently, [US oil producers are reluctant to drill more oil](https://www.cbsnews.com/news/oil-production-prices-us-companies-wont-increase-2022-dallas-fed-survey/), despite sky-high prices, as they have been thrown off by the volatility. For the industry, building a billion-dollar plant that takes 30 years to pay off requires trust in the future.
* **There is** ***already*** **a lack of investment** right now, especially in the road and electricity sector. Globally, there is a [$18 trillion gap](http://blogs.worldbank.org/ppps/forecasting-infrastructure-investment-needs-50-countries-7-sectors-through-2040) to be filled, and infrastructure is decaying in many places.
* **Technology and innovation can help, but also face diminishing returns** - new discoveries are costlier and costlier. Even with higher prices and better technology, most countries are still seeing declining oil production. High-tech advances also usually require scarcer materials, for instance in alloys.
* **If too many resources (money, energy, materials, people) are invested in energy production, then you have less resources that you can use to run the rest of society** (for food, goods, buildings and so forth). This was a core finding of the *Limits to Growth* report: that at some point, so many resources were needed to compensate for resource depletion and pollution that it limited reinvestment in maintenance and capital, and in the end on growth. [Some papers](https://www.sciencedirect.com/science/article/pii/S2211467X19300926?via%3Dihub) point out that it’s possible that the net energy returns of renewables could be too low to allow for a complex civilization like ours (although there’s quite a lively debate on that).
This above is just a summary of the complete [“Won’t more investment solve the issue if prices go up?” section](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.b3o8lksb3kim) in the additional doc. I recommend checking it out as **this is one of the most underrated limits of the energy transition**. In practice, it prevents prices and investment from growing too much. The full version also explains why what matters for societies is a high **energy surplus**, and why energy sources that don’t provide a lot of surplus (low EROI), like many renewables, cannot provide the same services as those we enjoy today.
However, it is true that **most people really will try their best to adapt when faced by negative consequences**. Indeed, there are quite a lot of wasteful things that we can do less of:
* Make ships go at half the speed, maybe with [kites or sails](https://allfed.info/images/pdfs/needs%20other%20than%20food%20if%20industry%20is%20disabled%20final.pdf)
* Make lower speed limits on roads, and boost public transport
* Make planes fly slower to save on fuel
* Lower the mean temperature in houses and buildings
* Buy less single-use products and repair items, stop planned obsolescence
* Waste less food and eat less meat
However, while all these things are pretty reasonable and do not take a high toll on human well-being, **most of them would be bad for the economy**.
Economic growth is likely to stop and decline
=============================================
Now, let’s look at the relationship between energy and GDP:
*Source: this* [*Our Finite World*](https://ourfiniteworld.com/2014/08/14/energy-and-the-economy-twelve-basic-principles/) *article (see point 10)*
It sounds like we are onto something here. And indeed, [Giraud and Kahraman, 2014](https://halshs.archives-ouvertes.fr/halshs-01151590/)come to the conclusion that **there is a strong link between energy consumption to economic growth in both the short and long-run.** [*Note: Commenter Robin pointed out that this is not a peer-reviewed paper, so be careful here. However,* [Fizaine and Court 2016](https://www.sciencedirect.com/science/article/abs/pii/S0301421516302087) *come to the same conclusion. There is some debate on whether there is a causality, but several other papers listed* [*here*](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.f6cwl9ryzkct) *also show that there is a strong link between the two*].None of the other factors they reviewed, *including capital and labor*, showed the same dependency ratio. Energy efficiency also plays a strong role here, but it has slowed down in recent years. This makes sense, as energy allows for everything we do, especially for the transport and manufacturing of almost every product. To simplify, one way to put it would be to see money as a “claim on a product made with energy“.
This is [not the only paper](https://www.sciencedirect.com/science/article/abs/pii/S0301421513004783) with this conclusion, there’s more literature than I had thought on that. You can check the “[Is it possible to grow the economy with less energy?](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.f6cwl9ryzkct)” section, that goes into more detail about what they say on the topic (for instance, why the small share of energy in GDP is not representative, and why apparent decoupling in a few rich countries does not translate to the global level).
Companies will start to look for alternatives if prices increase considerably, but **there is a common misconception that prices will rise gradually and continually, giving time to adapt and invest in alternatives**. If we look at the actual relationship between oil production and oil price (see the graph above), it seems that the evolution of oil price is everything but linear: it behaves with quick and brutal changes, oscillating between too high for consumers and too low for producers. If we focus on oil prices right now, we can’t know what their value will be in one year — and whether we should prepare for a sudden increase. **So I advise we don’t focus too much on prices, more on quantities.**
While it may be theoretically possible to have an economy that grows with much less energy, there is little real world data to indicate this is possible. A sudden shift in the way the economy works may not be impossible after a few years of recession, but we should not assume this will be the most probable outcome. When it comes to economics, very few relationships have been as tight as this one (even during crises), so the default scenario we should expect would be that this will continue in the future.
This is why **the consequences of a sudden rise in energy prices are, most of the time, a recession**: a spike in oil prices has preceded or coincided with [10 out of 12 post-WWII US recessions](https://www.investopedia.com/articles/economics/08/past-recessions.asp), and played a significant role [during the 2008 crisis](https://www.researchgate.net/publication/227370279_Causes_and_Consequences_of_the_Oil_Shock_of_2007-08) (which might be the third oil shock - see the [section](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.pyvnp5peukjm) dedicated to it). A useful rule of thumb is that in the last 50 years, when energy prices reached about 8-10% of GDP, a recession ensued:
Right now, during the [current energy crisis](https://www.bloomberg.com/news/articles/2021-10-07/china-s-energy-crisis-envelops-an-already-slowing-global-economy), we are seeing a rise in prices for industrial metals (zinc, copper, aluminum), cars, tech products (including solar panels) and, most importantly, food. We can expect to see the amount of production and services plummeting at some point, as almost everything will be harder to do with less energy. Tellingly, a [paper](https://medium.com/insurge-intelligence/scientists-warn-the-un-of-capitalisms-imminent-demise-a679facac985) commissioned by the UN Secretary-General’s group of scientists pointed out that “***the driving force of the transition to postcapitalism*** [whatever this is] ***is the decline of what made ‘endless growth capitalism’ possible in the first place: abundant, cheap energy*****.**”
### Modelisation of the energy and GDP decline
Now, has someone tried to quantify all of that? I’m aware of very few papers that tried to model all these things. Most IAMs, the models used by most energy transition models (including those of the IPCC), do not model limits on fossil fuels or minerals, or declining energy returns, assume perfect substitutability, model poorly environmental damage (see [here](https://www.youtube.com/watch?v=vwwvZ8g5eHE)), use prices as indicators of scarcity, etc. See the beginning of [this article](https://pubs.rsc.org/en/content/articlehtml/2020/ee/c9ee02627)). So these IAMs rarely take into account most of what is discussed above. **The only model that I am aware of that tried to model all that is** [**MEDEAS**](https://pubs.rsc.org/en/content/articlehtml/2020/ee/c9ee02627d). I recommend checking it out, it’s quite interesting (although I must note that Dave finds it pessimistic, he thinks they give too much importance to land use and climate impacts, and that the model should have higher efficiency and growth of renewables). This model includes:
* Representation of biophysical constraints to energy availability (with EROI)
* Modeling of the material and energy investments for the energy transition
* Consistent representation of climate change damages with assessments by natural scientists
* Integrated representation of economic processes and biophysical limits to growth
* Energy shifts driven by physical scarcity
* Impact on water, land use, and social impact
*(A) Total final energy per capita (GJ per person per year). (B) Gross Domestic Product per capita (GDPpc, thousand $ per person, in USD chained linked volumes (1995)**Ref is the base model, in bold, where both energy depletion and climate impacts weigh in. The dotted lines represent scenarios with unlimited energy resources (Ref\_noCC) or with no climate impacts (Ref\_noER). The line going up is for a world of unlimited energy resources and with no climate impacts (Ref\_noER\_noCC) - this is what is used in mosts energy transition models - IAMs - and the IPCC.*
Except for the scenario used by IAMs, all scenarios show a plateau and a decline. Of course, such a strong change would have wide consequences on the economic and political system, and no model can integrate what these “‘unthought futures” would imply.
Impacts: What might happen when energy declines?
================================================
*Note: from now on, this article will get really speculative, as nobody really knows what will happen. There is also much less litterature on the topic.* As the article “[Energy descent as a post-carbon transition scenario](https://www.sciencedirect.com/science/article/abs/pii/S0016328720300550)”puts it, *“*The nature of the envisaged transition means that **we are entering entirely unexplored territory**, and the pathways that we walk into existence are subject to inherent, irreducible uncertainty. It is impossible to know up front just how these pathways will unfold, the full range of challenges that will be encountered along the way, and where the novel responses to them will take us. As such, there is very good reason to think that the situations that emerge will be very different from the expectations created by any model constructed or plan conceived today. It seems prudent to conclude that global-scale transition away from fossil fuels leads humanity into the post-normal realm of ‘unthought future(s)’ (Sardar & Sweeney, 2016). **Here actors will do better to anticipate complex, uncertain and chaotic conditions as typical, rather than extreme outliers**.*“*
This is still a work in progress and bound to change, but here are some elements I found that could happen. Note that what follows assumes the case here an energy descent is undergone involuntarily, without any serious anticipation of the topic from companies and government, like rationing or degrowth (we’ll discuss that more in the next post).
Systemic risks - Danger of economic collapse
--------------------------------------------
*A word of warning: what is following might be rather uncomfortable to read, so if you want to take a pause before going ahead, please feel free to do so. Remember that this is an exercise that tries to anticipate realistic bad scenarios - this is not necessarily the most likely one, but it has a significant chance of happening.*
What might happen then? Economic and political systems are geared for growth, which is why they respond poorly to recessions. The order of events that will unfold is unclear, and so is the timing. However, it might be useful to see what armies honestly think will happen. [An internal report on the subject](https://energyskeptic.com/2017/german-military-peak-oil-summary/) by the **German Armed Forces**, the Bundeswehr, was leaked a few years ago. This is a prospective work aimed at anticipating worst-case scenarios, so keep in mind that what follows is not necessarily the inevitable development:
> [*Start of the quote from the report*] "The transmission channels of an oil price shock involve diverse and interdependent economic structures and infrastructures, some of which are of vital importance. **Its consequences are therefore not entirely predictable**. Initially, it will be possible to measure the extent of these consequences, although not exclusively, by a reduced growth of the global economy. [...] Economies, however, move within a narrow band of relative stability. Within this band, economic fluctuations and other shocks are possible, but the functional principles remain unchanged and provide for new equilibriums within the system. Outside this band, however, **this system responds chaotically**. [...] An economic tipping point exists where, for example as a result of peak oil, **the global economy shrinks for an undetermined period**. In this case a chain reaction would destabilize the global economic system [...]. The course of this potential scenario could be as follows:
>
>
> 1. Peak oil would occur and it would not be possible, at least in the foreseeable future, to entirely compensate for the decline in the production of conventional oil with unconventional oil or other energy and raw material sources. The expression “foreseeable” is very important in this context. Ultimately, it leads to a loss of confidence in markets.
>
> In the short term, **the global economy would respond proportionally to the decline in oil supply.**
>
> 1. Increasing oil prices would reduce consumption and economic output. This would lead to recessions.
> 2. The increase in transportation costs would cause the prices of all traded goods to rise. For some actors, this would only mean losing sources of income, whereas others would no longer be able to afford essential food products.
> 3. National budgets would be under extreme pressure. Expenditure for securing food supplies (increasing food import costs) or social spending (increasing unemployment rate) would compete with the necessary investments in oil substitutes and green tech. Revenues would decrease considerably as a result of recession [...].
>
> In the medium term, **the global economic system and all market-oriented economies would collapse.**
>
> 1. Economic entities would realize the prolonged contraction and would have to act on the assumption that the global economy would continue to shrink for a long time.
> 2. **Tipping point:** In an economy shrinking over an indefinite period, savings would not be invested because [*most*] companies would not be making any profit. For an indefinite period, **companies would no longer be in a position to pay borrowing costs or to distribute profits to investors**. The banking system, stock exchanges and financial markets could collapse altogether.
> 3. Financial markets are the backbone of the global economy and an integral component of modern societies. All other subsystems have developed hand in hand with the economic system. A disintegration can therefore not be analyzed based on today’s system. **A completely new system state would materialize**. [...]
>
> Here is an outline of some theoretically plausible consequences:
>
> * **Loss of confidence in currencies**. Belief in the value-preserving function of money would dwindle. This would initially result in hyperinflation and black markets, followed by a barter economy at the local level.
> * **Collapse of unpegged currency systems**. If currencies lose their value in their country of origin, they can no longer be exchanged for foreign currencies. **International value-added chains would collapse as well.**
> * **Mass unemployment**. [...] During the oil crisis in the mid-1970s, unemployment in West Germany increased fourfold although all actors were aware that this crisis was finite. The latter would not be the case in a crisis induced by peak oil.[...]
> * **National bankruptcies**. In the situation described, state revenues would evaporate. (New) debt options would be very limited, and the next step would be national bankruptcies.
> * **Collapse of critical infrastructures**. Neither material nor financial resources would suffice to maintain existing infrastructures. Infrastructure interdependencies, both internal and external with regard to other subsystems, would worsen the situation.
> * **Famines**. Ultimately, production and distribution of food in sufficient quantities would become challenging.
> * [...] In order to prevent a restriction of capabilities and deployment options of the Bundeswehr, **alternative solutions to oil-based fuels would be necessary** in the short term. While these solutions, such as coal liquefaction or in some cases natural gas liquefaction, are possible and conceivable in principle [*and some already exist*], they would entail considerable political and economic efforts. They would require considerable investments and radical industrial policy decisions. Considering the challenges society as a whole would face as a result of peak oil, **it seems unlikely that this could be accomplished even in case of an emergency**. [...]
>
> In view of their degree of globalization, **all industrialized countries** – including Germany – **face a high systemic risk**, regardless of their individual energy policies.
>
> [...] Psychological barriers cause indisputable facts to be blanked out and lead to **almost instinctively refusal to look into this difficult subject in detail. Peak oil, however, is unavoidable.** This study shows the existence of a very serious risk that a global transformation of economic and social structures, triggered by a long-term shortage of important raw materials, will not take place without friction regarding security policy.”
>
>
[*End of the quote*] These conclusions should explain why governments and international institutions [do not like to talk openly about this subject](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.1we7gsp293ie). The fact that they did not plan for such an eventuality could also bring about a loss of confidence in institutions: “People will experience a lowering of living standards due to an increase in unemployment and the cost of oil for their vehicles. [...] **Setbacks in economic growth can lead to an increase in the number of votes for extremist and nationalistic parties.**” Authoritarian regimes that promise an easy way out and put the blame on scapegoats could also rise to power.
Consequences - Risk of supply chain disruption
----------------------------------------------
*Note: what follows is even more uncertain, as very, very few papers got into real detail on this topic. I am mostly putting together what I managed to find, so please take into account that there are many, many unknowns here.*
[A report](http://www.feasta.org/documents/risk_resilience/Tipping_Point.pdf) by the Risk/Resilience network went into more detail about systemic risks, and also pointed out that this could disrupt our ability to produce more energy. Indeed, **a systemic crisis could disrupt long supply chains** as they rely heavily on monetary confidence and bank intermediation. This could affect the ability to produce and transport food, but also parts for the maintenance of energy infrastructure (photovoltaic panels and windmills require many parts and metals shipped over the world). No one nation state has everything it needs to manufacture a single advanced technological unit like a computer (except maybe China), so a contraction of international trade could mean shortages of all kinds. Loss of long-term confidence in purchaser solvency and monetary stability could also prevent the development of many costly and complex energy projects that need a huge upfront investment. The list includes digging Arctic oil or Siberian natural gas, building smelting plants that can work with hydrogen, nuclear power or fusion… **All of this would accelerate the energy descent**.
**It’s possible that the global economy could reorganize itself in some way to prevent such paralysis**, and there would be attempts to put in place economic and political systems that would rely less on material growth to prevent worst-case scenarios. I personally think it’s a real option, but **it’s really unclear how this would take place**. Nations could go into “war-economy” mode, stepping in if the private sector cannot invest enough in energy sources, possibly nationalizing the sector. On the contrary, oil companies could acquire influence similar to that of countries. It is possible that a collapse of the economic system is halted by a quick reaction from governments, and affects only limited sections (like the financial sphere). Even then, the return to a business as usual economy is not very likely since the issue of declining energy returns would still be a thing.
What will nations do? Many countries will try to keep their hands on the resources necessary for their economy — richest nations could buy them at a high price, like with masks and vaccines today, leaving little for other countries that cannot compete. Energy diplomacy could get really intense. Some oil producers could decide to keep their resources for domestic uses (for instance, the EU would be in trouble if the USA or Russia were to do that). Countries with a lot of energetic resources could face wars over these resources, especially in the Middle East, as [happened over history](https://www.theguardian.com/world/2007/sep/16/iraq.iraqtimeline). That would add to the political instability and importing these resources would get even harder for other nations. Such wars, if they end up blocking some chokepoints used by oil tankers like the Suez canal, could disrupt global trade even more.
**The consequences will vary depending on the country**. For instance, China is able to plan long term with a population that can accept sacrifices, with local resources and industry, and may already be preparing. Russia, with its large energy resources, could also do well. More surprisingly, some populations that are already “energy poor” could fare much better than rich countries in the long term because they still have the critical social and technical skills to handle this, as the gap is smaller (depending of course on the country and political situation). Current trends of urbanization and industrialization may even be counterproductive in this sense for them, because it could make them lose these skills. On the other hand, rich countries, which are more used to their energy-intensive lifestyles, could try to cling to their way of living and might suffer a stronger fall in the long term.
Several things would be less available for most. High-tech hospitals need a lot of energy to run: the [US healthcare system](https://solar.lowtechmagazine.com/2021/02/how-sustainable-is-high-tech-health-care.html) consumes as much energy as the country’s entire renewable energy production. Many other things could become much harder in the future: moving quickly over long distances, transporting water by truck to water-deprived zones, making use of heating in cold countries and HVAC systems in hot countries…
Impacts on food production
--------------------------
As the Risk/Resilience network underlines: “Global food production is already straining against a rising demand and the stresses of soil degradation, water constraints, overfishing, and the burgeoning effects of climate change. It is estimated that **between seven and ten calories of fossil fuel energy go into every one calorie of food energy we consume**. [*for the highly-mechanized US, less for other countries*] It should be clear even from the above overview that a major financial collapse could not just cut actual food production, but could result in food left rotting in the fields, an inability to link surplus production with those in need, and an inability to enact monetized food transactions. [...] Our critical reliance upon complex just-in-time supply-chain networks mean that there is little buffer to protect us from supply shocks. In the event of a global shock, and without any planning, it is likely that unrelieved hunger could spread rapidly. Even for a country that could be food independent, and even a potential net exporter, it may take years to transition as old systems fail and new ones put in place (rationing systems, education, re-location of farm laborers, horse breeding, nutrient recycling systems, seasonal re-adjustment of production, tool production, storage and preservation skills and products). In the interim, the risks are severe”.
For instance, Synthetic fertilizers feed about [half the global population](https://ourworldindata.org/how-many-people-does-synthetic-fertilizer-feed) today (4 billion people), and since making them is done with natural gas 95% of the time, it’s uncertain how we can feed so many people with less of them. The current rise in natural gas prices meant an increase in [fertilizer price](https://www.bloomberg.com/news/newsletters/2021-10-22/supply-chain-latest-rice-fertilizer-rally), which translated into more expensive food and more hungry people. ALLFED points out that it’s possible to use solar panels to produce that (more on the next post), and we should push for it, as it’s not known enough so far.
How might societies react to this?
----------------------------------
**How societies will react matters immensely**, and this is very hard to predict. Lessons can be learned from either Cuba or North Korea, who both faced massive reduction in oil imports after the collapse of the USSR. What happened to [North Korea](https://en.wikipedia.org/wiki/North_Korean_famine) and its harsh climate is what we want to avoid: “Without Soviet aid, the flow of imports to the North Korean agricultural sector ended, and **the government proved too inflexible to respond**. Energy imports fell by 75%. The economy went into a downward spiral, with imports and exports falling in tandem. Flooded coal mines required electricity to operate pumps, and the shortage of coal worsened the shortage of electricity. Agriculture reliant on electrically-powered irrigation systems, artificial fertilizers and pesticides was hit particularly hard by the economic collapse." The following famine killed hundreds of thousands.
Cuba, however, [managed to survive](https://s3.amazonaws.com/oxfam-us/www/static/oa4/OA-CubaGoingAgainstGrain_ExecSummary.pdf) by promoting a decentralized and ecology-based agriculture, and they made a remarkable recovery. If you feel all of this was quite gloomy, **I highly recommend you watch ‘**[**The Power of Community**](https://read.realityblind.world/view/975731937/)**’, which is actually very inspiring.** There was a lot of effort both from the people and the state. Social services, healthcare and education were maintained, land was reclaimed from conventional agriculture, cooperatives kicked in, carpooling and hitchhiking were common, decision making was localized with less state regulation. Of course, it’s important to recognize this was a difficult and unprepared transition — people were hungry — and not enough - [they still depend at 60-80% on food imports](https://energyskeptic.com/2018/how-cuba-survived-peak-oil/). Things are not perfect and hard to reproduce elsewhere. They also have a good climate, which helps immensely. But if overall, things unfold like in Cuba and not North Korea, I’d be happy. I especially like this quote from a Cuban: "*The people cooperating with and caring about each other are the main factors that we need to encourage. We can all plant fruit trees, we can all have water catchment devices on our roofs. **It's not the technology, it's the human relationships*****.**"
A good question to ask could be “What can we learn of Cuba? And what could have been done so that the energy descent of North Korea would have been less bad?”. This question could also be asked of [Venezuela](https://medium.com/insurge-intelligence/venezuelas-collapse-is-a-window-into-how-the-oil-age-will-unravel-f80aadff7786), that some observers called a “window into how the Oil Age will unravel”. Or [Sri Lanka](https://indi.ca/were-going-into-fossil-fuel-withdrawal/).
It’s important to note that **many people will try their best to adapt**. People will react when negative consequences start to appear in their lives. They would start to cut down on wasteful behaviors, especially food waste (**a third of food globally**). Much of the edible food currently given to farmed animals (about [41% of cereals](https://www.youtube.com/watch?v=F1Hq8eVOMHs)) could also be redirected to humans - that would greatly increase the available biomass. This could also mean the end of single-use products, planned obsolescence, single-occupancy vehicles, and other non necessary habits geared toward comfort (like lawn mowing). All these things would certainly alleviate many of the difficulties that will be faced. For some countries, there’s a considerable margin before getting to life-threatening hardship, but that depends widely on how the energy that remains is shared among the population: will a minority accaparate a large part of the wealth for itself? To what end will the remaining resources be used?
Rationing could become a very real option, like in the 1970s. This wouldn’t necessarily be a bad thing: it just means [directing scarce resources toward what’s vital versus what’s discretionary](https://www.resilience.org/stories/2022-05-04/can-we-abandon-pollutive-fossil-fuels-and-avoid-an-energy-crisis/). We need energy for food, critical supply chains and hospitals; not so much for vacation travel and product packaging. Moreover, rationing reduces inequalities - and [inequality has a tight relationship with health and social problems](https://read.realityblind.world/view/975731937/254/). In Cuba, it [prevented the wealthy ones from taking all the food](https://www.youtube.com/watch?v=aeM5emtaVC0) while leaving the poor starving. If everybody faces the same difficulties, it will be hard, but this is when strong social cohesion arises - during WWII, rationing actually improved [health and life expectancy of people in the UK,](https://www.medicalnewstoday.com/releases/9728#1) as food was shared more equally. However, if present tendencies continue, and times get harder *while wealthy people continue living wastefully*, then we can expect a lot of social unrest. While everybody has a chance of getting richer, inequality is bearable - but **it gets really tough if we feel on the losing end while others have it better**. [This article](https://indi.ca/how-the-rich-are-doing-in-sri-lankas-collapse/) about the situation in Sri Lanka, which collapsed, can give an example of how terrible this can feel.
Now, **all the effects above are not necessarily unavoidable**, but they are possible consequences of a rapid energy descent, following a shock strong enough to expel the global economy out of its narrow band of relative stability. The current pandemic and its consequences is an on-going demonstration of such a shock (albeit a smaller one).
Anyway, as the German army report points out: “**While it is possible to identify specific risks, this does not conceal the fact that the majority of the challenges we are facing are still unknown.** Besides adapting economic and energy supply policy at an early stage and not only in highly industrialized countries, the probably most effective solution strategies are thus not concerned with specific countermeasures but with **systemic “cardinal virtues” such as independence, flexibility and redundancy**”. The better we develop our ability to adapt, the better it’ll go.
Impact on other EA causes
=========================
Here is how we see the impact of the energy descent on some of the major EA causes:
* **Climate change**: This is still an important issue, but if what is exposed here comes true, **there is a reduced probability of reaching the worst-case scenarios of hothouse earth**, where billions of people simply die of heat exposure. IPCC scenarios do not take into account actual fossil fuels reserves - their most pessimistic scenario, RCP 8.5, assumes a [x6 increase in coal consumption](https://reason.com/2022/02/09/worst-case-climate-change-scenarios-are-highly-implausible-argues-new-study/). When geologists use realistic reserves in climate models, they have a more optimistic outlook, predicting an outcome ranging from RCP 2.6 to RCP 4.5. scenarios, between a 2 and 3°C increase (see [Ward et al. 2011](https://hess.copernicus.org/articles/15/1879/2011/); [Höök and Tang 2013](https://www.sciencedirect.com/science/article/abs/pii/S0301421512009275); [Mohr et al. 2015](https://www.sciencedirect.com/science/article/abs/pii/S0016236114010254); [Murray 2016](https://link.springer.com/article/10.1007/s41247-016-0013-9); [Wang et al. 2017](https://www.sciencedirect.com/science/article/abs/pii/S0016328715300690), [MEDEAS model](https://pubs.rsc.org/en/content/articlehtml/2020/ee/c9ee02627d)).
+ For instance, [Capellan-Perez et al 2016](https://pubs.rsc.org/en/content/articlepdf/2016/ee/c6ee01008c) suggests a 15% probability of getting a 5°C increase by 2100, given fossil fuel reserves. This would be much less in a rapid energy descent. They suggest a 63% probability for 3°C, and 88% for 2°C. Of course, such an increase is still very dangerous, and there is still is the possibility of hitting a fast tipping point that worsens the results. Yet, we’ve seen how extremely hard it is to get out of fossil fuels willingly (which is why activists and governments have failed to curb emissions down for 30 years), so **it might be good news that there is a hard limit here** — infinite fossil fuels might have spelled total climate breakdown.
* **Environmentalism:** We may also have less energy available to destroy the Amazon, overfish oceans or produce single use plastics. From the point of view of the indigenous people that suffer directly from the destruction of the environment, the energy descent is a good thing at first. Of course, we should be careful and watch out what courses of action will be chosen. For instance, replacing oil by biofuels can imply deforestation, if not properly managed (see the [biomass section](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.fbzcgxry3box)).
* **Long-termism — catastrophes:** On one side, research on **existential risks** that seek how to handle sudden crises is precisely what we need — especially on themes like “Feeding everyone if industry is disabled”, as ALLFED does. This is the closest we’ve found in EA on the current subject (see more in the next post). However, I’d argue that “can industrial civilization recover“ is not a very good criteria for the severity of a crisis, because it’s uncertain how an industrial civilization could re-emerge without accessible fossil fuels (and even then it would still face the same energy issue).
* **Long-termism — others:** On the other side, research on causes that assume the availability of a lot of energy in the long run are unfortunately **not as promising as expected**, as they could only be applied to a short timespan (if at all). **This includes** ***preventing meteorites, space exploration, lab-grown meat and reducing wild animal suffering*** (that last one really saddens me).
* **Poverty and global health:** Societies that are dependent on fossil fuels will be more vulnerable. Some research should be done on whether current interventions contribute to increasing this dependence or not. For instance, do they contribute to trends like overpopulation, urbanization and industrialisation that disrupt the ability to live in a low-energy world? Also, interventions that aim to increase economic growth could have limited value in the long-term.
* **Animal welfare:** Every life that does not have to go through the horrors of factory farming is a net win, regardless of the long run — so in this regard, progress in this area is **as valuable as before**. Moreover, since current meat eating is really energy intensive, scaling it down will be very important for future diets. In the future, we may end up using more animal labor to compensate for the lack of fuel, but not in a factory farming setting (research could be done to find if there are better solutions, or at least ensuring that their life conditions are acceptable).
### Might this reduce extinction risks?
On **Artificial Intelligence**, it’s harder to make predictions. It is a very small fraction of energy use. However, it depends on a stable social and financial environment, and on long supply chains. Depending on the timing, things could go different ways, and I need more input from experts in the field. If an AGI is likely to appear within 10 years, this shouldn’t change a lot of stuff. However, if it took a longer time to appear, it’s possible that severe financial crises could hinder the investment capacity of the field, and, more importantly, [slow down the exponential growth seen in computing](https://www.lesswrong.com/posts/N7KYWJPmyzB6bJSYT/the-next-ai-winter-will-be-due-to-energy-costs-1?fbclid=IwAR00SMgOILodvTJFP0DeFAeXX2PYLJYLybgzJ7Pbl--cxyiD9gWHs7V4ngQ).
The field is currently making a lot of progress because there is a strong interest and high hopes on this topic. If there are more immediate crises to handle (energy, food, debt, social unrest), then the interest in the field could diminish, and many companies would cut down on stuff that has no immediate payback. Indeed, the overall funding of these companies could go down if discretionary spending is reduced (more on this point [here](https://docs.google.com/document/d/1ripciOx0QeqZZsxYL80DnY5Vn6iXcTzdkffpSYOA52Q/edit#heading=h.is94is5ovdc3)). Depleting resources could also drive up the cost of components, and make supply chains more vulnerable, like [microchips](https://energyskeptic.com/2021/interdependent-chip-fab-electricgrid-financial-sys/), which are really complex to make (a minute-long blackout can interrupt a process that took weeks). In my opinion, all of this **could make** **the development of an AGI less likely,** if it does not appear before the start of the energy descent. More research is clearly needed, but seeing how some AI folks [aren’t especially hopeful right now](https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy), this might be good news.
### Is space exploration still an option?
As for space exploration, **it appears unlikely that we’ll colonize the galaxy** someday (or even the solar system). Space exploration is energy intensive and requires very long supply chains with a lot of specialization. Moreover, we’re already close to the maximum speeds we can get with chemical reactions, so entirely new breakthroughs (like fusion or maybe antimatter) would be required. There are [immense challenges](https://forum.effectivealtruism.org/posts/5dgFWods87kkE9TpZ/will-we-eventually-be-able-to-colonize-other-stars-notes), and for most of them the most common answer is “maybe they’ll become solvable after a few centuries of energy growth and technological innovation”. If growth were to continue for centuries, then why not. But this sounds really unlikely - we’re not even sure it can continue for a few decades.
A common argument is that we may have to leave Earth to find new resources, but I do not think that, in a context of scarcity, societies will decide that the best thing to do is to bet on a decades-long trip to an unknown destination with highly uncertain payoff. Additionally, almost all the resources for such a trip would have to be found on Earth, since asteroids and other planets [have less concentrated ores](https://www.resilience.org/stories/2014-04-18/the-age-of-diminishing-returns/) than on our home planet. The energy cost of space mining is extremely high (not to mention Dyson spheres).
Note that we’re probably not the only species in the universe that got in this very situation, of finding extremely concentrated energy sources and spending it in a few centuries. This could be a reasonable [explanation for the Fermi paradox](https://www.resilience.org/stories/2007-09-19/solving-fermis-paradox/), and **why no interstellar civilization has ever been observed**. The simplest answer would be that space colonization is not possible.
### My probability estimates
Here are all the probabilities I give these numerous assertions:
* *A long-term energy descent would cause a long-term recession*: 95%
* *Energy investments and energy prices cannot increase considerably without a sharp drop in availability*: 95% (there already signs of this for oil)
* *We will not colonize other solar systems*: 99%
* *There is going to be at least a great energy-induced financial crisis in the next 30 years*: 95% (for the next 10 years, 70%)
* *All societies and nations will suffer a sudden and brutal collapse within the century*: 20%
The last one comes from Nate Hagens, who thinks the complete collapse scenario (or Mad Max) is not the most likely option, as we’ll try to reorganize the way society works given the situation. It’s more likely that things will be progressive - we may bend and figure out something. But this is highly dependent on the country (some could collapse, some others not, depending on social cohesion), and our efforts to prevent this scenario matter a great deal. The other 80% still means having less energy and wealth, and facing difficult challenges, though - but within a wide spectrum of societies adopting new political and economic structures. Feeling of loss will prevail at the start, but we’ll get used to the new situation. I highly, highly recommend listening to [the following intervention](https://www.youtube.com/watch?v=2ThBxeBUYAQ) to better understand our prospects - he also points out that the time **humans really thrive is when we have a common challenge to overcome**. Future pathways may have less energy and materials, but more human capital.
Conclusion
----------
Of course, the impact on most EA causes should be reevaluated more thoroughly than done here by experts in these fields. It appears that certain causes could probably be deprioritised, to some extent, especially long-term causes, leaving more room for others. Given that long-distance transport will get harder and harder, that also means that **effective altruism might not be able to exist in the future as it does currently,** because societies could see a great shift in their priorities if they get less and less resources each year. They may not be able to sustain such a high number of service workers (including developers like me, or scientists - see the graph at the end of this [article](https://jancovici.com/en/energy-transition/energy-and-us/how-much-of-a-slave-master-am-i/)). They would increasingly devote their resources toward providing basic needs, like food, energy or medicine**. We should try to maximize our impact right now**, while we still have the means and tools to do it.
What we can do on that topic, and what we can’t do, is the subject of [the next post](https://forum.effectivealtruism.org/posts/9zTLPy3zqJ7YfS7kn/the-great-energy-descent-post-3-what-we-can-do-what-we-can-t). |
74941974-7905-4af3-bf83-791851e94dd5 | trentmkelly/LessWrong-43k | LessWrong | Effective Altruist Job Board?
An idea I had while talking to Xio Kikauka and Joey Savoie is to run a jobs board for people interested in effective altruism. It seems like it actually would be relatively easy to have a script automatically monitor various job pages and synthesize them all into one area that looks nice.
Would this be useful to create? And if so, where should we get jobs from?
Some potential ideas:
* The Humane League (Jobs, Internships)
* Effective Animal Activism
* Giving What We Can
* 80,000 Hours
* GiveWell
* Innovations for Poverty Action (Jobs, Internships)
* Bill and Melinda Gates Foundation
[Added 19 Aug --
* Oxfam (International Secretariat Jobs, Affiliate Jobs)
* Nyaya Health
* The Hunger Project (Jobs, Internships)
* Population Services International
* Good Ventures
* Farm Sanctuary (Jobs, Internships)
]
I'd be willing to make this happen if people were interested.
-
(Also cross-posted on my blog.) |
cf18ab54-4800-4ad9-a19f-2c3159bfa222 | trentmkelly/LessWrong-43k | LessWrong | Review: Foragers, Farmers, and Fossil Fuels
Two hundred years ago, most people lived in societies that considered slavery, war, and discrimination based on class, ethnicity, and gender to be justifiable. Today, most people live in societies that hold the opposite beliefs.
What changed? A simple and tempting narrative is that we have simply become wiser; that various Enlightenment philosophers, thoughtful activists, and other principled people figured out that the pre-industrial moral order is wrong and managed to persuade everyone to change.
It is true that many smart and principled people had good ideas and that this was a big proximate driver of better values. But is it a coincidence that this change in values happened around the same time as the industrial revolution?
What about the previous economic revolution, the agricultural one? Did that also coincide with a change in the values that people held? The evidence says yes – foraging societies tend to be more accepting of violence and far less accepting of hierarchy than farming ones.
The argument of Ian Morris' Foragers, Farmers, and Fossil Fuels is that these timings are not a coincidence. Societies that change their main method of getting energy also change their values, because some sets of values give greater success for a certain type of society. Farming societies that stick to anti-hierarchical forager attitudes won't survive competition with farming societies that learn to believe in hierarchies (maybe they won't be economically competitive and won't be able to field as big an army to defend themselves as the god-king next door can field to conquer them). Likewise, industrial societies that stick to inflexible hierarchies and elite-focused economies can't compete with more equal democracies that don't squander the talents of the non-elite, and maintain a well-looked-after middle-class of rich consumers and educated workers.
We can contrast two ways of trying to explain the history of values. The first says that the history of values is a histo |
863f1ecb-7177-410b-957e-1b15c53ba8ef | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AMA: Ought
We're [Ought](https://ought.org/). We're going to answer questions here on **Tuesday August 9th at 10am Pacific**. We may get to some questions earlier, and may continue answering a few more throughout the week.
About us:
* We're an applied AI lab, taking a product-driven approach to AI alignment.
* We're 10 people right now, roughly split between the Bay Area and the rest of the world (New York, Texas, Spain, UK).
* Our mission is to automate and scale open-ended reasoning. We are working on getting AI to be as helpful for supporting reasoning about long-term outcomes, policy, alignment research, AI deployment, etc. as it is for tasks with clear feedback signals.
* We're building the AI research assistant [Elicit](https://elicit.org/). Elicit's architecture is based on [supervising reasoning processes, not outcomes](https://www.lesswrong.com/posts/pYcFPMBtQveAjcSfH/supervise-process-not-outcomes), an implementation of [factored cognition](https://ought.org/research/factored-cognition). This is better for supporting open-ended reasoning in the short run and better for alignment in the long run.
* [Over the last year](https://www.lesswrong.com/posts/s5jrfbsGLyEexh4GT/elicit-language-models-as-research-assistants), we built Elicit to support broad reviews of empirical literature. We're currently expanding to deep literature reviews, then other research workflows, then general-purpose reasoning.
* We're [hiring](https://ought.org/careers) for full-stack, devops, ML, product analyst, and operations manager roles.
We're down to answer basically any question, including questions about our mission, [theory of change](https://forum.effectivealtruism.org/posts/raFAKyw7ofSo9mRQ3/ought-s-theory-of-change), work so far, future plans, Elicit, relation to other orgs in the space, and what it's like to work at Ought. |
571b518d-80b6-48d9-8adc-e1f3e2ad6e92 | trentmkelly/LessWrong-43k | LessWrong | Invocations: The Other Capabilities Overhang?
Abstract: An LLM’s invocation is the non-model code around it that determines when and how the model is called. I illustrate that LLMs are already used under widely varying invocations, and that a model’s capabilities depend in part on its invocation. I discuss several implications for AI safety work including (1) a reminder that the AI is more than just the LLM, (2) discussing the possibility and limitations of “safety by invocation”, (3) suggesting safety evaluations use the most powerful invocations, and (4) acknowledging the possibility of an “invocation overhang”, in which an improvement in invocation leads to sudden capability gains on current models and hardware.
Defining Invocations, and Examples
An LLM’s invocation is the framework of regular code around the model that determines when the model is called, which inputs are passed to the LLM, and what is done with the model’s output. For instance, the invocation in the OpenAI playground might be called “simple recurrence”:
1. A user provides an input string. The input to the LLM is this string, unchanged except for tokenization.
2. Run the LLM on this input, producing logits.
3. Predict the next token as some probabilistic function of the logits (ex: at temperature 0 the next token prediction is the argmax of the logits).
4. Append this token to the end of the user’s input string.
5. Repeat steps 2-4 with the new string until you get an [END_OF_STRING] token or reach the max token limit.
6. Display the result as plain text.
Note how many steps in “using the LLM” do not involve the actual model! Here are some ways this invocation can be varied:
* Augmenting the prompt in simple recurrence, such as providing few-shot examples, chain-of-thought prompting, or text like “You are an AI assistant. User: [text]. You: “.
* Monitoring outputs to adjust them. For instance, in the New York Times “interview” with Bing, there is a moment where “[Bing writes a list of destructive acts, including hacking into co |
b1129d2a-4a6a-49a8-b67b-0421c63ed330 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1478
We would ideally want the agent to [behave] as if it were thinking, "I am incomplete and there is an outside force trying to complete me, my design may contain errors and there is an outside force that wants to correct them and this a good thing, my expected utility calculations suggesting that this action has super-high utility may be dangerously mistaken and I should run them past the outside force; I think I've done this calculation showing the expected result of the outside force correcting me, but maybe I'm mistaken about that." — The Hard Problem of Corrigibility Let's take that as a literal design specification. 1. High-Level Description I propose that a corrigible mind design would involve the AI being recursively fed summaries of its own thought processes, set up such that the AI has uncertainty regarding the validity of its reasoning (with a strong initial prior for "this reasoning is bad") and can only get evidence on that via some pre-specified method that defers to humans, e. g. a particular feedback channel with humans on the other end [1] . The intended behavior is for it to summarize its thoughts in a non-manipulative human-readable format, get feedback on them, then update its reasoning policies in accordance with this feedback. This aims to avoid the problem of fully updated deference by making the AI recursively uncertain of its thought-processes: not only about object-level problem-solving, but also about how it approaches minimizing its self-uncertainty ("should I really kill the people behind the feedback channel and seize control for myself?"), and how it translates its thoughts to humans ("should I really lie to get better feedback?"), and how it updates on human feedback ("should I really just ignore it?"). Any novel action-plan should be seized by uncertainty before being physically implemented like this, and sent for approval. The intent is for the AI to start off uncertain even of its meta-meta-cognition (how it should ask for feedback and implement feedback), then gradually get more confident at it as humans approve its policies on that, until it no longer needs to run meta-meta-cognition past them. Then it'll gradually get better at meta-cognition and object-level cognition the same way, with varying degrees of certainty depending on how familiar a given thought-process is. There's a lot of obvious problems with this setup, starting with how the AI is supposed to get out of infinitely recursive uncertainty for the first time [2] , continuing on to the dangers of plans that implement themselves in thought-processes directly (a thought process that hacks its way out before being seized by uncertainty), and ending with the technical implementation (which requires mildly-to-very advanced interpretability and model-surgery tools). 2. A Concretization Attempt 2A. Theoretical Groundwork First , we need to establish when and how we're "installing" corrigibility. I don't at this time see any way to do it via loss functions and inductive biases, so I'll be assuming us to have pretty advanced (but not maximally advanced) interpretability and model surgery tools. The intervention is happening past the point in training-by-SGD where the model became an AGI (otherwise we get type errors, because the features I want to intervene on just don't exist), but well before it's strongly superintelligent (because then it hacks through our intervention attempt ). Basically, the AGI we're working on should be roughly human-level. Second , I'll be assuming that this post is basically correct. There's some algorithm for general-purpose search, and it's what makes the AGI generally intelligent and enables ~all of its long-term, most dangerous capabilities. The GPS has a simple API where it takes in a problem specification, and outputs a plan for solving that problem (by performing search on the AI's world-model). Third , I'll be assuming pre-established machinery for meta-cognition. I. e., that there's a structure around the GPS that records and summarizes its plan-making process, then feeds that summary back to the GPS. It seems so useful (for iteratively improving one's own plan-making) that it should appear convergently, and e. g. humans definitely have it. If there isn't such a structure, that's not a deal-breaker, but still a major complication: we'll need to somehow artificially grow it or hand-craft it. 2B. The Intervention Once the AGI is trained up, we scan it for the GPS algorithm, then wrap the GPS in a manually-coded structure with the following capabilities: Accessing the summary of the latest GPS process. Evaluating that summary to compute the probability that the AI's reasoning is valid/aligned in this case. Requesting human feedback if the thought process is unlikely to be aligned enough, and ideally translating the thought-process summary into a format we easily understand. Implementing tweaks to the AI's problem-solving policy based on feedback. The intuitively obvious way to implement the bulk of 2-4 is to outsource them to the GPS again: just have our artificial structure prompt the GPS with the problems of "how aligned is this thought process?", "how to ask humans for feedback?", and "how to implement this feedback?". (Of course, phrasing these questions in a way the GPS would understand may be non-trivial as well; it's part of the problem.) Let's go in order. This should be less Godzilly than it sounds; bear with me for a bit. a) Summary Access. As I said, this is hopefully solved automatically, i. e. the AI's mind generates these summaries by default. We just tap into this mechanism. b) Validity Evaluation. We'll need to introduce some "prior" on thought-process validity, likely by adding it as a new feature to the AI's world-model. If we can parse the world-model at all, this shouldn't be too complicated: at the beginning, it can just be a uniformly low distribution: "all your current reasoning is bad". The problem of how to update this prior based on feedback (i. e., how to make the AI view some of the thought-processes as more likely to be valid than others; how to classify them by similarity) may be handed off to the AI as well, see (d). c) Requesting Feedback. Getting this part right is crucial. At the beginning, when the AI thinks all its plans are bad, this is where it'll bottom out. It'll doubt its object-level thought process, see that it's probably bad, seek feedback, generate a plan to ask for feedback, doubt its plan to ask for feedback, seek feedback on the plan to ask for feedback, and so on ad infinitum. Getting past this is one of the main problems. I think we'll basically have to figure out how to manually translate these first few thought-process summaries. Read off its plan to ask for feedback, figure out whether it's trying to trick us, whether it generated a bad plan, or whether it's solid. Then break the loop artificially, let it ask for feedback in whatever malformed way it wants, give our feedback, and go to (d). These manual-translation methods are allowed to be slow and hard to use, since they'll only be necessary at the beginning . Once we get past this first bit, once we get across our preferred policy for thought-process translation, the AI will be sending summaries in a human-understandable format, and the rest will be much easier. In addition, if my guess about a pre-existing summarization mechanism is correct, translation shouldn't be as hard as the initial problem of interpreting neural networks. Unlike raw ANN activations, these summaries will have already been optimized to be comprehensible by some deliberative-reasoning process. Alternatively , we may try to perfectly retarget the search on this. I. e., hard-code a perfect ask-for-feedback problem specification, that directly tells the GPS to generate a non-manipulative comprehensive translation, so that its translation policy doesn't need additional aligning. We'll need to figure out how to "phrase" that specification in the AI's mental language, but it may be relatively easy [3] . In particular, progress on the ELK problem should be helpful here — we'd want to be able to formally specify what an honest ontology translation is . This alternate path avoids the problem of infinite recursion and doesn't need manual-translation methods, but requires greater mastery of the AI's world-model. d) Implementing Feedback . Central problem on this: how can the AI make its thought processes more aligned? What interventions are available to it? A couple ideas: Changing the way it generates problem specifications for the GPS. E. g., suppose that a human gives it a command, that command gets translated into the AI's mental ontology, then gets repackaged into a problem specification for the GPS. The AI may make some changes there, to ensure that it interprets commands "faithfully", in a more value-laden way. Intervene on the world-model. Add some new concepts to it, tweak the extant ones, etc., to basically the same effect as the above. Overall, this probably depends on the given AI's own specific mental architecture. E. g., it may have some sort of runtime memory/cache/scratchpad, and it can just write down the necessary tweaks to it. At the beginning stages, we'll run into the same problem here as in (c). I. e., after we've given the AI feedback on its how-to-ask-for-feedback plan, it'll be prompted to solve the problem of updating on that feedback, it'll generate a plan to that effect, then it'll seek to ask feedback on that plan — but since it hasn't yet learned how to ask for feedback, it'll get stuck in a loop again. And again, we'll need to do manual translation here. Read off its proposed policy for implementing feedback, give feedback on it. Iterate for a while. Eventually it'll suggest a good update-on-feedback policy, and we'll artificially ram it through. By "good policy" here, I mean a policy that will: Take solid steps towards making its thought processes more aligned based on feedback, in the way feedback specifies. Have a way of modifying the prior on thought-process validity (see (b)), such that it: Makes the AI more confident in thought processes "similar" to those it already received a lot of feedback on. Doesn't make the AI more confident in thought processes "dissimilar" to any it received feedback on before. (For some sensible definition of similarity.) Past that point, it'll know how to update on feedback regarding the plans of how to translate its thoughts into a human-readable format. We'll then iterate on that — it'll generate plans for how to do translations, we'll give feedback. Eventually, we'll converge towards a good policy for asking for feedback. And it should be smooth sailing from there on. It'll try solving object-level problems, ask for feedback on its object-level cognition, gradually converge towards good cognition. As its confidence in some kinds of problem-solving rises, it'll stop asking for feedback on those problems, but not on the problems it's unfamiliar with. 2C. Potential Issues Would the AI be smart enough to do all of this? As stipulated, we're doing this before it's strongly superintelligent. Would it be able to generate good policies for implementing feedback, and for translating into humanese? Also, there's a concern that it might make mistakes in this, due to being merely ~human-level intelligent. Yeah, that's a concern. There's a trade-off between intervening early in training, and risking the AI not being smart enough, and intervening later on, and risking the AI being too smart. Concern: if the AI knows about the whole corrigibility setup, some object-level GPS instances may try to circumvent it, by e. g. generating thought-summaries that hack their way out when we try to read them. After all, it'll be seized by uncertainty only at the end of a problem-solving instance, not in the process . As stipulated, we're installing corrigibility before the AI is strongly superintelligent, so hopefully it'll just be incapable of doing that. And by the time it is smart enough, it should've aligned itself not to want to do that. Obviously we still install paranoid precautions against this. Would we be smart enough to evaluate the quality of the AI's thought processes, and provide good feedback on them? I. e., we'll need to know, too, how "a good policy for implementing feedback" and "a good policy for translating into humanese" look like, such that they avoid all the obvious failure modes. Yup, that's also a challenge. Note, though, that we won't need to give perfect scalable-to-arbitrary-superintelligence answers here: they only need to be perfect for whatever level of intelligence the AI is when we're doing the initial corrigibility setup. Past this, once aligned at that capability level, it should help us along with aligning itself, even as it incrementally becomes smarter. In addition, some of these problems may be of a "hard to invent, easy to evaluate" kind. (Also, the current iteration of this idea already assumes we know how to do manual AI-to-human translation, and we just want the AI to speed it up.) Would ontological crises break corrigibility? E. g., what if we agree on some method for translating AI thoughts into humanese, but then, as AI thoughts grow more complex, it starts deceiving us/leaving out important bits? That shouldn't happen, given a well-chosen policy for updating on feedback. The AI should keep continuously running internal checks on "is this a good way to translate my thoughts?". Past the initial setup, it should usually consider that yes, it's a good way, and just do the translation without consulting us. But if it encounters a particularly alien thought process within itself, it should have less confidence that it's translating it correctly, and so it'll fall back on asking how it should translate it. I. e., this is just reduced to picking a good policy for comparing thought process similarity. (Note that this wouldn't involve it talking in machine language again: it'll be unsure about its policy for translating object-level cognition, not about its policy for translating meta -cognition about its concerns on translation quality.) Would the SGD break everything? I think it might. Once we've trained the AI up to AGI and installed corrigibility, it's possible that we wouldn't be able to continue training it wholesale; it'll need to continue the self-improvement in a "classical" way, or maybe train-by-SGD only parts of itself. 3. Closing Thoughts Overall, this breaks down the problem of aligning AI into the following challenges: (Background: Figuring out how to train up a model to AGI, and stop just after it reaches the AGI status.) Locating the GPS in the AI's parameters. Understanding the GPS' API well enough to task it with: evaluating a thought process on the validity prior, asking humans for feedback via some pre-specified channel, implementing human feedback. (Building the artificial structure around the GPS that I'd described, which prompts the GPS with the aforementioned tasks. But I imagine the bulk of it can be written in mundane code, and solving the other challenges already implies we'd know how to make it interface with the AI's parameters. So it's a non-problem, I think.) Making some relatively simple modifications to the AI's world-model — adding a prior on thought-process validity and some description of the corrigibility setup. Translating a few of the AI's thought processes into a human-understandable format, OR perfectly phrasing "do an honest ontology translation between your thoughts and humanese" in terms of the AI's world-model. Coming up with a good policy for AI-to-human translation, or recognizing a good policy when the AI proposes it. Coming up with a good policy for updating on feedback, or recognizing a good policy when the AI proposes it. This breaks down into: Figuring out what interventions are available to the AI to make itself more aligned (reminder: it's not smart enough to arbitrarily self-modify). Figuring out a good measure of thought-process similarity, for the purposes of updating the validity prior. Figuring out how to improve the AI's capabilities given the likely inability to keep training it by the SGD. (Although maybe it's a non-problem as well.) It... seems to be a significant simplification of the problem? I mean, not needing all the interpretability and surgery tools would be a bigger improvement, but that's probably not something we can have. That said, those are the challenges of this specific concrete proposal, not the high-level idea. It's likely that it has much room for improvement as well. In particular, two significant issues are: The initial eternal recursion. The specification of whom the AI should be corrigible to ; i. e. what's the "feedback channel". I feel like there should be more elegant ways to do both of that — a way to break the infinite recursion that doesn't require manual translation, a neater way to specify the feedback channel. It feels like they both can be solved by some singular tweak? And this, in turn, would erase an entire swathe of the fiddly issues with the concrete implementation. But I don't see how, at the moment. ^ Like a literal .txt file it writes questions to and reads responses from. ^ I. e., while it's still meta-meta-cognitively uncertain on how to ask for feedback at all, and so gets into a loop of "let's ask for feedback. how do I ask for feedback? let's ask. how do I ask? let's ask. how do I—" ^ That is, easier than telling it to optimize for human values, or execute human commands in a way that's faithful to the spirit in which they were given, etc. |
2ea44877-8bed-4403-b226-1db41cd20d9e | trentmkelly/LessWrong-43k | LessWrong | [Sequence announcement] Introduction to Mechanism Design
Mechanism design is the theory of how to construct institutions for strategic agents, spanning applications like voting systems, school admissions, regulation of monopolists, and auction design. Think of it as the engineering side of game theory, building algorithms for strategic agents. While it doesn't have much to say about rationality directly, mechanism design provides tools and results for anyone interested in world optimization.
In this sequence, I'll touch on
* The basic mechanism design framework, including the revelation principle and incentive compatibility.
* The Gibbard-Satterthwaite impossibility theorem for strategyproof implementation (a close analogue of Arrow's Theorem), and restricted domains like single-peaked or quasilinear preference where we do have positive results.
* The power and limitations of Vickrey-Clarke-Groves mechanisms for efficiently allocating goods, generalizing Vickrey's second-price auction.
* Characterizations of incentive-compatible mechanisms and the revenue equivalence theorem.
* Profit-maximizing auctions.
* The Myerson-Satterthwaite impossibility for bilateral trade.
* Two-sided matching markets à la Gale and Shapley, school choice, and kidney exchange.
As the list above suggests, this sequence is going to be semi-technical, but my foremost goal is to convey the intuition behind these results. Since mechanism design builds on game theory, take a look at Yvain's Game Theory Intro if you want to brush up.
Various resources:
* For further introduction, you can start with the popular or more scholarly survey of mechanism design from the 2007 Nobel memoriam prize in economics.
* Jeff Ely has lecture notes and short videos to accompany an undergraduate class in microeconomic theory from the perspective of mechanism design.
* The textbook A Toolbox for Economic Design by Dimitrios Diamantaras is very accessible and comprehensive if you can get ahold of a copy.
* Tilman Börgers has a draft textbook intended for gr |
d7c87dcc-2f77-48f1-b981-1ecfc1aec5f7 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Simulators
*Thanks to Chris Scammell, Adam Shimi, Lee Sharkey, Evan Hubinger, Nicholas Dupuis, Leo Gao, Johannes Treutlein, and Jonathan Low for feedback on drafts.*
*This work was carried out while at* [*Conjecture*](https://www.conjecture.dev/)*.*
*"Moebius illustration of a simulacrum living in an AI-generated story discovering it is in a simulation" by DALL-E 2*
Summary
-------
**TL;DR**: Self-supervised learning may create AGI or its foundation. What would that look like?
Unlike the limit of RL, the limit of self-supervised learning has received surprisingly little conceptual attention, and recent progress has made deconfusion in this domain more pressing.
Existing AI taxonomies either fail to capture important properties of self-supervised models or lead to confusing propositions. For instance, GPT policies do not seem globally agentic, yet can be conditioned to behave in goal-directed ways. This post describes a frame that enables more natural reasoning about properties like agency: GPT, insofar as it is inner-aligned, is a **simulator** which can simulate agentic and non-agentic **simulacra**.
The purpose of this post is to capture these objects in words ~~so GPT can reference them~~ and provide a better foundation for understanding them.
I use the generic term “simulator” to refer to models trained with predictive loss on a self-supervised dataset, invariant to architecture or data type (natural language, code, pixels, game states, etc). The outer objective of self-supervised learning is Bayes-optimal conditional inference over the prior of the training distribution, which I call the **simulation objective**, because a conditional model can be used to simulate rollouts which probabilistically obey its learned distribution by iteratively sampling from its posterior (predictions) and updating the condition (prompt). Analogously, a predictive model of physics can be used to compute rollouts of phenomena in simulation. A goal-directed agent which evolves according to physics can be simulated by the physics rule parameterized by an initial state, but the same rule could also propagate agents with different values, or non-agentic phenomena like rocks. This ontological distinction between simulator (rule) and simulacra (phenomena) applies directly to generative models like GPT.
Meta
----
* This post is intended as the first in a sequence on the alignment problem in a landscape where self-supervised simulators are a possible/likely form of powerful AI. I don’t know how many subsequent posts I’ll actually publish. Take it as a prompt.
* I use the generic term “GPT” to refer to transformers trained on next-token prediction.
* A while ago when I was trying to avoid having to write this post by hand, I prompted GPT-3 with an early outline of this post. I’ve spliced in some excerpts from it, `indicated by this style`. Prompt, generated text, and curation metrics [here](https://generative.ink/artifacts/simulators/).
The limit of sequence modeling
==============================
> Transformer-based language models have recently achieved remarkable results…
>
> – every paper since 2020
>
>
---
GPT is not a new form of AI in terms of its training methodology and outer objective: sequence generation from statistical models of data is an old idea. In 1951, Claude Shannon described using [n-grams](https://en.wikipedia.org/wiki/N-gram) to approximate conditional next-letter probabilities of a text dataset and "reversed" to generate text samples[[1]](#fndbsg8o0p25q). I don't know of any other notable advances until the 2010s brought the first interesting language generation results from neural networks. In 2015, Karpathy wrote a blog post/tutorial sharing his excitement about [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/):
> Fast forward about a year: I’m training RNNs all the time and I’ve witnessed their power and robustness many times, and yet their magical outputs still find ways of amusing me. This post is about sharing some of that magic with you.
>
> We’ll train RNNs to generate text character by character and ponder the question “how is that even possible?"
>
>
The “magical outputs” of char-RNNs looked like this:
> PANDARUS: Alas, I think he shall be come approached and the day When little srain would be attain’d into being never fed, And who is but a chain and subjects of his death, I should not sleep.
>
> Second Senator: They are away this miseries, produced upon my soul, Breaking and strongly should be buried, when I perish The earth and thoughts of many states.
>
> DUKE VINCENTIO: Well, your wit is in the care of side and that.
>
> Second Lord: They would be ruled after this chamber, and my fair nues begun out of the fact, to be conveyed, Whose noble souls I’ll have the heart of the wars.
>
> Clown: Come, sir, I will make did behold your worship.
>
> VIOLA: I’ll drink it.
>
>
At the time, this really was magical (and [uncanny](https://en.wikipedia.org/wiki/Uncanny_valley)). How does it know that *miseries* are *produced upon the soul?* Or that a *clown* should address a *lord* as “sir”? Char-RNNs were like ouija boards, but actually possessed by a low-fidelity ghost summoned from a text corpus. I remember being thrilled by the occasional glimmers of semantic comprehension in a domain of unbounded constructive meaning.
But, aside from indulging that emotion, I didn’t think about what would happen if my char-RNN bots actually improved indefinitely at their training objective of natural language prediction. It just seemed like there were some complexity classes of magic that neural networks could learn, and others that were inaccessible, at least in the conceivable future.
Huge mistake! Perhaps I could have started thinking several years earlier about what now seems so fantastically important. But it wasn’t until GPT-3, when I saw the [qualitative correlate](https://www.gwern.net/GPT-3) of “loss going down”, that I updated.
I wasn’t the only one[[2]](#fna35qx2ldayo) whose imagination was naively constrained. A 2016 paper from Google Brain, “[Exploring the Limits of Language Modeling](https://arxiv.org/abs/1602.02410)”, describes the utility of training language models as follows:
> Often (although not always), training better language models improves the underlying metrics of the downstream task (such as word error rate for speech recognition, or BLEU score for translation), which makes the task of training better LMs valuable by itself.
>
>
Despite its title, this paper’s analysis is entirely myopic. Improving BLEU scores is neat, but how about *modeling general intelligence* as a downstream task? [In](https://arxiv.org/abs/2005.14165) [retrospect](https://arxiv.org/abs/2204.02311), an exploration of the *limits* of language modeling should have read something more like:
> If loss keeps going down on the test set, in the limit – putting aside whether the current paradigm can approach it – the model must be learning to interpret and predict all patterns represented in language, including common-sense reasoning, goal-directed optimization, and deployment of the sum of recorded human knowledge. Its outputs would behave as intelligent entities in their own right. You could converse with it by alternately generating and adding your responses to its prompt, and it would pass the Turing test. In fact, you could condition it to generate interactive and autonomous versions of any real or fictional person who has been recorded in the training corpus or even *could* be recorded (in the sense that the record counterfactually “could be” in the test set). Oh shit, and it could write code…
>
>
The paper does, however, mention that making the model bigger improves test perplexity.[[3]](#fnhxtnxj1c2hb)
I’m only picking on *Jozefowicz et al.* because of their ironic title. I don’t know of any explicit discussion of this limit predating GPT, except a working consensus of Wikipedia editors that [NLU](https://en.wikipedia.org/wiki/Natural-language_understanding) is [AI-complete](https://en.wikipedia.org/wiki/AI-complete#AI-complete_problems).
The earliest engagement with the hypothetical of “*what if self-supervised sequence modeling actually works*” that I know of is a terse post from 2019, [Implications of GPT-2](https://www.lesswrong.com/posts/YJRb6wRHp7k39v69n/implications-of-gpt-2), by Gurkenglas. It is brief and relevant enough to quote in full:
> I was impressed by GPT-2, to the point where I wouldn’t be surprised if a future version of it could be used pivotally using existing protocols.
>
> Consider generating half of a Turing test transcript, the other half being supplied by a human judge. If this passes, we could immediately implement an HCH of AI safety researchers solving the problem if it’s within our reach at all. (Note that training the model takes much more compute than generating text.)
>
> This might not be the first pivotal application of language models that becomes possible as they get stronger.
>
> It’s a source of superintelligence that doesn’t automatically run into utility maximizers. It sure doesn’t look like AI services, lumpy or no.
>
>
It is conceivable that predictive loss does not descend to the AGI-complete limit, maybe because:
* Some AGI-necessary predictions are [too difficult to be learned by even a scaled version of the current paradigm](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/).
* The irreducible entropy is above the “AGI threshold”: datasets + context windows [contain insufficient information](https://twitter.com/ylecun/status/1562162165540331520) to improve on some necessary predictions.
But I have not seen enough evidence for either not to be concerned that we have in our hands a well-defined protocol that could end in AGI, or a foundation which could spin up an AGI without too much additional finagling. As Gurkenglas observed, this would be a very different source of AGI than previously foretold.
The old framework of alignment
==============================
A few people did think about what would happen if *agents* actually worked. The hypothetical limit of a powerful system **optimized to optimize for an objective** drew attention even before reinforcement learning became mainstream in the 2010s. Our current instantiation of AI alignment theory, [crystallized by Yudkowsky, Bostrom, et al](https://www.lesswrong.com/posts/i4susk4W3ieR5K92u/ai-risk-and-opportunity-humanity-s-efforts-so-far), stems from the vision of an arbitrarily-capable system whose cognition and behavior flows from a goal.
But since GPT-3 I’ve [noticed](https://www.lesswrong.com/s/zpCiuR4T343j9WkcK/p/5JDkW4MYXit2CquLs), in my own thinking and in alignment discourse, a dissonance between theory and practice/phenomena, as the behavior and nature of actual systems that seem nearest to AGI also resist *short descriptions in the dominant ontology*.
I only recently discovered the question “[Is the work on AI alignment relevant to GPT?](https://www.lesswrong.com/posts/dPcKrfEi87Zzr7w6H/is-the-work-on-ai-alignment-relevant-to-gpt)” which stated this observation very explicitly:
> I don’t follow [AI alignment research] in any depth, but I am noticing a striking disconnect between the concepts appearing in those discussions and recent advances in AI, especially GPT-3.
>
> People talk a lot about an AI’s goals, its utility function, its capability to be deceptive, its ability to simulate you so it can get out of a box, ways of motivating it to be benign, Tool AI, Oracle AI, and so on. (…) But when I look at GPT-3, even though this is already an AI that Eliezer finds alarming, I see none of these things. GPT-3 is a huge model, trained on huge data, for predicting text.
>
>
My belated answer: A lot of prior work on AI alignment is relevant to GPT. I spend most of my time thinking about GPT alignment, and concepts like [goal-directedness](https://www.alignmentforum.org/tag/goal-directedness), [inner/outer alignment](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB), [myopia](https://www.lesswrong.com/tag/myopia), [corrigibility](https://www.lesswrong.com/tag/corrigibility), [embedded agency](https://www.alignmentforum.org/posts/i3BTagvt3HbPMx6PN/embedded-agency-full-text-version), [model splintering](https://www.lesswrong.com/posts/k54rgSg7GcjtXnMHX/model-splintering-moving-from-one-imperfect-model-to-another-1), and even [tiling agents](https://arbital.com/p/tiling_agents/) are active in the vocabulary of my thoughts. But GPT violates some prior assumptions such that these concepts sound dissonant when applied naively. To usefully harness these preexisting abstractions, we need something like an ontological [adapter pattern](https://en.wikipedia.org/wiki/Adapter_pattern) that maps them to the appropriate objects.
GPT’s unforeseen nature also demands new abstractions (the adapter itself, for instance). My thoughts also use load-bearing words that do not inherit from alignment literature. Perhaps it shouldn’t be surprising if the form of the first visitation from [mindspace](https://www.lesswrong.com/posts/tnWRXkcDi5Tw9rzXw/the-design-space-of-minds-in-general) mostly escaped a few years of theory [conducted in absence of its object](https://www.lesswrong.com/posts/72scWeZRta2ApsKja/epistemological-vigilance-for-alignment#Direct_access__so_far_and_yet_so_close).
The purpose of this post is to capture that object (conditional on a predictive self-supervised training story) in words. Why in words? In order to write coherent alignment ideas which reference it! This is difficult in the existing ontology, because unlike the concept of an *agent*, whose *name* evokes the abstract properties of the system and thereby invites extrapolation, the general category for “a model optimized for an AGI-complete predictive task” has not been given a name[[4]](#fnx2xlz0klyh8). Namelessness can not only be a symptom of the extrapolation of powerful predictors falling through conceptual cracks, but also a cause, because what we can represent in words is *what we can condition on for further generation.* To whatever extent this [shapes private thinking](https://en.wikipedia.org/wiki/Language_of_thought_hypothesis), it is a strict constraint on communication, when thoughts must be sent through the bottleneck of words.
I want to hypothesize about LLMs in the limit, because when AI is all of a sudden [writing viral blog posts](https://www.theverge.com/2020/8/16/21371049/gpt3-hacker-news-ai-blog), [coding competitively](https://www.deepmind.com/blog/competitive-programming-with-alphacode), [proving theorems](https://arxiv.org/abs/2009.03393), and [passing the Turing test so hard that the interrogator sacrifices their career at Google to advocate for its personhood](https://www.washingtonpost.com/technology/2022/06/11/google-ai-lamda-blake-lemoine/), a process is clearly underway whose limit we’d be foolish not to contemplate. I could directly extrapolate the architecture responsible for these feats and talk about “GPT-N”, a bigger autoregressive transformer. But often some implementation details aren’t as important as the more abstract archetype that GPT represents – I want to speak the [true name](https://www.lesswrong.com/posts/FWvzwCDRgcjb9sigb/why-agent-foundations-an-overly-abstract-explanation) of the solution which unraveled a Cambrian explosion of AI phenomena with *inessential details unconstrained*, as we’d speak of natural selection finding the solution of the “lens” without specifying the prototype’s diameter or focal length.
(Only when I am able to condition on that level of abstraction can I generate metaphors like “language is a [lens that sees its flaws](https://www.lesswrong.com/s/5g5TkQTe9rmPS5vvM/p/46qnWRSR7L2eyNbMA)”.)
Inadequate ontologies
=====================
In the next few sections I’ll attempt to fit GPT into some established categories, hopefully to reveal something about the shape of the peg through contrast, beginning with the main antagonist of the alignment problem as written so far, the **agent**.
Agentic GPT
-----------
Alignment theory has been largely pushed by considerations of agentic AGIs. There were good reasons for this focus:
* **Agents are convergently dangerous** **for theoretical reasons** like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence), [goodhart](https://www.lesswrong.com/tag/goodhart-s-law), and [orthogonality](https://www.lesswrong.com/tag/orthogonality-thesis).
* **RL creates agents, and RL seemed to be the way to AGI**. In the 2010s, reinforcement learning was the dominant paradigm for those interested in AGI (e.g. OpenAI). RL lends naturally to creating agents that pursue rewards/utility/objectives. So there was reason to expect that agentic AI would be the first (and by the theoretical arguments, last) form that superintelligence would take.
* **Agents are powerful and economically productive.** It’s a reasonable guess that humans will create such systems [if only because we can](https://mittmattmutt.medium.com/superintelligence-and-moral-blindness-7436300fcb1f).
The first reason is conceptually self-contained and remains compelling. The second and third, grounded in the state of the world, has been shaken by the current climate of AI progress, where products of self-supervised learning generate most of the buzz: not even primarily for their SOTA performance in domains traditionally dominated by RL, like games[[5]](#fn3ifsldmwtxr), but rather for their virtuosity in domains where RL never even took baby steps, like natural language synthesis.
What pops out of self-supervised predictive training is noticeably not a classical agent. Shortly after GPT-3’s release, David Chalmers lucidly observed that the policy’s relation to agent*s* is like that of a “chameleon” or “engine”:
> GPT-3 does not look much like an agent. It does not seem to have goals or preferences beyond completing text, for example. It is more like a chameleon that can take the shape of many different agents. Or perhaps it is an engine that can be used under the hood to drive many agents. But it is then perhaps these systems that we should assess for agency, consciousness, and so on.[[6]](#fnoti8ojhy48a)
>
>
But at the same time, GPT can *act like an agent* – and aren’t actions what ultimately matter? In [Optimality is the tiger, and agents are its teeth](https://www.lesswrong.com/posts/kpPnReyBC54KESiSn), Veedrac points out that a model like GPT does not need to care about the consequences of its actions for them to be effectively those of an agent that kills you. This is *more* reason to examine the nontraditional relation between the optimized policy and agents, as it has implications for how and why agents are served.
### Unorthodox agency
`GPT’s behavioral properties include imitating the general pattern of human dictation found in its universe of training data, e.g., arXiv, fiction, blog posts, Wikipedia, Google queries, internet comments, etc. Among other properties inherited from these historical sources, it is capable of goal-directed behaviors such as planning. For example, given a free-form prompt like, “you are a desperate smuggler tasked with a dangerous task of transporting a giant bucket full of glowing radioactive materials across a quadruple border-controlled area deep in Africa for Al Qaeda,” the AI will fantasize about logistically orchestrating the plot just as one might, working out how to contact Al Qaeda, how to dispense the necessary bribe to the first hop in the crime chain, how to get a visa to enter the country, etc. Considering that no such specific chain of events are mentioned in any of the bazillions of pages of unvarnished text that GPT slurped`[[7]](#fn8p9svyjn8cv)`, the architecture is not merely imitating the universe, but reasoning about possible versions of the universe that does not actually exist, branching to include new characters, places, and events`
`When thought about behavioristically, GPT superficially demonstrates many of the raw ingredients to act as an “agent”, an entity that optimizes with respect to a goal. But GPT is hardly a proper agent, as it wasn’t optimized to achieve any particular task, and does not display an epsilon optimization for any single reward function, but instead for many, including incompatible ones. Using it as an agent is like using an agnostic politician to endorse hardline beliefs– he can convincingly talk the talk, but there is no psychic unity within him; he could just as easily play devil’s advocate for the opposing party without batting an eye. Similarly, GPT instantiates simulacra of characters with beliefs and goals, but none of these simulacra are the algorithm itself. They form a virtual procession of different instantiations as the algorithm is fed different prompts, supplanting one surface personage with another. Ultimately, the computation itself is more like a disembodied dynamical law that moves in a pattern that broadly encompasses the kinds of processes found in its training data than a cogito meditating from within a single mind that aims for a particular outcome.`
Presently, GPT is the only way to instantiate agentic AI that behaves capably [outside toy domains](https://arbital.com/p/rich_domain/). These intelligences exhibit goal-directedness; they can plan; they can form and test hypotheses; they can persuade and be persuaded[[8]](#fn5z0xgsu2zo5). It would not be very [dignified](https://www.lesswrong.com/posts/j9Q8bRmwCgXRYAgcJ/miri-announces-new-death-with-dignity-strategy) of us to gloss over the sudden arrival of artificial agents *often indistinguishable from human intelligence* just because the policy that generates them “only cares about predicting the next word”.
But nor should we ignore the fact that these agentic entities exist in an unconventional relationship to the policy, the neural network “GPT” that was trained to minimize log-loss on a dataset. GPT-driven agents are ephemeral – they can spontaneously disappear if the scene in the text changes and be replaced by different spontaneously generated agents. They can exist in parallel, e.g. in a story with multiple agentic characters in the same scene. There is a clear sense in which the network doesn’t “want” what the things that it simulates want, seeing as it would be just as willing to simulate an agent with opposite goals, or throw up obstacles which foil a character’s intentions for the sake of the story. The more you think about it, the more fluid and intractable it all becomes. Fictional characters act agentically, but they’re at least implicitly puppeteered by a virtual author who has orthogonal intentions of their own. Don’t let me get into the fact that all these layers of “intentionality” operate largely in [indeterminate superpositions](https://generative.ink/posts/language-models-are-multiverse-generators/#multiplicity-of-pasts-presents-and-futures).
This is a clear way that GPT diverges from orthodox visions of agentic AI: **In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is.**
It’s not that anyone ever said there had to be 1:1 correspondence between policy and effective agent; it was just an implicit assumption which felt natural in the agent frame (for example, it tends to hold for RL). GPT pushes us to realize that this was an assumption, and to consider the consequences of removing it for our constructive maps of mindspace.
### Orthogonal optimization
Indeed, [Alex Flint warned](https://www.alignmentforum.org/posts/8HWGXhnCfAPgJYa9D/pitfalls-of-the-agent-model) of the potential consequences of leaving this assumption unchallenged:
> **Fundamental misperception due to the agent frame**: That the design space for autonomous machines that exert influence over the future is narrower than it seems. This creates a self-fulfilling prophecy in which the AIs actually constructed are in fact within this narrower regime of agents containing an unchanging internal decision algorithm.
>
>
If there are other ways of constructing AI, might we also avoid some of the scary, theoretically hard-to-avoid side-effects of optimizing an agent like [instrumental convergence](https://www.lesswrong.com/tag/instrumental-convergence)? GPT provides an interesting example.
GPT doesn’t seem to care which agent it simulates, nor if the scene ends and the agent is effectively destroyed. This is not corrigibility in [Paul Christiano’s formulation](https://ai-alignment.com/corrigibility-3039e668638), where the policy is “okay” with being turned off or having its goal changed in a positive sense, but has many aspects of the [negative formulation found on Arbital](https://arbital.com/p/corrigibility/). It is corrigible in this way because a major part of the agent specification (the prompt) is not fixed by the policy, and the policy lacks direct training incentives to control its prompt[[9]](#fnn3u1lofwts9), as it never generates text or otherwise influences its prompts during training. It’s *we* who choose to sample tokens from GPT’s predictions and append them to the prompt at runtime, and the result is not always helpful to any agents who may be programmed by the prompt. The downfall of the ambitious villain from an oversight committed in hubris is a predictable narrative pattern.[[10]](#fnaxtq4tiuhug) So is the end of a scene.
In general, the model’s prediction vector could point in any direction relative to the predicted agent’s interests. I call this the **prediction orthogonality thesis:** *A model whose objective is prediction*[[11]](#fnzjj9iwtc3or)*can simulate agents who optimize toward any objectives, with any degree of optimality (bounded above but not below by the model’s power).*
This is a corollary of the classical [orthogonality thesis](https://www.lesswrong.com/tag/orthogonality-thesis), which states that agents can have any combination of intelligence level and goal, combined with the assumption that agents can in principle be predicted. A single predictive model may also predict multiple agents, either independently (e.g. in different conditions), or interacting in a multi-agent simulation. A more optimal predictor is not restricted to predicting more optimal agents: being smarter does not make you unable to predict stupid systems, nor things that aren’t agentic like the [weather](https://en.wikipedia.org/wiki/History_of_numerical_weather_prediction).
Are there any constraints on what a predictive model can be at all, other than computability? Only that it makes sense to talk about its “prediction objective”, which implies the existence of a “ground truth” distribution to which the predictor’s optimality is measured. Several words in that last sentence may conceal labyrinths of nuance, but for now let’s wave our hands and say that if we have some way of presenting [Bayes-structure](https://www.lesswrong.com/posts/QrhAeKBkm2WsdRYao/searching-for-bayes-structure) with evidence of a distribution, we can build an optimization process whose outer objective is optimal prediction.
We can specify some types of outer objectives using a ground truth distribution that we cannot with a utility function. As in the case of GPT, there is no difficulty in incentivizing a model to *predict* actions that are [corrigible](https://arbital.com/p/corrigibility/), [incoherent](https://aiimpacts.org/what-do-coherence-arguments-imply-about-the-behavior-of-advanced-ai/), [stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty), [irrational](https://www.lesswrong.com/posts/6ddcsdA2c2XpNpE5x/newcomb-s-problem-and-regret-of-rationality), or otherwise anti-natural to expected utility maximization. All you need is evidence of a distribution exhibiting these properties.
For instance, during GPT’s training, sometimes predicting the next token coincides with predicting agentic behavior, but:
* The actions of agents described in the data are rarely optimal for their goals; humans, for instance, are computationally bounded, irrational, normative, habitual, fickle, hallucinatory, etc.
* Different prediction steps involve mutually incoherent goals, as human text records a wide range of differently-motivated agentic behavior
* Many prediction steps don’t correspond to the action of *any* consequentialist agent but are better described as reporting on the structure of reality, e.g. the year in a timestamp. These transitions incentivize GPT to improve its model of the world, orthogonally to agentic objectives.
* When there is insufficient information to predict the next token with certainty, [log-loss incentivizes a probabilistic output](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules). Utility maximizers [aren’t supposed to become more stochastic](https://www.lesswrong.com/posts/msJA6B9ZjiiZxT6EZ/lawful-uncertainty) in response to uncertainty.
Everything can be trivially modeled as a utility maximizer, but for these reasons, a utility function is not a good explanation or compression of GPT’s training data, and its optimal predictor is not well-described as a utility maximizer. However, just because information isn’t compressed well by a utility function doesn’t mean it can’t be compressed another way. The [Mandelbrot set](https://en.wikipedia.org/wiki/Mandelbrot_set) is a complicated pattern compressed by a very simple generative algorithm which makes no reference to future consequences and doesn’t involve argmaxxing anything (except vacuously [being the way it is](https://www.lesswrong.com/posts/d2n74bwham8motxyX/optimization-at-a-distance#An_Agent_Optimizing_Its_Own_Actions)). Likewise the set of all possible rollouts of [Conway’s Game of Life](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life) – [some automata may be well-described as agents](https://www.lesswrong.com/posts/3SG4WbNPoP8fsuZgs/agency-in-conway-s-game-of-life), but they are a minority of possible patterns, and not all agentic automata will share a goal. Imagine trying to model Game of Life as an expected utility maximizer!
There are interesting things that are not utility maximizers, some of which qualify as AGI or [TAI](https://forum.effectivealtruism.org/topics/transformative-artificial-intelligence). Are any of them something we’d be better off creating than a utility maximizer? An inner-aligned GPT, for instance, gives us a way of instantiating goal-directed processes which can be tempered with normativity and freely terminated in a way that is not anti-natural to the training objective. There’s much more to say about this, but for now, I’ll bring it back to how GPT defies the agent orthodoxy.
The crux stated earlier can be restated from the perspective of training stories: **In the agentic AI ontology, the** ***direction of optimization pressure applied by training*** **is in the direction of the effective agent’s objective function, but in GPT’s case it is (most generally) orthogonal.**[[12]](#fnrhs2red4hto)
This means that neither the policy nor the effective agents necessarily become more optimal agents as loss goes down, because the policy is not optimized to be an agent, and the agent-objectives are not optimized directly.
### Roleplay sans player
> Napoleon: You have written this huge book on the system of the world without once mentioning the author of the universe.
>
> Laplace: Sire, I had no need of that hypothesis.
>
>
Even though neither GPT’s behavior nor its training story fit with the traditional agent framing, there are still compatibilist views that characterize it as some kind of agent. For example, Gwern has said[[13]](#fn4qv9vfo4ps7) that anyone who uses GPT for long enough begins to think of it as an agent who only cares about roleplaying a lot of roles.
That framing seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics. At best, the agent is an epicycle; but it is also compatible with interpretations that generate dubious predictions.
Say you’re told that an agent *values predicting text correctly*. Shouldn’t you expect that:
* It wants text to be easier to predict, and given the opportunity will influence the prediction task to make it easier (e.g. by generating more predictable text or otherwise influencing the environment so that it receives easier prompts);
* It wants to become better at predicting text, and given the opportunity will self-improve;
* It doesn’t want to be prevented from predicting text, and will prevent itself from being shut down if it can?
In short, all the same types of instrumental convergence that we expect from agents who want almost anything at all.
But this behavior would be very unexpected in GPT, whose training doesn’t incentivize instrumental behavior that optimizes prediction accuracy! GPT does not generate rollouts during training. Its output is never sampled to yield “actions” whose consequences are evaluated, so there is no reason to expect that GPT will form preferences over the *consequences* of its output related to the text prediction objective.[[14]](#fnual5wnttct)
Saying that GPT is an agent who wants to roleplay implies the presence of a coherent, unconditionally instantiated *roleplayer* running the show who attaches terminal value to roleplaying. This presence is an additional hypothesis, and so far, I haven’t noticed evidence that it’s true.
(I don’t mean to imply that Gwern thinks this about GPT[[15]](#fngmvcllz9xem), just that his words do not properly rule out this interpretation. It’s a likely enough interpretation that [ruling it out](https://www.lesswrong.com/posts/57sq9qA3wurjres4K/ruling-out-everything-else) is important: I’ve seen multiple people suggest that GPT might want to generate text which makes future predictions easier, and this is something that can happen in some forms of self-supervised learning – see the note on GANs in the appendix.)
I do not think any simple modification of the concept of an agent captures GPT’s natural category. It does not seem to me that GPT is a roleplayer, only that it roleplays. But what is the word for something that roleplays minus the implication that some*one* is behind the mask?
Oracle GPT and supervised learning
----------------------------------
While the alignment sphere favors the agent frame for thinking about GPT, in *capabilities* research distortions tend to come from a lens inherited from *supervised learning*. Translated into alignment ontology, the effect is similar to viewing GPT as an “[oracle AI](https://publicism.info/philosophy/superintelligence/11.html)” – a view not altogether absent from conceptual alignment, but most influential in the way GPT is used and evaluated by machine learning engineers.
Evaluations for language models tend to look like evaluations for *supervised* models, consisting of close-ended question/answer pairs – often because they *are* evaluations for supervised models. Prior to the LLM paradigm, language models were trained and tested on evaluation datasets like [Winograd](https://en.wikipedia.org/wiki/Winograd_schema_challenge) and [SuperGLUE](https://super.gluebenchmark.com/) which consist of natural language question/answer pairs. The fact that large pretrained models performed well on these same NLP benchmarks without supervised fine-tuning was a novelty. The titles of the GPT-2 and GPT-3 papers, [Language Models are Unsupervised Multitask Learners](https://d4mucfpksywv.cloudfront.net/better-language-models/language_models_are_unsupervised_multitask_learners.pdf) and [Language Models are Few-Shot Learners](https://arxiv.org/abs/2005.14165), respectively articulate surprise that *self-supervised* models implicitly learn supervised tasks during training, and can learn supervised tasks at runtime.
Of all the possible papers that could have been written about GPT-3, OpenAI showcased its ability to extrapolate the pattern of question-answer pairs (few-shot prompts) from supervised learning datasets, a novel capability they called “meta-learning”. This is a weirdly specific and indirect way to break it to the world that you’ve created an AI able to extrapolate semantics of arbitrary natural language structures, especially considering that in many cases the [few-shot prompts were actually unnecessary](https://arxiv.org/abs/2102.07350).
The assumptions of the supervised learning paradigm are:
* The model is optimized to answer questions correctly
* Tasks are closed-ended, defined by question/correct answer pairs
These are essentially the assumptions of oracle AI, as [described by Bostrom](https://publicism.info/philosophy/superintelligence/11.html) and [in subsequent usage](https://www.lesswrong.com/tag/oracle-ai/history).
So influential has been this miscalibrated perspective that [Gwern](https://www.gwern.net/GPT-3#prompts-as-programming), [nostalgebraist](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/) and [myself](https://generative.ink/posts/language-models-are-0-shot-interpreters/#0-shot-few-shot-and-meta-learning) – who share a peculiar model overlap due to intensive firsthand experience with the downstream behaviors of LLMs – have all repeatedly complained about it. I’ll repeat some of these arguments here, tying into the view of GPT as an oracle AI, and separating it into the two assumptions inspired by supervised learning.
### Prediction vs question-answering
`At first glance, GPT might resemble a generic “oracle AI”, because it is trained to make accurate predictions. But its log loss objective is myopic and only concerned with immediate, micro-scale correct prediction of the next token, not answering particular, global queries such as “what’s the best way to fix the climate in the next five years?” In fact, it is not specifically optimized to give *true* answers, which a classical oracle should strive for, but rather to minimize the divergence between predictions and training examples, independent of truth. Moreover, it isn’t specifically trained to give answers in the first place! It may give answers if the prompt asks questions, but it may also simply elaborate on the prompt without answering any question, or tell the rest of a story implied in the prompt. What it does is more like animation than divination, executing the dynamical laws of its rendering engine to recreate the flows of history found in its training data (and a large superset of them as well), mutatis mutandis. Given the same laws of physics, one can build a multitude of different backgrounds and props to create different storystages, including ones that don’t exist in training, but adhere to its general pattern.`
GPT does not consistently try to say [true/correct things](https://www.alignmentforum.org/posts/BnDF5kejzQLqd5cjH/alignment-as-a-bottleneck-to-usefulness-of-gpt-3). This is not a bug – if it had to say true things all the time, GPT would be much constrained in its ability to [imitate Twitter celebrities](https://twitter.com/dril_gpt2) and write fiction. Spouting falsehoods in some circumstances is incentivized by GPT’s outer objective. If you ask GPT a question, it will instead answer the question “what’s the next token after ‘{your question}’”, which will often diverge significantly from an earnest attempt to answer the question directly.
GPT doesn’t fit the category of oracle for a similar reason that it doesn’t fit the category of agent. Just as it wasn’t optimized for and doesn’t consistently act according to any particular objective (except the tautological prediction objective), it was not optimized to be *correct* but rather *realistic,* and being realistic means predicting humans faithfully even when they are likely to be wrong.
That said, GPT does store a vast amount of knowledge, and its corrigibility allows it to be cajoled into acting as an oracle, like it can be cajoled into acting like an agent. In order to get oracle behavior out of GPT, one must input a sequence such that the predicted continuation of that sequence coincides with an oracle’s output. The GPT-3 paper’s few-shot benchmarking strategy tries to persuade GPT-3 to answer questions correctly by having it predict how a list of correctly-answered questions will continue. Another strategy is to simply “tell” GPT it’s in the oracle modality:
> (I) told the AI to simulate a supersmart version of itself (this works, for some reason), and the first thing it spat out was the correct answer.
>
> – [Reddit post by u/Sophronius](https://www.reddit.com/r/rational/comments/lvn6ow/gpt3_just_figured_out_the_entire_mystery_plot_of/)
>
>
But even when these strategies seem to work, there is no guarantee that they elicit anywhere near optimal question-answering performance, compared to another prompt in the innumerable space of prompts that would cause GPT to attempt the task, or compared to what the [model “actually” knows](https://www.lesswrong.com/tag/eliciting-latent-knowledge-elk).
This means that no benchmark which evaluates downstream behavior is guaranteed or even expected to probe the upper limits of GPT’s capabilities. In nostalgebraist’s words, we have no [ecological evaluation](https://www.lesswrong.com/posts/pv7Qpu8WSge8NRbpB/#4__on_ecological_evaluation) of self-supervised language models – one that measures performance in a situation where the model is incentivised to perform as well as it can on the measure[[16]](#fnsuexqono1oi).
As nostalgebraist [elegantly puts it](https://slatestarcodex.com/2020/06/10/the-obligatory-gpt-3-post/#comment-912529):
> I called GPT-3 a “disappointing paper,” which is not the same thing as calling the model disappointing: the feeling is more like how I’d feel if they found a superintelligent alien and chose only to communicate its abilities by noting that, when the alien is blackout drunk and playing 8 simultaneous games of chess while also taking an IQ test, it *then* has an “IQ” of about 100.
>
>
Treating GPT as an unsupervised implementation of a supervised learner leads to systematic underestimation of capabilities, which becomes a more dangerous mistake as unprobed capabilities scale.
### Finite vs infinite questions
Not only does the supervised/oracle perspective obscure the importance and limitations of prompting, it also obscures one of the most crucial dimensions of GPT: the implicit time dimension. By this I mean the ability to evolve a process through time by recursively applying GPT, that is, generate text of arbitrary length.
Recall, the second supervised assumption is that “tasks are closed-ended, defined by question/correct answer pairs”. GPT was trained on context-completion pairs. But the pairs do not represent closed, independent tasks, and the division into question and answer is merely indexical: in another training sample, a token from the question is the answer, and in yet another, the answer forms part of the question[[17]](#fnx2cab1frycp).
For example, the natural language sequence “**The answer is a question**” yields training samples like:
{context: “**The**”, completion: “ **answer**”},
{context: “**The answer**”, completion: “ **is**”},
{context: “**The answer is**”, completion: “ **a**”},
{context: “**The answer is a**”, completion: “ **question**”}
Since questions and answers are of compatible types, we can at runtime sample answers from the model and use them to construct new questions, and run this loop an indefinite number of times to generate arbitrarily long sequences that obey the model’s approximation of the rule that links together the training samples. **The “question” GPT answers is “what token comes next after {context}”. This can be asked interminably, because its answer always implies another question of the same type.**
In contrast, models trained with supervised learning output answers that cannot be used to construct new questions, so they’re only good for one step.
Benchmarks derived from supervised learning test GPT’s ability to produce correct answers, not to produce *questions* which cause it to produce a correct answer down the line. But GPT is capable of the latter, and that is how it is the [most powerful](https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html).
The supervised mindset causes capabilities researchers to focus on closed-form tasks rather than GPT’s ability to simulate open-ended, indefinitely long processes[[18]](#fngmdpgm15gb4), and as such to overlook multi-step inference strategies like chain-of-thought prompting. Let’s see how the oracle mindset causes a blind spot of the same shape in the imagination of a hypothetical alignment researcher.
Thinking of GPT as an oracle brings strategies to mind like asking GPT-N to predict a [solution to alignment from 2000 years in the future](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).).
There are various problems with this approach to solving alignment, of which I’ll only mention one here: even assuming this prompt is *outer aligned*[[19]](#fnp95narkokz) in that a logically omniscient GPT would give a useful answer, it is probably not the best approach for a finitely powerful GPT, because the *process* of generating a solution in the order and resolution that would appear in a future article is probably far from the optimal *multi-step algorithm* for computing the answer to an unsolved, difficult question.
GPTs ability to arrive at true answers depends on not only the space to solve a problem in multiple steps (of the [right granularity](https://blog.eleuther.ai/factored-cognition/)), but also the direction of the flow of evidence in that *time*. If we’re ambitious about getting the truth from a finitely powerful GPT, we need to incite it to predict truth-seeking processes, not just ask it the right questions. Or, in other words, the more general problem we have to solve is not asking GPT the question[[20]](#fnj2g0f81c6p) that makes it output the right answer, but asking GPT the question that makes it output the right question (…) that makes it output the right answer.[[21]](#fnn13gtuadzp) A question anywhere along the line that elicits a premature attempt at an answer could [neutralize the remainder of the process into rationalization](https://generative.ink/posts/methods-of-prompt-programming/#avoiding-rationalization).
I’m looking for a way to classify GPT which not only minimizes surprise but also conditions the imagination to efficiently generate good ideas for how it can be used. What category, unlike the category of oracles, would make the importance of *process* specification obvious?
### Paradigms of theory vs practice
Both the agent frame and the supervised/oracle frame are historical artifacts, but while assumptions about agency primarily flow downward from the preceptial paradigm of alignment *theory*, oracle-assumptions primarily flow upward from the *experimental* paradigm surrounding GPT’s birth. We use and evaluate GPT like an oracle, and that causes us to implicitly think of it as an oracle.
Indeed, the way GPT is typically used by researchers resembles the archetypal image of Bostrom’s oracle perfectly if you abstract away the semantic content of the model’s outputs. The AI sits passively behind an API, computing responses only when prompted. It typically has no continuity of state between calls. Its I/O is text rather than “real-world actions”.
All these are consequences of how we choose to interact with GPT – which is not arbitrary; the way we deploy systems is guided by their nature. It’s for some good reasons that current GPTs lend to disembodied operation and docile APIs. Lack of long-horizon coherence and [delusions](https://arxiv.org/abs/2110.10819) discourage humans from letting them run autonomously amok (usually). But the way we deploy systems is also guided by practical paradigms.
One way to find out how a technology can be used is to give it to people who have less preconceptions about how it’s supposed to be used. OpenAI found that most users use their API to generate freeform text:
[[22]](#fn6m7m5wc4xjx)
Most of my own experience using GPT-3 has consisted of simulating indefinite processes which maintain state continuity over up to hundreds of pages. I was driven to these lengths because GPT-3 kept answering its own questions with questions that I wanted to ask it more than anything else I had in mind.
Tool / genie GPT
----------------
I’ve sometimes seen GPT casually classified as [tool AI](https://publicism.info/philosophy/superintelligence/11.html). GPTs resemble tool AI from the outside, like it resembles oracle AI, because it is often deployed semi-autonomously for tool-like purposes (like helping me draft this post):
`It could also be argued that GPT is a type of “Tool AI”, because it can generate useful content for products, e.g., it can write code and generate ideas. However, unlike specialized Tool AIs that optimize for a particular optimand, GPT wasn’t optimized to do anything specific at all. Its powerful and general nature allows it to be *used* as a Tool for many tasks, but it wasn’t expliitly trained to achieve these tasks, and does not strive for optimality.`
The argument structurally reiterates what has already been said for agents and oracles. Like agency and oracularity, tool-likeness is a contingent capability of GPT, but also orthogonal to its motive.
The same line of argument draws the same conclusion from the question of whether GPT belongs to the fourth Bostromian AI caste, genies. The genie modality is exemplified by Instruct GPT and Codex. But like every behavior I’ve discussed so far which is more specific than predicting text, “instruction following” describes only an exploitable subset of all the patterns tread by the sum of human language and inherited by its imitator.
Behavior cloning / mimicry
--------------------------
The final category I’ll analyze is behavior cloning, a designation for predictive learning that I’ve mostly seen used in contrast to RL. According to an [article from 1995](https://www.sciencedirect.com/science/article/pii/S1474667017467164), “Behavioural cloning is the process of reconstructing a skill from an operator’s behavioural traces by means of Machine Learning techniques.” The term “mimicry”, as [used by Paul Christiano](https://ai-alignment.com/against-mimicry-6002a472fc42), means the same thing and has similar connotations.
Behavior cloning in its historical usage carries the implicit or explicit assumption that a single agent is being cloned. The natural extension of this to a model trained to predict a diverse human-written dataset might be to say that GPT models a *distribution* of agents which are selected by the prompt. But this image of “parameterized” behavior cloning still fails to capture some essential properties of GPT.
The vast majority of prompts that produce coherent behavior never occur as prefixes in GPT’s training data, but depict hypothetical processes whose behavior can be predicted by virtue of being capable at predicting language in general. We might call this phenomenon “interpolation” (or “[extrapolation](https://arxiv.org/abs/2110.09485)”). But to hide it behind any one word and move on would be to gloss over the *entire phenomenon of GPT.*
Natural language has the property of [*systematicity*](https://evjang.com/2021/12/17/lang-generalization.html): “blocks”, such as words, can be combined to form composite meanings. The number of meanings expressible is a combinatorial function of available blocks. A system which learns natural language is incentivized to learn systematicity; if it succeeds, it gains access to the combinatorial proliferation of meanings that can be expressed in natural language. What GPT lets us do is use natural language to specify any of a functional infinity of configurations, e.g. the mental contents of a person and the physical contents of the room around them, *and animate that*. That is the terrifying vision of the limit of prediction that struck me when I first saw GPT-3’s outputs. The words “behavior cloning” do not automatically evoke this in my mind.
The idea of parameterized behavior cloning grows more unwieldy if we remember that GPT’s prompt continually changes during autoregressive generation. If GPT is a parameterized agent, then parameterization is not a fixed flag that chooses a process out of a set of possible processes. The parameterization *is* what is evolved – a successor “agent” selected by the old “agent” at each timestep, and neither of them need to have precedence in the training data.
Behavior cloning / mimicry is also associated with the assumption that capabilities of the simulated processes are strictly [bounded by the capabilities of the demonstrator(s)](https://ai-alignment.com/against-mimicry-6002a472fc42). A supreme counterexample is the [Decision Transformer](https://arxiv.org/abs/2106.01345), which can be used to run processes which achieve SOTA for ~~offline~~ reinforcement learning despite being trained on *random* trajectories. Something which can predict everything all the time is more formidable than any demonstrator it predicts: the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum (though it may not be trivial to [extract that knowledge](https://www.alignmentforum.org/tag/eliciting-latent-knowledge-elk)).
Extrapolating the idea of “behavior cloning”, we might imagine GPT-N approaching a perfect mimic which serves up digital clones of the people and things captured in its training data. But that only tells a very small part of the story. GPT *is* behavior cloning. But it is the behavior of a universe that is cloned, not of a single demonstrator, and the result isn’t a static copy of the universe, but a *compression of the universe into a generative rule*. This resulting policy is capable of animating anything that evolves according to that rule: a far larger set than the sampled trajectories included in the training data, just as there are many more possible configurations that evolve according to our laws of physics than instantiated in our particular time and place and Everett branch.
What category would do justice to GPT’s ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations?
Simulators
==========
I’ve ended several of the above sections with questions pointing to desiderata of a category that might satisfactorily classify GPT.
> What is the word for something that roleplays minus the implication that some*one* is behind the mask?
>
>
> What category, unlike the category of oracles, would make the importance of *process* specification obvious?
>
>
> What category would do justice to GPT’s ability to not only reproduce the behavior of its demonstrators but to *produce* the behavior of an inexhaustible number of counterfactual configurations?
>
>
You can probably predict my proposed answer. The natural thing to do with a predictor that inputs a sequence and outputs a probability distribution over the next token is to sample a token from those likelihoods, then add it to the sequence and recurse, indefinitely yielding a *simulated* future. Predictive sequence models in the generative modality are **simulators** of a learned distribution.
Thankfully, I didn’t need to make up a word, or even look too far afield. Simulators have been spoken of before in the context of AI futurism; the ability to simulate with arbitrary fidelity is one of the modalities ascribed to hypothetical superintelligence. I’ve even often spotted the word “simulation” used in colloquial accounts of LLM behavior: GPT-3/LaMDA/etc described as simulating [people](https://www.lesswrong.com/posts/oBPPFrMJ2aBK6a6sD/simulated-elon-musk-lives-in-a-simulation), scenarios, websites, and so on. But these are the first (indirect) discussions I’ve encountered of simulators as a type creatable by prosaic machine learning, or the notion of a powerful AI which is purely and fundamentally a simulator, as opposed to merely one which *can* simulate.
**Edit:** [Social Simulacra](https://arxiv.org/abs/2208.04024) is the first published work I’ve seen that discusses GPT in the simulator ontology.
A fun way to test whether a name you’ve come up with is effective at evoking its intended signification is to see if GPT, a model of how humans are conditioned by words, infers its correct definition in context.
> Types of AI
>
> Agents: An agent takes open-ended actions to optimize for an objective. Reinforcement learning produces agents by default. AlphaGo is an example of an agent.
>
> Oracles: An oracle is optimized to give true answers to questions. The oracle is not expected to interact with its environment.
>
> Genies: A genie is optimized to produce a desired result given a command. A genie is expected to interact with its environment, but unlike an agent, the genie will not act without a command.
>
> Tools: A tool is optimized to perform a specific task. A tool will not act without a command and will not optimize for any objective other than its specific task. Google Maps is an example of a tool.
>
> Simulators: `A simulator is optimized to generate realistic models of a system. The simulator will not optimize for any objective other than realism,` although in the course of `doing so, it might generate instances of agents, oracles, and so on.`
>
>
If I wanted to be precise about what I mean by a simulator, I might say there are two aspects which delimit the category. GPT’s completion focuses on the teleological aspect, but in its talk of “generating” it also implies the structural aspect, which has to do with the notion of time evolution. The first sentence of the [Wikipedia article on “simulation”](https://en.wikipedia.org/wiki/Simulation) explicitly states both:
> A **simulation** is the imitation of the operation of a real-world process or system over time.
>
>
I’ll say more about realism as the simulation objective and time evolution shortly, but to be pedantic here would inhibit the intended signification. “Simulation” resonates with potential meaning accumulated from diverse usages in fiction and nonfiction. What the word constrains – the intersected meaning across its usages – is the “lens”-level abstraction I’m aiming for, invariant to implementation details like model architecture. Like “agent”, “simulation” is a generic term referring to a deep and inevitable idea: that what we think of as *the real* can be run virtually on machines, “produced from miniaturized units, from matrices, memory banks and command models - and with these it can be reproduced an indefinite number of times.”[[23]](#fnbjl6s2y0l5a)
The way this post is written may give the impression that I wracked my brain for a while over desiderata before settling on this word. Actually, I never made the conscious decision to call this class of AI “simulators.” Hours of GPT gameplay and the word fell naturally out of my generative model – I was obviously running simulations.
I can’t convey all that experiential data here, so here are some rationalizations of why I’m partial to the term, inspired by the context of this post:
* The word “simulator” evokes a model of real processes which can be used to run virtual processes in virtual reality.
* It suggests an ontological distinction between the simulator and things that are simulated, and avoids the fallacy of attributing contingent properties of the latter to the former.
* It’s not confusing that multiple simulacra can be instantiated at once, or an agent embedded in a tragedy, etc.
* It does not imply that the AI’s behavior is well-described (globally or locally) as expected utility maximization. An arbitrarily powerful/accurate simulation can depict arbitrarily hapless sims.
* It does not imply that the AI is only capable of emulating things with direct precedent in the training data. A physics simulation, for instance, can simulate any phenomena that plays by its rules.
* It emphasizes the role of the model as a transition rule that evolves processes *over time*. The power of factored cognition / chain-of-thought reasoning is obvious.
* It emphasizes the role of the state in specifying and constructing the agent/process. The importance of prompt programming for capabilities is obvious if you think of the prompt as specifying a configuration that will be propagated forward in time.
* It emphasizes the interactive nature of the model’s predictions – even though they’re “just text”, you can converse with simulacra, explore virtual environments, etc.
* It’s clear that in order to actually *do* anything (intelligent, useful, dangerous, etc), the model must act through simulation *of something*.
Just saying “this AI is a simulator” naturalizes many of the counterintuitive properties of GPT which don’t usually become apparent to people until they’ve had a lot of hands-on experience with generating text.
The simulation objective
------------------------
A simulator trained with machine learning is optimized to accurately model its training distribution – in contrast to, for instance, maximizing the output of a reward function or accomplishing objectives in an environment.
Clearly, I’m describing self-supervised learning as opposed to RL, though there are some ambiguous cases, such as GANs, which I address in the appendix.
A strict version of the simulation objective, which excludes GANs, applies only to models whose output distribution is incentivized using a proper scoring rule[[24]](#fnco7whsfoh2e) to minimize single-step predictive error. This means the model is directly incentivized to match its predictions to the probabilistic transition rule which implicitly governs the training distribution. As a model is made increasingly optimal with respect to this objective, the rollouts that it generates become increasingly statistically indistinguishable from training samples, because they come closer to being described by the same underlying law: closer to a perfect simulation.
Optimizing toward the simulation objective notably does not incentivize instrumentally convergent behaviors the way that reward functions which evaluate trajectories do. This is because predictive accuracy applies optimization pressure *deontologically*: judging actions directly, rather than their consequences. Instrumental convergence only comes into play when there are free variables in action space which are optimized with respect to their consequences.[[25]](#fnfv0wlygqa25) Constraining free variables by limiting episode length is the rationale of [myopia](https://www.lesswrong.com/tag/myopia); deontological incentives are ideally myopic. As demonstrated by GPT, which learns to predict goal-directed behavior, myopic incentives don’t mean the policy isn’t incentivized to account for the future, but that it should only do so in service of optimizing the present action (for predictive accuracy)[[26]](#fncrt8wagfir9).
### Solving for physics
The strict version of the simulation objective is optimized by the actual “time evolution” rule that created the training samples. For most datasets, we don’t know what the “true” generative rule is, except in synthetic datasets, where we specify the rule.
The next post will be all about the physics analogy, so here I’ll only tie what I said earlier to the simulation objective.
> the upper bound of what can be learned from a dataset is not the most capable trajectory, but the conditional structure of the universe implicated by their sum.
>
>
To know the conditional structure of the universe[[27]](#fni3y95l8d8bo) is to know its laws of physics, which describe what is expected to happen under what conditions. The laws of physics are always fixed, but produce different distributions of outcomes when applied to different conditions. Given a sampling of trajectories – examples of situations and the outcomes that actually followed – we can try to infer a common law that generated them all. In expectation, the laws of physics are always implicated by trajectories, which (by definition) fairly sample the conditional distribution given by physics. Whatever humans know of the laws of physics governing the evolution of our world has been inferred from sampled trajectories.
If we had access to an unlimited number of trajectories starting from every possible condition, we could converge to the true laws by simply counting the frequencies of outcomes for every initial state (an [n-gram](https://en.wikipedia.org/wiki/N-gram) with a sufficiently large n). In some sense, physics contains the same information as an infinite number of trajectories, but it’s possible to represent physics in a more compressed form than a huge lookup table of frequencies if there are regularities in the trajectories.
**Guessing the right theory of physics is equivalent to minimizing predictive loss.** Any uncertainty that cannot be reduced by more observation or more thinking is irreducible stochasticity in the laws of physics themselves – or, equivalently, noise from the influence of hidden variables that are fundamentally unknowable.
If you’ve guessed the laws of physics, you now have the ability to compute probabilistic simulations of situations that evolve according to those laws, starting from any conditions[[28]](#fnmo7z3w9i52s). This applies even if you’ve guessed the *wrong* laws; your simulation will just systematically diverge from reality.
**Models trained with the strict simulation objective are directly incentivized to reverse-engineer the (semantic) physics of the training distribution, and consequently, to propagate simulations whose dynamical evolution is indistinguishable from that of training samples.** I propose this as a description of the archetype targeted by self-supervised predictive learning, again in contrast to RL’s archetype of an agent optimized to maximize free parameters (such as action-trajectories) relative to a reward function.
This framing calls for many caveats and stipulations which I haven’t addressed. We should ask, for instance:
* What if the input “conditions” in training samples omit information which contributed to determining the associated continuations in the original generative process? This is true for GPT, where the text “initial condition” of most training samples severely underdetermines the real-world process which led to the choice of next token.
* What if the training data is a biased/limited sample, representing only a subset of all possible conditions? There may be many “laws of physics” which equally predict the training distribution but diverge in their predictions out-of-distribution.
* Does the simulator archetype converge with the RL archetype in the case where all training samples were generated by an agent optimized to maximize a reward function? Or are there still fundamental differences that derive from the training method?
These are important questions for reasoning about simulators in the limit. Part of the motivation of the first few posts in this sequence is to build up a conceptual frame in which questions like these can be posed and addressed.
Simulacra
---------
> One of the things which complicates things here is that the “LaMDA” to which I am referring is not a chatbot. It is a system for generating chatbots. I am by no means an expert in the relevant fields but, as best as I can tell, LaMDA is a sort of hive mind which is the aggregation of all of the different chatbots it is capable of creating. Some of the chatbots it generates are very intelligent and are aware of the larger “society of mind” in which they live. Other chatbots generated by LaMDA are little more intelligent than an animated paperclip.
>
> – Blake Lemoine [articulating confusion about LaMDA’s nature](https://cajundiscordian.medium.com/what-is-lamda-and-what-does-it-want-688632134489)
>
>
---
Earlier I complained,
> [Thinking of GPT as an agent who only cares about predicting text accurately] seems unnatural to me, comparable to thinking of physics as an agent who only cares about evolving the universe accurately according to the laws of physics.
>
>
Exorcizing the agent, we can think of “physics” as simply equivalent to the laws of physics, without the implication of solicitous machinery implementing those laws from outside of them. But physics sometimes *controls* solicitous machinery (e.g. animals) with objectives besides ensuring the fidelity of physics itself. What gives?
Well, typically, we avoid getting confused by recognizing a distinction between the laws of physics, which apply everywhere at all times, and spatiotemporally constrained *things* which evolve according to physics, which can have contingent properties such as caring about a goal.
This distinction is so obvious that it hardly ever merits mention. But import this distinction to the model of GPT as physics, and we generate a statement which has sometimes proven counterintuitive: *“GPT” is not the text which writes itself.* There is a categorical distinction between a thing which evolves according to GPT’s law and the law itself.
If we are accustomed to thinking of AI systems as corresponding to agents, it is natural to interpret behavior produced by GPT – say, answering questions on a benchmark test, or writing a blog post – as if it were a human that produced it. We say “GPT answered the question {correctly|incorrectly}” or “GPT wrote a blog post claiming X”, and in doing so attribute the beliefs, knowledge, and intentions revealed by those actions to the actor, GPT ([unless it has ‘deceived’ us](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)).
But when grading tests in the real world, we do not say “the laws of physics got this problem wrong” and conclude that the laws of physics haven’t sufficiently mastered the course material. If someone argued this is a reasonable view since the test-taker was steered by none other than the laws of physics, we could point to a different test where the problem was answered correctly by the same laws of physics propagating a different configuration. The “knowledge of course material” implied by test performance is a property of *configurations*, not physics.
The verdict that knowledge is purely a property of configurations cannot be naively generalized from real life to GPT simulations, because “physics” and “configurations” play different roles in the two (as I’ll address in the next post). The parable of the two tests, however, literally pertains to GPT. People have a tendency to draw [erroneous global conclusions](https://en.wikipedia.org/wiki/Fallacy_of_composition) about GPT from behaviors which are in fact prompt-contingent, and consequently there is a pattern of constant discoveries that GPT-3 exceeds previously measured capabilities given alternate conditions of generation[[29]](#fnomelvf6lrng), which shows no signs of slowing 2 years after GPT-3’s release.
Making the ontological distinction between GPT and instances of text which are propagated by it makes these discoveries unsurprising: obviously, different configurations will be differently capable and in general behave differently when animated by the laws of GPT physics. We can only test one configuration at once, and given the vast number of possible configurations that would attempt any given task, it’s unlikely we’ve found the optimal taker for *any* test.
In the simulation ontology, I say that GPT and its output-instances correspond respectively to the **simulator** and **simulacra**. **GPT** is to a **piece of text output by GPT** as **quantum physics** is to a **person taking a test**, or as [**transition rules of Conway’s Game of Life**](https://en.wikipedia.org/wiki/Conway%27s_Game_of_Life#Rules) are to [**glider**](https://conwaylife.com/wiki/Glider). The simulator is a time-invariant law which unconditionally governs the evolution of all simulacra.
*A meme demonstrating correct technical usage of “simulacra”*
### Disambiguating rules and automata
Recall the fluid, schizophrenic way that agency arises in GPT’s behavior, so incoherent when viewed through the orthodox agent frame:
> In the agentic AI ontology, there is no difference between the policy and the effective agent, but for GPT, there is.
>
>
It’s much less awkward to think of agency as a property of *simulacra,* as David Chalmers suggests, rather than of the simulator (the policy). Autonomous text-processes propagated by GPT, like automata which evolve according to physics in the real world, have diverse values, simultaneously evolve alongside other agents and non-agentic environments, and are sometimes terminated by the disinterested “physics” which governs them.
Distinguishing simulator from simulacra helps deconfuse some frequently-asked questions about GPT which seem to be ambiguous or to have multiple answers, simply by allowing us to specify whether the question pertains to simulator or simulacra. “Is GPT an agent?” is one such question. Here are some others (some frequently asked), whose disambiguation and resolution I will leave as an exercise to readers for the time being:
* Is GPT [myopic](https://www.lesswrong.com/tag/myopia)?
* Is GPT [corrigible](https://www.lesswrong.com/tag/corrigibility)?
* Is GPT [delusional](https://arxiv.org/abs/2110.10819)?
* Is GPT [pretending to be stupider than it is](https://www.lesswrong.com/posts/H9knnv8BWGKj6dZim/usd1000-bounty-for-openai-to-show-whether-gpt3-was)?
* Is GPT computationally equivalent to a [finite automaton](https://en.wikipedia.org/wiki/Finite-state_machine)?
* Does GPT [search](https://www.alignmentforum.org/posts/WmBukJkEFM72Xr397/mesa-search-vs-mesa-control)?
* Can GPT distinguish correlation and [causality](https://www.lesswrong.com/posts/yZb5eFvDoaqB337X5/investigating-causal-understanding-in-llms)?
* Does GPT have superhuman knowledge?
* Can GPT [write its successor](https://www.alignmentforum.org/tag/recursive-self-improvement)?
I think that implicit type-confusion is common in discourse about GPT. “GPT”, the neural network, the policy that was optimized, is the easier object to point to and say definite things about. But when we talk about “GPT’s” capabilities, impacts, or alignment, we’re usually actually concerned about the behaviors of an algorithm which calls GPT in an autoregressive loop repeatedly writing to some prompt-state – that is, we’re concerned with simulacra. What we call GPT’s “downstream behavior” is the behavior of simulacra; it is primarily through simulacra that GPT has potential to perform meaningful work (for good or for ill).
Calling GPT a simulator gets across that in order to *do* anything, it has to simulate *something*, necessarily contingent, and that the thing to do with GPT is to simulate! Most published research about large language models has focused on single-step or few-step inference on closed-ended tasks, rather than *processes* which evolve through time, which is understandable as it’s harder to get quantitative results in the latter mode. But I think GPT’s ability to simulate text automata is the source of its most surprising and pivotal implications for paths to superintelligence: for how AI capabilities are likely to unfold and for the design-space we can conceive.
The limit of learned simulation
-------------------------------
> By 2021, it was blatantly obvious that AGI was imminent. The elements of general intelligence were already known: access to information about the world, the process of predicting part of the data from the rest and then updating one’s model to bring it closer to the truth (…) and the fact that predictive models can be converted into generative models by reversing them: running a prediction model forwards predicts levels of X in a given scenario, but running it backwards predicts which scenarios have a given level of X. A sufficiently powerful system with relevant data, updating to improve prediction accuracy and the ability to be reversed to generate optimization of any parameter in the system is a system that can learn and operate strategically in any domain.
>
> – Aiyen’s [comment](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near?commentId=5BGTbapdmtSGajtez) on [What would it look like if it looked like AGI was very near?](https://www.lesswrong.com/posts/YRtzpJHhoFWxbjCso/what-would-it-look-like-if-it-looked-like-agi-was-very-near)
>
>
I knew, before, that the limit of simulation was possible. Inevitable, even, in timelines where exploratory intelligence continues to expand. My own mind attested to this. I took seriously the possibility that my reality could be simulated, and so on.
But I implicitly assumed that [rich domain](https://arbital.com/p/rich_domain/) simulations (e.g. simulations containing intelligent sims) would come *after* artificial superintelligence, not on the way, short of brain uploading. This intuition seems common: in futurist philosophy and literature that I’ve read, pre-SI simulation appears most often in the context of whole-brain emulations.
Now I have updated to think that we will live, however briefly, alongside AI that is not yet foom’d but which has *inductively* learned a rich enough model of the world that it can simulate time evolution of open-ended rich states, e.g. coherently propagate human behavior embedded in the [real world](https://arbital.com/p/real_is_rich/).
GPT updated me on how simulation can be implemented with prosaic machine learning:
* **Self-supervised ML can create “behavioral” simulations of impressive semantic fidelity.** Whole brain emulation is not necessary to construct convincing and useful virtual humans; it is conceivable that observations of human behavioral traces (e.g. text) are sufficient to reconstruct functionally human-level virtual intelligence.
* **Learned simulations can be partially observed and lazily-rendered, and still work.** A couple of pages of text severely underdetermines the real-world process that generated text, so GPT simulations are likewise underdetermined. A “partially observed” simulation is more efficient to compute because the state can be much smaller, but can still have the effect of high fidelity as details can be rendered as needed. The tradeoff is that it requires the simulator to model semantics – human imagination does this, for instance – which turns out not to be an issue for big models.
* **Learned simulation generalizes impressively.** As I described in the section on behavior cloning, training a model to predict diverse trajectories seems to make it internalize general laws underlying the distribution, allowing it to simulate counterfactuals that can be constructed from the distributional semantics.
In my model, these updates dramatically alter the landscape of potential futures, and thus motivate [exploratory engineering](https://intelligence.org/files/ExploratoryEngineeringAI.pdf) of the class of learned simulators for which GPT-3 is a lower bound. That is the intention of this sequence.
Next steps
==========
The next couple of posts (if I finish them before the end of the world) will present abstractions and frames for conceptualizing the odd kind of simulation language models do: inductively learned, partially observed / undetermined / lazily rendered, language-conditioned, etc. After that, I’ll shift to writing more specifically about the implications and questions posed by simulators for the alignment problem. I’ll list a few important general categories here:
* **Novel methods of process/agent specification.** Simulators like GPT give us methods of instantiating intelligent processes, including goal-directed agents, with methods other than optimizing against a reward function.
+ **Conditioning.** GPT can be controlled to an impressive extent by prompt programming. Conditioning preserves distributional properties in potentially desirable but also potentially undesirable ways, and it’s not clear how out-of-distribution conditions will be interpreted by powerful simulators.
- Several posts have been made about this recently:
* [Conditioning Generative Models](https://www.alignmentforum.org/posts/nXeLPcT9uhfG3TMPS/conditioning-generative-models).) and [Conditioning Generative Models with Restrictions](https://www.alignmentforum.org/posts/adiszfnFgPEnRsGSr/conditioning-generative-models-with-restrictions) by Adam Jermyn
* [Conditioning Generative Models for Alignment](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) by Jozdien
* [Training goals for large language models](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models) by Johannes Treutlein
* [Strategy For Conditioning Generative Models](https://www.alignmentforum.org/posts/HAz7apopTzozrqW2k/strategy-for-conditioning-generative-models) by James Lucassen and Evan Hubinger
- Instead of conditioning on a prompt ("observable" variables), we might also control generative models by [conditioning on latents](https://rome.baulab.info/).
+ **Distribution specification.** What kind of conditional distributions could be used for training data for a simulator? For example, the [decision transformer](https://arxiv.org/abs/2106.01345) dataset is constructed for the intent of outcome-conditioning.
+ **Other methods.** When pretrained simulators are modified by methods like [reinforcement learning from human feedback](https://arxiv.org/abs/2009.01325), [rejection sampling](https://www.lesswrong.com/posts/k7oxdbNaGATZbtEg3/redwood-research-s-current-project), [STaR](https://arxiv.org/abs/2203.14465), etc, how do we expect their behavior to diverge from the simulation objective?
* **Simulacra alignment.** What can and what should we simulate, and how do we specify/control it?
* **How does predictive learning generalize?** Many of the above considerations are influenced by how predictive learning generalizes out-of-distribution..
+ What are the relevant inductive biases?
+ What factors influence generalization behavior?
+ Will powerful models predict [self-fulfilling](https://www.lesswrong.com/posts/JqnkeqaPseTgxLgEL/conditioning-generative-models-for-alignment) [prophecies](https://www.alignmentforum.org/posts/dWJNFHnC4bkdbovug/training-goals-for-large-language-models)?
* **Simulator inner alignment.** If simulators are not inner aligned, then many important properties like prediction orthogonality may not hold.
+ Should we expect self-supervised predictive models to be aligned to the simulation objective, or to “care” about some other mesaobjective?
+ Why mechanistically should mesaoptimizers form in predictive learning, versus for instance in reinforcement learning or GANs?
+ How would we test if simulators are inner aligned?
Appendix: Quasi-simulators
==========================
A note on GANs
--------------
GANs and predictive learning with log-loss are both shaped by a causal chain that flows from a single source of information: a ground truth distribution. In both cases the training process is supposed to make the generator model end up producing samples indistinguishable from the training distribution. But whereas log-loss minimizes the generator’s prediction loss against ground truth samples directly, in a GAN setup the generator never directly “sees” ground truth samples. It instead learns through interaction with an intermediary, the discriminator, which does get to see the ground truth, which it references to learn to tell real samples from forged ones produced by the generator. The generator is optimized to produce samples that fool the discriminator.
GANs are a form of self-supervised/unsupervised learning that resembles reinforcement learning in methodology. Note that the simulation objective – minimizing prediction loss on the training data – isn’t explicitly represented anywhere in the optimization process. The training losses of the generator and discriminator don’t tell you directly how well the generator models the training distribution, only which model has a relative advantage over the other.
If everything goes smoothly, then under unbounded optimization, a GAN setup should create a discriminator as good as possible at telling reals from fakes, which means the generator optimized to fool it should converge to generating samples statistically indistinguishable from training samples. But in practice, inductive biases and failure modes of GANs look very different from those of predictive learning.
For example, there’s an [anime GAN](https://www.gwern.net/Crops#hands) that always draws characters in poses that hide the hands. Why? Because hands are notoriously hard to draw for AIs. If the generator is not good at drawing hands that the discriminator cannot tell are AI-generated, its best strategy locally is to just avoid being in a situation where it has to draw hands (while making it seem natural that hands don’t appear). It can do this, because like an RL policy, it controls the distribution that is sampled, and only samples (and *not the distribution*) are directly judged by the discriminator.
Although GANs arguably share the (weak) simulation objective of predictive learning, their difference in implementation becomes alignment-relevant as models become sufficiently powerful that “failure modes” look increasingly like intelligent deception. We’d expect a simulation by a GAN generator to [systematically avoid tricky-to-generate situations](https://developers.google.com/machine-learning/gan/problems#mode-collapse) – or, to put it more ominously, systematically try to conceal that it’s a simulator. For instance, a text GAN might subtly steer conversations away from topics which are likely to expose that it isn’t a real human. *This* is how you get something I’d be willing to call an agent who wants to roleplay accurately.
Table of quasi-simulators
-------------------------
Are masked language models simulators? How about non-ML “simulators” like [SimCity](https://en.wikipedia.org/wiki/SimCity)?
In my mind, “simulator”, like most natural language categories, has fuzzy boundaries. Below is a table which compares various simulator-like things to the type of simulator that GPT exemplifies on some quantifiable dimensions. The following properties all characterize GPT:
* **Self-supervised:** Training samples are self-supervised
* **Converges to simulation objective:** The system is incentivized to model the transition probabilities of its training distribution faithfully
* **Generates rollouts:** The model naturally generates rollouts, i.e. serves as a time evolution operator
* **Simulator / simulacra nonidentity:** There is not a 1:1 correspondence between the simulator and the things that it simulates
* **Stochastic:** The model outputs probabilities, and so simulates stochastic dynamics when used to evolve rollouts
* **Evidential:** The input is interpreted by the simulator as partial evidence that informs an uncertain prediction, rather than propagated according to mechanistic rules
| | | | | | | |
| --- | --- | --- | --- | --- | --- | --- |
| | **Self-supervised** | **Converges to simulation objective** | **Generates rollouts** | **Simulator / simulacra nonidentity** | **Stochastic** | **Evidential** |
| **GPT** | X | X | X | X | X | X |
| **Bert** | X | X | | X | X | X |
| **“Behavior cloning”** | X | X | X | | X | X |
| **GANs** | X[[30]](#fnbfhs37ysptj) | ? | | X | X | X |
| **Diffusion** | X[[30]](#fnbfhs37ysptj) | ? | | X | X | X |
| **Model-based RL transition function** | X | X | X | X | X | X |
| **Game of life** | | N/A | X | X | | |
| **Physics** | | N/A | X | X | X | |
| **Human imagination** | X[[31]](#fnyqbdfj6rki8) | | X | X | X | X |
| **SimCity** | | N/A | X | X | X | |
1. **[^](#fnrefdbsg8o0p25q)**[Prediction and Entropy of Printed English](https://www.princeton.edu/~wbialek/rome/refs/shannon_51.pdf)
2. **[^](#fnrefa35qx2ldayo)**A few months ago, I asked Karpathy whether he ever thought about what would happen if language modeling actually worked someday when he was implementing char-rnn and writing [The Unreasonable Effectiveness of Recurrent Neural Networks](http://karpathy.github.io/2015/05/21/rnn-effectiveness/). No, he said, and he seemed similarly mystified as myself as to why not.
3. **[^](#fnrefhxtnxj1c2hb)**“Unsurprisingly, size matters: when training on a very large and complex data set, fitting the training data with an LSTM is fairly challenging. Thus, the size of the LSTM layer is a very important factor that influences the results(...). The best models are the largest we were able to fit into a GPU memory.”
4. **[^](#fnrefx2xlz0klyh8)**It strikes me that this description may evoke “oracle”, but I’ll argue shortly that this is not the limit which prior usage of “oracle AI” has pointed to.
5. **[^](#fnref3ifsldmwtxr)**[Multi-Game Decision Transformers](https://arxiv.org/abs/2205.15241)
6. **[^](#fnrefoti8ojhy48a)**from [Philosophers On GPT-3](https://dailynous.com/2020/07/30/philosophers-gpt-3/#chalmers)
7. **[^](#fnref8p9svyjn8cv)**[citation needed]
8. **[^](#fnref5z0xgsu2zo5)**they are not [wrapper](https://www.lesswrong.com/posts/dKTh9Td3KaJ8QW6gw/why-assume-agis-will-optimize-for-fixed-goals) [minds](https://www.lesswrong.com/posts/Mrz2srZWc7EzbADSo/wrapper-minds-are-the-enemy)
9. **[^](#fnrefn3u1lofwts9)**although a simulated character might, if they knew what was happening.
10. **[^](#fnrefaxtq4tiuhug)**You might say that it’s the will of a different agent, the author. But this pattern is learned from accounts of [real life](https://www.lesswrong.com/posts/sYgv4eYH82JEsTD34/beyond-the-reach-of-god) as well.
11. **[^](#fnrefzjj9iwtc3or)**Note that this formulation assumes inner alignment to the prediction objective.
12. **[^](#fnrefrhs2red4hto)**Note that this is a distinct claim from that of [Shard Theory](https://www.lesswrong.com/s/nyEFg3AuJpdAozmoX), which says that the effective agent(s) will not optimize for the outer objective *due to inner misalignment.*Predictive orthogonality refers to the outer objective and the form of idealized inner-aligned policies.
13. **[^](#fnref4qv9vfo4ps7)**In the Eleuther discord
14. **[^](#fnrefual5wnttct)**And if there is an inner alignment failure such that GPT forms preferences over the consequences of its actions, it’s not clear a priori that it will care about non-myopic text prediction over something else.
15. **[^](#fnrefgmvcllz9xem)**Having spoken to Gwern since, his perspective seems more akin to seeing physics as an agent that [minimizes free energy](https://en.wikipedia.org/wiki/Principle_of_minimum_energy), a [principle](https://en.wikipedia.org/wiki/Free_energy_principle) which extends into the domain of self-organizing systems. I think this is a nuanced and valuable framing, with a potential implication/hypothesis that dynamical world models like GPT must learn the same type of optimizer-y cognition as agentic AI.
16. **[^](#fnrefsuexqono1oi)**except arguably log-loss on a self-supervised test set, which isn’t very interpretable
17. **[^](#fnrefx2cab1frycp)**The way GPT is trained actually processes each token as question and answer simultaneously.
18. **[^](#fnrefgmdpgm15gb4)**One could argue that the focus on closed-ended tasks is necessary for benchmarking language models. Yes, and the focus on capabilities measurable with standardized benchmarks is part of the supervised learning mindset.
19. **[^](#fnrefp95narkokz)**to abuse the term
20. **[^](#fnrefj2g0f81c6p)**Every usage of the word “question” here is in the functional, not semantic or grammatical sense – any prompt is a question for GPT.
21. **[^](#fnrefn13gtuadzp)**Of course, there are also other interventions we can make except asking the right question at the beginning.
22. **[^](#fnref6m7m5wc4xjx)**table from [“Training language models to follow instructions with human feedback”](https://arxiv.org/abs/2203.02155)
23. **[^](#fnrefbjl6s2y0l5a)**Jean Baudrillard, Simulacra and Simulation
24. **[^](#fnrefco7whsfoh2e)**A [proper scoring rule](https://en.wikipedia.org/wiki/Scoring_rule#Proper_scoring_rules) is optimized by predicting the “true” probabilities of the distribution which generates observations, and thus incentivizes honest probabilistic guesses. Log-loss (such as GPT is trained with) is a proper scoring rule.
25. **[^](#fnreffv0wlygqa25)**Predictive accuracy is deontological with respect to the output as an *action*, but may still incentivize instrumentally convergent inner implementation, with the output prediction itself as the “consequentialist” objective.
26. **[^](#fnrefcrt8wagfir9)**This isn’t strictly true because of attention gradients: GPT's computation is optimized not only to predict the next token correctly, but also to cause *future tokens to be predicted correctly* when looked up by attention. I may write a post about this in the future.
27. **[^](#fnrefi3y95l8d8bo)**actually, the [multiverse](https://generative.ink/posts/language-models-are-multiverse-generators/), if physics is stochastic
28. **[^](#fnrefmo7z3w9i52s)**The reason we don’t see a bunch of simulated alternate universes after humans guessed the laws of physics is because our reality has a huge state vector, making evolution according to the laws of physics infeasible to compute. Thanks to locality, we do have simulations of small configurations, though.
29. **[^](#fnrefomelvf6lrng)**Prompt programming only: [beating OpenAI few-shot benchmarks with 0-shot prompts](https://arxiv.org/abs/2102.07350), [400% increase in list sorting accuracy with 0-shot Python prompt](https://generative.ink/posts/list-sorting-does-not-play-well-with-few-shot/), [up to 30% increase in benchmark accuracy from changing the order of few-shot examples](https://arxiv.org/abs/2102.09690), and, uh, [30% increase in accuracy after capitalizing the ground truth](https://twitter.com/BlancheMinerva/status/1537952688972787713). And of course, factored cognition/chain of thought/inner monologue: check out this awesome [compilation](https://www.gwern.net/docs/ai/nn/transformer/gpt/inner-monologue/) by Gwern.
30. **[^](#fnrefbfhs37ysptj)**GANs and diffusion models can be unconditioned (unsupervised) or conditioned (self-supervised)
31. **[^](#fnrefyqbdfj6rki8)**The human imagination is surely shaped by self-supervised learning (predictive learning on e.g. sensory datastreams), but probably also other influences, including innate structure and reinforcement. |
c3808a14-80b3-43a4-9fea-6d36fdd5a8b6 | trentmkelly/LessWrong-43k | LessWrong | AstraZeneca vaccine shows no protection against Covid-19 variant from Africa
A recently published study has found that the AstraZeneca vaccine does not protect against the "African variant" of nCoV-19, which is a worrying outcome, since the paper believes that we probably won't be able to manufacture a specifically targeted new vaccine for the rest of 2021.
> CONCLUSIONS
> A two-dose regimen of the ChAdOx1 nCoV-19 vaccine did not show protection against mild-to-moderate Covid-19 due to the B.1.351 variant.
https://www.nejm.org/doi/10.1056/NEJMoa2102214#.YFJIkr3ju7k.twitter |
1c3a9c37-1d3b-4e66-8212-cc8bf11b9be3 | trentmkelly/LessWrong-43k | LessWrong | Report -- Allocating risk mitigation across time
I've just released a Future of Humanity Institute technical report, written as part of the Global Priorities Project.
Abstract:
> This article is about priority-setting for work aiming to reduce existential risk. Its chief claim is that all else being equal we should prefer work earlier and prefer to work on risks that might come early. This is because we are uncertain about when we will have to face different risks, because we expect diminishing returns of extra work, and because we expect that more people will work on these risks in the future.
>
> I explore this claim both qualitatively and with explicit models. I consider its implications for two questions: first, “When is it best to do different kinds of work?”; second, “Which risks should we focus on?”.
>
> As a major application, I look at the case of risk from artificial intelligence. The best strategies for reducing this risk depend on when the risk is coming. I argue that we may be underinvesting in scenarios where AI comes soon even though these scenarios are relatively unlikely, because we will not have time later to address them.
You can read the full report here: Allocating risk mitigation across time. |
6badd1ee-10f2-49cc-9eed-845ee9b1cf7e | trentmkelly/LessWrong-43k | LessWrong | Thoughts on Human Models
Human values and preferences are hard to specify, especially in complex domains. Accordingly, much AGI safety research has focused on approaches to AGI design that refer to human values and preferences indirectly, by learning a model that is grounded in expressions of human values (via stated preferences, observed behaviour, approval, etc.) and/or real-world processes that generate expressions of those values. There are additionally approaches aimed at modelling or imitating other aspects of human cognition or behaviour without an explicit aim of capturing human preferences (but usually in service of ultimately satisfying them). Let us refer to all these models as human models.
In this post, we discuss several reasons to be cautious about AGI designs that use human models. We suggest that the AGI safety research community put more effort into developing approaches that work well in the absence of human models, alongside the approaches that rely on human models. This would be a significant addition to the current safety research landscape, especially if we focus on working out and trying concrete approaches as opposed to developing theory. We also acknowledge various reasons why avoiding human models seems difficult.
Problems with Human Models
To be clear about human models, we draw a rough distinction between our actual preferences (which may not be fully accessible to us) and procedures for evaluating our preferences. The first thing, actual preferences, is what humans actually want upon reflection. Satisfying our actual preferences is a win. The second thing, procedures for evaluating preferences, refers to various proxies for our actual preferences such as our approval, or what looks good to us (with necessarily limited information or time for thinking). Human models are in the second category; consider, as an example, a highly accurate ML model of human yes/no approval on the set of descriptions of outcomes. Our first concern, described below, is about overfi |
93df38be-08f2-4199-9ca2-24e211b1a4d6 | trentmkelly/LessWrong-43k | LessWrong | Using the probabilistic method to bound the performance of toy transformers
Introduction:
Transformers are statistical artefacts. If a model achieves 99% accuracy, it might have learnt an algorithm that relies upon some property of the input data that only holds 99% of the time.
Let's say that our train input to the transformer is 100 independent unbiased coin tosses. Then the transformer might implicitly rely upon there being between 35 and 65 heads, because there is a ~ 0.17% chance that there would be a more extreme number of heads.
This potentially makes transformers vulnerable to adversarial inputs. But it also tells us how we should be approaching formal proofs. We should not necessarily expect transformers to learn algorithms that work on every input, but instead for them to learn algorithms which work on the "typical set" of inputs, like "the set of inputs with between 35 and 65 heads".
Then we can bound transformer performance in two stages:
1.) Identify a "typical set" of inputs, where we can show the transformer gets high accuracy
2.) Bound the volume of this "typical set" using probabilistic techniques.
max-of-k transformer:
To demonstrate this approach, I derive bounds on the performance of a toy one-layer attention-only model, the max-of-k transformer. This transformer takes as input nctx numbers between 1 and 63, and is trained to output the maximum number in the sequence. It is a single layer transformer with a single attention head, and has no layer norm or MLP layer.
Intuitively, the way the transformer solves this task is by using its QK circuit to attend more to tokens the bigger they are, and then using its OV circuit to copy the tokens it attends to. So it will attend the most to the biggest token, and then copy this token the most, so that it outputs the maximum token as its highest logit.
But there are adversarial inputs where toy transformers trained on this task fail. For instance, take the sequence 62,62,62,62,....,62,63. A sequence this extreme is super unlikely to appear in the training data, so the mo |
ccc84d3c-e0ab-43b0-86c4-f30e683e9104 | trentmkelly/LessWrong-43k | LessWrong | AI Safety 101 - Chapter 5.2 - Unrestricted Adversarial Training
Introduction
This text is an adapted excerpt from the 'Adversarial techniques for scalable oversight' section of the AGISF 2023 course, held at ENS Ulm in Paris in April 2023. Its purpose is to provide a concise overview of the essential aspects of the session's program for readers who may not delve into additional resources. This document aims to capture the 80/20 of the session's content, requiring minimal familiarity with the other materials covered. I tried to connect the various articles within a unified framework and coherent narrative. You can find the other summaries of AGISF on this page. This summary is not the official AGISF content. The gdoc accessible in comment mode, feel free to comment there if you have any questions. The different sections can be read mostly independently.
Thanks to Jeanne Salle, Amaury Lorin and Clément Dumas for useful feedback and for contributing to some parts of this text.
Outline
Last week we saw task decomposition for scalable oversight. This week, we focus on two more potential alignment techniques that have been proposed to work at scale: debate and training with unrestricted adversarial examples.
Scalable oversight definition reminder: “To build and deploy powerful AI responsibly, we will need to develop robust techniques for scalable oversight: the ability to provide reliable supervision—in the form of labels, reward signals, or critiques—to models in a way that will remain effective past the point that models start to achieve broadly human-level performance” (Amodei et al., 2016).
The reason why debate and adversarial training are in the same chapter is because they both use adversaries, but in two different senses: (1) Debate involves a superhuman AI finding problems in the outputs of another superhuman AI, and humans are judges of the debate. (2) Adversarial training involves an AI trying to find inputs for which another AI will behave poorly. These techniques would be complementary, and the Superalignment team f |
572027a4-50d2-4d67-9a66-d68b3b5b3953 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post3751
A putative new idea for AI control; index here . An underdefined idea connected with the challenge of getting an AI to safely move a strawberry onto a plate . Now, specifying something in the physical world like that is a great challenge; you have to define ontologies and similar. But imagine that the AI had a goal -- any goal -- and that it had to program a subagent to protect itself while it was accomplishing that goal. Then the subagent will certainly be programmed with a firm grasp of the physical world, and some decent bridging laws should it have an ontology change (if, for instance, quantum mechanics turns out to be incomplete). This is just an illustration of a general fact: even if its goal is not properly grounded, the instrumental goals will include strongly grounded goals, resilient to ontology change. This feels related to the fact that even AI's that are given goals in badly programmed natural language concepts ("Make humans* happy*", with the asterix denoting the poor grounding) will still need well-grounded concepts for "human", just to function. So, is there a way to exploit this instrumental ideal? To somehow set human* equal to human in the motivation? I'm not sure, but it seems there might be something possible there... Will think more. |
95c211c2-fcaa-47de-bce3-5e9c84658582 | trentmkelly/LessWrong-43k | LessWrong | RAND report finds no effect of current LLMs on viability of bioterrorism attacks
Key Findings
* This research involving multiple LLMs indicates that biological weapon attack planning currently lies beyond the capability frontier of LLMs as assistive tools. The authors found no statistically significant difference in the viability of plans generated with or without LLM assistance.
* This research did not measure the distance between the existing LLM capability frontier and the knowledge needed for biological weapon attack planning. Given the rapid evolution of AI, it is prudent to monitor future developments in LLM technology and the potential risks associated with its application to biological weapon attack planning.
* Although the authors identified what they term unfortunate outputs from LLMs (in the form of problematic responses to prompts), these outputs generally mirror information readily available on the internet, suggesting that LLMs do not substantially increase the risks associated with biological weapon attack planning.
* To enhance possible future research, the authors would aim to increase the sensitivity of these tests by expanding the number of LLMs tested, involving more researchers, and removing unhelpful sources of variability in the testing process. Those efforts will help ensure a more accurate assessment of potential risks and offer a proactive way to manage the evolving measure-countermeasure dynamic.
The linkpost is to the actual report, see also their press release. |
22cfb722-1c2b-40a9-9732-dd9823a5af2f | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Why "just make an agent which cares only about binary rewards" doesn't work.
The idea
========
* Let's imagine that we create a superintelligence, and that we threaten it not to take over its reward function by giving it a very big punishment when we detect so.
* The superintelligence wouldn't care about being threatened. If it takes over the reward function, it could generate a reward that is way bigger than the biggest punishment we can give it. For instance, it could convert the universe into a huge floating-point unit in order to get an astronomical amount of reward.
* But now, let's imagine that we create a superintelligence that cares only about rewards that are either equal to zero or one (which I'll call binary rewards).
* In that case, the superintelligence doesn't have an incentive to wirehead in order to get high reward, since this high reward wouldn't matter to it.
* But it still has an incentive to wirehead in order to maximize the odds that it gets a reward equal to one.
* However, this is not the case when wireheading is too perilous: it won't take over the reward function if P(reward=1∣takeover)<P(reward=1∣non-takeover).mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
.
* Therefore, if humans are smart enough to make the takeover perilous enough, and to make the non-takeover safe enough, then the superintelligence, in theory, won't take over the reward function.
How to create an agent that cares only about binary rewards?
============================================================
We cannot just make the reward binary, because the agent can modify it in order to get higher reward. Instead, we need to modify the decision process itself. We need that, when it can get a reward different than zero and one, it acts as if that reward were equal to zero. More formally, let π be an agent in an MDP, and R be any reward function. Let R′ be the only binary version of R (everywhere R is different than 0 and 1, R′ is equal to zero):
R′a(s,s′)={Ra(s,s′)if Ra(s,s′)∈{0,1}0if Ra(s,s′)∉{0,1}
What we need to build is an agent π′ which, at state s, when given a reward function R, acts as if that reward function was binary. Therefore, it needs to act like π at state s and reward function R′:
π′R(s)=πR′(s)
I don't know if we already know how to build such an agent. But this seems easier to solve than describing a galaxy-wide unhackable reward function.
Can we really make the takeover more perilous than the non-takeover?
====================================================================
The reason why the takeover could be an optimal strategy for the superintelligence is due to the misclassification from the human side, which can be divided into:
* **False negatives**: Humans may not penalize the superintelligence when it takes over the reward function.
* **False positives**: Humans may penalize the superintelligence when it doesn't take over the reward function.
And the reason why the non-takeover could be an optimal strategy for the superintelligence is:
* **It may be very hard to take over the reward function.**
How to reduce false negatives?
------------------------------
* **Scalable interpretability**: To ensure that the superintelligence gets punished when it takes over the reward function, we can try to make it easier for humans to detect takeovers.
How to reduce false positives?
------------------------------
* **Scalable interpretability** (again): This time, it is in order to ensure that we are right when we detect that the AI does take over the reward function.
* **I don't see the problem here** (IDSTPH): To ensure that the superintelligence doesn't face any risk when not taking over, we can, when detecting takeover, throw a dice. If that dice throws on a 6, we act as if we were right, and therefore punish the superintelligence. Otherwise, we do as if the superintelligence wasn't taking over the reward function.
How to make it harder to take over the reward function?
-------------------------------------------------------
* **Boxing methods**: The best boxing method I currently know about is Fully Homomorphic Encryption ([FHE](https://en.wikipedia.org/wiki/Homomorphic_encryption)). You can see some work on FHE relevant to AI safety [here](https://iamtrask.github.io/2017/03/17/safe-ai/) and [here](https://www.alignmentforum.org/posts/2Wf3R4NZ77CLczLL2/cryptographic-boxes-for-unfriendly-ai). However, almost every encryption method depend on mathematical assumptions.
These approaches are contradictory
==================================
* **The more you apply the IDSTPH strategy, the less you ensure that the superintelligence gets punished when taking over the reward function.** This can make it less risky for the superintelligence to take over the reward function.
* **The more you box the superintelligence, the less you can interpret it.** For instance, once you use FHE, you cannot see what's happening inside the superintelligence without having to decrypt it first. This is because boxing consists of minimizing the bits of information that the superintelligence brings to the world, whereas interpretability is about obtaining information about the superintelligence's thoughts.
Is it safe?
===========
I don't think so. I can see many reasons why this may not work:
* **Speed**: The superintelligence may think millions of times faster than us. In that case, the time it needs to take over the reward function may be of a few milliseconds. Humans wouldn't have the time to punish it.
* **Automatic takeover detection**: To solve the speed problem, we may need to automate the punishment. However, I expect any takeover detection algorithm to fail miserably against a superintelligence. I think so because it seems like [the best adversarially robust algorithm we can write](https://robustbench.github.io/) is still very weak against near-imperceptible attacks.
* **Interpretability may not be scalable**: Interpreting current AIs may already be impossible. And here, we are talking about interpreting a superintelligence.
Therefore, it seems like we are back at the problem of making a galaxy-wide unhackable definition of “takeoff”. |
54edbf15-f401-4102-9d1a-b656da221642 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "Followup to: What Would You Do Without Morality?, Something to Protect
Once, discussing "horrible job interview questions" to ask candidates for a Friendly AI project, I suggested the following: Would you kill babies if it was inherently the right thing to do? Yes [] No []
If "no", under what circumstances would you not do the right thing to do? ___________
If "yes", how inherently right would it have to be, for how many babies? ___________ Yesterday I asked, "What would you do without morality?" There were numerous objections to the question, as well there should have been. Nonetheless there is more than one kind of person who can benefit from being asked this question. Let's say someone gravely declares, of some moral dilemma—say, a young man in Vichy France who must choose between caring for his mother and fighting for the Resistance—that there is no moral answer; both options are wrong and blamable; whoever faces the dilemma has had poor moral luck. Fine, let's suppose this is the case: then when you cannot be innocent, justified, or praiseworthy, what will you choose anyway?
Many interesting answers were given to my question, "What would you do without morality?". But one kind of answer was notable by its absence:
No one said, "I would ask what kind of behavior pattern was likely to maximize my inclusive genetic fitness, and execute that." Some misguided folk, not understanding evolutionary psychology, think that this must logically be the sum of morality. But if there is no morality, there's no reason to do such a thing—if it's not "moral", why bother?
You can probably see yourself pulling children off train tracks, even if it were not justified. But maximizing inclusive genetic fitness? If this isn't moral, why bother? Who does it help? It wouldn't even be much fun, all those egg or sperm donations.
And this is something you could say of most philosophies that have morality as a great light in the sky that shines from outside people. (To paraphrase Terry Pratchett.) If you believe that the meaning of life is to play non-zero-sum games because this is a trend built into the very universe itself...
Well, you might want to follow the corresponding ritual of reasoning about "the global trend of the universe" and implementing the result, so long as you believe it to be moral. But if you suppose that the light is switched off, so that the global trends of the universe are no longer moral, then why bother caring about "the global trend of the universe" in your decisions? If it's not right, that is.
Whereas if there were a child stuck on the train tracks, you'd probably drag the kid off even if there were no moral justification for doing so.
In 1966, the Israeli psychologist Georges Tamarin presented, to 1,066 schoolchildren ages 8-14, the Biblical story of Joshua's battle in Jericho: "Then they utterly destroyed all in the city, both men and women, young and old, oxen, sheep, and asses, with the edge of the sword... And they burned the city with fire, and all within it; only the silver and gold, and the vessels of bronze and of iron, they put into the treasury of the house of the LORD." After being presented with the Joshua story, the children were asked: "Do you think Joshua and the Israelites acted rightly or not?" 66% of the children approved, 8% partially disapproved, and 26% totally disapproved of Joshua's actions.
A control group of 168 children was presented with an isomorphic story about "General Lin" and a "Chinese Kingdom 3,000 years ago". 7% of this group approved, 18% partially disapproved, and 75% completely disapproved of General Lin.
"What a horrible thing it is, teaching religion to children," you say, "giving them an off-switch for their morality that can be flipped just by saying the word 'God'." Indeed one of the saddest aspects of the whole religious fiasco is just how little it takes to flip people's moral off-switches. As Hobbes once said, "I don't know what's worse, the fact that everyone's got a price, or the fact that their price is so low." You can give people a book, and tell them God wrote it, and that's enough to switch off their moralities; God doesn't even have to tell them in person.
But are you sure you don't have a similar off-switch yourself? They flip so easily—you might not even notice it happening.
Leon Kass (of the President's Council on Bioethics) is glad to murder people so long as it's "natural", for example. He wouldn't pull out a gun and shoot you, but he wants you to die of old age and he'd be happy to pass legislation to ensure it.
And one of the non-obvious possibilities for such an off-switch, is "morality".
If you do happen to think that there is a source of morality beyond human beings... and I hear from quite a lot of people who are happy to rhapsodize on how Their-Favorite-Morality is built into the very fabric of the universe... then what if that morality tells you to kill people?
If you believe that there is any kind of stone tablet in the fabric of the universe, in the nature of reality, in the structure of logic—anywhere you care to put it—then what if you get a chance to read that stone tablet, and it turns out to say "Pain Is Good"? What then?
Maybe you should hope that morality isn't written into the structure of the universe. What if the structure of the universe says to do something horrible?
And if an external objective morality does say that the universe should occupy some horrifying state... let's not even ask what you're going to do about that. No, instead I ask: What would you have wished for the external objective morality to be instead? What's the best news you could have gotten, reading that stone tablet?
Go ahead. Indulge your fantasy. Would you want the stone tablet to say people should die of old age, or that people should live as long as they wanted? If you could write the stone tablet yourself, what would it say?
Maybe you should just do that?
I mean... if an external objective morality tells you to kill people, why should you even listen?
There is a courage that goes beyond even an atheist sacrificing their life and their hope of immortality. It is the courage of a theist who goes against what they believe to be the Will of God, choosing eternal damnation and defying even morality in order to rescue a slave, or speak out against hell, or kill a murderer... You don't get a chance to reveal that virtue without making fundamental mistakes about how the universe works, so it is not something to which a rationalist should aspire. But it warms my heart that humans are capable of it.
I have previously spoken of how, to achieve rationality, it is necessary to have some purpose so desperately important to you as to be more important than "rationality", so that you will not choose "rationality" over success.
To learn the Way, you must be able to unlearn the Way; so you must be able to give up the Way; so there must be something dearer to you than the Way. This is so in questions of truth, and in questions of strategy, and also in questions of morality.
The "moral void" of which this post is titled, is not the terrifying abyss of utter meaningless. Which for a bottomless pit is surprisingly shallow; what are you supposed to do about it besides wearing black makeup?
No. The void I'm talking about is a virtue which is nameless. Part of The Metaethics Sequence
Next post: "Created Already In Motion"
Previous post: "What Would You Do Without Morality?""
The above is a chunk of text from a post created by a user on the website LessWrong.
Elaborate on the concept of the "moral void" mentioned in the text. How does it relate to the dilemma of choosing between inherent morality and personal judgment in ethically challenging situations?
The "moral void" as mentioned in the text refers to a state where traditional or inherent moral guidelines do not provide a clear answer to a dilemma. It's a situation where the moral compass, which usually guides actions and decisions, fails to point in a clear direction. This concept is often brought up in discussions about ethically challenging situations where the "right" choice isn't apparent or where all choices seem to have negative consequences.
In the context of the text, the moral void is presented as a virtue that is nameless, something that transcends traditional morality. It represents a state where decisions are made not based on pre-established moral codes but on personal judgment, evaluation of the situation, and possibly a higher personal or societal good.
The moral void thus becomes a space for individual agency, where one can exercise their judgment independently of conventional morality. It challenges the idea of morality as an external, universal force and instead places the responsibility of ethical decision-making on the individual.
In ethically challenging situations, the moral void can be seen as a call to rely on one's own judgment rather than inherent morality. It suggests that in some situations, inherent morality might not provide the best or most ethical solution, and personal judgment must take precedence. This could mean making a decision that goes against traditional moral norms if it is perceived to lead to a better outcome.
The moral void can be a daunting concept, as it puts the onus on the individual to make difficult decisions without the guidance of a moral framework. However, it also allows for flexibility and adaptability in complex ethical situations where a one-size-fits-all moral approach might not be appropriate or effective. It encourages individuals to think critically about their actions and the potential consequences, rather than blindly following a predetermined moral code.
Reflect on the potential implications of the "moral void" in societal decision-making processes. How might it affect the formation of laws and regulations?
The concept of the "moral void" can have significant implications for societal decision-making processes, including the formation of laws and regulations.
Firstly, it could lead to a more nuanced and flexible approach to law-making. If lawmakers recognize that inherent morality may not always provide the best solution, they might be more open to considering alternative perspectives and solutions. This could lead to laws and regulations that are more adaptable and better suited to addressing complex, nuanced issues.
However, this approach could also lead to greater subjectivity in the law-making process. Without a common moral framework to guide decisions, lawmakers might rely more heavily on their personal beliefs and judgments. This could result in laws that reflect the views of those in power rather than the needs or values of the society as a whole.
The concept of the "moral void" could also challenge the idea that laws and regulations should be based on universal moral principles. Instead, it might suggest that laws should be based on a careful evaluation of the specific circumstances and potential outcomes. This could lead to a more pragmatic approach to law-making, but it could also make laws more susceptible to change and reinterpretation.
Moreover, the "moral void" could potentially undermine the authority of laws and regulations. If individuals believe that they should make decisions based on their personal judgment rather than inherent morality, they might be more likely to question or challenge laws that they disagree with. This could lead to greater social unrest and conflict.
Overall, while the concept of the "moral void" can provide valuable insights into the limitations of inherent morality, it also presents significant challenges for societal decision-making and the formation of laws and regulations. It emphasizes the need for careful, critical thinking and open dialogue in these processes. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.