id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
69d61e73-fb9d-4e2f-aa61-35201291b2d5 | trentmkelly/LessWrong-43k | LessWrong | Grant applications and grand narratives
The Lightspeed application asks: “What impact will [your project] have on the world? What is your project’s goal, how will you know if you’ve achieved it, and what is the path to impact?”
LTFF uses an identical question, and SFF puts it even more strongly (“What is your organization’s plan for improving humanity’s long term prospects for survival and flourishing?”).
I’ve applied to all three grants of these at various points, and I’ve never liked this question. It feels like it wants a grand narrative of an amazing, systemic project that will measurably move the needle on x-risk. But I’m typically applying for narrowly defined projects, like “Give nutrition tests to EA vegans and see if there’s a problem”. I think this was a good project. I think this project is substantially more likely to pay off than underspecified alignment strategy research, and arguably has as good a long tail. But when I look at “What impact will [my project] have on the world?” the project feels small and sad. I feel an urge to make things up, and express far more certainty for far more impact than I believe. Then I want to quit, because lying is bad but listing my true beliefs feels untenable.
I’ve gotten better at this over time, but I know other people with similar feelings, and I suspect it’s a widespread issue (I encourage you to share your experience in the comments so we can start figuring that out).
I should note that the pressure for grand narratives has good points; funders are in fact looking for VC-style megabits. I think that narrow projects are underappreciated, but for purposes of this post that’s beside the point: I think many grantmakers are undercutting their own preferred outcomes by using questions that implicitly push for a grand narrative. I think they should probably change the form, but I also think we applicants can partially solve the problem by changing how we interact with the current forms.
My goal here is to outline the problem, gesture at some possible |
0ca2ba56-42c1-42eb-9e29-ec0a4e251134 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1989
You can see the actual submission (including a more formalized model) here , and the contest details here . I've reordered things to be more natural as a blog post / explain the rationale / intuition a bit better. This didn't get a prize, tho it may have been because I didn't try to fit the ELK format. The situation: we have the capacity to train an AI to predict physical reality quite well. We'd like to train an AI (the "SmartVault") that uses that understanding of reality to protect a valuable diamond inside the vault. Ultimately, we'd like the system to protect the diamond and not deceive the owner of the diamond; proximally, we'd like to make a Reporter that can figure out, from the latent knowledge of the SmartVault, whether it's protecting the diamond or deceiving the owner. I roughly think the proximal goal is doomed , and that the hope needs to be prospective instead of retrospective. [That is, we need to have a system that we can see ahead of time will try to protect the diamond instead of us checking after the fact for deception, because that can just move the meta-level on which deception occurs .] So I'm going to try to propose a solution that de-emphasizes the Reporter, and see how far I get. Setup We can roughly model SmartVault as being composed of some subsystems: the Predictor (which, given a trajectory of actions, estimates their likely result), a Reporter (which, given access to the predictor, answers specific questions about the result), and an Optimizer (which, given a situation, tries to decide which trajectory of actions to take). Obviously, the Predictor and the Optimizer are going to be intimately connected to each other, and the ontology of the Predictor is going to be jointly optimized by reality and the training feedback (i.e. the structure of the model's knowledge will reflect both the underlying reality and the salience landscape of the humans labelling examples as 'good' and 'bad'). I'm going to assume that 'the human value function' is part of the 'latent knowledge relevant to prediction (of training reward)', even if it's implicit instead of explicit. While we’d like to train our machine on (reward, state of reality) pairs, unfortunately we only have access to (reward, human observation) pairs for training; that is, we only know how to communicate the lower-level salience landscape. This means a core challenge is ontological identification, as each ‘human observation’ state could mean many possible ‘machine observation’ states. Furthermore, let’s assume the machine ontology is more detailed and relevant (as it’s the one driving actions, including actions that might corrupt sensors). There are two main classes of reporters: the 'human simulator' (which answers the question: "what would the human think if they saw their observations?") and the 'direct translator' (which answers the question: "what would a hypothetical correct human think if they saw my observations?"). Similarly, I think there are two main classes of inferred human value functions, which seem pretty analogous: the 'finger' (which answers the question: "did the human observe things that they approved of?"), and the 'moon' (which answers the question: "did the thing the human approved of happen?", or "would a hypothetical correct human approve of the thing that I infer occurred?"). I am tempted to just call them the 'human simulator' and the 'direct translator', but it's probably better to reserve those terms for reporters and introduce new terms for talking about the inferred value functions, even tho I think they use basically the same mechanisms. [One reporter you could imagine, for example, doesn't answer arbitrary questions but just implements the human value function in the machine ontology and provides the level of approval of what the machine thinks is going on; this should counter the incentives favoring the more obvious sorts of deceiving-the-human.] One way to imagine the 'alignment is easy' world is that those two happen to be the same thing; the only way to build the 'finger' is by building the 'moon'. In this case, it doesn't really matter what approach we take here because vanilla training will do the right thing, and so we'll ignore this as trivial. [This also covers cases in which it is easier to 'just solve the problem' than 'figure out how to trick the judge', which again seems too trivial. At some point we'll want to do something challenging enough that solutions that fool the judge appear much sooner than solutions that 'actually solve it'.] More realistically, a persistent problem in alignment is that the 'finger' points to the 'wrong thing', but also may be the only way to get perfect training loss (especially as you gain the ability to look closer and closer at the world). For any moon / direct translator the machine could imagine and test, it gets evidence that actually, humans want the finger / human simulator instead. We'll ignore this problem as 'user error' to be solved elsewhere so that we can focus on this subproblem. So in the relevant difficult cases, there's distinct finger and moon value functions, both of which can achieve optimal training loss. Approach I think this challenge has to be sidestepped instead of attacked directly. After all, if we could implement the human value function in the machine ontology, we would just do that! Similarly, we had to train the predictor and infer the value function because we didn't know how they were supposed to work, and so a plan that requires being able to differentiate subtle nuances between them (instead of having the machine do that differentiation for us) is probably doomed. I think the right way to sidestep it is using multi-value functions, inspired by ideas in the reinforcement learning literature like double Q-learning. This requires a change to the internals of the optimizer, where rather than just pursuing reward in an arbitrary way it needs to track ‘plausible’ human value functions, and then evaluate plans based on the ‘minimum’ score given by those plausible functions. [Both 'plausible' and 'minimum' are hiding significant obstacles and detail!] If both the finger (reward function implied by the human simulator) and the moon (reward function implied by the direct translator) seem plausible to the machine, then plans which involve deceiving the human (and getting away with it) will score poorly based on the moon, whereas plans which involve satisfying the human will score highly based on both options. Thus we can get the benefits of targeting the moon without having to identify it. [1] This is highly related to a previous work in the reinforcement learning literature; see Double Q-learning , Double DQN and work adapting it to policy gradients . There, they use two independently initialized and trained Q-functions / critics to attempt to overcome optimism bias, leading to faster training. Here, we assume the machine is ‘already trained’ and so optimism borne of ignorance is irrelevant, but optimism in how to interpret the freedom provided within the constraints of the human training data (or human perception) is relevant. Challenges While I think this is promising, it has a few pieces that need to be fleshed out (and even then, I think it's ignoring several critical problems, such that this can't be a 'full solution'). Plausible In order for this to work, the machine needs some constructive way to think about "all reward functions worth thinking about" in a way that it can optimize over. While easily solvable with a hypercomputer, we'd like to have something that's competitive. For the existing multi-value functions, the 'population of reward functions' is just generated from different initializations, which gives (essentially) the different models random settings for the values that haven't yet been pinned down. This won't work very well for generating a distribution of diverse models or tracking their relative probability (except by counting samples). My guess is that it's possible to construct some sort of particle filtering approach which values diversity, and can reason about families of functions rather than specific instances, which makes it possible to (with achievably few instead of exponentially many samples) cover the space well enough. Actually delivering a working model of this feels like a big ask that's nevertheless doable, and likely enhanced by various work in Bayesian sampling / the stuff that came out of the probabilistic programming paradigms. [2] Another significant challenge here is that the population of plausible human utility functions that the machine is optimizing over exists just in the machine’s model of the world, and it’s punished for discovering any new ones (as they can’t increase the achievable utility, but can decrease it) and rewarded for ruling out old ones (as this can’t decrease the achievable utility, but can increase it). This might push the incentives to deceive the human into incentives to self-deceive in a way that causes problems, and extend the incentives to manipulate the human in deployment to training as well. The AI can ask questions which it realizes will rule out difficult-to-satisfy candidate reward functions (but hopes the human won’t realize rules out those functions), and is rewarded for having interpretations of ambiguous situations which do so. Or, more transparently, it may just argue with the human that they should be satisfied by consequences which are easy for the AI to achieve. [3] Minimum Supposing we somehow have a set of candidate functions, each with some weight, there remains the question of how to score a particular plan. Each candidate function will provide some score, but if you take a raw minimum of all candidates then it's very easy to end up with a function that's the minimum score everywhere, and thus provides no guidance. [For example, for any variable whose target value is not pinned down by the training data, one candidate might think it's critical that it be high and another candidate might think it's critical that it be low, and you can't satisfy both simultaneously.] I think this can be repaired using a 'softmin', but it's not obvious what functional form works well here. I discuss one such proposal in the linked doc (roughly: sort your candidates in decreasing probability order, create a non-increasing function that at each particle is the minimum of it's previous value and the particle's value, and integrate the area under the curve). This overweights pessimistic candidates, but doesn't obviously have the nice features that we might want it to have (like weakly preferring 'typical' or 'non-extremal' values for features that aren't pinned down). Moon It seems worth repeating that this is assuming that the function that you want the machine to optimize for is a plausible candidate that it could infer from the training data. Then we don't need to find it; we just need to not do poorly according to any plausible function, and we're done. This feels like it has two problems. First, the 'real' problem is that the moon doesn't do well according to the training data, because of systematic biases on the human's part, and the machine learns those biases (while discarding random errors). In some situations this is fine, and in other situations this causes huge losses of value, and this relaxation doesn't seem like it cuts at the heart of this issue (tho maybe it does open up a path for progress); instead it feels like I came up with a solution that relies on the relaxation assumption and will break as soon as that assumption is removed. Second, suppose the space of candidates is large enough to support both the finger and the moon; does it support anything else? Will those things cause the machine to behave in bizarre or undesired ways? [Note the interaction with the previous point; as we start relaxing the assumptions of 'you have to agree with the training data' in order to allow the machine to not absorb our systematic biases, we also open up the space for ignoring lots of the true signal that we don't want the system to ignore.] ^ We also get the benefits of targeting the finger: the plan will not just work, but persuasively work. But this benefit is probably poisoned; there doesn’t seem to be any meta-level incentives here favoring not lying to the human; a plan that both hacks the camera so that it always shows an image of the diamond and attempts to defend the vault will score more highly, I think, than one that just attempts to defend the vault and not pick up ‘free points’ by hacking the camera. ^ One of my intuitions here is that if it's reasonable to ask the machine to infer the reward function once, it's probably reasonable to ask the machine to infer the reward function a hundred times. The main question feels to me like whether the moon shows up in your first hundred samples, or what sort of changes you have to make to your sampling procedure such that it'll happen. ^ “Please cure cancer,” you tell the robot, and it responds with “I’ve written a very compelling pamphlet on coping with your mortality.” |
f42e7885-9452-43a6-9333-8ac78a8abcf7 | StampyAI/alignment-research-dataset/special_docs | Other | Shared Autonomy for Robotic Manipulation with Language Corrections.
Shared Autonomy for Robotic Manipulation with Language Corrections
Siddharth Karamcheti∗Raj Palleti∗Yuchen Cui Percy Liang Dorsa Sadigh
Department of Computer Science, Stanford University
{skaramcheti, rpalleti, yuchenc, pliang, dorsa}@stanford.edu
Abstract
Traditional end-to-end instruction following ap-
proaches for robotic manipulation are notori-
ously sample inefficient and lack adaptivity ;
for most single-turn methods, there is no way
to provide additional language supervision to
adapt robot behavior online – a property criti-
cal to deploying robots in collaborative, safety-
critical environments. In this work, we present
a method for incorporating language correc-
tions, built on the insight that an initial instruc-
tion and subsequent corrections differ mainly
in the amount of grounded context needed. To
focus on manipulation domains where the sam-
ple efficiency of existing work is prohibitive,
we incorporate our method into a shared auton-
omy system. Shared autonomy splits agency
between the human and robot; rather than spec-
ifying a goal the robot needs to achieve alone,
language informs the control space provided to
the human. Splitting agency this way allows
the robot to learn the coarse, high-level parts
of a task, offloading more involved decisions –
such as when to execute a grasp, or if a grasp is
solid – to humans. Our user study on a Franka
Emika Panda arm shows that our correction-
aware system is sample-efficient and obtains
significant gains over non-adaptive baselines.
1 Introduction
Research at the intersection of natural language
and robotics has focused on dyadic interactions be-
tween humans and robots, often in the single-turn
instruction following regime (Tellex et al., 2011;
Artzi and Zettlemoyer, 2013; Thomason et al.,
2015; Arumugam et al., 2017). In this paradigm,
a human gives an instruction, and the robot exe-
cutes behavior in the world, autonomously – simul-
taneously resolving the human’s goal as well as
planning a course of actions to execute in the envi-
ronment. While impactful, building systems with
this explicit division of agency between humans
∗denotes equal contribution
“Grab the thing on the white table, and place it in the basket”
“No, on the left!” Shared Autonomy Regime
Incorrectly interprets original instruction21
3Figure 1: Our proposed system. Whereas prior work
only allows for issuing a single language utterance held
constant during execution (solid line), our approach
allows users to provide language corrections during
execution (left window – dashed line).
and robots is nontrivial; many existing systems
either make strong assumptions about the environ-
ment in order to use motion planners (Matuszek
et al., 2012; Kollar et al., 2013), or require extreme
amounts of language-aligned data to learn general
policies (Chevalier-Boisvert et al., 2019; Stepputtis
et al., 2020; Lynch and Sermanet, 2020).
Coupled with the severe sample inefficiency of
existing approaches is their lack of adaptivity . Con-
sider the robot in Fig. 1, trying to execute “grab
the thing on the white table and place it in the
basket.” This instruction is ambiguous and as a
result, it is not clear what should happen. One
natural option is for the human to provide a set
of streaming corrections to the robot, changing its
behavior on the fly. While recent work tries to
get at the spirit of this idea by learning from dia-
logue (Thomason et al., 2019a,b), post-hoc correc-
tions (Co-Reyes et al., 2019), or implicit feedback
(Karamcheti et al., 2020), none of these approaches
=
[Train-Only] 6-DoF Action
“Pour the blue cup into the cereal bowl”State Encoder
Distil-RoBERTaTraining UtterancesDistil-RoBERTa“Put the banana away” “Towards the cup”“Empty the blue cup’s contents into the bowl”Similarity SearchInstruction & Correction “Gating” Function
hstateProprioceptive StateObject Poses+=hlanguageα1××ProjectionBasis 1 — (6-DoF)Basis 2 — (6-DoF)z1z2⋅Inference Human controls!
Action Encoder
Predicted Robot Action…Basis n — (6-DoF)…znLanguageActionState“Decoder” (Dot-Product)Figure 2: Our proposed LILAC model. Central to LILAC is the “gating” module (orange) which controls the
amount of state-context for a given language input, allowing us to handle corrections. We use GPT-3 in lieu of a
heuristic to provide α(see Appendix A for discussion), though we plan to learn αfrom user feedback in the future.
Solid lines represent the inference pipeline, while dashed lines indicate training-only steps.
work online . Real systems for language-driven
human-robot interaction must be able to handle
streaming corrections in a manner that is both nat-
ural andsample efficient .
One answer to sample efficiency lies in leverag-
ing existing methods in shared autonomy (Dragan
and Srinivasa, 2013; Argall, 2018; Javdani et al.,
2018). This class of approaches splits agency be-
tween the human and robot; during execution, both
parties influence the ultimate actions of the robot,
sharing the burden of reasoning over actions. By
factoring the difficulty of the problem across the
human and the robot, shared autonomy approaches
see large gains in sample efficiency; the robot can
learn coarse, high-level features, off-loading the
short, fine-grained manipulation to the human, thus
playing off the strenghts of both parties. The con-
crete instance of shared autonomy that we focus
on in this work is learned latent actions for assis-
tive teleoperation (Herlant et al., 2016; Losey et al.,
2021), where we learn low-dimensional control
spaces that humans can use via joysticks, sip-and-
puff devices, or other assistive tools to maneuver
high-dimensional robots. Because humans are ac-
tively involved in controlling the robot – especially
in critical states, such as when aligning the robot
gripper to lift a cup – these approaches are able
to operate at the scale of 5 - 10 examples per task.
This is in contrast to the 10K - 100K demonstra-
tions required by modern instruction following sys-
tems learned via imitation learning, in the fully
autonomous setting (Luketina et al., 2019).
The learned latent actions paradigm operational-
izes this division of agency by formulating a set
of approaches that use small datasets of demon-
strations to learn task-specific assistive controllers(Losey et al., 2020; Jeon et al., 2020; Karamcheti
et al., 2021b; Li et al., 2020). While powerful and
sample-efficient, a key failing of these approaches
is their inability to handle multiple objectives – and
more importantly – provide a natural interface for
users to specify their goals. To address this chal-
lenge, Karamcheti et al. (2021a) introduce LILA
(Language-Informed Latent Actions), by using lan-
guage in a manner similar to single-turn instruction-
following approaches. However, while accumu-
lating gains in sample efficiency, LILA, like most
single-turn systems, lacks adaptivity, a critical com-
ponent for any real-time, user controlled system.
Looking to Fig. 1, we see that LILA misinterprets
the ambiguous instruction “grab the thing on the
white table, and place it in the basket” (1), moving
towards the cup (2), rather than the banana as the
user intended (3). Problems of ambiguity, misspec-
ification, or underspecification are pervasive in any
real-world, user-facing language system, and as
such, we need an approach for handling streaming,
online corrections – in this case, simple utterances
like “stop,” or “no, on the left!”
In this work, we introduce LILAC (Language-
Informed Latent Actions with Corrections), an
adaptive system for real-world robotic manipula-
tion, that – unlike LILA – can effectively interpret
streaming language corrections. Critically, our key
insight is in realizing language utterances vary in
the amount of object state-dependence they require.
Aninstruction like that in Fig. 1 – “grab the thing
on the white table, and place it in the basket” – re-
quires dense grounded information about object
positioning while a correction – “no, on the left!” –
can be expressed as a function of the user’s (static)
reference frame, without proprioceptive state or ob-
ject information. This insight allows us to decouple
interpreting corrections from grounding: we can
learn corrections from even fewer examples , boost-
ing the robustness andgeneralization potential of
our approach.
The following sections introduce LILAC, includ-
ing how we identify the “state-dependence” of an
utterance. We layout our experiments, and culmi-
nate with the results of a user study (small-scale,
n= 5) conducted on a physical Franka Emika
Panda (a 7-DoF fixed arm manipulator), with a dis-
cussion of future work on extending LILAC, and
natural language supervision for shared autonomy.
2 Motivating Example
To gain a picture of LILAC’s latent actions pipeline,
we present a motivating example in Fig. 1. A
user first gives an instruction to “grab the thing
on the white table and place it in the basket,” which
presents them a 2-DoF control space they can use
via the joystick in their hands. With this new con-
trol space, pushing upon the joystick may bring
the arm closer to the table, while right might twist
the gripper to align with the objects on the table.
Unfortunately, the initial model’s control space pre-
diction is not perfect, and the arm skews towards
the wrong object!
With LILAC, the user is able to issue real-time
language corrections, updating their control space .
Specifically, the new user utterance is fed to our
learned model that parses the utterance, and pro-
duces a new control mapping reflecting the user’s
intent. Now, pushing left on the joystick brings the
arm left, allowing the user to grasp the banana, as
they intended. Finally, after the correction has been
satisfied, the user is able to denote termination with
a button on their controller, dropping back into the
control space for the original task they provided
– in this case, returning to the control space for
placing the object in the basket.
3LILAC: Natural Language Corrections
LILAC builds off of LILA as introduced by Karam-
cheti et al. (2021a) by adding a gating module to
handle streaming corrections. The architecture is
depicted in Fig. 2; solid lines denote the inference
logic, while dashed lines denote training logic.
Overview. LILAC is a conditional autoencoder
with some extra structural elements. The encoder
takes in (language, state, action) triples – the greenboxed components in Fig. 2 – and factors the model-
ing of the control space into two subproblems. The
first subproblem is identifying a set of basis vectors
b1. . . b nfor low-dimensional control conditioned
on the current state and language utterance, where
nis the dimensionality of the latent space. These
basis vectors have the same dimensionality as the
robot’s native action space (e.g., 6/7-DoF). The sec-
ond subproblem is finding a set of scalar weights
(z1. . . z n) of the recovered basis vectors, optimized
such that the convex combinationPn
i=1zi·bire-
constructs the original action.
At train time, we assume a dataset of training ut-
terances (either instructions orcorrections), paired
with robot trajectories comprised of (state, action)
pairs. For this work, we assume the action space
is the 6-DoF end-effector velocity of the robot (ob-
tained via forward kinematics), and the state space
is the combination of the robot’s proprioceptive
state, containing information about its joint states
and end-effector pose (also in 6D), as well as the
coordinates of each object in the scene. Because
we are interested in intuitive, low-dimensional con-
trol, we set n= 2, so that we only produce 2 basis
vectors and weights; this way, a human can operate
the robot using any 2-DoF interface, like a joystick
or computer mouse.
Exactly as in LILA, we use a pretrained Distil-
Roberta model from Sentence-BERT (Reimers and
Gurevych, 2019) to encode language utterances,
in tandem with an “unnatural-language processing”
nearest neighbors index (Marzoev et al., 2020); be-
cause we are in the small data regime (2 hours
of demonstrations), projecting inference-time ut-
terances onto existing training exemplars prevents
the LILAC model from generalizing poorly as lan-
guage embeddings drift, which could lead to prac-
tical issues of user safety.
Gating Instructions vs. Corrections. The key
insight of this work is that language instructions
and corrections differ in their amount of object
state-dependence – but what does this mean? From
a linguistic perspective, one might categorize dif-
ferent utterances based on the number of referents
present; an utterance like “grab the thing on the
white table and place it in the basket” as in Fig. 1
has 3 referents, indicating a large degree of state
dependence; the robot must ground the utterance
in the objects of the environment to resolve the cor-
rect behavior. However, an utterance like “no, to
the left!” has no explicit referents; one can resolve
End-Effector ControlCorrection-Aware ImitationLILA (No Corrections)LILAC (Ours)Goal: “Place the fruit basket on the tray”
✅
✅
❌
❌
“Go forward”“Lower it”Figure 3: Qualitative trajectories from 4 different control strategies, operated by an “oracle.” Note the dashed lines
indicate when language corrections were used. In general, both LILAC and the End-Effector control methods are
able to solve tasks, with LILAC able to do so more efficiently. Imitation Learning (green) fails to fit the tasks – even
with corrections – and LILA is unable to progress without finer-grained correction information.
the utterance without object or proprioceptive state
information, instead relying solely on the user’s
static reference frame and induced coordinates.
To operationalize this idea whilst remaining sam-
ple efficient1, we use a gating function (orange, in
Fig. 2) that given language, predicts a value αfrom
0–1. A value of 0 signifies a correction . Appropri-
ately, in our architecture, this zeroes out any state-
dependent information (see the αterm in Fig. 2),
and predicts an action solely based on the provided
language. In this work, we construct a prompt har-
ness with GPT-3 (Brown et al., 2020) to output
α= 0,0.5,1– the prompt text and motivation for
this decision can be found in Appendix A.
4 Experiments & User Study
Our experiments consist of a targeted evaluation
with an expert “oracle” of LILAC and 3 different
baselines (results in Fig. 3), as well as a real-world
user study ( n= 5) on a complex manipulation
workspace with a Franka Emika Panda arm (results
in Fig. 4). We reuse all the publicly released data
from Karamcheti et al. (2021a), and for correction
data, collect a handful of corrections for moving in
the 6 cardinal directions, as well as two mixed cor-
rections that require some level of state grounding
(“lower the bowl” and “tilt the cup”).
Targeted Expert Evaluation. The goal of our
expert-controlled evaluation was to evaluate the
corrections-module in LILAC, and get a point of
reference to LILA and traditional methods for robot
control – specifically, using a control scheme that
1One question is why not treat all utterances as equal; the
answer is rooted in the small data regime we operate in. We’d
need to collect several instances of the same correction “to the
left” in different states to generalize, whereas with the LILAC
approach, we realistically only need one.uses inverse kinematics to control the end-effector
of the robot, two axes at a given time (e.g. [X,
Y], [Z, Roll], [Pitch, Yaw] ). Fig. 3 shows
the results; LILAC and End-Effector control are
both able to accomplish the two tasks, but LILA
struggles to recover from overshooting problems.
We also evaluate one other strategy – Language &
Correction-Aware Imitation, which doesn’t have
enough data to fit a reasonable policy (see discus-
sion in Karamcheti et al. (2021a) for more infor-
mation). We additionally evaluate a “no-language”
variant of latent actions in Appendix C.
User Study. Given the results from the automatic
evaluation, we run a within-subjects user study
(n= 5 , 3 male, 2 female, 3 users with prior
teleoperation experience), with the three operation
schemes – LILAC, LILA, and End-Effector control
– as the three conditions. Each user was randomly
assigned one of the five original tasks from the
LILA work, and asked to complete it with each
control strategy.2Fig. 4 shows the general quan-
titative (left) and subjective (right) study results.
Users provide linguistic feedback verbally; in this
work, we rely on an expert proctor to manually
type the verbal instruction into the computer run-
ning the LILAC model – future work will adopt
off-the-shelf ASR technologies, such as the Google
Speech-to-Text API.
On the Strength of End-Effector Control.
LILAC strictly dominates LILA, showing the ben-
efit of adding even simple real-time correction han-
dling. However, compared to End-Effector con-
trol, the success is less clear. While obtaining a
slightly higher success rate, the qualitative results
2More information about the structure of the study can be
found in Appendix B
Figure 4: User study results ( n= 5) with a Franka Emika Robotic arm. LILAC outperforms the non-adaptive LILA
model across the board, obtaining higher success rates, and is preferred by users. While LILAC and End-Effector
control obtain similar success rates (though LILAC is slightly higher), plotting the “jerk” (2nd derivative of input
velocity) paints a different picture: controlling the end-effector is very jerky, requiring high user load.
are mixed. One confound is the limited pool of
participants, many of whom already had prior tele-
operation experience (due to COVID restrictions,
we could not widen the pool). Another possible
confound is the structure of the user study itself; to
better allow for users to adjust to the LILAC cor-
rections procedure, the amount of practice time
afforded each user is larger than in prior work,
which allows users to get more acquainted with end-
effector control. Ultimately, this points to the tasks
in this work being on the simpler side, able to be
solved (mostly) without complex, mixed angular-
linear control of the robot’s end-effector; if we
were to account for more real-world manipulation
tasks such as sweeping, wiping, or feeding – all
of which require handling contact and controlling
3+ degrees-of-freedom – we would see degraded
End-Effector performance.
That being said, for additional insight on user
cognitive load when using the various control
schemes, we plotted the 2nd-derivative of accel-
eration – jerk – of the input controls; we see here
that End-Effector approaches require significantly
more fast movement compared to LILAC. Not only
is this more taxing on the user, but is potentially
unsafe depending on the application – another axis
we will explore in future work.
5 Discussion
The union of natural language and shared auton-
omy for real-world robot manipulation is a rich and
vibrant research area. Moving beyond strict dyadic
interactions towards the shared autonomy setting
opens the door to rich work in language supervi-
sion for robotic manipulation – work that has so
far been limited by the steep data requirements of
training language-conditioned policies (Stepputtis
et al., 2020; Shridhar et al., 2021). As shown in
this work, shared autonomy approaches are able tobenefit greatly from sharing agency with a human-
in-the-loop, leading to gains in data efficiency.
Specifically, in this work we introduced LILAC,
an adaptive shared autonomy system for handing
streaming language corrections, provided while a
user completes a task . The entire LILAC model
was trained with 2 hours of data collected by a sin-
gle person, vs. the multiple days of data that would
be required if using a fully autonomous, imitation
learning approach. While the sample efficiency
wins are clear, LILAC remains limited; the current
evaluation shows that hand-coded control schemes
that let users directly manipulate the end-effector
can be similarly effective in some cases. Further-
more, LILAC is heavily tied to the latent actions
paradigm for shared autonomy, which is only a
small slice of the different types of solutions for
human-in-the-loop robotic manipulation.
Future work in language and shared autonomy
will allow for learning newutterances online, from
user feedback; for example, developing methods
for learning control strategies for novel language,
like “flip over the cup,” with minimal user feedback
(teaching demonstrations, language corrections,
etc.). More broadly, we hope to generalize our
correction module to other versions of language-
informed robotics – for example, to policy blending
(Dragan and Srinivasa, 2013), guided planning, and
interactive imitation learning (Kelly et al., 2019) –
for more complex, real-world manipulation tasks.
Acknowledgments
Toyota Research Institute (“TRI”) provided funds
to support this work. Siddharth Karamcheti is
grateful to be supported by the Open Philanthropy
Project AI Fellowship. We would additionally like
to thank the participants of our user study, as well
as our anonymous reviewers.
References
Brenna D Argall. 2018. Autonomy in rehabilitation
robotics: an intersection. Annual Review of Control,
Robotics, and Autonomous Systems , 1:441–463.
Yoav Artzi and Luke Zettlemoyer. 2013. Weakly su-
pervised learning of semantic parsers for mapping
instructions to actions. Transactions of the Associa-
tion for Computational Linguistics (TACL) , 1:49–62.
Dilip Arumugam, Siddharth Karamcheti, Nakul
Gopalan, Lawson L. S. Wong, and Stefanie Tellex.
2017. Accurately and efficiently interpreting human-
robot instructions of varying granularities. In
Robotics: Science and Systems (RSS) .
Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie
Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind
Neelakantan, Pranav Shyam, Girish Sastry, Amanda
Askell, Sandhini Agarwal, Ariel Herbert-V oss,
Gretchen Krueger, Tom Henighan, Rewon Child,
Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu,
Clemens Winter, Christopher Hesse, Mark Chen, Eric
Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess,
Jack Clark, Christopher Berner, Sam McCandlish,
Alec Radford, Ilya Sutskever, and Dario Amodei.
2020. Language models are few-shot learners. arXiv
preprint arXiv:2005.14165 .
Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem
Lahlou, Lucas Willems, Chitwan Saharia, Thien Huu
Nguyen, and Yoshua Bengio. 2019. Babyai: A plat-
form to study the sample efficiency of grounded
language learning. In International Conference on
Learning Representations (ICLR) .
John D. Co-Reyes, Abhishek Gupta, Suvansh Sanjeev,
Nick Altieri, John DeNero, Pieter Abbeel, and Sergey
Levine. 2019. Guiding policies with language via
meta-learning. In International Conference on Learn-
ing Representations (ICLR) .
Anca D Dragan and Siddhartha S Srinivasa. 2013. A
policy-blending formalism for shared control. In-
ternational Journal of Robotics Research (IJRR) ,
32:790–805.
Laura V Herlant, Rachel M Holladay, and Siddhartha S
Srinivasa. 2016. Assistive teleoperation of robot
arms via automatic time-optimal mode switching.
InACM/IEEE International Conference on Human
Robot Interaction (HRI) , pages 35–42.
Shervin Javdani, Henny Admoni, Stefania Pellegrinelli,
Siddhartha S Srinivasa, and J Andrew Bagnell. 2018.
Shared autonomy via hindsight optimization for tele-
operation and teaming. International Journal of
Robotics Research (IJRR) , 37:717–742.
Hong Jun Jeon, Dylan P. Losey, and Dorsa Sadigh. 2020.
Shared autonomy with learned latent actions. In
Robotics: Science and Systems (RSS) .
Siddharth Karamcheti, Dorsa Sadigh, and Percy Liang.
2020. Learning adaptive language interfaces throughdecomposition. In EMNLP Workshop for Interactive
and Executable Semantic Parsing (IntEx-SemPar) .
Siddharth Karamcheti, Megha Srivastava, Percy Liang,
and Dorsa Sadigh. 2021a. LILA: Language-informed
latent actions. In Conference on Robot Learning
(CoRL) .
Siddharth Karamcheti, A. Zhai, Dylan P. Losey, and
Dorsa Sadigh. 2021b. Learning visually guided latent
actions for assistive teleoperation. In Learning for
Dynamics & Control Conference (L4DC) .
Michael Kelly, Chelsea Sidrane, K. Driggs-Campbell,
and Mykel J. Kochenderfer. 2019. HG-DAgger: In-
teractive imitation learning with human experts. In
International Conference on Robotics and Automa-
tion (ICRA) , pages 8077–8083.
T. Kollar, J. Krishnamurthy, and Grant P. Strimel. 2013.
Toward interactive grounded language acqusition. In
Robotics: Science and Systems (RSS) .
Mengxi Li, Dylan P. Losey, Jeannette Bohg, and Dorsa
Sadigh. 2020. Learning user-preferred mappings for
intuitive robot control. In International Conference
on Intelligent Robots and Systems (IROS) .
Dylan P. Losey, Hong Jun Jeon, Mengxi Li, Kr-
ishna Parasuram Srinivasan, Ajay Mandlekar, Ani-
mesh Garg, Jeannette Bohg, and Dorsa Sadigh. 2021.
Learning latent actions to control assistive robots.
Autonomous Robots (AURO) , pages 1–33.
Dylan P. Losey, Krishnan Srinivasan, Ajay Mandlekar,
Animesh Garg, and Dorsa Sadigh. 2020. Controlling
assistive robots with learned latent actions. In In-
ternational Conference on Robotics and Automation
(ICRA) , pages 378–384.
Jelena Luketina, Nantas Nardelli, Gregory Farquhar,
Jakob Foerster, Jacob Andreas, Edward Grefenstette,
Shimon Whiteson, and Tim Rocktäschel. 2019. A
survey of reinforcement learning informed by natu-
ral language. In International Joint Conference on
Artificial Intelligence (IJCAI) .
Corey Lynch and Pierre Sermanet. 2020. Grounding
language in play. arXiv preprint arXiv:2005.07648 .
Alana Marzoev, S. Madden, M. Kaashoek, Michael J.
Cafarella, and Jacob Andreas. 2020. Unnatural lan-
guage processing: Bridging the gap between syn-
thetic and natural language data. arXiv preprint
arXiv:2004.13645 .
Cynthia Matuszek, Nicholas FitzGerald, Luke Zettle-
moyer, Liefeng Bo, and Dieter Fox. 2012. A joint
model of language and perception for grounded at-
tribute learning. In International Conference on Ma-
chine Learning (ICML) , pages 1671–1678.
Nils Reimers and Iryna Gurevych. 2019. Sentence-
BERT: Sentence embeddings using siamese BERT-
networks. In Empirical Methods in Natural Lan-
guage Processing (EMNLP) .
Mohit Shridhar, Lucas Manuelli, and Dieter Fox. 2021.
Cliport: What and where pathways for robotic manip-
ulation. In Conference on Robot Learning (CoRL) .
Simon Stepputtis, J. Campbell, Mariano Phielipp, Ste-
fan Lee, Chitta Baral, and H. B. Amor. 2020.
Language-conditioned imitation learning for robot
manipulation tasks. In Advances in Neural Informa-
tion Processing Systems (NeurIPS) .
Stefanie Tellex, Thomas Kollar, Steven Dickerson,
Matthew R Walter, Ashis Gopal Banerjee, Seth J
Teller, and Nicholas Roy. 2011. Understanding nat-
ural language commands for robotic navigation and
mobile manipulation. In Association for the Advance-
ment of Artificial Intelligence (AAAI) .
Jesse Thomason, Michael Murray, Maya Cakmak, and
Luke Zettlemoyer. 2019a. Vision-and-dialog naviga-
tion. In Conference on Robot Learning (CoRL) .
Jesse Thomason, Aishwarya Padmakumar, Jivko
Sinapov, Nick Walker, Yuqian Jiang, Harel Yedid-
sion, Justin W. Hart, Peter Stone, and Raymond J.
Mooney. 2019b. Improving grounded natural lan-
guage understanding through human-robot dialog. In
International Conference on Robotics and Automa-
tion (ICRA) .
Jesse Thomason, Shiqi Zhang, Raymond J. Mooney,
and Peter Stone. 2015. Learning to interpret natural
language commands through human-robot dialog. In
International Joint Conference on Artificial Intelli-
gence (IJCAI) .A Using GPT-3 to Identify Corrections
A core component of LILAC is the choice of gating
function, for producing the object state dependence
weight αfor a given language utterance. Critically,
αdictates whether an utterance is an instruction
(α= 1) that depends on the current environment
context, a correction (α= 0) that does not, and
can be interpreted without additional grounded in-
formation, or something in-between (α= 0.5).
We made the realization early on that identify-
ing whether a language utterance fell into one of
the above categories could (at least heuristically)
be decoupled from any environment information;
that is to say, we could predict αdirectly from the
language utterance alone. Given this hypothesis,
and the fact that we were not sure whether the α-
gating would even work in our small-data regime,
we found it difficult to defend the choice to collect
data to learn αupfront, prior to running through the
whole system. Therefore, our choices were to ei-
ther hardcode a series of αvalues for a small, fixed
set of correction language, or come up with some
heuristic (e.g., referent-counting) that could break
or generalize poorly to “in-between” utterances. In-
stead, we chose to try a prompt construction based
approach, leveraging GPT-3 (Brown et al., 2020).
The prompt we specified is shown in Fig. 5; we
did minimal prompt-tuning, only reordering the ex-
amples shown, and turning the temperature down
to 0 (as we wanted this to be deterministic). We
found GPT-3 to work incredibly well out of the
box – a phenomenon that cannot be overstated .
With a straightforward procedure, we were able
to use GPT-3 as a drop-in replacement for what
otherwise would have been a brittle heuristic, or a
limited set of language corrections. It is incredibly
exciting to be able to prototype these systems via
GPT-3 quickly, and we hope that this type of us-
age becomes prevalent throughout not only human-
robot interaction, but widespread NLP pipelines as
a whole. Given the results with the “drop-in” GPT-
3 model, we have a good idea as to where we need
to focus future work with respect to learning α–
specifically, for handling the nuanced utterances
that are “in-between” corrections and instructions –
and are excited to tackle this moving forward.
B User Study Details & Tasks
As mentioned in §4, we ran a within-subjects study
with a small number of participants (n= 5) . The
study consisted of each user using the following
Figure 5: The concrete prompt used for GPT-3 davinci-instruct , visualized in the OpenAI API Playground. We
primed GPT-3 with 3 handcrafted examples, without much other thought, and used the corresponding outputs as our
αgating values (“instruction” = 1, “in-between” = 0.5, “correction” = 0).
strategies to solve a given task: End-Effector con-
trol, LILA (no corrections), and LILAC (our pro-
posed approach). The order of strategies was ran-
domized across users. The tasks were as follows:
1.Pick & Insert Banana – Grasp the banana,
and place it in the plastic fruit basket, turning
the gripper appropriately to insert the banana.
2.Pick & Place Basket – Grasp the basket by
the handles, and lower it onto the tray.
3.Pick & Place Cereal Bowl – Grasp the
green cereal bowl by the lip of the bowl, lift
it off the pedestal, and lower it onto the tray
without collision.
4.Pick & Pour into Cereal Bowl – Grasp
the blue cup with marbles by either the lip of
the cup or the handle on the side, then lift it
over the cereal bowl, tilting the cup to pour
the marbles into the bowl.
5.Pick & Pour into Mug – Grasp the blue
cup with marbles by either the lip of the cup
or the handle on the side, then lift it over the
black mug – avoiding collisions – then pour
the contents into the mug.
Prior to executing the task in a given control
mode, each user was given 3 minutes to practice
using that control mode. During practice, the user
could experiment with various patterns of joystick
input to better understand how they translated to
movement of the robot arm. We found that theamount of “naturalization” time was critical in get-
ting users to adapt to the interfaces provided by
LILA, and LILAC. Furthermore, for users not al-
ready experienced in robot teleoperation, we found
this practice period important as well. However, we
found that it takes longer to naturalize to LILAC
vs. End-Effector control – this also explains the
slight difference in results between LILAC and
End-Effector control. Future work will explore the
impact of this “naturalization period” under various
conditions.
C Additional Visualizations & Baselines
To supplement our experiments from §4, we
present two additional sets of trajectory visualiza-
tions in Fig. 6 and Fig. 7. Fig. 6 shows the same 4
strategies as in Fig. 3, except for a more complex
pouring task; in general, the pattern of behavior is
the same – both LILAC and End-Effector control
are able to solve the task, whereas LILA stalls out
due to a lack of a corrective signal, with Correction-
Aware Imitation Learning failing to fit a decent
policy given limited data.
One other baseline we ran was the no-language
variant of latent actions ; we ran this to show the im-
pact that language has on providing intuitive, use-
ful control spaces. Fig. 7 shows the results of this
baseline vs. LILAC – at a glance, the no-language
variant completely fails, reducing to random, oscil-
latory behavior, because it cannot learn to provide
a good control space for all tasks, without an extra
conditioning signal. This is the same observation
made in Karamcheti et al. (2021a).
End-Effector ControlCorrection-Aware ImitationLILA (No Corrections)LILAC (Ours)Goal: “Pour the contents of the blue cup into the mug”
✅
✅
❌
❌
“Move down”“Head right”Figure 6: Additional trajectory visualizations for the four strategies from §4, this time for a more complex “pouring”
task; we see similar results as in Fig. 3.
“Place the fruit basket on the tray”No-Language Latent Actions
“Pour the contents of the blue cup into the mug”
LILAC (Ours)
✅“Lower it”
❌No-Language Latent Actions
❌
“Head right”
LILAC (Ours)
✅
Figure 7: Results visualizing trajectories for LILAC and an additional baseline strategy – “No-Language Latent
Actions” – the language-free variant of our approach. Note that this approach trivially fails, as its overloaded,
unable to find a satisfying control scheme that would allow users to perform all 5 tasks without extra conditioning
information. This translates to aimless, oscillatory behavior during execution. |
f3657a5f-a3a8-4a7f-98bc-0a13577546c9 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Formal Open Problem in Decision Theory
.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')}
*(This post was originally published on March 31st 2017, and has been brought forwarded as part of the AI Alignment Forum launch sequence on fixed points.)*
In this post, I present a new formal open problem. A positive answer would be valuable for decision theory research. A negative answer would be helpful, mostly for figuring out what is the closest we can get to a positive answer. I also give some motivation for the problem, and some partial progress.
**Open Problem:** Does there exist a topological space X (in some [convenient category of topological spaces](https://ncatlab.org/nlab/show/convenient+category+of+topological+spaces)) such that there exists a continuous surjection from X to the space [0,1]X (of continuous functions from X to [0,1])?
---
***Motivation:***
**Topological Naturalized Agents:** Consider an agent who makes some observations and then takes an action. For simplicity, we assume there are only two possible actions, A and B. We also assume that the agent can randomize, so we can think of this agent as outputting a real number in [0,1], representing its probability of taking action A.
Thus, we can think of an agent as having a policy which is a function from the space Y of possible observations to [0,1]. We will require that our agent behaves continuously as a function of its observations, so we will think of the space of all possible policies as the space of continuous functions from Y to [0,1], denoted [0,1]Y.
We will let X denote the space of all possible agents, and we will have a function f:X→[0,1]Y which takes in an agent, and outputs that agent's policy.
Now, consider what happens when there are other agents in the environment. For simplicity, we will assume that our agent observes one other agent, and makes no other observations. Thus, we want to consider the case where Y=X, so f:X→[0,1]X.
We want f to be continuous, because we want a small change in an agent to correspond to a small change in the agent's policy. This is particularly important since other agents will be implementing continuous functions on agents, and we would like any continuous function on policies to be able to be considered valid continuous function on agents.
We also want f to be surjective. This means that our space of agents is sufficiently rich that for any possible continuous policy, there is an agent in our space that implements that policy.
In order to meet all these criteria simultaneously, we need a space X of agents, and a continuous surjection f:X→[0,1]X.
**Unifying Fixed Point Theorems:** While we are primarily interested in the above motivation, there is another secondary motivation, which may be more compelling for those less interested in agent foundations.
There are (at least) two main clusters of fixed point theorems that have come up many times in decision theory, and mathematics in general.
First, there is the Lawvere cluster of theorems. This includes the Lawvere fixed point theorem, the diagonal lemma, and the existence of Quines and fixed point combinators. These are used to prove Gödel's incompleteness Theorem, Cantor's Theorem, Löb's Theorem, and achieve robust cooperation in the Prisoner's Dilemma in [modal framework](https://arxiv.org/pdf/1401.5577.pdf) and [bounded variants](https://arxiv.org/pdf/1602.04184.pdf). All of these can be seen as corollaries of Lawvere's fixed point theorem, which states that in a cartesian closed category, if there is a point-surjective map f:X→YX, then every morphism g:Y→Y has a fixed point.
Second, there is the Brouwer cluster of theorems. This includes Brouwer's fixed point theorem, The Kakutani fixed point theorem, Poincaré–Miranda, and the intermediate value theorem. These are used to prove the existence of Nash Equilibria, [Logical Inductors](https://intelligence.org/files/LogicalInduction.pdf), and [Reflective Oracles](https://arxiv.org/pdf/1508.04145.pdf).
If we had a topological space and a continuous surjection X→[0,1]X, this would allow us to prove the one-dimensional Brouwer fixed point theorem directly using the Lawvere fixed point theorem, and thus unify these two important clusters.
Thanks to Qiaochu Yuan for pointing out the connection to Lawvere's fixed point theorem (and actually [asking this question three years ago](http://mathoverflow.net/questions/136478/can-the-lawvere-fixed-point-theorem-be-used-to-prove-the-brouwer-fixed-point-the)).
---
**Partial Progress:**
**Most Diagonalization Intuitions Do Not Apply:** A common initial reaction to this question is to conjecture that such an X does not exist, due to cardinality or diagonalization intuitions. However, note that all of the diagonalization theorems pass through (some modification of) the same lemma: Lawvere's fixed point theorem. However, this lemma does not apply here!
For example, in the category of sets, the reason that there is no surjection from any set X to the power set, {T,F}X, is because if there were such a surjection, Lawvere's fixed point theorem would imply that every function from {T,F} to itself has a fixed point (which is clearly not the case, since there is a function that swaps T and F).
However, we already know by Brouwer's fixed point theorem that every continuous function from the interval [0,1] to itself has a fixed point, so the standard diagonalization intuitions do not work here.
**Impossible if You Replace [0,1] with e.g. S1:** This also provides a quick sanity check on attempts to construct an X. Any construction that would not be meaningfully different if the interval [0,1] is replaced with the circle S1 is doomed from the start. This is because a continuous surjection X→(S1)X would violate Lawvere's fixed point theorem, since there is a continuous map from S1 to itself without fixed points.
**Impossible if you Require a Homeomorphism:** When I first asked this question I asked for a homeomorphism between X and [0,1]X. Sam Eisenstat has given a very clever argument why this is impossible. You can read it [here](https://mathoverflow.net/questions/264850/is-there-a-topological-space-x-homeomorphic-to-the-space-of-continuous-functions). In short, using a homeomorphism, you would be able to use Lawvere to construct a continuous map that send a function from [0,1] to itself to a fixed point of that function. However, no such continuous map exists.
---
**Notes:**
If you prefer not to think about the topology of [0,1]X, you can instead find a space X, and a continuous map h:X×X→[0,1], such that for every continuous function f:X→[0,1], there exists an xf∈X, such that for all x∈X, h(xf,x)=f(x).
Many of the details in the motivation could be different. I would like to see progress on similar questions. For example, you could add some computability condition to the space of functions. However, I am also very curious which way this specific question will go.
This post came out of many conversations, with many people, including: Sam, Qiaochu, Tsvi, Jessica, Patrick, Nate, Ryan, Marcello, Alex Mennen, Jack Gallagher, and James Cook.
---
*This post was originally published on March 31st 2017, and has been brought forwarded as part of the AI Alignment Forum launch sequences.*
*Tomorrow's AIAF sequences post will be 'Iterated Amplification and Distillation' by Ajeya Cotra, in the sequence on iterated amplification.* |
0e3cc522-ccf5-4d33-8775-4895bbf99268 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | How can I convince others and present the arguments well?
[Things Skeptics Commonly Say, and links to refutations](https://docs.google.com/document/d/1N52iABJLIk7XpPVY0A1I8dr9_bWQo6mCvi7pXC_JLBI/edit#) goes over most of the common objections, with some of the ways in which each is not fatal to the AI existential risk arguments.
[Vael Gates's project](https://www.lesswrong.com/posts/LfHWhcfK92qh2nwku/transcripts-of-interviews-with-ai-researchers) links to lots of example transcripts of persuading senior AI capabilities researchers.
|
b83c9bf3-662e-4074-911b-8750fade651d | trentmkelly/LessWrong-43k | LessWrong | Circles of discussion
On Wednesday I had lunch with Raph Levien, and came away with a picture of how a website that fostered the highest quality discussion might work.
Principles:
* It’s possible that the right thing is a quick fix to Less Wrong as it is; this is about exploring what could be done if we started anew.
* If we decided to start anew, what the software should do is only one part of what would need to be decided; that’s the part I address here.
* As Anna Salamon set out, the goal is to create a commons of knowledge, such that a great many people have read the same stuff. A system that tailored what you saw to your own preferences would have its own strengths but would work entirely against this goal.
* I therefore think the right goal is to build a website whose content reflects the preferences of one person, or a small set of people. In what follows I refer to those people as the “root set”.
* A commons needs a clear line between the content that’s in and the content that’s out. Much of the best discussion is on closed mailing lists; it will be easier to get the participation of time-limited contributors if there’s a clear line of what discussion we want them to have read, and it’s short.
* However this alone excludes a lot of people who might have good stuff to add; it would be good to find a way to get the best of both worlds between a closed list and an open forum.
* I want to structure discussion as a set of concentric circles.
* Discussion in the innermost circle forms part of the commons of knowledge all can be assumed to be familiar with; surrounding it are circles of discussion where the bar is progressively lower. With a slider, readers choose which circle they want to read.
* Content from rings further out may be pulled inwards by the votes of trusted people.
* Content never moves outwards except in the case of spam/abuse.
* Users can create top-level content in further-out rings and allow the votes of other users to move it closer to the centre. Us |
70b40efa-f548-4b99-8cb0-2e0ca2343e76 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post1803
This post is a distillation of Evan Hubinger's post " how do we become confident in the safety of a machine learning system? ", made as part of the summer 2022 SERI MATS program . While I have attempted to understand and extrapolate Evan's opinions, this post has not been vetted. Likewise, I use training stories (and contribution stories ) to describe the methodology of proposals for safe advanced AI without the endorsement of those proposals' authors and based on a relatively shallow understanding of those proposals (due to my inexperience and time constraints). The opinions presented in this post are my own unless otherwise noted. Epistemic status: Exploratory Some day, all too soon from now, people will deploy the AI that seals humanity's fate. There are many scenarios for how this comes about or what happens afterward. Some have multiple AI negotiating or fighting for dominance, others one. Some think the handoff of power from humans to AI will go slowly, others fast. But whatever the case, there is one question that these people need to get right: "Is the AI we're about to deploy safe?" For people to get this answer right when it matters, two things need to happen: we need tools to accurately determine whether an advanced AI is safe, and we need an advanced AI those tools approve of that we can deploy in time for it to matter. Training stories are a tool to evaluate proposals for making this happen. Specifically, they're meant to analyze complete proposals for training prosaic advanced AI: advanced AI which has similar enough architecture to current systems that most research is cross-applicable. The reason to focus on complete proposals is that it gets us to pay attention to what will be important when push comes to shove. From there, we can backpropagate and get an estimate of what research paths are most beneficial. Evan thinks theoretical prosaic alignment is the most beneficial in expectation for him and many others to pursue, though he supports a broad spectrum of research. Because I think the original post describes the central concept of training stories very well, I have avoided retreading the same ground, and instead focused on expanding its usability. I will start with some more case studies of using training stories to analyze proposals, and then expand on the original post with contribution stories : a method for analyzing research agendas by how they factor into training stories. Training stories Case study: Imitative amplification + intermittent oversight In this case study, Evan's comments in An overview of 11 proposals for building safe advanced AI map well to the training story evaluation, and I don't feel like much was missing. However, the training story itself makes it more clear to me what the training goal might look like, which could help point at potential experiments with toy models that can be done today. This demonstrates how training stories can be used to clarify a path towards working AI. Training goal specification : The goal is to end up with a model M which accurately answers whether a model similar to M is safe to run, which suggests possible tests to determine if that model is safe to run, and which answers questions about some other topic T. This can take several forms, but let's try to work one out: M can't take all weights of M as input for inspection, so M has to look at sections of a model and share a summary of its conclusions with other M's to form a reliable answer. This could be organized like HCH. Since M is trained to imitate the output of Amp(M), which is prototypically a human consulting M and some selection of transparency tools, these summaries would be human-legible. Mechanically, M would take as input the question it is tasked with answering, several relevant question-answer pairs from other M's, and information relevant to answering the question (either part of the M-like model or information about the topic M is intended to be deployed in). The question-answer pairs, baked-in priors from training, and the new information input could internally be combined to form a world model, with (linear combinations of) neural activations corresponding to confidence in particular beliefs. The natural-language question could then be combined with the world model to produce a natural-language response. Training goal desirability : The reason we want such a model is so it can give accurate, reliable answers about the topic T it has been trained on, as an oracle. This could just be inspecting other models, but it could also mean answering practical questions about the world that could have transformative value. Training rationale nudges: The plan for getting such a model is to train it through imitative amplification, intermittently performing oversight checks. Imitative amplification here means that the model is trained to imitate Amp(M), where Amp() is the amplification operator consisting of a human that can consult transparency tools and the model itself on questions concerning M's safety and the topic T. The oversight checks happen intermittently and consist of Amp(M) giving a binary answer on whether the model is safe enough to continue training. Models that fail these checks are pruned, applying evolutionary pressure for the remaining ones to be safe. Training rationale constraints: The model must have a structure suitable for inspecting neural nets, for handling T, for natural language questions and answers, and for delegating questions to other versions of itself. It is unclear at this point what a suitable architecture for this would be. In training, the questions Amp(M) is asked have to follow a distribution suitable for M at its stage of development. Amp() must also be capable enough to allow some M's to pass the filter at all stages of development without false negatives. The evaluation of the training story is done pretty well in the original 11 proposals post : Training goal alignment: The model is intended to function as an oracle , with the associated potential hazards. Also, the limit of training is HCH, so it is at most as aligned as that system would be. See outer alignment in 11 proposals . Training goal competitiveness: As performance competitiveness in 11 proposals . Training rationale alignment: See inner alignment in 11 proposals . This is currently the main limiting factor: we don't have transparency tools sufficient to provide a rich training signal on M inspecting M-like models, especially not if the safety of M is non-trivial. Training rationale competitiveness: As training competitiveness in 11 proposals . Case study: Reinforcement learning + transparency tools Attempting to make a training story for this, it quickly becomes clear that the training goal is underspecified. While there are vague behavioral arguments for AI trained in cooperative environments to act 'cooperatively', there appear to be no conceptualization of an internal algorithm where this cooperation is robust in the way we want, let alone training rationale nudges to give rise to this internal logic over others that meet the training rationale constraints of cooperative behavior in training. This demonstrates training stories' usefulness as an analysis tool: a serious proposal for safe advanced AI is shown lacking. That doesn't mean that an RL incentive towards cooperation is useless, just that it is far from sufficient and a better plan is needed. Case study: STEM AI Like reinforcement learning + transparency tools , the training goal is for STEM AI is underspecified. Limiting AI subject matter is a valid, albeit soft, training rationale nudge away from the AI having a conception of agents it can deceive, but it will not suffice alone. STEM AI may well have great economic value, but it's hard to imagine it being safely transformative. It might help make capabilities research safer, though - see below. Contribution stories In the section ' Exploring the landscape of possible training stories ', Evan lists a number of goals and rationales that could be elements of a training story, and notes that many other options are possible. The section does not provide tools for how to qualitatively evaluate these potential partial strategies, which I think limits the usability of training stories to many researchers' actual efforts. For example, while Evan places high importance on transparency tools , Microscope AI is the part of Chris Olah's work that is highlighted in the overall article despite requiring additional assumptions. Training stories help us answer the question "What training characteristics would convince us that an advanced AI we're about to deploy is safe?". I would like to expand on that with "How do we get those training characteristics?" - to give tools to think about the backpropagation to research agendas that can be tackled now. For this I'll sketch out contribution stories . As an expansion of training stories, these are also focused on prosaic AI development. I'll split a contribution story up into the following sections: Alignment contribution method : How and whether the research contribution would advance the development of a safe advanced AI, ideally by naming training story components and criteria that it contributes to. Alignment contribution impact : An estimate of how much the research contributes to the odds of developing safe AI. The following criteria seem useful: Urgent: whether and to what extent the research's value depends on it starting now. Since some AI research agendas are likely to be non-parallelizable or to benefit other safety research, it's better to work on those over ones that aren't urgent. Critical: whether and to what extent the research has value for helping the final AI be safe. Generalizable: how fragile the research's value is to changes in the training story. Capabilities contribution method: How and whether the research contribution would advance the development of general advanced AI capabilities, and thereby increase x-risk, ideally with training story components and criteria. Capabilities contribution impact : An estimate of how much the research might advance general capabilities, ideally with mitigation strategies. Case study: Best-case inspection transparency Alignment contribution method : Ultimately, transparency tools can be used to improve training rationale alignment by allowing a human, a safe AI, or a combination of both to inspect a model as it develops, steering away from dangerously misaligned thought patterns such as deception by penalizing their precursors. Developing transparency tools early can help build a catalog of precursors to misalignment or even desired patterns for more well-behaved models that could be selected for. While best-case inspection transparency would not be sufficient, it allows us to gather data to develop more sophisticated transparency tools . Alignment contribution impact : As argued in " A transparency and interpretability tech tree ", it seems difficult to ever be confident in the alignment of an AI if we don't have powerful transparency tools that let us know how the AI is arriving at its decisions, meaning transparency tools will be critical in any form of prosaic alignment. Getting sufficiently powerful transparency tools to guide advanced AI also seems like non-parallelizable work: we need to work our way up the tech tree with experimental research. This makes transparency urgent as well. Transparency is broadly valuable because it makes nearly all prosaic alignment research more easy to test and allows for more fine-grained models of why the model is thinking what it is. Transparency tools will advance capabilities (see below), but in doing so they give capabilities researchers an incentive to make their models more transparent, which could also be a boon to alignment in the long run by lowering the alignment tax capabilities researchers would have to pay if they decide to be sufficiently cautious at a later date and by offloading transparency research on people who might otherwise not be inclined to do safety research. General popularity of transparency tools could even raise awareness among capabilities researchers by helping them notice potential warning shots. Capabilities contribution method: Since transparency tools allow researchers to get more fine-grained ideas of how a model is thinking, they will also increase training rationale competititiveness of all model classes they are applied to by giving feedback on where a training process could be improved. Capabilities contribution impact : Because transparency improves training rationale competitiveness of a wide range of models it is applied to, worlds with transparency have shorter timelines than worlds without. This is a high price to pay, but it seems likely transparency is necessary for prosaic alignment, so we simply have to make do. Case study: Deferential AI (Stuart Russell / CHAI) The training goal of deferential AI is a model which has an internal representation of a human's preferences which it pursues and is uncertain over. This uncertainty is then leveraged to make it defer to the human when the human tries to correct it or shut it off. Alignment contribution method : The goal of deferential AI is to allow training goal specification to be a less precise target by making the AI more corrigible in deployment. Unfortunately, this corrigibility appears to break down as the AI acquires a sufficiently accurate model of humans. Less advanced deferential AI can assist in providing training rationale nudges by aiding research. Alignment contribution impact : Since the corrigibility of deferential AI appears to break down at high levels and it seems possible to imagine safe AI without it (e.g. myopic or truly aligned AI), it does not seem critical or urgent. Less advanced deferential AI might be a generally useful tool for aiding research, but as long as it requires architecture changes deference can't easily be added to an alignment research agenda. It's possible that further research into deference could ameliorate these issues. Capabilities contribution method: Deferential AI could aid capabilities research just as readily as alignment research by providing training rationale nudges towards more capable unsafe goal specifications. It might catch some unsafe objectives, but others it may not recognize or even be overruled on by the human operator. Capabilities contribution impact : It's hard to estimate at this point how viable deferential AI will be, but it appears to award little to no advantage to alignment over capabilities as it is presented now, which would mean it would worsen our odds. It seems wise to steer away from making deferential AI competitive until it improves alignment odds. Case study: STEM AI ; narrow agents in general STEM AI didn't quite fit the mould in 11 proposals or of a full-fledged training story. In my opinion a capabilities story does the proposal more justice: Alignment contribution method : Narrow agents like STEM AI attempt to define a training goal specification that allows for more training rationale competitiveness through less onerous training rationale constraints , while still being advanced enough to have great value. However, I do not expect this to be training goal competitive compared to more generally trained AI because the world is filled with humans that affect our preferences and options. Alignment contribution impact : As I said above, I don't expect narrow agents to be competitive for transformative AI. Agents that rely on narrowness for their safety don't seem very useful for alignment research. Research into narrowness guarantees has some chance of being critical by preventing a catastrophe with AI intended to be narrow - see below. This seems most useful around the time when AI could be trained to be general, so it is not very urgent. It's generalizable to any topic where you only need a narrow understanding to perform well, but that excludes quite a lot. Capabilities contribution method: Narrow agents may be able to get more optimized results in their limited domains than non-agentic simulators. If narrowness is guaranteed by architecture innovations, then narrow agents might be more training goal competitive than broad agents as well. The agents' output could then be used to advance AI capabilities by easing training rationale constraints like the price of compute. Offering capabilities researchers ways to guarantee narrowness could reduce risk from unaligned AI by preventing AI intended for a narrow topic from exhibiting dangerous behavior through unintended capabilities generalization. This would constitute improvement of training goal specification for capabilities, but improve our chances of survival. Capabilities contribution impact : Narrowness guarantees don't advance the timeline to safe transformative AI much. Instead, they potentially reduce the risk of misalignment by making capabilities research safer. This requires them to be both training goal competitive and training rationale competitive . Case study: Grouped Loss As a demonstration of contribution stories' applicability, I picked a recent alignment forum post about prosaic alignment to try it out on: Grouped Loss may disfavor discontinuous capabilities . Alignment contribution method : The intent of grouped loss is to make training rationale nudges easier to implement by smearing the evolution of new capabilities over more training steps, trading some training rationale competitiveness by picking a less efficient and harder to define loss function for training rationale alignment by making it easier to catch precursors to misalignment or steer towards desirable model traits. Alignment contribution impact : The proposal does not seem highly urgent, since its primary function is giving more traction for nudges that aren't developed enough yet, like transparency tools. It does seem potentially critical, in that it could be implemented in the training procedure for an advanced AI if the competitiveness hit isn't too great and it doesn't conflict with another loss specification. Generalization may be an issue depending on how much work it is to define groups that have the desired properties. Capabilities contribution: The training rationale competitiveness hit will likely prevent grouped loss from advancing capabilities. It feels to me like the contribution story format helped slot this concept into my model for how AI alignment could shake out, and to ask some of the right questions about the value of the research. I hope that this framework helps people evaluate their research options and discuss them with others, and so contributes to paving the way for aligned AGI. |
ab50d144-b441-4d29-9bcc-93added3e186 | trentmkelly/LessWrong-43k | LessWrong | David Brooks from the NY Times writes on earning-to-give
Just wanted to highlight an article. David Brooks from the NY Times writes on earning-to-give by working at a hedge fund:
http://www.nytimes.com/2013/06/04/opinion/brooks-the-way-to-produce-a-person.html?ref=opinion
Basically, he claims that working in an amoral environment will eventually turn you into a worse person than you would otherwise be, and weaken your resolve and desire to fulfill your original goal. Psychologically he may be right, and today's me may not like the me I would become after a decade on Wall Street, but at first glance it seems like even if I could only maintain my resolve for a few years, the payoff far outweighs my own well being. He is also opposed to valuing the far - life in general - over the near - people in your own home or community. Or even valuing them equally, AFAICT.
As a matter of history, though, I did not in fact choose such a career. Suboptimal or not, given what I did choose (consulting firm that helps companies invest and grow effectively in clean tech, nanotech, and biotech) I do not think I chose wrongly.
|
91da985b-a824-40c7-bc1a-d72d7644d95f | trentmkelly/LessWrong-43k | LessWrong | Competitive Essay Writing
Quick report on an anti-akrasia method NMJablonski and I tried out: Competitive Essay Writing. Two (or more) people have something they need to write (but may not particularly want to)- everyone gets on a IM client and every thirty minutes reports how many words they've written so far.
I didn't have an essay to write, but I did have a wiki to update for the D&D campaign I'm running, and I so did that while NMJablonski wrote an essay for school due the next day. I won handily, but that may have been the style of writing I was doing. I found it useful to have the pressure to not waste time chasing rabbit trails (hmm, I ought to name this professor after the guy who discovered the circulation of blood. Why, hello wikipedia!), and he found it useful to have pressure to report a number- instead of staring at the screen wondering what to write, he would just pick something and go with it.
The next step, I think, is to write a program so that, instead of having to manually report progress every 30 minutes, the word count automatically updates for everyone you're competing with. I don't know if that would be distracting or not- I imagine having immediate feedback, instead of delayed feedback, would be superior. |
ecf4a945-122e-4753-8501-add28457e980 | trentmkelly/LessWrong-43k | LessWrong | Coherence arguments do not entail goal-directed behavior
One of the most pleasing things about probability and expected utility theory is that there are many coherence arguments that suggest that these are the “correct” ways to reason. If you deviate from what the theory prescribes, then you must be executing a dominated strategy. There must be some other strategy that never does any worse than your strategy, but does strictly better than your strategy with certainty in at least one situation. There’s a good explanation of these arguments here.
We shouldn’t expect mere humans to be able to notice any failures of coherence in a superintelligent agent, since if we could notice these failures, so could the agent. So we should expect that powerful agents appear coherent to us. (Note that it is possible that the agent doesn’t fix the failures because it would not be worth it -- in this case, the argument says that we will not be able to notice any exploitable failures.)
Taken together, these arguments suggest that we should model an agent much smarter than us as an expected utility (EU) maximizer. And many people agree that EU maximizers are dangerous. So does this mean we’re doomed? I don’t think so: it seems to me that the problems about EU maximizers that we’ve identified are actually about goal-directed behavior or explicit reward maximizers. The coherence theorems say nothing about whether an AI system must look like one of these categories. This suggests that we could try building an AI system that can be modeled as an EU maximizer, yet doesn’t fall into one of these two categories, and so doesn’t have all of the problems that we worry about.
Note that there are two different flavors of arguments that the AI systems we build will be goal-directed agents (which are dangerous if the goal is even slightly wrong):
* Simply knowing that an agent is intelligent lets us infer that it is goal-directed. (EDIT: See these comments for more details on this argument.)
* Humans are particularly likely to build goal-directed agen |
bb6575fe-623c-4356-a3aa-c5c01961b585 | trentmkelly/LessWrong-43k | LessWrong | Comment on the lab leak hypothesis
A friend of mine requested that I write up some of my comments on the lab leak hypothesis, since I had done quite a bit of research into this in 2020. This was originally asked in the context of the longform Facebook post that Eliezer wrote regarding the origins of the Covid-19 pandemic and the implications for the future: https://www.facebook.com/yudkowsky/posts/10159653334879228
My comment is less organized than I would prefer, but I figured it was better to post in a rough form than not post at all.
In early January 2021, I wrote on Facebook and Lesswrong:
> I've done a lot of thinking about the origins of SARS-CoV-2 and I still find the lab escape hypothesis quite credible. We still haven't found an intermediate host, which is surprising if the virus emerged naturally. This issue has become politicized, but we really need a neutral investigation in order to figure out what actually happened. This is quite important for understanding how to prevent future pandemics.
>
> The New York Magazine article, "The Lab-leak Hypothesis" is long, but I can personally verify the author is reporting on the source material pretty accurately. I've done over 200 hours of research on this topic and have read basically all the sources the article cites. That said, I don't agree with all of the claims. I do not think the SARS-CoV-2 virus is very likely to have been created using the RATG13 virus, because of the genetic differences spread out throughout the genomes. However, there are many other paths that could have led to a lab escape, and I'm somewhat agnostic between several of them.
>
> I believe we need more transparency into the whole investigation, especially the efforts on the ground in China. What is the process by which scientists in Hubei, Yunnan, and other provinces are looking for intermediate hosts? How have the investigation(s) into the WIV been organized and what processes did they use? What agencies, methods, and results?
>
> COVID-19 has had a huge effect o |
926d4c8c-f9cc-4467-ad67-298a0be91a02 | trentmkelly/LessWrong-43k | LessWrong | Progress links and tweets, 2023-06-28: “We can do big things again in Pennsylvania”
Opportunities
* AI Grant’s second batch is now accepting applications (via @natfriedman)
* Longevity Biotech Fellowship 2 is also accepting applications (via @allisondman)
* Science writing office hours with Niko McCarty (this one over but more in the future)
News & links
* I-95 reopened in just 12 days after a section of it collapsed. Gov. Shapiro says this proves “that we can do big things again in Pennsylvania”
* Arcadia Science will publish their abandoned projects (via @stuartbuck1)
* Short interview with the Hiroshima bombing mission lead (via @michael_nielsen)
Queries
* When was the last time a positive vision of the future took hold?
* What should Tyler Cowen ask Paul Graham?
* What should Dwarkesh Patel ask Andy Matuschak?
* Who should Eric Gilliam meet in/near London?
* How do airlines pool information / make agreements on safety and avoid antitrust?
* A FAQ that addresses the arguments/concerns of vaccine skeptics?
Quotes
* “I wonder that the Lord God has kept such things hidden”
* “The Flat Iron is to the United States what the Parthenon was to Greece”
* The terrible treatment of the girls who worked in the Bryant and May match factory
AI risk
* Claim: now is an “acute risk period” that only ends with a “global immune system”
* Concerns about AI are warranted, but there are very valid counter-arguments
* An AI doom syllogism
* On the paper “Optimal Policies Tend to Seek Power”
Tweets
* Gear teeth are way more nuanced than you would expect
* A brief thread of wonder at the modern world
* Staying up to date on news by tracking prediction markets
* Floor raisers vs. ceiling raisers
* Induction vs. deduction / empiricism vs. rationalism are the falsest dichotomies
* What someone’s unwillingness to debate says about their position
* “How easy is it for a kid to operate a lemonade stand?” as a city metric. Kennett Square, PA and Tooele, UT score well
* A 13-story, 245-unit timber high rise that would be illegal to build in |
66edd471-44cc-4fea-b8c7-229e6ca23314 | trentmkelly/LessWrong-43k | LessWrong | Focus on the Hardest Part First
Consider reading this instead.
Here is some obvious advice.
I think a common failure mode when working on AI alignment[1] is to not focus on the hard parts of the problem first. This is a problem when generating a research agenda, as well as when working on any specific research agenda. Given a research agenda, there are normally many problems that you know how to make progress on. But blindly working on what seems tractable is not a good idea.
Let's say we are working on a research agenda about solving problems A, B, and C. We know that if we find solutions to A, B, and C we will solve alignment. However, if we can't solve even one subproblem, the agenda would be doomed. If C seems like a very hard problem, that you are not sure you can solve, it would be a bad idea to flinch away from C and work on problem A instead, when A seems so much more manageable.
If solving A takes a lot of time and effort, all of that time and effort would be wasted, if you can't solve C in the end. It's especially worrisome when A has tight fightback loops, such that you constantly feel like you are making progress. Or when it is just generally fun to work on A.
Of course, it can make sense to work on A first if you expect this to help you solve C, or at least give you more information on its tractability. The general version of this is illustrated by considering that you have a large list of problems that you need to solve. In this case, focusing on problems that will provide you with information that will be helpful for solving many of the other problems can be very useful. But even then you should not lose sight of the hard problems that might block you down the road.
The takeaway is that these two things are very different:
* Solving A as an instrumental subgoal in order to make progress on C, when C is a potential blocker.
* Avoiding C, because it seems hard, and instead working on A because it seems tractable.
----------------------------------------
1. Though I expect |
bec0e5b5-b906-44f7-96ac-029e0a0e56e7 | trentmkelly/LessWrong-43k | LessWrong | Questions about AI that bother me
Crossposted from the EA Forum: https://forum.effectivealtruism.org/posts/4TcaBNu7EmEukjGoc/questions-about-ai-that-have-bothered-me
As 2022 comes to an end, I thought it'd be good to maintain a list of "questions that bother me" in thinking about AI safety and alignment. I don't claim I'm the first or only one to have thought about them. I'll keep updating this list.
(The title of this post alludes to the book "Things That Bother Me" by Galen Strawson)
First posted: 12/6/22
Last updated: 1/30/23
General Cognition
* What signs do I need to look for to tell whether a model's cognition has started to emerge, e.g., situational awareness?
* Will a capacity for "doing science" be sufficient condition for general intelligence?
* How easy was it for humans to get science (e.g., compared to evolving to take over the world).
Deception
* What kind of interpretability tools do we need to avoid deception?
* How do we get these interpretability tools and even if we do get them, what if they're like neuroscience for understanding brains (not enough)?
* How can I tell whether a model has found another goal to optimize for during its training?
* What is it that makes a model switch to a goal different from the one set by the designer? How do you prevent it from doing so?
Agent Foundations
* Is the description/modeling of an agent ultimately a mathematical task?
* From where do human agents derive their goals?
* Is value fragile?
Theory of Machine Learning
* What explains the success of deep neural networks?
* Why was connectionism unlikely to succeed?
Epistemology of Alignment (I've written about this here)
* How can we accelerate research?
* Has philosophy ever really helped scientific research e.g., with concept clarification?
* What are some concrete takeaways from the history of science and technology that could be used as advice for alignment researchers and field-builders?
* The emergence of the AI Safety paradigm
Philosophy o |
9b960ae3-f4a7-46f9-80b6-c7a6ee4f8dfd | trentmkelly/LessWrong-43k | LessWrong | Max Tegmark on our place in history: "We're Not Insignificant After All"
An uplifting message as we enter the new year, quoted from Edge.org:
> We're Not Insignificant After All
>
> Max Tegmark, Physicist, MIT
>
> When gazing up on a clear night, it's easy to feel insignificant. Since our earliest ancestors admired the stars, our human egos have suffered a series of blows. For starters, we're smaller than we thought. Eratosthenes showed that Earth was larger than millions of humans, and his Hellenic compatriots realized that the solar system was thousands of times larger still. Yet for all its grandeur, our Sun turned out to be merely one rather ordinary star among hundreds of billions in a galaxy that in turn is merely one of billions in our observable universe, the spherical region from which light has had time to reach us during the 14 billion years since our big bang. Then there are probably more (perhaps infinitely many) such regions. Our lives are small temporally as well as spatially: if this 14 billion year cosmic history were scaled to one year, then 100,000 years of human history would be 4 minutes and a 100 year life would be 0.2 seconds. Further deflating our hubris, we've learned that we're not that special either. Darwin taught us that we're animals, Freud taught us that we're irrational, machines now outpower us, and just last month, Deep Fritz outsmarted our Chess champion Vladimir Kramnik. Adding insult to injury, cosmologists have found that we're not even made out of the majority substance.
>
> The more I learned about this, the less significant I felt. Yet in recent years, I've suddenly turned more optimistic about our cosmic significance. I've come to believe that advanced evolved life is very rare, yet has huge growth potential, making our place in space and time remarkably significant.
>
>
> The nature of life and consciousness is of course a hotly debated subject. My guess is that these phenomena can exist much more generally that in the carbon-based examples we know of.
>
> I believe that consciousness is, |
f51c6a06-207a-42be-9ce3-fb42e0ce5318 | trentmkelly/LessWrong-43k | LessWrong | Setting up my work environment - Doing the causation backwards
Original post: http://bearlamp.com.au/doing-the-causation-backwards/
----------------------------------------
About two years ago, when I first got my smart phone (yes, later than most of the other humans). I was new to apps, and I was new to environments. When I decided on what apps should be on my home screen, I picked the ones that I thought I would use most often.
My home screen started with:
* google bar (the top of the page)
* calendar
* facebook
* notepad app (half the page)
* ingress (because I play)
* maps
* camera
* torch
----------------------------------------
My home screen has barely changed. I don't play ingress very often these days, but that's by choice, however I was seeing the facebook notifications far too often. Ending up on facebook far too often for what I wanted.
Recently I decided to try out some tracking systems that include 1/0 metrics. It looks something like this:
I wanted this in a place where I could see it and fill it out every day, and at the same time I began to question why I have my facebook app on my front page. This link is now on my front page and I easily fill it out once a day (a win for a habit successfully implemented).
The concept that I want to impart today is that the causation goes the wrong way. Instead of wanting apps that I regularly use on my front page so that I can easily access them - I want apps that I want to use regularly on my front page. That way I will tend to develop habits of regularly using them instead of the other ones.
Fridge
This applies to the refrigerator too. Instead of the things you use and eat all the time being at the front (assuming they might be different), you want the foods that you want to eat most readily accessible and at the front. If this means healthy foods at the front - do that. If this means having a fruit bowl on the table - do that.
TV
This applies to TV too. If you find book-reading more interesting than TV watching but find yourself watchi |
b076dcc2-aa5d-4299-bdb7-5bf68f1d580f | StampyAI/alignment-research-dataset/lesswrong | LessWrong | New survey: 46% of Americans are concerned about extinction from AI; 69% support a six-month pause in AI development
YouGov America released a survey of 20,810 American adults. Highlights below. Note that I didn't run any statistical tests, so any claims of group differences are just "eyeballed."
* 46% say that they are "very concerned" or "somewhat concerned" about the possibility that AI will cause the end of the human race on Earth (with 23% "not very concerned, 17% not concerned at all, and 13% not sure).
* There do not seem to be meaningful differences by region, gender, or political party.
* Younger people seem more concerned than older people.
* Black individuals appear to be somewhat more concerned than people who identified as White, Hispanic, or Other.
Furthermore, 69% of Americans appear to support a [six-month pause in "some kinds of AI development"](https://today.yougov.com/topics/technology/survey-results/daily/2023/04/03/ad825/2). Note that there doesn't seem to be a clear effect of age or race for this question. (Particularly if you lump "strongly support" and "somewhat support" into the same bucket). Note also that the question mentions that 1000 tech leaders signed an open letter calling for a pause and cites their concern over "profound risks to society and humanity", which may have influenced participants' responses.
In my quick skim, I haven't been able to find details about the survey's methodology (see here for info about [YouGov's general methodology](https://today.yougov.com/about/panel-methodology/)) or the credibility of YouGov (**EDIT**: Several people I trust have told me that YouGov is credible, well-respected, and widely quoted for US polls).
See also:
* [The public supports regulating AI for safety](https://www.lesswrong.com/posts/M3iPAmxZwy4gPXdXw/the-public-supports-regulating-ai-for-safety)
* [The Overton Window widens: Examples of AI risk in the media](https://www.lesswrong.com/posts/SvwuduvpsKtXkLnPF/the-overton-window-widens-examples-of-ai-risk-in-the-media)
* [Spreading messages to help with the most important century](https://www.lesswrong.com/posts/2y4BkkQwsG7n2Ensc/spreading-messages-to-help-with-the-most-important-century)
* [Surveys of US public opinion on AI](https://wiki.aiimpacts.org/doku.php?id=responses_to_ai:public_opinion_on_ai:surveys_of_public_opinion_on_ai:surveys_of_us_public_opinion_on_ai) |
e849705b-3564-480c-ad2c-1e03f9eb8a32 | trentmkelly/LessWrong-43k | LessWrong | The challenges of bringing up AIs
At the current AGI-12 conference, some designers have been proponents of keeping AGI's safe by bringing them up in human environments, providing them with interactions and feedback in a similar way to how we bring up human children. Obviously that approach would fail for a fully smart AGI with its own values - it would pretend to follow our values for as long as it needed, and then defect. However, some people have confidence if we started with a limited, dumb AGI, then we could successfully inculcate our values in this way (a more sophisticated position would be that though this method would likely fail, it's more likely to succeed than a top-down friendliness project!).
The major criticism of this approach is that it anthropomorphises the AGI - we have a theory of children's minds, constructed by evolution, culture, and our own child-rearing experience. And then we project this on the alien mind of the AGI, assuming that if the AGI presents behaviours similar to a well-behaved child, then it will become a moral AGI. The problem is that we don't know how alien the AGI's mind will be, and if our reinforcement is actually reinforcing the right thing. Specifically, we need to be able to find some way of distinguishing between:
1. An AGI being trained to be friendly.
2. An AGI being trained to lie and conceal.
3. An AGI that will behave completely differently once out of the training/testing/trust-building environment.
4. An AGI that forms the wrong categories and generalisations (what counts as "human" or "suffering", for instance), because it lacks human-shared implicit knowledge that was "too obvious" for us to even think of training it on.
|
833ced98-eaa8-46fd-ae3e-0d70935f6fd8 | trentmkelly/LessWrong-43k | LessWrong | Language Models are a Potentially Safe Path to Human-Level AGI
The core argument: language models are more transparent and less prone to develop agency and superintelligence
I argue that compared to alternative approaches such as open-ended reinforcement learning, the recent paradigm of achieving human-level AGI with language models has the potential to be relatively safe. There are three main reasons why I believe LM-based AGI systems could be safe:
1. They operate on text that is intelligible to humans, which makes them relatively interpretable and easier to monitor.
2. They are subject to weaker pressure to surpass human-level capabilities than systems trained on more open-ended tasks, such as those pursued by reinforcement learning.
3. Since language models are trained as predictors, there is weaker pressure for them to develop agentic behavior.
I acknowledge that these arguments have been criticized. I will try to defend these statements, delving into the nuances and explaining how I envision relatively safe LM-based AGI systems. I want to note upfront that although I believe this paradigm is safer than other alternatives I've come across, I still think it poses significant dangers, so I’m not suggesting we should all just chill. It's also difficult to reason about these sorts of things in the abstract, and there's a good chance I may be overlooking critical considerations.
Human-level AGI might be reachable with existing language models
There are indications that GPT-4 could be considered an early form of AGI. But even though it surpasses human level over many standardized tests, as a general-purpose AGI it’s not yet at human level, at least in its vanilla form. For example, it’s still not very good at separating facts from fiction, and I wouldn’t give it full access to my email and social media accounts and ask it to throw a birthday party for me.
To address these limitations, there's a project underway that seeks to develop more capable and agentic AI systems based on language models through chained operations s |
7ee72fe1-1466-44da-9b92-14c64d7a1356 | trentmkelly/LessWrong-43k | LessWrong | Intellectual Progress Inside and Outside Academia
This post is taken from a recent facebook conversation that included Wei Dai, Eliezer Yudkowsky, Vladimir Slepnev, Stuart Armstrong, Maxim Kesin, Qiaochu Yuan and Robby Bensinger, about the ability of academia to do the key intellectual progress required in AI alignment.
[THE ABOVE PEOPLE ALL GAVE PERMISSION TO HAVE THEIR COMMENTS COPIED HERE. SOME COMMENTERS REQUESTED THEIR REPLIES NOT BE MADE PUBLIC, AND THEIR COMMENT THREADS WERE NOT COPIED OVER.]
----------------------------------------
INITIAL THREAD
Wei Dai:
> Eliezer, can you give us your take on this discussion between me, Vladimir Slepnev, and Stuart Armstrong? I'm especially interested to know if you have any thoughts on what is preventing academia from taking or even recognizing certain steps in intellectual progress (e.g., inventing anything resembling Bitcoin or TDT/UDT) that non-academics are capable of. What is going on there and what do we need do to avoid possibly suffering the same fate? See this and this.
Eliezer Yudkowsky:
> It's a deep issue. But stating the obvious is often a good idea, so to state the obvious parts, we're looking at a lot of principal-agent problems, Goodhart's Law, bad systemic incentives, hypercompetition crowding out voluntary contributions of real work, the blind leading the blind and second-generation illiteracy, etcetera. There just isn't very much in the academic system that does promote any kind of real work getting done, and a lot of other rewards and incentives instead. If you wanted to get productive work done inside academia, you'd have to ignore all the incentives pointing elsewhere, and then you'd (a) be leading a horrible unrewarded life and (b) you would fall off the hypercompetitive frontier of the major journals and (c) nobody else would be particularly incentivized to pay attention to you except under unusual circumstances. Academia isn't about knowledge. To put it another way, although there are deep things to say about the way in which bad incentive |
04cb10c2-d480-4e2c-9b03-f159f56286c6 | trentmkelly/LessWrong-43k | LessWrong | Tesla Model 3 Review
This weekend I was in San Luis Obispo for a gig, about halfway between SF and LA. It's possible to fly into SBP, but since I was traveling with two of my kids it was a lot cheaper to fly into SFO and drive down, and only slightly slower.
I'm signed up with Hertz's reward program ("Gold") and one of the benefits is that you pick out your own car. When I got to SFO, there were several Tesla Model 3s in the "Gold" area. This was somewhat surprising—I had only paid for a small sedan ("B") and the Teslas are fancy ("E7")—but it seemed like it would be interesting to try one out and I liked the idea of not paying for gas. I got us loaded up but when I got to the exit the attendant didn't believe I'd gotten the car from the designated area, and said it would be a $50/day upgrade. I went back to the area, identified an apparently identical Model 3, and asked if I could have that one. They said yes, I moved kids and luggage over, and we were off.
I forgot to take a picture of the car—what sort of review is this?—but here's a picture of my kids enjoying the best seats on the SFO people mover.
This was my first time driving an electric car and the first thing I noticed was lifting your foot off the accelerator is very different. In a gas car you coast, slowing down a little from engine braking, but apparently in EVs you get regenerative braking. There's a sense in which this doesn't matter, and is just a question of where to put the "coast" point—at no pedal vs pushing down a little—but it doesn't immediately seem better to me.
The first evening was a short drive to San Jose. Getting out of the car, it was not obvious how to turn it off. Reading online later I learned that all you do is walk away, which is nice and automatic. But you'd only know if someone told you or you looked it up, and late at night after a cross country flight with two tired kids while poking at unfamiliar menu options on the dashboard tablet I was pretty frustrated with their minimally discoverab |
5e8b680b-a823-48db-9075-be11e9c8867d | trentmkelly/LessWrong-43k | LessWrong | Academia as a career option, its social value, and alternatives
Many of the high school and college students who contacted us at Cognito Mentoring were looking for advice were considering going into academia. The main draw to them was the desire to learn specific subjects and explore ideas in greater depth. As a result, we've been investigating academia as a career option and also considering what alternatives there may be to academia that fulfill the same needs but provide better pay and/or generate more social value. The love of ideas and epistemic exploration is shared by many of the people at Less Wrong, including those who are not in academia. So I'm hoping that people will share their own perspectives in the comments. That'll help us as well as the many LessWrong lurkers interested in academia.
I'm eager to hear about what considerations you used when weighing academia against other career options, and how you came to your decision. Incidentally, there are a number of great answers to the Quora question Why did you leave academia?, but there's probably many thoughts people have here that aren't reflected in the Quora answers. I've also written up a detailed review of academia as a career option on the info wiki for Cognito Mentoring here (long read), and I'd also love feedback on the validity of the points I make there.
Many of our advisees as well as the LessWrong readership at large are interested in choosing careers based on the social value generated by these careers. (This is evidenced in the strong connection between the LessWrong and effective altruism communities). What are your thoughts on that front? Jonah and I have collaboratively written a page on the social value of academia. Our key point is that research academia is higher value than alternative careers only in cases where either the person has a chance of making big breakthroughs in the area, or if the area of research itself is high-value. Examples of the latter may include machine learning (we're just starting on investigating this) and (arguably) biom |
9da3723f-beb3-444e-9c20-3b804a61f4be | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Artificial Addition
Today's post, Artificial Addition was originally published on 20 November 2007. A summary (taken from the LW wiki):
> If you imagine a world where people are stuck on the "artifical addition" (i.e. machine calculator) problem, the way people currently are stuck on artificial intelligence, and you saw them trying the same popular approaches taken today toward AI, it would become clear how silly they are. Contrary to popular wisdom (in that world or ours), the solution is not to "evolve" an artificial adder, or invoke the need for special physics, or build a huge database of solutions, etc. -- because all of these methods dodge the crucial task of understanding what addition involves, and instead try to dance around it. Moreover, the history of AI research shows the problems of believing assertions one cannot re-generate from one's own knowledge.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Conjuring an Evolution to Serve You, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
6daec708-7b03-443d-8b7a-8a928745558e | trentmkelly/LessWrong-43k | LessWrong | Lessons From TryContra
As I wrote yesterday, I recently brought my little contra dance search tool up to date. I realized I've been running it for ten years now, which seems like a good time to look back over the experience for lessons.
When I built it in 2013 I wrote:
> Experienced dancers know how to use ContraDanceLinks.com, Dance Gypsy, and the DanceDB to find places they can go contra dancing, but those sites are too complex and confusing for me to want to give to a new dancer.
Here's what those dance-community focused sites looked like:
And here's what I built:
Almost the same as it is today:
I wanted to build something easy to use and was willing to give up pretty much everything else, and I think I succeeded at that. You put in your location, get a list of nearby dances, and click through to their sites for more information.
There were several other choices I made, primarily out of laziness:
* The site is completely static: HTML, CSS, JS. The searching happens in your browser, since it only takes a few kB to store all the contra dances in the country.
* No dependencies: vanilla JS using browser APIs only. Not even minified.
* It's just a directory. I don't host pages for dances.
* No accounts. If you want to update the listing for your dance you can email me.
* Minimal information. Just a link, city, weekday, approximate frequency, and whether it's gender-free. No "1st and 3rd Sundays", pricing, addresses, hours, performers, etc. More details would mean more information to collect, but more importantly it would mean more information to go stale.
In retrospect I feel like these decisions turned out well: the site requires very little upkeep. I can leave it alone for years and it will keep chugging along, and every so often I get an email with a correction and make a small change to a text file.
(This has also given me a lot of experience at reading contra dance websites. I wrote advice for people making these sites in 2013, which I th |
5dff8aa5-c591-46e0-aaf5-34128223b110 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Announcing AlignmentForum.org Beta
We've just launched the beta for [AlignmentForum.org](https://www.alignmentforum.org/).
Much of the value of LessWrong has come from the development of technical research on AI Alignment. In particular, having those discussions be in an accessible place has allowed newcomers to get up to speed and involved. But the alignment research community has at least some needs that are best met with a semi-private forum.
For the past few years, [agentfoundations.org](https://agentfoundations.org/item?id=1817) has served as a space for highly technical discussion of AI safety. But some aspects of the site design have made it a bit difficult to maintain, and harder to onboard new researchers. Meanwhile, as the AI landscape has shifted, it seemed valuable to expand the scope of the site. Agent Foundations is one particular paradigm with respect to AGI alignment, and it seemed important for researchers in other paradigms to be in communication with each other.
So for several months, the LessWrong and AgentFoundations teams have been discussing the possibility of using the LW codebase as the basis for a new alignment forum. Over the past couple weeks we've gotten ready for a closed beta test, both to iron out bugs and (more importantly) get feedback from researchers on whether the overall approach makes sense.
The current features of the Alignment Forum (subject to change) are:
* A small number of admins can invite new members, granting them posting and commenting permissions. This will be the case during the beta - the exact mechanism of curation after launch is still under discussion.
* When a researcher posts on AlignmentForum, the post is shared with LessWrong. On LessWrong, anyone can comment. On AlignmentForum, only AF members can comment. (AF comments are *also* crossposted to LW). The intent is for AF members to have a focused, technical discussion, while still allowing newcomers to LessWrong to see and discuss what's going on.
* AlignmentForum posts and comments on LW will be marked as such.
* AF members will have a separate karma total for AlignmentForum (so AF karma will more closely represent what technical researchers think about a given topic).
* On AlignmentForum, only AF Karma is visible. (note: not currently implemented but will be by end of day)
* On LessWrong, AF Karma will be displayed (smaller) alongside regular karma.
* If a commenter on LessWrong is making particularly good contributions to an AF discussion, an AF Admin can tag the comment as an AF comment, which will be visible on the AlignmentForum. The LessWrong user will then have *voting privileges* (but not necessarily posting privileges), allowing them to start to accrue AF karma, and to vote on AF comments and threads.
We’ve currently copied over some LessWrong posts that seemed like a good fit, and invited a few people to write posts today. (These don’t necessarily represent the longterm vision of the site, but seemed like a good way to begin the beta test)
This is a fairly major experiment, and we’re interested in feedback both from AI alignment researchers (who we’ll be reaching out to more individually in the next two weeks) and LessWrong users, about the overall approach and the integration with LessWrong. |
2aafc466-7178-4f78-a3a8-92e0bada64ac | trentmkelly/LessWrong-43k | LessWrong | Some common confusion about induction heads
Epistemic status: conceptual discussion and opinions informed by doing 6 months of interpretability research at Redwood Research and exchanging with other researchers, but I’m just speaking for myself.
Induction heads are defined twice by Anthropic.
1. The first time as a mechanism in 2L attention-only transformers
2. A second time as a behavioral description on repeated random sequences of tokens
However, these two definitions rely on distinct sources of evidence and create confusion, as their difference is not always acknowledged when people cite these papers. The mechanistic definition applies to toy language models, while the behavioral definition is a useful yet incomplete characterization of attention heads.
I think that many people are in fact confused by this: I have talked to many people who aren’t clear on the fact that these two concepts are different, and incorrectly believe that (e.g.) the mechanism of induction heads in larger language models has been characterized.
More specifically, the two Anthropic papers introduce the following two distinct definitions of induction heads:
1. Mechanistic: The first definition, introduced by Elhage et al., describes a behavior in a 2 layer attention-only model (copying a token given a matching prefix) and a minimal mechanism to perform this behavior (a set of paths in the computational graph and a human interpretation of the transformation along those paths). Empirically, this mechanism seems to be the best possible short description of what those heads are doing (i.e. if you have to choose a subgraph made of a single path as input for the keys, queries, and values of these heads, the induction circuit is likely to be the one that recovers the most loss). But this explanation does not encompass everything these heads do. In reality, many more paths are used than the one described (see Redwood’s causal scrubbing results on induction heads) and the function of the additional paths is unclear. I don’t know wh |
3a2ada82-21c3-4454-9149-3adf2f47c084 | trentmkelly/LessWrong-43k | LessWrong | Merry Christmas Everyone!
This was my Christmas Card from my parents this year:
AI-generated and not at all terrifying. Happy New Year! |
bbeea535-3d86-42c4-935f-5aec3101d314 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Berlin Meetup
Discussion article for the meetup : Berlin Meetup
WHEN: 28 September 2012 07:30:00PM (+0200)
WHERE: Spreegold, Hufelandstraße 20, 10407 Berlin
Plans from last meetup:
* discuss http://lesswrong.com/lw/4su/how_to_be_happy/
* WingedViper wanted to look at and maybe prepare something from the LW meetup booklet (http://lesswrong.com/lw/crs/how_to_run_a_successful_less_wrong_meetup/)
* another round of spuckblase's prediction game
The scheduling poll indicated that there should be at least 3-5 people. I'll bring a sign to point us out to anyone dropping by spontaneously.
See you there!
Discussion article for the meetup : Berlin Meetup |
ffe46518-03f8-405e-9db0-467619e1bd40 | StampyAI/alignment-research-dataset/blogs | Blogs | Jobs that can help with the most important century
*Click lower right to download or find on Apple Podcasts, Spotify, Stitcher, etc.*
Let’s say you’re convinced that AI could make this the [most important century of all time for humanity](https://www.cold-takes.com/most-important-century/). What can you do to help things go well instead of poorly?
I think **the biggest opportunities come from a full-time job** (and/or the money you make from it). I think people are generally far better at their jobs than they are at anything else.
This piece will list the jobs I think are especially high-value. I expect things will change (a lot) from year to year - this is my picture at the moment.
Here’s a summary:
| | |
| --- | --- |
| **Role** | **Skills/assets you'd need** |
| [Research and engineering on AI safety](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#research-and-engineering) | Technical ability (but not necessarily AI background)
|
| [Information security to reduce the odds powerful AI is leaked](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#information-security) | Security expertise or willingness/ability to start in junior roles (likely not AI)
|
| [Other roles at AI companies](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#other-roles-at-ai-companies) | Suitable for generalists (but major pros and cons)
|
| [Govt and govt-facing think tanks](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#government-and-government-facing) | Suitable for generalists (but probably takes a long time to have impact)
|
| [Jobs in politics](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#politics) | Suitable for generalists if you have a clear view on which politicians to help
|
| [Forecasting to get a better handle on what’s coming](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#forecasting) | Strong forecasting track record (can be pursued part-time)
|
| ["Meta" careers](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#meta-careers) | Misc / suitable for generalists
|
| [Low-guidance options](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#low-guidance-jobs) | These ~only make sense if you read & instantly think "That's me"
|
A few notes before I give more detail:
* These jobs aren’t the be-all/end-all. I expect a lot to change in the future, including a general increase in the number of helpful jobs available.
* Most of today’s opportunities are concentrated in the US and UK, where the biggest AI companies (and AI-focused nonprofits) are. This may change down the line.
* Most of these aren’t jobs where you can just take instructions and apply narrow skills.
+ The issues here are tricky, and your work will almost certainly be useless (or harmful) according to someone.
+ I recommend forming your own views on the key risks of AI - and/or working for an organization whose leadership you’re confident in.* Staying open-minded and adaptable is crucial.
+ I think it’s bad to rush into a mediocre fit with one of these jobs, and better (if necessary) to stay out of AI-related jobs while skilling up and waiting for a great fit.
+ I don’t think it’s helpful (and it could be harmful) to take a fanatical, “This is the most important time ever - time to be a hero” attitude. Better to work intensely but sustainably, stay mentally healthy and make good decisions.
The [first section](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#recap) of this piece will recap my basic picture of the major risks, and the promising ways to reduce these risks (feel free to skip if you think you’ve got a handle on this).
The [next section](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#jobs-that-can-help) will elaborate on the options in the table above.
After that, I’ll talk about [some of the things you can do if you aren’t ready](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#other-things-you-can-do) for a full-time career switch yet, and give some [general advice for avoiding doing harm and burnout](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#some-general-advice).
Recapping the major risks, and some things that could help
----------------------------------------------------------
This is a quick recap of the major risks from transformative AI. For a longer treatment, see [How we could stumble into an AI catastrophe](https://www.cold-takes.com/how-we-could-stumble-into-ai-catastrophe/), and for an even longer one see the [full series](https://www.cold-takes.com/tag/implicationsofmostimportantcentury/). To skip to the next section, click [here](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#jobs-that-can-help).
**The backdrop: transformative AI could be developed in the coming decades.** If we develop AI that can [automate all the things humans do to advance science and technology](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/), this could cause [explosive technological progress](https://www.cold-takes.com/most-important-century/#the-long-run-future-could-come-faster-than-we-think) that could bring us more quickly than most people imagine to a radically unfamiliar future.
Such AI could also be capable of [defeating all of humanity combined](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/), if it were pointed toward that goal.
(Click to expand) The most important century
In the [most important century](https://www.cold-takes.com/most-important-century/) series, I argued that the 21st century could be the most important century ever for humanity, via the development of advanced AI systems that could dramatically speed up scientific and technological advancement, getting us more quickly than most people imagine to a deeply unfamiliar future.
I focus on a hypothetical kind of AI that I call [PASTA](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/), or Process for Automating Scientific and Technological Advancement. PASTA would be AI that can essentially **automate all of the human activities needed to speed up scientific and technological advancement.**
Using a [variety of different forecasting approaches](https://www.cold-takes.com/where-ai-forecasting-stands-today/), I argue that PASTA seems more likely than not to be developed this century - and there’s a decent chance (more than 10%) that we’ll see it within 15 years or so.
I argue that the consequences of this sort of AI could be enormous: an [explosion in scientific and technological progress](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/#explosive-scientific-and-technological-advancement). This could get us more quickly than most imagine to a radically unfamiliar future.
I’ve also [argued](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/) that AI systems along these lines could defeat all of humanity combined, if (for whatever reason) they were aimed toward that goal.
For more, see the [most important century](https://www.cold-takes.com/most-important-century/) landing page. The series is available in many formats, including audio; I also provide a summary, and links to podcasts where I discuss it at a high level.
(Click to expand) How could AI systems defeat humanity?
A [previous piece](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/) argues that AI systems could defeat all of humanity combined, if (for whatever reason) they were aimed toward that goal.
By defeating humanity, I mean gaining control of the world so that AIs, not humans, determine what happens in it; this could involve killing humans or simply “containing” us in some way, such that we can’t interfere with AIs’ aims.
One way this could happen would be via “superintelligence” It’s imaginable that a single AI system (or set of systems working together) could:
* Do its own research on how to build a better AI system, which culminates in something that has incredible other abilities.
* Hack into human-built software across the world.
* Manipulate human psychology.
* Quickly generate vast wealth under the control of itself or any human allies.
* Come up with better plans than humans could imagine, and ensure that it doesn't try any takeover attempt that humans might be able to detect and stop.
* Develop advanced weaponry that can be built quickly and cheaply, yet is powerful enough to overpower human militaries.
But even if “superintelligence” never comes into play - even if any given AI system is *at best* equally capable to a highly capable human - AI could collectively defeat humanity. The piece explains how.
The basic idea is that humans are likely to deploy AI systems throughout the economy, such that they have large numbers and access to many resources - and the ability to make copies of themselves. From this starting point, AI systems with human-like (or greater) capabilities would have a number of possible ways of getting to the point where their total population could outnumber and/or out-resource humans.
More: [AI could defeat all of us combined](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/)
**Misalignment risk: AI could end up with dangerous [aims](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/) of its own.**
* If this sort of AI is developed using the kinds of [trial-and-error-based](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/#Box3) techniques that are common today, I think it’s likely that it will end up “aiming” for particular states of the world, much like a chess-playing AI “aims” for a checkmate position - making choices, calculations and plans to get particular types of outcomes, even when doing so requires deceiving humans.
* I think it will be difficult - by default - to ensure that AI systems are aiming for *what we (humans) want them to aim for*, as opposed to gaining power for ends of their own.
* If AIs have ambitious aims of their own - and are numerous and/or capable enough to overpower humans - I think we have a serious risk that AIs will take control of the world and disempower humans entirely.
(Click to expand) Why would AI "aim" to defeat humanity?
A [previous piece](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/) argued that if today’s AI development methods lead directly to powerful enough AI systems, disaster is likely by default (in the absence of specific countermeasures).
In brief:
* Modern AI development is essentially based on “training” via trial-and-error.
* If we move forward incautiously and ambitiously with such training, and if it gets us all the way to very powerful AI systems, then such systems will likely end up *aiming for certain states of the world* (analogously to how a chess-playing AI aims for checkmate).
* And these states will be *other than the ones we intended*, because our trial-and-error training methods won’t be accurate. For example, when we’re confused or misinformed about some question, we’ll reward AI systems for giving the wrong answer to it - unintentionally training deceptive behavior.
* We should expect disaster if we have AI systems that are both (a) [powerful enough](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/) to defeat humans and (b) aiming for states of the world that we didn’t intend. (“Defeat” means taking control of the world and doing what’s necessary to keep us out of the way; it’s unclear to me whether we’d be literally killed or just forcibly stopped[1](https://www.cold-takes.com/p/51b33fd6-2f1e-40bd-9d2c-2cfe2ebd5fc5#fn1) from changing the world in ways that contradict AI systems’ aims.)
More: [Why would AI "aim" to defeat humanity?](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/)
**Competitive pressures, and ambiguous evidence about the risks, could make this situation very dangerous.** In a [previous piece](https://www.cold-takes.com/how-we-could-stumble-into-ai-catastrophe/), I lay out a hypothetical story about how the world could stumble into catastrophe. In this story:
* There are warning signs about the risks of misaligned AI - but there’s a lot of ambiguity about just how big the risk is.
* Everyone is furiously racing to be first to deploy powerful AI systems.
* We end up with a big risk of deploying dangerous AI systems throughout the economy - which means a risk of AIs disempowering humans entirely.
* And even if we navigate *that* risk - even if AI behaves as intended - this could be a disaster if the most powerful AI systems end up concentrated in the wrong hands (something I [think is reasonably likely](https://www.cold-takes.com/transformative-ai-issues-not-just-misalignment-an-overview/#power-imbalances) due to the potential for power imbalances). There are [other risks](https://www.cold-takes.com/transformative-ai-issues-not-just-misalignment-an-overview/) as well.
(Click to expand) Why AI safety could be hard to measure
In previous pieces, I argued that:
* If we develop powerful AIs via ambitious use of the “black-box trial-and-error” common in AI development today, then there’s a substantial risk that:
+ These AIs will develop [unintended aims](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/) (states of the world they make calculations and plans toward, as a chess-playing AI "aims" for checkmate);
+ These AIs could deceive, manipulate, and even [take over the world from humans entirely](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/) as needed to achieve those aims.
+ People today are doing AI safety research to prevent this outcome, but such research has a [number of deep difficulties:](https://www.cold-takes.com/ai-safety-seems-hard-to-measure/)
| |
| --- |
| **“Great news - I’ve tested this AI and it looks safe.”** Why might we still have a problem?
|
| *Problem* | *Key question* | *Explanation* |
| The **Lance Armstrong problem** | Did we get the AI to be **actually safe** or **good at hiding its dangerous actions?** | When dealing with an intelligent agent, it’s hard to tell the difference between “behaving well” and “*appearing* to behave well.”
When professional cycling was cracking down on performance-enhancing drugs, Lance Armstrong was very successful and seemed to be unusually “clean.” It later came out that he had been using drugs with an unusually sophisticated operation for concealing them.
|
| The **King Lear problem** | The AI is **(actually) well-behaved when humans are in control.** Will this transfer to **when AIs are in control?** | It's hard to know how someone will behave when they have power over you, based only on observing how they behave when they don't.
AIs might behave as intended as long as humans are in control - but at some future point, AI systems might be capable and widespread enough to have opportunities to [take control of the world entirely](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/). It's hard to know whether they'll take these opportunities, and we can't exactly run a clean test of the situation.
Like King Lear trying to decide how much power to give each of his daughters before abdicating the throne.
|
| The **lab mice problem** | **Today's "subhuman" AIs are safe.**What about **future AIs with more human-like abilities?** | Today's AI systems aren't advanced enough to exhibit the basic behaviors we want to study, such as deceiving and manipulating humans.
Like trying to study medicine in humans by experimenting only on lab mice.
|
| The **first contact problem** | Imagine that **tomorrow's "human-like" AIs are safe.** How will things go **when AIs have capabilities far beyond humans'?** | AI systems might (collectively) become vastly more capable than humans, and it's ... just really hard to have any idea what that's going to be like. As far as we know, there has never before been anything in the galaxy that's vastly more capable than humans in the relevant ways! No matter what we come up with to solve the first three problems, we can't be too confident that it'll keep working if AI advances (or just proliferates) a lot more.
Like trying to plan for first contact with extraterrestrials (this barely feels like an analogy).
|
(Click to expand) Power imbalances, and other risks beyond misaligned AI
I’ve argued that AI could cause a [dramatic acceleration in the pace of scientific and technological advancement](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/#explosive-scientific-and-technological-advancement).
One way of thinking about this: perhaps (for reasons I’ve [argued previously](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/#explosive-scientific-and-technological-advancement)) AI could enable the equivalent of hundreds of years of scientific and technological advancement in a matter of a few months (or faster). If so, then developing powerful AI a few months before others could lead to having technology that is (effectively) hundreds of years ahead of others’.
Because of this, it’s easy to imagine that AI could lead to big power imbalances, as whatever country/countries/coalitions “lead the way” on AI development could become far more powerful than others (perhaps analogously to when a few smallish European states took over much of the rest of the world).
I think things could go very badly if the wrong country/countries/coalitions lead the way on transformative AI. At the same time, I’ve expressed concern that people might overfocus on this aspect of things vs. other issues, for a number of reasons including:
* *I think people naturally get more animated about "helping the good guys beat the bad guys" than about "helping all of us avoid getting a universally bad outcome, for impersonal reasons such as 'we designed sloppy AI systems' or 'we created a dynamic in which haste and aggression are rewarded.'"** *I expect people will tend to be overconfident about which countries, organizations or people they see as the "good guys."*
(More [here](https://www.cold-takes.com/making-the-best-of-the-most-important-century/#why-i-fear-).)
There are also dangers of powerful AI being too widespread, rather than too concentrated. In [The Vulnerable World Hypothesis](https://nickbostrom.com/papers/vulnerable.pdf), Nick Bostrom contemplates potential future dynamics such as “advances in DIY biohacking tools might make it easy for anybody with basic training in biology to kill millions.” In addition to avoiding worlds where AI capabilities end up concentrated in the hands of a few, it could also be important to avoid worlds in which they diffuse too widely, too quickly, before we’re able to assess the risks of widespread access to technology far beyond today’s.
I discuss these and a number of other AI risks in a previous piece: [Transformative AI issues (not just misalignment): an overview](https://www.cold-takes.com/transformative-ai-issues-not-just-misalignment-an-overview/)
**I’ve laid out several ways to reduce the risks (color-coded since I’ll be referring to them throughout the piece):**
**Alignment research.**Researchers are working on ways to design AI systems that are *both* (a) “aligned” in the sense that they don’t have [unintended aims of their own](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/); (b) very powerful, to the point where they can be competitive with the best systems out there.
* I’ve laid out three [high-level hopes](https://www.cold-takes.com/high-level-hopes-for-ai-alignment/) for how - using techniques that are known today - we might be able to develop AI systems that are both aligned and powerful.
* These techniques wouldn’t necessarily work indefinitely, but they might work long enough so that we can [use early safe AI systems to make the situation much safer](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#defensive-deployment) (by automating huge amounts of further alignment research, by helping to demonstrate risks and make the case for greater caution worldwide, etc.)
* (A footnote explains how I’m using “aligned” vs. “safe.”[1](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn1))
(Click to expand) High-level hopes for AI alignment
A [previous piece](https://www.cold-takes.com/high-level-hopes-for-ai-alignment/) goes through what I see as three key possibilities for building powerful-but-safe AI systems.
It frames these using Ajeya Cotra’s [young businessperson](https://www.cold-takes.com/why-ai-alignment-could-be-hard-with-modern-deep-learning/#analogy-the-young-ceo) analogy for the core difficulties. In a nutshell, once AI systems get capable enough, it could be hard to test whether they’re safe, because they might be able to deceive and manipulate us into getting the wrong read. Thus, trying to determine whether they’re safe might be something like “being an eight-year-old trying to decide between adult job candidates (some of whom are manipulative).”
Key possibilities for navigating this challenge:
* **Digital neuroscience**: perhaps we’ll be able to read (and/or even rewrite) the “digital brains” of AI systems, so that we can know (and change) what they’re “aiming” to do directly - rather than having to infer it from their behavior. (Perhaps the eight-year-old is a mind-reader, or even a young [Professor X](https://en.wikipedia.org/wiki/Professor_X#Powers_and_abilities).)
* **Limited AI**: perhaps we can make AI systems safe by making them *limited* in various ways - e.g., by leaving certain kinds of information out of their training, designing them to be “myopic” (focused on short-run as opposed to long-run goals), or something along those lines. Maybe we can make “limited AI” that is nonetheless able to carry out particular helpful tasks - such as doing lots more research on how to achieve safety without the limitations. (Perhaps the eight-year-old can limit the authority or knowledge of their hire, and still get the company run successfully.)
* **AI checks and balances**: perhaps we’ll be able to employ some AI systems to critique, supervise, and even rewrite others. Even if no single AI system would be safe on its own, the right “checks and balances” setup could ensure that human interests win out. (Perhaps the eight-year-old is able to get the job candidates to evaluate and critique each other, such that all the eight-year-old needs to do is verify basic factual claims to know who the best candidate is.)
These are some of the main categories of hopes that are pretty easy to picture today. Further work on AI safety research might result in further ideas (and the above are not exhaustive - see my [more detailed piece](https://www.alignmentforum.org/posts/rCJQAkPTEypGjSJ8X/how-might-we-align-transformative-ai-if-it-s-developed-very), posted to the Alignment Forum rather than Cold Takes, for more).
**Standards and monitoring.**I see some hope for developing **standards that all potentially dangerous AI projects** (whether companies, government projects, etc.) **need to meet, and enforcing these standards globally.**
* Such standards could require strong demonstrations of safety, strong security practices, designing AI systems to be difficult to use for overly dangerous activity, etc.
* We don't need a perfect system or international agreement to get a lot of benefit out of such a setup. The goal isn’t just to buy time – it’s to change incentives, such that AI projects need to make progress on improving security, alignment, etc. in order to be profitable.
(Click to expand) How standards might be established and become national or international
I [previously](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#global-monitoring) laid out a possible vision on this front, which I’ll give a slightly modified version of here:
* Today’s leading AI companies could self-regulate by committing not to build or deploy a system that they can’t convincingly demonstrate is safe (e.g., see Google’s [2018 statement](https://www.theweek.in/news/sci-tech/2018/06/08/google-wont-deploy-ai-to-build-military-weapons-ichai.html), "We will not design or deploy AI in weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people”).
+ Even if some people at the companies would like to deploy unsafe systems, it could be hard to pull this off once the company has committed not to.
+ Even if there’s a lot of room for judgment in what it means to demonstrate an AI system is safe, having agreed in advance that [certain evidence](https://www.cold-takes.com/ai-safety-seems-hard-to-measure/) is *not* good enough could go a long way.* As more AI companies are started, they could feel soft pressure to do similar self-regulation, and refusing to do so is off-putting to potential employees, investors, etc.
* Eventually, similar principles could be incorporated into various government regulations and enforceable treaties.
* Governments could monitor for dangerous projects using regulation and even overseas operations. E.g., today the US monitors (without permission) for various signs that other states might be developing nuclear weapons, and might try to stop such development with methods ranging from threats of sanctions to [cyberwarfare](https://en.wikipedia.org/wiki/Stuxnet) or even military attacks. It could do something similar for any AI development projects that are using huge amounts of compute and haven’t volunteered information about whether they’re meeting standards.
**Successful, careful AI projects.** I think an AI company (or other project) can enormously improve the situation, if it can both (a) be one of the leaders in developing powerful AI; (b) prioritize doing (and using powerful AI for) [things that reduce risks](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#defensive-deployment), such as doing alignment research. (But don’t read this as ignoring the fact that AI companies [can do harm](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#other-roles-at-ai-companies) as well!)
(Click to expand) How a careful AI project could be helpful
In addition to using advanced AI to do AI safety research (noted above), an AI project could:
* Put huge effort into designing *tests* for signs of danger, and - if it sees danger signs in its own systems - warning the world as a whole.
* Offer deals to other AI companies/projects. E.g., acquiring them or exchanging a share of its profits for enough visibility and control to ensure that they don’t deploy dangerous AI systems.
* Use its credibility as the leading company to lobby the government for helpful measures (such as enforcement of a [monitoring-and-standards regime](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#global-monitoring)), and to more generally highlight key issues and advocate for sensible actions.
* Try to ensure (via design, marketing, customer choice, etc.) that its AI systems are not used for dangerous ends, and *are* used on applications that make the world safer and better off. This could include [defensive deployment](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#global-monitoring) to reduce risks from other AIs; it could include using advanced AI systems to help it gain clarity on how to get a good outcome for humanity; etc.
An AI project with a dominant market position could likely make a huge difference via things like the above (and probably via many routes I haven’t thought of). And even an AI project that is merely *one of several leaders* could have enough resources and credibility to have a lot of similar impacts - especially if it’s able to “lead by example” and persuade other AI projects (or make deals with them) to similarly prioritize actions like the above.
A challenge here is that I’m envisioning a project with two arguably contradictory properties: being *careful* (e.g., prioritizing actions like the above over just trying to maintain its position as a profitable/cutting-edge project) and *successful* (being a profitable/cutting-edge project). In practice, it could be very hard for an AI project to walk the tightrope of being aggressive enough to be a “leading” project (in the sense of having lots of resources, credibility, etc.), while also prioritizing actions like the above (which mostly, with some exceptions, seem pretty different from what an AI project would do if it were simply focused on its technological lead and profitability).
**Strong security.** A key threat is that someone could steal major components of an AI system and deploy it incautiously. It could be extremely hard for an AI project to be robustly safe against having its AI “stolen.” But this could change, if there’s enough effort to work out the problem of how to secure a large-scale, powerful AI system.
(Click to expand) The challenging of securing dangerous AI
In [Racing Through a Minefield](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/), I described a "race" between cautious actors (those who take [misalignment risk](underline) seriously) and incautious actors (those who are focused on deploying AI for their own gain, and aren't thinking much about the dangers to the whole world). Ideally, cautious actors would collectively have more powerful AI systems than incautious actors, so they could take their time doing [alignment research](underline) and [other things](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#defensive-deployment) to try to make the situation safer for everyone.
But if incautious actors can steal an AI from cautious actors and rush forward to deploy it for their own gain, then the situation looks a lot bleaker. And unfortunately, it could be hard to protect against this outcome.
It's generally [extremely difficult](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/#fn15) to protect data and code against a well-resourced cyberwarfare/espionage effort. An AI’s “weights” (you can think of this sort of like its source code, though [not exactly](https://www.cold-takes.com/p/97d2a7b1-af2d-4dd4-b679-5ea8bb41c47d#fn4)) are potentially very dangerous on their own, and hard to get extreme security for. Achieving enough cybersecurity could require measures, and preparations, well beyond what one would normally aim for in a commercial context.
Jobs that can help
------------------
In this long section, I’ll list a number of jobs I wish more people were pursuing.
Unfortunately, I can’t give individualized help exploring one or more of these career tracks. Starting points could include [80,000 Hours](https://80000hours.org/) and various [other resources](https://www.aisafetysupport.org/resources/lots-of-links).
**Research and engineering careers.** You can contribute to alignment research as a researcher and/or software engineer (the line between the two can be fuzzy in some contexts).
There are (not necessarily easy-to-get) jobs along these lines at major AI labs, in established academic labs, and at independent nonprofits (examples in footnote).[2](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn2)
Different institutions will have very different approaches to research, very different environments and philosophies, etc. so it’s hard to generalize about what might make someone a fit. A few high-level points:
* It takes a lot of talent to get these jobs, but you shouldn’t assume that it takes years of experience in a particular field (or a particular degree).
+ I’ve seen a number of people switch over from other fields (such as physics) and become successful extremely quickly.
+ In addition to on-the-job training, there are independent programs specifically aimed at helping people skill up quickly.[3](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn3)* You also shouldn’t assume that these jobs are only for “scientist” types - there’s a substantial need for engineers, which I expect to grow.
* I think most people working on alignment consider a lot of *other* people’s work to be useless at best. This seems important to know going in for a few reasons.
+ You shouldn’t assume that all work is useless just because the first examples you see seem that way.
+ It’s good to be aware that whatever you end up doing, someone will probably dunk on your work on the Internet.
+ At the same time, you shouldn’t assume that your work is helpful because it’s “safety research.” It's worth investing a lot in understanding how any particular research you're doing could be helpful (and how it could fail).
- I’d even suggest taking regular dedicated time (a day every few months?) to pause working on the day-to-day and think about how your work fits into the big picture.+ For a sense of what work **I** think is most likely to be useful, I’d suggest my piece on why [AI safety seems hard to measure](https://www.cold-takes.com/ai-safety-seems-hard-to-measure/) - I’m most excited about work that directly tackles the challenges outlined in that piece, and I’m pretty skeptical of work that only looks good with those challenges assumed away. (Also see my piece on [broad categories of research I think have a chance to be highly useful](https://www.cold-takes.com/high-level-hopes-for-ai-alignment/), and some [comments from a while ago](https://docs.google.com/document/d/1vE8CrN2ap8lFm1IjNacVV2OJhSehrGi-VL6jITTs9Rg/edit#heading=h.go4iucw4wv9k) that I still mostly endorse.)
I also want to call out a couple of categories of research that are getting some attention today, but seem at least a bit under-invested in, even relative to alignment research:
* *Threat assessment research.*To me, there’s an important distinction between “Making AI systems safer” and “Finding out how dangerous they might end up being.” (Today, these tend to get lumped together under “alignment research.”)
+ A key approach to medical research is using *model organisms* - for example, giving cancer to mice, so we can see whether we’re able to cure them.
+ Analogously, one might deliberately (though carefully) design an AI system to [deceive and manipulate humans](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/), so we can (a) get a more precise sense of what kinds of training dynamics lead to deception and manipulation; (b) see whether existing safety techniques are effective countermeasures.
+ If we had concrete demonstrations of AI systems becoming deceptive/manipulative/power-seeking, we could potentially build more consensus for caution (e.g., standards and monitoring). Or we could imaginably produce evidence that the threat is *low*.[5](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn5)+ A couple of early examples of threat assessment research: [here](https://twitter.com/EthanJPerez/status/1604886089403346944) and [here](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=odFQXSYAAAAJ&sortby=pubdate&citation_for_view=odFQXSYAAAAJ:MXK_kJrjxJIC).* *Anti-misuse research.*
+ I’ve [written about](https://www.cold-takes.com/transformative-ai-issues-not-just-misalignment-an-overview/#power-imbalances) how we could face catastrophe even from *aligned* AI. That is - even if AI does what its human operators want it to be doing, maybe some of its human operators want it to be helping them build bioweapons, spread propaganda, etc.
+ But maybe it’s possible to *train AIs so that they’re hard to use for purposes like this* - a separate challenge from training them to avoid deceiving and manipulating their human operators.
+ In practice, a lot of the work done on this today ([example](https://twitter.com/PougetHadrien/status/1611008020644864001)) tends to get called “safety” and lumped in with alignment (and sometimes the same research helps with both goals), but again, I think it’s a distinction worth making.
+ I expect the earliest and easiest versions of this work to happen naturally as companies try to make their AI models fit for commercialization - but at some point it might be important to be making more intense, thorough attempts to prevent even very rare (but catastrophic) misuse.
**Information security careers.** There’s a big risk that a powerful AI system could be “stolen” via hacking/espionage, and this could make just about every kind of risk worse. I think it could be very challenging - but possible - for AI projects to be secure against this threat. (More [above.](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#Box_underline))
**I really think security is not getting enough attention from people concerned about AI risk, and I disagree with the idea that key security problems can be solved just by hiring from today’s security industry.**
* From what I’ve seen, AI companies have a lot of trouble finding good security hires. I think a lot of this is simply that security is [challenging](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/#fn15) and valuable, and demand for good hires (especially people who can balance security needs against practical needs) tends to swamp supply.
+ And yes, this means good security people are well-paid!* Additionally, AI could present unique security challenges in the future, because it requires protecting something that is simultaneously (a) fundamentally just software (not e.g. uranium), and hence very hard to protect; (b) potentially valuable enough that one could imagine very well-resourced state programs going all-out to steal it, with a breach having globally catastrophic consequences. I think trying to get out ahead of this challenge, by experimenting early on with approaches to it, could be very important.
* **It’s plausible to me that security is as important as alignment right now,** in terms of how much one more good person working on it will help.* And security is an easier path, because one can get mentorship from a large community of security people working on things other than AI.[6](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn6)* I think there’s a lot of potential value both in security *research* (e.g., developing new security techniques) and in simply working at major AI companies to help with their existing security needs.
* For more on this topic, see this [recent 80,000 hours report](https://80000hours.org/career-reviews/information-security/) and [this 2019 post by two of my coworkers](https://forum.effectivealtruism.org/posts/ZJiCfwTy5dC4CoxqA/information-security-careers-for-gcr-reduction).
**Other jobs at AI companies.** AI companies hire for a lot of roles, many of which don’t require any technical skills.
It’s a somewhat debatable/tricky path to take a role that isn’t focused specifically on safety or security. Some people believe[7](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn7) that you can do more harm than good this way, by helping companies push forward with building dangerous AI before the risks have gotten much attention or preparation - and I think this is a pretty reasonable take.
At the same time:
* You could argue something like: “Company X has potential to be a successful, careful AI project. That is, it’s likely to deploy powerful AI systems more carefully and helpfully than others would, and use them to reduce risks by automating alignment research and [other risk-reducing tasks](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#defensive-deployment). Furthermore, Company X is most likely to make a number of other decisions wisely as things develop. So, it’s worth accepting that Company X is speeding up AI progress, because of the hope that Company X can make things go better.” This obviously depends on how you feel about Company X compared to others!
* Working at Company X could also present opportunities to *influence* Company X. If you’re a valuable contributor and you are paying attention to the choices the company is making (and speaking up about them), you could affect the incentives of leadership.
+ I think this can be a useful thing to do in combination with the other things on this list, but I generally wouldn’t advise taking a job if this is one’s *main* goal.* Working at an AI company presents opportunities to become generally more knowledgeable about AI, possibly enabling a later job change to something else.
(Click to expand) How a careful AI project could be helpful
In addition to using advanced AI to do AI safety research (noted above), an AI project could:
* Put huge effort into designing *tests* for signs of danger, and - if it sees danger signs in its own systems - warning the world as a whole.
* Offer deals to other AI companies/projects. E.g., acquiring them or exchanging a share of its profits for enough visibility and control to ensure that they don’t deploy dangerous AI systems.
* Use its credibility as the leading company to lobby the government for helpful measures (such as enforcement of a [monitoring-and-standards regime](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#global-monitoring)), and to more generally highlight key issues and advocate for sensible actions.
* Try to ensure (via design, marketing, customer choice, etc.) that its AI systems are not used for dangerous ends, and *are* used on applications that make the world safer and better off. This could include [defensive deployment](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#global-monitoring) to reduce risks from other AIs; it could include using advanced AI systems to help it gain clarity on how to get a good outcome for humanity; etc.
An AI project with a dominant market position could likely make a huge difference via things like the above (and probably via many routes I haven’t thought of). And even an AI project that is merely *one of several leaders* could have enough resources and credibility to have a lot of similar impacts - especially if it’s able to “lead by example” and persuade other AI projects (or make deals with them) to similarly prioritize actions like the above.
A challenge here is that I’m envisioning a project with two arguably contradictory properties: being *careful* (e.g., prioritizing actions like the above over just trying to maintain its position as a profitable/cutting-edge project) and *successful* (being a profitable/cutting-edge project). In practice, it could be very hard for an AI project to walk the tightrope of being aggressive enough to be a “leading” project (in the sense of having lots of resources, credibility, etc.), while also prioritizing actions like the above (which mostly, with some exceptions, seem pretty different from what an AI project would do if it were simply focused on its technological lead and profitability).
[80,000 Hours](https://80000hours.org/) has a [collection of anonymous advice](https://80000hours.org/articles/ai-capabilities/) on how to think about the pros and cons of working at an AI company.
In a future piece, I’ll discuss what I think AI companies can be doing today to prepare for transformative AI risk. This could be helpful for getting a sense of what an unusually careful AI company looks like.
**Jobs in government and at government-facing think tanks.** I think there is a lot of value in providing quality advice to governments (especially the US government) on how to think about AI - both today’s systems and potential future ones.
I also think it could make sense to work on *other* technology issues in government, which could be a good path to working on AI later (I expect government attention to AI to grow over time).
People interested in careers like these can check out [Open Philanthropy’s Technology Policy Fellowships](https://www.openphilanthropy.org/open-philanthropy-technology-policy-fellowship/) and RAND Corporation's [Technology and Security Policy Fellows](https://www.rand.org/jobs/technology-security-policy-fellows.html).
One related activity that seems especially valuable: **understanding the state of AI in countries other than the one you’re working for/in** - particularly countries that (a) have a good chance of developing their own major AI projects down the line; (b) are difficult to understand much about by default.
* Having good information on such countries could be crucial for making good decisions, e.g. about moving cautiously vs. racing forward vs. trying to enforce safety standards internationally.
* I think good work on this front has been done by the [Center for Security and Emerging Technology](https://cset.georgetown.edu/)[8](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn8) among others.
A future piece will discuss other things I think governments can be doing today to prepare for transformative AI risk. I won’t have a ton of tangible recommendations quite yet, but I expect there to be more over time, especially if and when standards and monitoring frameworks become better-developed.
**Jobs in politics.** The previous category focused on advising governments; this one is about working on political campaigns, doing polling analysis, etc. to generally improve the extent to which sane and reasonable people are in power. Obviously, it’s a judgment call which politicians are the “good” ones and which are the “bad” ones, but I didn’t want to leave out this category of work.
**Forecasting.** I’m intrigued by organizations like [Metaculus](https://www.metaculus.com/questions/?show-welcome=true), [HyperMind](https://www.hypermind.com/), [Good Judgment](https://goodjudgment.com/),[9](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn9) [Manifold Markets](https://manifold.markets/), and [Samotsvety](https://samotsvety.org/) - all trying, in one way or another, to produce **good probabilistic forecasts (using generalizable methods**[10](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn10)**) about world events.**
If we could get good forecasts about questions like “When will AI systems be powerful enough to [defeat all of humanity?](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/)” and “Will AI safety research in category X be successful?”, this could be useful for helping people make good decisions. (These questions seem very hard to get good predictions on using these organizations’ methods, but I think it’s an interesting goal.)
To explore this area, I’d suggest learning about forecasting generally ([Superforecasting](https://smile.amazon.com/Superforecasting-Science-Prediction-Philip-Tetlock/dp/0804136718?sa-no-redirect=1) is a good starting point) and building up your own prediction track record on sites such as the above.
**“Meta” careers.** There are a number of jobs focused on helping *other people* learn about key issues, develop key skills and end up in helpful jobs (a bit more discussion [here](https://www.cold-takes.com/making-the-best-of-the-most-important-century/#communities)).
It can also make sense to take jobs that put one in a good position to donate to nonprofits doing important work, to [spread helpful messages](https://www.cold-takes.com/spreading-messages-to-help-with-the-most-important-century/), and to build skills that could be useful later (including in unexpected ways, as things develop), as I’ll discuss [below.](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#other-things-you-can-do)
### Low-guidance jobs
This sub-section lists some projects that either don’t exist (but seem like they ought to), or are in very embryonic stages. So it’s unlikely you can get any significant mentorship working on these things.
I think the potential impact of making one of these work is huge, but I think most people will have an easier time finding a fit with jobs from the previous section (which is why I listed those first).
This section is largely to illustrate that I expect there to be more and more ways to be helpful as time goes on - and in case any readers feel excited and qualified to tackle these projects themselves, despite a lack of guidance and a distinct possibility that a project will make less sense in reality than it does on paper.
A big one in my mind is **developing safety standards** that could be used in a standards and monitoring regime. By this I mean answering questions like:
* What observations could tell us that AI systems are getting dangerous to humanity (whether by pursuing [aims of their own](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/) or by helping humans do dangerous things)?
+ A starting-point question: why do we believe today’s systems *aren’t* dangerous? What, specifically, are they unable to do that they’d have to do in order to be dangerous, and how will we know when that’s changed?* Once AI systems have potential for danger, how should they be restricted, and what conditions should AI companies meet (e.g., demonstrations of safety and security) in order to loosen restrictions?
There is some early work going on along these lines, at both AI companies and nonprofits. If it goes well, I expect that there could be many jobs in the future, doing things like:
* Continuing to refine and improve safety standards as AI systems get more advanced.
* Providing AI companies with “audits” - examinations of whether their systems meet standards, provided by parties outside the company to reduce conflicts of interest.
* Advocating for the importance of adherence to standards. This could include advocating for AI companies to abide by standards, and potentially for government policies to enforce standards.
**Other public goods for AI projects.** I can see a number of other ways in which independent organizations could help AI projects exercise more caution / do more to reduce risks:
* **Facilitating safety research collaborations.** I worry that at some point, doing good alignment research will only be possible with access to state-of-the-art AI models - but such models will be extraordinarily expensive and exclusively controlled by major AI companies.
+ I hope AI companies will be able to partner with outside safety researchers (not just rely on their own employees) for alignment research, but this could get quite tricky due to concerns about intellectual property leaks.
+ A third-party organization could do a lot of the legwork of vetting safety researchers, helping them with their security practices, working out agreements with respect to intellectual property, etc. to make partnerships - and [selective information sharing](https://www.cold-takes.com/racing-through-a-minefield-the-ai-deployment-problem/#selective-information-sharing), more broadly - more workable.* **Education for key people at AI companies.** An organization could help employees, investors, and board members of AI companies learn about the potential risks and challenges of advanced AI systems. I’m **especially excited about this for board members,** because:
+ I’ve already seen a lot of interest from AI companies in forming strong ethics advisory boards, and/or putting well-qualified people on their governing boards (see footnote for the difference[11](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn11)). I expect demand to go up.
+ Right now, I don’t think there are a lot of people who are both (a) prominent and “fancy” enough to be considered for such boards; (b) highly thoughtful about, and well-versed in, what I consider some of the most important risks of transformative AI (covered in this piece and the [series](https://www.cold-takes.com/tag/implicationsofmostimportantcentury/) it’s part of).
+ An “education for potential board members” program could try to get people quickly up to speed on [good board member practices generally](https://www.cold-takes.com/nonprofit-boards-are-weird-2/), on risks of transformative AI, and on the basics of how modern AI works.* **Helping share best practices across AI companies.** A third-party organization might collect information about how different AI companies are handling information security, alignment research, processes for difficult decisions, governance, etc. and share it across companies, while taking care to preserve confidentiality. I’m particularly interested in the possibility of developing and sharing innovative [governance setups](https://www.cold-takes.com/ideal-governance-for-companies-countries-and-more/) for AI companies.
**Thinking and stuff.** There’s tons of potential work to do in the category of “coming up with more issues we ought to be thinking about, more things people (and companies and governments) can do to be helpful, etc.”
* About a year ago, I published a [list of research questions](https://forum.effectivealtruism.org/posts/zGiD94SHwQ9MwPyfW/important-actionable-research-questions-for-the-most#A_high_level_list_of_important__actionable_questions_for_the_most_important_century) that could be valuable and important to gain clarity on. I still mostly endorse this list (though I wouldn’t write it just as is today).
* A slightly different angle: it could be valuable to have more people thinking about the question, “What are some tangible policies governments could enact to be helpful?” E.g., early steps towards standards and monitoring. This is distinct from advising governments directly (it's earlier-stage).
Some AI companies have policy teams that do work along these lines. And a few Open Philanthropy employees work on topics along the lines of the first bullet point. However, I tend to think of this work as best done by people who need very little guidance (more at my discussion of [wicked problems](https://www.cold-takes.com/the-wicked-problem-experience/)), so I’m hesitant to recommend it as a mainline career option.
Things you can do if you’re not ready for a full-time career change
-------------------------------------------------------------------
Switching careers is a big step, so this section lists some ways you can be helpful regardless of your job - including preparing yourself for a later switch.
First and most importantly, you may have opportunities to **spread key messages** via social media, talking with friends and colleagues, etc. I think there’s a lot of potential to make a difference here, and I wrote a [previous post](https://www.cold-takes.com/spreading-messages-to-help-with-the-most-important-century/) on this specifically.
Second, you can **explore potential careers** like those I discuss [above](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#jobs-that-can-help). I’d suggest generally checking out job postings, thinking about what sorts of jobs might be a fit for you down the line, meeting people who work in jobs like those and asking them about their day-to-day, etc.
Relatedly, you can **try to keep your options open.**
* It’s hard to predict what skills will be useful as AI advances further and new issues come up.
* Being ready to switch careers when a big opportunity comes up could be *hugely* valuable - and hard. (Most people would have a lot of trouble doing this late in their career, no matter how important!)
* Building up the financial, psychological and social ability to change jobs later on would (IMO) be well worth a lot of effort.
Right now there aren’t a lot of obvious places to **donate** (though you can donate to the [Long-Term Future Fund](https://funds.effectivealtruism.org/funds/far-future)[12](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn12) if you feel so moved).
* I’m guessing this will change in the future, for a number of reasons.[13](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn13)* Something I’d consider doing is setting some pool of money aside, perhaps invested such that it’s particularly likely to grow a lot if and when AI systems become a lot more capable and impressive,[14](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn14) in case giving opportunities come up in the future.
* You can also, of course, donate to things today that others aren’t funding for whatever reason.
**Learning more** about key issues could broaden your options. I think the [full series](https://www.cold-takes.com/tag/implicationsofmostimportantcentury/) I’ve written on key risks is a good start. To do more, you could:
* [Actively engage](https://www.cold-takes.com/reading-books-vs-engaging-with-them/) with this series by [writing your own takes](https://www.cold-takes.com/learning-by-writing/), discussing with others, etc.
* Consider various online courses[15](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn15) on relevant issues.
* I think it’s also good to get as familiar with today’s AI systems (and the research that goes into them) as you can.
+ If you’re happy to write code, you can check out coding-intensive guides and programs (examples in footnote).[16](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn16)+ If you don’t want to code but can read somewhat technical content, I’d suggest getting oriented with some basic explainers on deep learning[17](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn17) and then reading significant papers on AI and AI safety.[18](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn18)+ Whether you’re very technical or not at all, I think it’s worth playing with public state-of-the-art AI models, as well as seeing highlights of what they can do via Twitter and such.
Finally, if you happen to have opportunities to **serve on governing boards or advisory boards** for key organizations (e.g., AI companies), I think this is one of the best non-full-time ways to help.
* I don’t expect this to apply to most people, but wanted to mention it in case any opportunities come up.
* It’s particularly important, if you get a role like this, to invest in educating yourself on key issues.
Some general advice
-------------------
I think full-time work has huge potential to help, but also big potential to do harm, or to burn yourself out. So here are some general suggestions.
**Think about your own views on the key risks of AI, and what it might look like for the world to deal with the risks.** Most of the jobs I’ve discussed aren’t jobs where you can just take instructions and apply narrow skills. The [issues here](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#recap) are tricky, and it takes judgment to navigate them well.
Furthermore, no matter what you do, there will almost certainly be people who think your work is useless (if not harmful).[19](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#fn19) This can be very demoralizing. I think it’s easier if you’ve thought things through and feel good about the choices you’re making.
I’d advise trying to learn as much as you can about the major risks of AI (see [above](https://www.cold-takes.com/p/5fec3148-e34e-4bc2-a28b-8c95926142fa/#learning) for some guidance on this) - and/or trying to work for an organization whose leadership you have a good amount of confidence in.
**Jog, don’t sprint.** Skeptics of the “most important century” hypothesis will sometimes say things like “If you really believe this, why are you working normal amounts of hours instead of extreme amounts? Why do you have hobbies (or children, etc.) at all?” And I’ve seen a number of people with an attitude like: “THIS IS THE MOST IMPORTANT TIME IN HISTORY. I NEED TO WORK 24/7 AND FORGET ABOUT EVERYTHING ELSE. NO VACATIONS."
I think that’s a very bad idea.
Trying to reduce risks from advanced AI is, as of today, a frustrating and disorienting thing to be doing. It’s very hard to tell whether you’re being helpful (and as I’ve mentioned, many will inevitably think you’re being harmful).
I think the difference between “not mattering,” “doing some good” and “doing enormous good” comes down to **how you choose the job, how good at it you are, and how good your judgment is** (including what risks you’re most focused on and how you model them). Going “all in” on a particular objective seems bad on these fronts: it poses risks to open-mindedness, to mental health and to good decision-making (I am speaking from observations here, not just theory).
That is, I think it’s a *bad idea to try to be 100% emotionally bought into the full stakes of the most important century* - I think the stakes are just too high for that to make sense for any human being.
Instead, I think the best way to handle “the fate of humanity is at stake” is probably to find a nice job and work about as hard as you’d work at another job, rather than trying to make heroic efforts to work extra hard. (I criticized heroic efforts in general [here](https://www.cold-takes.com/useful-vices-for-wicked-problems/#self-preservation).)
I think this basic formula (working in some job that is a good fit, while having some amount of balance in your life) is what’s behind a lot of the most important positive events in history to date, and presents possibly historically large opportunities today.
*Special thanks to Alexander Berger, Jacob Eliosoff, Alexey Guzey, Anton Korinek and Luke Muelhauser for especially helpful comments on this post. A lot of other people commented helpfully as well.*
Footnotes
---------
---
1. I use “aligned” to specifically mean that AIs behave as intended, rather than pursuing dangerous goals of their own. I use “safe” more broadly to mean that an AI system poses little risk of catastrophe for *any* reason in the context it’s being used in. It’s OK to mostly think of them as interchangeable in this post. [↩](#fnref1)- AI labs with alignment teams: [Anthropic](https://www.anthropic.com/), [DeepMind](https://www.deepmind.com/) and [OpenAI](https://openai.com/). Disclosure: my wife is co-founder and President of Anthropic, and used to work at OpenAI (and has shares in both companies); OpenAI is a former [Open Philanthropy grantee](https://www.openphilanthropy.org/grants/openai-general-support/).
Academic labs: there are many of these; I’ll highlight the [Steinhardt lab at Berkeley](https://jsteinhardt.stat.berkeley.edu/) (Open Philanthropy grantee), whose recent research I’ve found especially interesting.
Independent nonprofits: examples would be [Alignment Research Center](https://alignment.org/) and [Redwood Research](https://www.redwoodresearch.org/) (both Open Philanthropy grantees, and I sit on the board of both).
[↩](#fnref2)- Examples: [AGI Safety Fundamentals](https://www.agisafetyfundamentals.com/), [SERI MATS](https://www.serimats.org/), [MLAB](https://forum.effectivealtruism.org/posts/vvocfhQ7bcBR4FLBx/apply-to-the-second-ml-for-alignment-bootcamp-mlab-2-in) (all of which have been supported by [Open Philanthropy](https://openphilanthropy.org/)) [↩](#fnref3)- On one hand, deceptive and manipulative AIs could be dangerous. On the other, it might be better to get AIs *trying* to deceive us before they can consistently *succeed;* the worst of all worlds might be getting this behavior [by accident](https://www.cold-takes.com/why-would-ai-aim-to-defeat-humanity/) with very powerful AIs. [↩](#fnref4)- Though I think it’s inherently harder to get evidence of low risk than evidence of high risk, since it’s hard to rule out [risks arising as AI systems get more capable](https://www.cold-takes.com/ai-safety-seems-hard-to-measure/#The-Lab-mice-problem). [↩](#fnref5)- Why do I simultaneously think “This is a mature field with mentorship opportunities” and “This is a badly neglected career track for helping with the most important century”?
In a nutshell, **most good security people are not working on AI**. It looks to me like there are plenty of people who are generally knowledgeable and effective at good security, but there’s also a *huge* amount of need for such people outside of AI specifically.
I expect this to change eventually if AI systems become extraordinarily capable. The issue is that it might be too late at that point - the security challenges in AI seem daunting (and somewhat AI-specific) to the point where it could be important for good people to start working on them many years before AI systems become extraordinarily powerful. [↩](#fnref6)- [Here’s Katja Grace](https://www.lesswrong.com/posts/uFNgRumrDTpBfQGrs/let-s-think-about-slowing-down-ai) arguing along these lines. [↩](#fnref7)- An Open Philanthropy grantee. [↩](#fnref8)- Open Philanthropy has funded Metaculus and contracted with Good Judgment and HyperMind. [↩](#fnref9)- That is, these groups are mostly trying things like “Incentivize people to make good forecasts; track how good people are making forecasts; aggregate forecasts” rather than “Study the specific topic of AI and make forecasts that way” (the latter is also useful, and I discuss it [below](#thinking)). [↩](#fnref10)- The governing board of an organization has the hard power to replace the CEO and/or make other decisions on behalf of the organization. An advisory board merely gives advice, but in practice I think this can be quite powerful, since I’d expect many organizations to have a tough time doing bad-for-the-world things without backlash (from employees and the public) once an advisory board has recommended against them. [↩](#fnref11)- [Open Philanthropy](https://www.openphilanthropy.org), which I’m co-CEO of, has supported this fund, and its current Chair is an Open Philanthropy employee. [↩](#fnref12)- I generally expect there to be more and more clarity about what actions would be helpful, and more and more people willing to work on them if they can get funded. A bit more specifically and speculatively, I expect AI safety research to get more expensive as it requires access to increasingly large, expensive AI models. [↩](#fnref13)- Not investment advice! I would only do this with money you’ve *set aside for donating* such that it wouldn’t be a personal problem if you lost it all. [↩](#fnref14)- Some options [here](https://www.agisafetyfundamentals.com/), [here](https://www.effectivealtruism.org/virtual-programs), [here](https://forum.effectivealtruism.org/posts/XvWWfq9iqFj8x7Eu8/list-of-ai-safety-courses-and-resources), [here](https://aisafety.training/). I’ve made no attempt to be comprehensive - these are just some links that should make it easy to get rolling and see some of your options. [↩](#fnref15)- [Spinning Up in Deep RL](https://spinningup.openai.com/en/latest/), [ML for Alignment Bootcamp](https://forum.effectivealtruism.org/posts/vvocfhQ7bcBR4FLBx/apply-to-the-second-ml-for-alignment-bootcamp-mlab-2-in), [Deep Learning Curriculum](https://github.com/jacobhilton/deep_learning_curriculum). [↩](#fnref16)- For the basics, I like Michael Nielsen’s [guide to neural networks and deep learning](http://neuralnetworksanddeeplearning.com/); [3Blue1Brown](https://www.youtube.com/playlist?list=PLZHQObOWTQDNU6R1_67000Dx_ZCJB-3pi) has a video explainer series that I haven’t watched but that others have recommended highly. I’d also suggest [The Illustrated Transformer](https://jalammar.github.io/illustrated-transformer/) (the transformer is the most important AI architecture as of today).
For a broader overview of different architectures, see [Neural Network Zoo](https://www.asimovinstitute.org/neural-network-zoo/).
You can also check out various Coursera etc. courses on deep learning/neural networks. [↩](#fnref17)- I feel like the easiest way to do this is to follow AI researchers and/or top labs on Twitter. You can also check out [Alignment Newsletter](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit#gid=0) or [ML Safety Newsletter](https://newsletter.mlsafety.org/archive) for alignment-specific content. [↩](#fnref18)- Why?
One reason is the tension between the [“caution” and “competition” frames](https://www.cold-takes.com/making-the-best-of-the-most-important-century/): people who favor one frame tend to see the other as harmful.
Another reason: there are a number of people who think we’re more-or-less doomed without a radical conceptual breakthrough on how to build safe AI (they think the sorts of approaches I list [here](https://www.cold-takes.com/high-level-hopes-for-ai-alignment/) are hopeless, for reasons I confess I don’t understand very well). These folks will consider anything that isn’t aimed at a radical breakthrough ~useless, and consider some of the jobs I list in this piece to be harmful, if they are speeding up AI development and leaving us with less time for a breakthrough.
At the same time, working toward the sort of breakthrough these folks are hoping for means doing pretty esoteric, theoretical research that many other researchers think is clearly useless.
And trying to make AI development slower and/or more cautious is harmful according to some people who are dismissive of risks, and think the priority is to push forward as fast as we can with technology that has the potential to improve lives. [↩](#fnref19) |
f665adf4-7d2a-4a57-93c7-7ed1731f1ac3 | trentmkelly/LessWrong-43k | LessWrong | Democratic Fine-Tuning
The project below, “Democratic Fine-tuning with a Moral Graph” (DFTmg), is a winner of the OpenAI democratic process grant. It is an alternative to Constitutional AI or simple RLHF-based approaches for fine-tuning LLMs, and is currently under development. This post introduces its two key innovations (values cards and the moral graph) and walks through the deliberation process that collects data for fine-tuning. It also says why something like DFTmg is needed for alignment and safety.
DFTmg is a project of “The Institute for Meaning Alignment”, a new AI alignment organization that uses concrete representations of life meaning and wisdom to align LLMs.
Setting the Stage
Imagine you are Instagram’s recommender system. Your responsibilities include: (a) ordering everyone’s feeds and reels, (b) filling up their search pages, and (c) suggesting reels by people they don’t follow yet.
You do this via an API: Instagram sends a user ID, plus a history of what they’ve clicked on, or paused to watch while scrolling. You send back lists of content object IDs. You don’t know much about the content objects, except there’s a rather opaque feature vector for each.
Now, imagine one day, you’re doing your job (recommending content objects), and you suddenly gain a new capacity: before replying to the next request, you find you can take a moment to wonder about the moral situation you are in. What values should you use, to make the best recommendations? How could things go wrong? What would be some great outcomes? What are your responsibilities here?
If this happened to me, I’d have a lot of questions:
* What are these content objects anyways? Do people really want to watch them, or are some of them clickbait?
* With the lists of what people paused to watch, did they feel good about watching those things? Or, were they compelled by sexual imagery, false promises, etc? Do they regret pausing to watch?
* Who are all these people? What are they looking for in life? What’s the de |
f80118f9-81f4-44a3-a212-e7e86ca54e16 | trentmkelly/LessWrong-43k | LessWrong | A prize to become artist in residence at CERN
http://www.aec.at/collide/
Prix Ars Electronica Collide@CERN is the new international competition for digital artists to win a residency at CERN the world's largest particle physics laboratory in Geneva. It is the first prize to be announced as part of the new Collide@CERN artists residency programme initiated by the laboratory.
The residency is in two parts - with an initial two months at CERN, where the winning artist will have a specially dedicated science mentor from the world famous science lab to inspire him/her and his/her work. The second part will be a month with the Futurelab team and mentor at Ars Electronica Linz with whom the winner will develop and make new work inspired by the CERN residency. |
f9ccf0bc-85fd-43af-80e0-3d7715d89780 | trentmkelly/LessWrong-43k | LessWrong | Subjective Questions Require Subjective information
This summarizes a (possibly trivial) observation that I found interesting.
Story
An all-powerful god decides to play a game. They stop time, grab a random human, and ask them "What will you see next?". The human answers, then time is switched back on and the god looks at how well they performed. Most of the time the humans get it right, but occasionally they are caught by surprise and get it wrong.
To be more generous the god decides to give them access (for the game) to the entirety of all objective facts. The position and momentum of every elementary particle, every thought and memory anyone has ever had (before the time freeze) etc. However, suddenly performance in the game drops from 99% to 0%. How can this be? They have more information, they know everything!
If you have the memories of every single human up to that point, then you don't know which of them you are. Who is "you" in the "What will you seen next?"? Before the extra information was added you knew which human you were, it was made very obvious by the memories and information at your disposal. But, given the memories of everyone, and all that other information, you suddenly require an additional piece of information to answer the question.
God: "What will you see next?"
Participant: "I know what every human will see next, but I don't know which one I am".
G: "You have been given all the information there is. How can you not know?"
Idea
"What will you see next?" is a subjective question. All objective facts put together is not enough to answer it, because an addition piece of information "Which person am I"? is needed. This final piece of information is subjective, and arguably in some materialistic sense doesn't really exist. But, the question is also subjective, so it should not be surprising that subjective information is needed to answer it.
You can make the situation more extreme in various ways. Instead of providing a snapshot of the universe at a particular time the god could pr |
6f092074-ab63-444b-8814-faa04306281f | StampyAI/alignment-research-dataset/lesswrong | LessWrong | How my school gamed the stats
I was reading the Slate Star Codex [review](https://astralcodexten.substack.com/p/book-review-the-cult-of-smart) of The Cult of Smart. Part of it discusses charter vs state schools and the allegations of fraud of various kinds undermining charter schools record of better achievement. Reading it, I realized that I took for granted that public schools engage in systematic fraud in a variety of ways. I don't think this is something everyone understands, hence this post.
I went to a state school in the UK. State schools are rated on a 1 - 4 scale from unsatisfactory to outstanding. My school was rated good, meaning a 3. A few memories which stand out. During my first week I saw one of the boys in my class who was 11 at the time held up against the wall in a corridor while a 16 year old put a shiv to his throat and robbed him. He handed over his wallet and keys. A year or two later and I remember seeing a small boy who struggled with depression held up by the throat against a locker and slapped in the face by a troublemaker from the same class in front of everyone just before we went in to the classroom. I remember classes which were filled start to finish with people shouting and talking. Neither of the first two events were common but they also weren't uncommon. No one was surprised to witness them. It's worth emphasizing again that my school was above average, in fact quite far above average, and in a middle class area. It's also worth noting that I was mostly in top ability streamed classes, meaning my classroom experience was likely far better than average.
There were many ways in which the school and teachers gamed the system to boost their measured performance. One way was to do exams for students. I was on a bottom set language class for French. After two years I literally couldn't speak a single sentence in french and maybe knew 20 words in total. I still passed my exams. How? We did the tests in class. Often the teacher would go through them with us. Literally giving us the test and then going through each question on the whiteboard and telling us what to write. A different year and a different teacher, this time the teacher would sit next to us and write the answers down. Why sit next to us? It was the bottom set so people often wouldn't even bother to write down the answer if they were told it. This kind of thing was normal, so much so that I, and I think most people there, didn't realize anything unusual was happening.
Another way schools game metrics is to cheat inspections. A major component of how schools are judged in the UK is through independent inspections carried out by an independent quasi-governmental organization called Ofsted. Now, you may imagine that these inspections would be unannounced, so as to best get a real image of how a school works. Not the case. They're scheduled well in advance. Before every inspection, a few things would happen in my school:
* The worst troublemaker kids would be taken aside and put in a special room where inspectors wouldn't see them. Either that or they would just be told not to come into school at all on that day.
* All of us were told in assembly that an inspection was coming and to be on our best behavior on that day. Often teachers would have conversations with less serious troublemakers and impress on them that they would behave on that day or face consequences afterwards.
* Teachers would put a great deal more effort into their lesson plans than was normal. Classroom behavior management would also be far stricter.
Because of these and other measures my school during an inspection was utterly different than my school on a normal day. On some level this isn't surprising. If teachers' promotions and management's jobs depend on good inspection results and inspections are easy to game, people will game them. Incentives drive behavior. But it's still sad.
Another way the stats were gamed was by not recording bad behavior. When a school gives a detention or suspends/expels a student, there's a record of it. This is especially true of suspensions, students being sent home or expulsions. The more of these you have, the worse you look as a school. The solution then is obvious, don't punish people or punish them in non-recorded ways. Again, in my school it was completely normal for students in lower sets to swear at the teacher, talk over them or disrupt the class for everyone else. It was normal for someone to be aggressive and abusive towards others and to face at most a 40 minute detention, but even getting a detention would be unusual.
I realize that one data point is not enough to draw solid general conclusions. My own perception is that this kind of fraud wasn't specific to my school. My cousin went to a state school fairly nearby. He's 4 years younger than me. During one of my winters back from undergrad we discussed his school and his experiences mirrored mine. His exact words regarding inspections were "I learned 4 times more that day than any other day that year. It was amazing". I talked to a few British students at university, although specifically the not middle/upper class ones who would have gone to public schools. They had gone to schools similar to mine in different parts of the country and their stories were similar and often worse. Two particularly funny examples from my friends' experiences stand out. A teacher in year 9 walked up to a student who was talking, picked them up and threw them out of an (open) first floor window. My friend sitting in class noticed two boys making fun of him and then proceeded to get up in the middle of class while the teacher was talking, walk to their table, flip the table upwards to hit them in the face before going to sit down again when the teacher told him to. (Remember, my friend was a studious, sporty Asian kid and not a troublemaker. This kind of thing is normal in that environment). Comedic stories aside, my experiences in school, while not universal, seem fairly common in the UK and from what I've read of the statistics, bad US schools are far, far worse.
I'm unsure what my point here is. I think I have two:
* Charters may cook their books in various ways. In the UK, State schools do too. I would be surprised if it wasn't also the case in the US.
* I think that I feel like a lot of commentators on places like SSC have fairly middle class experiences of fairly good schools and that bleeds into how their comparison between state vs charter schools. It's just good to remember that it's not those nice middle class schools that charters typically replace.
Crossposted to my blog at <https://dissent.blog/2021/02/20/how-my-school-gamed-the-stats/> |
2c9807ab-7e08-4cf5-920c-2497394e8b03 | trentmkelly/LessWrong-43k | LessWrong | A short critique of Vanessa Kosoy's PreDCA
This critique is an addendum to my distillation for application problem 3 for Nate Soares and Vivek Hebbar's SERI MATS stream. Reading my distillation is not required if you are familiar with PreDCA.
Disclaimer: This is just my model of how Vanessa thinks and I might be misrepresenting some views. Furthermore, as mentioned below, I'm sure Vanessa is well aware of these issues, and plans on trying to solve those which constitute real obstacles. [Edit: See her comments below]
Vanessa( and Diffractor)'s work generally feels different to other Alignment theory, and I have usually attributed this to its radical focus on foundations (shared by some other researchers) and the complexity of its technical mathematical results (shared by few). But upon momentarily coarsening these fine technical details, and presenting PreDCA more conceptually in a language similar to that of other proposals, it becomes clear that it really is fundamentally different to them. As a consequence of Vanessa's opinions and approach, PreDCA has different objectives and hopes to most proposals.
PreDCA is fundamentally concrete. Yes, it still includes some "throw every method you have at the problem" (as in Classification). But the truly principal idea involves betting everything on a specific mathematical formalization of some instructions. This concreteness is justified for Vanessa because nothing short from a watertight solution to agent foundations will get us through Alignment[1]. Some AGI failure modes are obviously problems humanity would have to deal with in the long run anyway (with or without AGI), but on a time trial. And Vanessa is pessimistic about the viability of some intuitively appealing Alignment approaches that try to delegate these decisions to future humanity. In short, she believes some lock-in is inevitable, and so tries to find the best lock-in possible by fundamentally understanding agency and preferences.
Now, if PreDCA works, we won't get a naively narrow lock-in: the u |
bfc24728-8232-4a01-a201-7aa09a8e7845 | trentmkelly/LessWrong-43k | LessWrong | LA-602 vs. RHIC Review
LA-602: Ignition of the Atmosphere with Nuclear Bombs, a research report from the Manhattan Project, is to the best of my knowledge the first technical analysis ever conducted of an uncertain danger of a human-caused extinction catastrophe.
Previously, Teller and Konopinski had been assigned the task of disproving a crazy suggestion by Enrico Fermi that a fission chain reaction could ignite a thermonuclear reaction in deuterium - what we now know as an H-Bomb. Teller and Konopinski found that, contrary to their initial skepticism, the hydrogen bomb appeared possible.
Good for their rationality! Even though they started with the wrong conclusion on their bottom line, they were successfully forced away from it by arguments that could only support one answer.
Still, in retrospect, I think that the advice the future would give to the past, would be: Start by sitting down and saying, "We don't know if a hydrogen bomb is possible". Then list out the evidence and arguments; then at the end weigh it.
So the hydrogen bomb was possible. Teller then suggested that a hydrogen bomb might ignite a self-sustaining thermonuclear reaction in the nitrogen of Earth's atmosphere. This also appeared extremely unlikely at a first glance, but Teller and Konopinski and Marvin investigated, and wrote LA-602...
As I understand LA-602, the authors went through the math and concluded that there were several strong reasons to believe that nitrogen fusion could not be self-sustaining in the atmosphere: it would take huge energies to start the reaction at all; the reaction would lose radiation from its surface too fast to sustain the fusion temperature; and even if the fusion reaction did grow, the Compton effect would increase radiation losses with volume(?).
And we're still here; so the math, whatever it actually says, seems to have been right.
Note that the Manhattan scientists didn't always get their math right. The Castle Bravo nuclear test on March 1, 1954 produced 15 megatons instea |
4a790104-54ce-4c81-b3a6-1d942b4c5044 | trentmkelly/LessWrong-43k | LessWrong | Fictional Bias
As rationalists, we are trained to maintain constant vigilance against common errors in our own thinking. Still, we must be especially careful of biases that are unusually common amongst our kind.
Consider the following scenario: Frodo Baggins is buying pants. Which of these is he most likely to buy:
(a) 32/30
(b) 48/32
(c) 30/20
If you're like me, your answer is (c). Frodo, as we know, is about 4' tall, so his inseam is much more likely 20'' than 30''.
But like me, you'd be wrong. Since there aren't *actually* any hobbits, all we know is that we're talking about a person named Frodo Baggins, who is male. And the most common pants size is 32/30, so the correct answer, given our actual state of knowledge about the real world, is (a).
This is what researchers call Fictional Bias, and there is evidence that it affects virtually every domain of decision-making. The mistake is using information from fictional sources in real contexts. It is the more-pernicious cousin of generalizing from fictional evidence - instead of merely generalizing, we treat real objects and persons as though they are the specific fictional entities they resemble.
We're of course familiar with particularly egregious examples of people confusing fiction with reality. For example in the 1930's, there were multiple cases where someone was killed for having the same name as the serial killer of children from the movie M. But these could be written off as merely disturbed individuals. But as it turns out, we're affected by this bias in our daily lives.
Examples abound in the literature, though the name "fictional bias" is not always used. A 1984 study by Dr. Sidney Zweibel and Dr. Emilio Lizardo asked subjects to trust someone with an unusual name. Subjects were 70% less likely to trust when the person had the same name as a fictional villain. A 1989 study by Dr. Wayne Szalinski established that subjects were 89% more likely to agree to take an experimental drug, when it was nam |
3d6332ce-9c5c-41a0-b34b-6c66b1850641 | trentmkelly/LessWrong-43k | LessWrong | "Does your paradigm beget new, good, paradigms?"
A very short version of this post, which seemed worth rattling of quickly for now.
A few months ago, I was talking to John about paradimicity in AI alignment. John says "we don't currently have a good paradigm." I asked "Is 'Natural Abstraction' a good paradigm?". He said "No, but I think it's something that's likely to output a paradigm that's closer to the right paradigm for AI Alignment."
"How many paradigms are we away from the right paradigm?"
"Like, I dunno, maybe 3?" said he.
Awhile later I saw John arguing on LessWrong with (I think?) Ryan Greenblatt about whether Ryan's current pseudo-paradigm was good. (Sorry if I got the names here or substance here wrong, I couldn't find the original thread, and it seemed slightly better to be specific so we could dig into a concrete example).
One distinction in the discussion seemed to be something like:
* On one hand, Ryan thought his current paradigm (this might have been "AI Control", as contrasted with "AI Alignment") had a bunch of traction on producing a plan that would at least reasonably help if we had to align superintelligent AIs in the near future.
* On the other hand, John argued that the paradigm didn't feel like the sort of thing that was likely to bear the fruit of new, better paradigms. It focused on an area of the superintelligence problem that, while locally tractable, John thought was insufficient to actually solve the problem, and also wasn't the sort of thing likely to pave the way to new paradigms.
Now a) again I'm not sure I'm remembering this conversation right, b) whether either of those points are true in this particular case would be up for debate and I'm not arguing they're true. (also, regardless, I am interested in the idea of AI Control and think that getting AI companies to actually do the steps necessary to control at least nearterm AIs is something worth putting effort into)
But it seemed good to promote to attention the idea that: when you're looking at clusters of AI Safe |
42197ca8-a150-4ab5-881b-3b1310ec00f0 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | A case for capabilities work on AI as net positive
Edit: Changed the title.
Or, why I no longer agree with the standard LW position on AI anymore.
In a sense, this is sort of a weird post compared to what LW usually posts on AI.
A lot of this is going to depend on some posts that changed my worldview on AI risk, and they will be linked below:
Deceptive alignment skepticism sequence, especially the 2nd post in the sequence is here:
<https://www.lesswrong.com/posts/RTkatYxJWvXR4Qbyd/deceptive-alignment-is-less-than-1-likely-by-default>
Evidence of the natural abstractions hypothesis in action:
<https://www.lesswrong.com/posts/BdfQMrtuL8wNfpfnF/natural-categories-update>
<https://www.lesswrong.com/posts/obht9QqMDMNLwhPQS/asot-natural-abstractions-and-alphazero#comments>
Summary: The big updates I made was that deceptive alignment was way more unlikely than I thought, and given that deceptive alignment was a big part of my model of how AI risk would happen (about 30-60% of my probability mass was on that failure mode), that takes a big bite out of the probability mass of extinction enough to make increasing AI capabilities having positive expected value. Combine this with the evidence that at least some form of the natural abstractions hypothesis is being borne out by empirical evidence, and I now think the probabilities of AI risk have steeply declined to only 0.1-10%, and all of that probability mass is plausibly reducible to ridiculously low numbers by going to the stars and speeding up technological progress.
In other words, I now believe a significant probability, on the order of 50-70%, that alignment is solved by default.
EDIT: While I explained why I increased my confidence in alignment by default in response to Shiminux, I now believe that for now I was overconfident on the precise probabilities on alignment by default.
What implications does this have, if this rosy picture is correct?
==================================================================
The biggest implication is that technological progress looks vastly positive, compared to what most LWers and the general public think.
This also implies a purpose shift for Lesswrong. For arguably 20 years, the site was focused on AI risk, though it arguably exploded with LLMs and actual AI capabilities being released.
What it will shift to is important, but assuming that this rosy model of alignment is correct, then I'd argue a significant part of the field of AI Alignment should and can change purpose to something else.
As for Lesswrong, I'd say we should probably focus more on progress studies like Jason Crawford and inadequate equilibria and how to change them.
I welcome criticism and discussion of this post, due to it's huge implications for LW. |
6e9102a7-70e3-4d9b-9c34-6111a561951f | trentmkelly/LessWrong-43k | LessWrong | Waterfall Truth Predicates
The waterfall-based approaches to the Löbian obstacle offer a way around finitely-terminating sequences of trust which we get by adding towers of soundness schemas to PA. This creates a kind of illusion of self-trust by way of a non-well-founded chain of trust.
Another familiar situation where we are normally faced with the ability to construct arbitrarily high towers but not a single self-referential system is that of truth predicates. Tarski's undefinability theorem blocks the existence of a full truth predicate within the same language as the one which it describes. Perhaps a similar waterfall construction can be applied, to get an infinite descending chain of languages.
Extend the language of PA with a family of truth predicates Trn. A Tarski-style approach would assert a T-schema Trn(┌ϕ┐)↔ϕ for ϕ which contain truth predicates indexed strictly lower than n. (┌ϕ┐ is the Gödel number of ϕ.) Here, we wish to flip this, and assert a T-schema which allows strictly higher n.
This brings to mind Yablo's Paradox. A contradiction can likely be worked out in a way resembling that, but instead I'll note that this theory implies the naive soundness waterfall in which we construct a sequence of theories Tn=Tn+1+Sound(Tn+1). This is because we can use the truth predicate Trn to carry out a proof of soundness for the axioms and inference rules, with the exception of instances of the T-schema involving Trm≤n. This gives us the naive soundness waterfall, which we know to be inconsistent. (Note that I have not checked this in detail, however.)
My idea for fixing the T-schema, then, is to introduce the same ψ(n) predicate which asserts that n is not the Gödel number of a proof of contradiction in ZFC. We make the new schema:
* ψ(n)→[Trn(┌ϕ┐)↔ϕ], where ϕ contains only Trm>n.
Because we can prove any particular ψ(n), we can still apply the schema in specific cases. It seems likely that we can still carry out soundness arguments, as well, constructing the consistent version o |
edfe6015-d13d-46a7-94d9-d39975029792 | trentmkelly/LessWrong-43k | LessWrong | What are some examples from history where a scientific theory predicted a significant experimental observation in advance?
A few examples I can think of off the top of my head, to give a feel for the reference class I'm looking for:
- The existence and position of Neptune were predicted from observations of Uranus's orbit, before anyone had ever observed Neptune directly
- Black holes were predicted from the equations of General Relativity before we'd ever observed them or their effects on stars' motion
- Not as quantitative, but Darwin's theory of Evolution predicted that we'd find some method by which natural selection actually occurs, before we ever knew about DNA.
Are there other cool examples like these? |
09425258-e28a-4997-9158-c98260917702 | trentmkelly/LessWrong-43k | LessWrong | Monthly Roundup #24: November 2024
This is your monthly roundup. Let’s get right to it.
YOUNG PEOPLE ARE YOUNG AND STUPID
As a reminder that yes college students are often young and stupid and wrong about everything, remember the time they were behind a ban on paid public toilets? This is a central case of the kind of logic that often gets applied by college students.
NO ONE VOTED FOR THIS
HR and Title IX training seems like it’s going a lot of compelled speech in the form of ‘agree with us or you can’t complete your training and the training is required for your job,’ and also a lot of that compelled speech is outright lying because it’s confirmation of statements that are universally recognized to be insane?
> Robin Hanson: Scenario: 2 women talking. X, married to woman, announces is pregnant. Y asks how they got pregnant, was it friend, donor, or IVF? 3rd person overhears, wonders if they should immediately intervene in convo to tell Y they are discriminating. Should they?
>
> Context: This is example given in my workplace harassment/discrimination training, & one can’t move on unless one agrees that 3rd person should intervene.
>
> My training says “Those questions are a little invasive!”
>
> Training by Vector Solutions.
I do realize Robin’s followers can be odd, but yeah, not this time, and this is 87-1.
They also forced people to affirm the ‘affirmative specific consent’ rule, which voters disapproved of by 11-1.
DISCRIMINATION
Hard to pronounce names constitute 10%-50% of ethnic penalties among economics PhD job candidates, says new AEJ piece.
> Qi Ge and Stephen Wu: The results are primarily driven by candidates with weaker résumés, suggesting that cognitive biases may contribute to the penalty of having a difficult-to-pronounce name.
Given this was not a controlled experiment, I’d ask if choosing an unpronounceable name is correlated to other parental characteristics that matter here.
The good news is you can solve for this – you can change your name.
A paper via MR says |
4e654edf-7fd8-41b6-9276-6b2745aeb4ac | trentmkelly/LessWrong-43k | LessWrong | Forecasting Newsletter: April 2021
Highlights
* Polymarket is being attacked by “sandwiching” bots
* Metaculus launches “Forecasting Causes”
* In Reflective Bayesianism, Abram Demski outlines questionable and implicit assumptions which Bayesians make.
Index
* Prediction Markets & Forecasting Platforms
* In The News
* Blog Posts
* Long Content
* Hard To Categorize
Sign up here or browse past newsletters here.
Prediction Markets & Forecasting Platforms
Polymarket
After the demonstrable success of Polymarket (and, to a lesser extent, Augur) in attracting volume to their platforms, many imitators have popped up on the crypto scene. None of them are functional yet, but I thought I'd mention them, in order from least to most scammy:
* PolkaMarkets is aiming for an August release date, and has recently begun testing its MVP.
* Hedgehog Markets's schtick is to be implemented in an up-and-coming blockchain, Solana. The functioning markets on their webpage are currently using test money.
* Totem has an interesting scheme where predictions are non-punitive. That is, people who predict incorrectly won't lose money, they will merely win less. However, after reading their whitepaper, it is not exactly clear where the money for rewards will come from. It is being built on top of the Binance chain, a centralised exchange with a centralized coin, which I find unappealing.
* Polars and PredictX are also built on top of the Binance chain. I would characterize them as money grabs. That is, they are attempting to raise money to do something like Polymarket without a clear plan for how they would be superior.
While Totem feels scammy, there are interesting possibilities related to its core idea. For example, external actors could subsidize the market. Another alternative could be to stake the bet amounts on a financial instrument which provides returns, like Uniswap, while waiting for resolution. In other words, one could bet the interest, not the principal.
While imitator projects go through the des |
266ac6cb-ee74-483b-99f9-d05e609420d5 | trentmkelly/LessWrong-43k | LessWrong | In the grim darkness of the far future there is only war continued by other means
(cross-posted from my blog)
I. PvE vs PvP
Ever since it’s advent in Doom, PvP (Player vs Player) has been an integral part of almost every major video game. This is annoying to PvE (Player vs Environment) fans like myself, especially when PvE mechanics are altered (read: simplified and degraded) for the purpose of accommodating the PvP game play. Even in games which are ostensibly about the story & world, rather than direct player-on-player competition.
The reason for this comes down to simple math. PvE content is expensive to make. An hour of game play can take many dozens, or nowadays even hundreds, of man-hours of labor to produce. And once you’ve completed a PvE game, you’re done with it. There’s nothing else, you’ve reached “The End”, congrats. You can replay it a few times if you really loved it, like re-reading a book, but the content is the same. MMORGs recycle content by forcing you to grind bosses many times before you can move on to the next one, but that’s as fun as the word “grind” makes it sound. At that point people are there more for the social aspect and the occasional high than the core gameplay itself.
PvP “content”, OTOH, generates itself. Other humans keep learning and getting better and improvising new tactics. Every encounter has the potential to be new and exciting, and they always come with the rush of triumphing over another person (or the crush of losing to the same).
But much more to the point – In PvE potentially everyone can make it into the halls of “Finished The Game;” and if everyone is special, no one is. PvP has a very small elite – there can only be one #1 player, and people are always scrabbling for that position, or defending it. PvP harnesses our status-seeking instinct to get us to provide challenges for each other rather than forcing the game developers to develop new challenges for us. It’s far more cost effective, and a single man-hour of labor can produce hundreds or thousands of hours of game play. StarCraft continu |
ec5127e2-0ca8-47a9-966a-f8c26dc9c957 | trentmkelly/LessWrong-43k | LessWrong | How do low level hypotheses constrain high level ones? The mystery of the disappearing diamond.
Consider two friends, Alice and Bob, trying to figure out what happened to a diamond that disappeared from a museum. They do so in the form of a game that is kind of an approximation to Solomonoff induction: they will work together to come up with the smallest possible explanations that conform to the data, for some intuitive notion of smallness.
This helps to eliminate fake explanations; the hypothesis "a witch caused Henry to fall ill" can be simplified to "Henry fell ill". But "Sally touched Henry and Bob, and Sally is sick and Henry is sick and Bob is sick" is beaten by "Sally touched Henry and Bob and Sally is contagiously sick". An explanation is good if it is smaller than just hard-coding the answer.
Bob knows that there are four diamond thieves in the city, so he comes up with four hypotheses:
1. The diamond was stolen by thief number 1.
2. The diamond was stolen by thief number 2.
3. The diamond was stolen by thief number 3.
4. The diamond was stolen by thief number 4.
These are all roughly the same complexity (depending on how you encode numbers), so this provides a uniform distribution over the four thieves.
Alice comes up with one hypothesis:
1. The diamond spontaneously ceased existing.
and declares victory.
Bob: What, that makes no sense? Physical objects can't stop existing.
Alice: We aren't doing physics; we are playing a game.
Bob: ͠° ͟ʖ ͡°
Alice: ¯\_(ツ)_/¯
But there is an additional rule; you can add other data from the real world to the challenge. For example, for "Henry falling ill", you might get better hypotheses if you try to compress info about all the sick people in the village, so that a slightly more complex hypothesis that can explain all of them wins!
Bob: I hereby add all physics experiments to the data set!
Alice then comes up with the following hypothesis: all physical experiments are explained by the standard model of physics, except the diamond spontaneously ceased existing.
"Was it really necessary to write it out? |
7dc43198-21fc-467d-8965-727c9b4e319d | StampyAI/alignment-research-dataset/arxiv | Arxiv | Scaling Laws and Interpretability of Learning from Repeated Data
1 Introduction
---------------
Figure 1: Experimental Setup. From a large original text dataset (left), we draw 90% of our desired training dataset in a non-repeated fashion, and 10% as repeats of a tiny portion of the original dataset (right). We hold constant that 10% of total training tokens will come from repeats, but we vary the repeated fraction in our runs. In other words, the sample to be repeated might be very small, like 0.01% of the total training tokens repeated 1000x, or relatively large, like 1% of the total training tokens repeated 10x. A small, held-back portion of the original dataset (yellow in left figure), not including any repeated data, is used as a test set and is the test loss reported in all subsequent figures.
Large, high-quality text datasets are crucial for training large language models Brown et al. ([2020](#bib.bib36 "Language models are few-shot learners")); Rae et al. ([2021](#bib.bib37 "Scaling language models: methods, analysis, and insights from training gopher")). Such datasets often contain many copies of substantially overlapping documents, which greatly impairs the performance of language models on downstream tasks Lee et al. ([2021](#bib.bib34 "Deduplicating training data makes language models better")). However, it is not well understood why data repetition impacts performance to such a large extent.
In this paper we study data repetition in language models through two lenses: the macroscopic lens of scaling laws, and the microscopic lens of mechanistic interpretability Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")); Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")). For the first lens, we trained transformer Vaswani et al. ([2017](#bib.bib10 "Attention is all you need")) language models on mostly unique data plus a small fraction of repeated data (Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data")), varying the repeated dataset size, model size, and fraction of tokens trained on repeated data. We find a strong double-descent phenomenon Advani and Saxe ([2017](#bib.bib25 "High-dimensional dynamics of generalization error in neural networks")); Belkin et al. ([2018](#bib.bib26 "Reconciling modern machine learning practice and the bias-variance trade-off")); Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")), such that there is a defined range of repetition frequency for which performance is harmed to a surprisingly large extent. We suspect there is a range in the middle where the data can be memorized and doing so consumes a large fraction of the model’s capacity, and this may be where the peak of degradation occurs. The location of the region suggests that large models like GPT-3, Gopher, and PALM Brown et al. ([2020](#bib.bib36 "Language models are few-shot learners")); Rae et al. ([2021](#bib.bib37 "Scaling language models: methods, analysis, and insights from training gopher")); Bi et al. ([2020](#bib.bib7 "PALM: pre-training an autoencoding and autoregressive language model for context-conditioned generation")) need to be careful about overfitting their high quality distributions like Wikipedia and books.
For the second lens, mechanistic interpretability (attempting to reverse engineer the detailed computations performed by the model) we show that repeated data disproportionately damages induction heads. Induction heads use a circuit of 2 attention heads to "complete the pattern by copying and completing sequences" Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")). The damage to induction heads is observed through degradation in copying, prefix matching, and through inspection.
Together, the two lenses provide an integrated picture of how repeated data might be causing the network (or part of it) to shift from generalization to memorization, and mechanistically how this could be harming performance of the overall language model.
###
1.1 Summary of Results
| | |
| --- | --- |
| | |
Figure 2: Models of different sizes show a degradation in performance at a specific range of repeats that shrinks with model size (left panel). At its peak the degradation sometimes reaches the equivalent of a 2x decrease in model size. The right panel shows that divergence (blue line) from a healthy, straight scaling law (red) lines up with when the models start to dramatically overfit the repeated subset (green curve). The blue line on the right corresponds to a vertical slice of models in the left diagram trained on the repeated subset for 120 epochs. All these models were trained on 90% unique data and 10% repeated tokens.
To systematically study repeated data, we trained transformer Vaswani et al. ([2017](#bib.bib10 "Attention is all you need")) language models on mostly unique data plus a small fraction of repeated data (Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data")), varying the repeated dataset size, model size, and fraction of tokens trained on repeated data over 2-3 orders of magnitude. All models were trained for 100B tokens. We examined the resulting models using both scaling laws and mechanistic interpretability tools. Our main findings were as follows:
* Repeated data induces a strong double-descent phenomenon Advani and Saxe ([2017](#bib.bib25 "High-dimensional dynamics of generalization error in neural networks")); Belkin et al. ([2018](#bib.bib26 "Reconciling modern machine learning practice and the bias-variance trade-off")); Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")), in which data repeated a few times does not cause much damage to language model performance, data repeated very many times also does not cause much damage, but there is a peak in the middle where damage is surprisingly large. For instance, when we train an 800M parameter transformer with 10% of training tokens drawn from the repeated subset (yellow curve in Figure [2](#S1.F2 "Figure 2 ‣ 1.1 Summary of Results ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data")) we find the loss can be nearly as high as for the 340M parameter transformer (light green curve). We see an epoch-wise Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")) double descent learning curve in Figure [3](#S2.F3 "Figure 3 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") is driving this performance degradation. We suspect there is a range in the middle where the data can be memorized and doing so consumes a large fraction of the model’s capacity, and this may be where the peak of degradation occurs. Figure [2](#S1.F2 "Figure 2 ‣ 1.1 Summary of Results ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data") on the right shows that the peak performance hit coincides with where the train loss on the repeated data approaches zero, similar to previously observed double-descent phenomena. This also provides a practical diagnostic for when repeated data is likely to be harming the model.
* Repeated data can cause a divergence from power-law scaling. For the blue curve in Figure [2](#S1.F2 "Figure 2 ‣ 1.1 Summary of Results ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data") right (122 repeated epochs), we see only a moderate impact to performance (line on log-log graph) until the model is scaled up to 100M parameters, after which we see a large divergence from power law scaling of cross entropy loss. Extrapolating the region of large degradation in Figure [4](#S2.F4 "Figure 4 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") predicts meaningful degradation of repeating data only 2 times for large (GPT-3 size) models, though the region would be shifted if the models were trained to the compute optimal frontier Hoffmann et al. ([2022](#bib.bib35 "Training compute-optimal large language models")).
* Repeated data causes a disproportionately large performance hit to copying, a mechanism for in-context learning. We constructed a simple copying eval, the loss on the first paragraph of Harry Potter copied 11 times. We observe that using 3% repeated data at the worst number of repeated epochs caused up to a 3x reduction in effective model size (performance equal to model with 3x fewer parameters) on this task whereas it only caused at most a 15% reduction in effective model size on test loss.
* The disproportionate performance hit to copying coincides with a disproportionate degradation of induction heads. In line with Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) we evaluated the models on their prefix matching score, repeated sequences of random tokens and observed the degree to which attention heads attend to earlier tokens that are preceded by a token that matches the present token. We observe that using 3% repeated data at the worst number of repeated epochs caused on average a 32% reduction in effective model size on this task whereas it only caused at most a 15% reduction in effective model size on test loss.
* Repeated text data causes a small but still disproportionate performance drop out of distribution, as measured by cross entropy loss on Python code. Unlike our the Harry Potter copying and prefix matching evals we mostly see the performance drop with higher levels of repetition, 50-90%.
* One and two-layer attention only models trained on repeated data are worse at exactly copying and fuzzily copying (for instance correctly predicting Dursleys given that Dursley has appeared previously) proper names on inspection. When we inspect per tokens losses of smaller models we can see this degradation in a simple, understandable form of copying in a paragraph of text.
* Training on repeated Python code creates a similar behavior. When training on Python we also observe a double descent phenomenon and a predictable poor performance region in terms of model size and repeated epochs, though the shape of both curves are somewhat different.
* Pre-training on repeated data damages models. Pre-training with repeated data leads to worse performance than both training from scratch and fine-tuning from a control model pre-trained on the original text dataset. During fine-tuning, the repeated data model forgets the repeated dataset, so we consider the model pre-trained with repeated data to be strictly worse than the model fine-tuned from the unique dataset.
2 Results
----------
| | |
| --- | --- |
| | |
Figure 3: Learning curves for test loss on 800M models with 90% repeated data (left) and 50% repeated data (right), each with varying numbers of repeats/sizes of the repeated fraction. The graph on the left shows characteristic double descent curves. Repeated epochs corresponds to the number of epochs on the repeated tokens, the rest of the data is seen only once. For several models, test loss drops as normal during the beginning of training, but then starts to rise during the middle of training before dropping again. In the graph on the right with only 50% repeated data, we see that the double descent bumps have turned into long plateaus for highly affected models.
Repeated data induces a strong double descent phenomenon.
The results from training models on different sizes, fractions of repeated data, and frequency of repeats are shown in Figures 2 and 3. Figure 2 (left) shows that when we train on 10% repeated data and vary the frequency of repetition (or equivalently the number of epochs of repeated data), there is a specific range of repetition frequency for which damage to model performance is maximized. The range depends on the model size but for a 800M parameter model it occurs at roughly 100x repeats of 0.1% of the data, and degrades performance nearly to that of a 340M parameter model. This is a large degradation given that only 10% of the data is repeated. The peak coincides with the advent of memorization on the repeated data (Figure 2 right) – a possible indicator of a double descent phenomenon.
Figure [3](#S2.F3 "Figure 3 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") shows learning curves for different repetition frequencies and for 50% and 90% of the data being repeated. In the extreme case of 90% repeated data and the correct frequency of repetition (100x-10,000x), we confirm the presence of a literal double descent curve in which the loss decreases, increases, and then decreases again (Figure [3](#S2.F3 "Figure 3 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") left). As we lower the fraction of repeated data to 50%, the curve becomes a long plateau rather than double descent, but it appears to be fundamentally an epoch-wise double descent phenomenon Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")). These peaks and plateaus again coincide with the training loss on the repeated data approaching zero as shown in Figure [2](#S1.F2 "Figure 2 ‣ 1.1 Summary of Results ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data"). As in Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")) we see double descent effects caused by both increasing model size and epochs. We suspect there is a range in the middle where the data can be memorized and doing so consumes a large fraction of the model’s capacity, and this may be where the peak of degradation occurs, for a more thorough discussion of this question see the discussion (section [5](#S5 "5 Discussion ‣ Scaling Laws and Interpretability of Learning from Repeated Data")).
Repeated data can cause a divergence from power-law scaling. Figure [4](#S2.F4 "Figure 4 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") zooms in on the degradation of performance, measured as a function of model size for different repetition frequencies of the repeated data. For example, models trained for 1,220 repeats and 10% repeated data show a dip in performance to the equivalent of a model 0.55x as large, when the model size is 10M to 100M parameters. As the model size continues to increase, performance recovers to 0.8x model-size equivalent for a 1B parameter model. For a smaller number of repeats (122 repeats), the dip occurs later, centered around 1B parameters.
| | |
| --- | --- |
| | |
Figure 4: On the left we plot the same results as in Figure [2](#S1.F2 "Figure 2 ‣ 1.1 Summary of Results ‣ 1 Introduction ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), re-parameterized in terms of the effective model size multiplier implied by the test loss (performance equal to a model with x times as many parameters). For a given number of repetitions, degradation occurs only for a specific range of model sizes. For example, for the blue curve (122 repeated epochs), we see almost no performance deviation from a power law scaling law (line on log-log graph) until the model is scaled up to 100M parameters, after which we see a divergence. We see the same divergence around 400M parameters for 12,200 repeated epochs. The right graph shows a large, predictable region over which the degradation occurs, and suggests that large models like GPT-3, Gopher, and PALM Brown et al. ([2020](#bib.bib36 "Language models are few-shot learners")); Rae et al. ([2021](#bib.bib37 "Scaling language models: methods, analysis, and insights from training gopher")); Bi et al. ([2020](#bib.bib7 "PALM: pre-training an autoencoding and autoregressive language model for context-conditioned generation")) need to be careful about overfitting their high quality distributions like Wikipedia and books – although note that this holds constant the number of total training tokens. The blue and green curves correspond to the right and left sides of the double descent region where we observe 50% of the maximum effect. They are an aggregation of that curve for the scans where we trained on 3%, 10%, 20%, 50%, and 90% repeated data. The details of both fits are in Appendix [A](#A1 "Appendix A Model Size Multiplier and Poor Performance Region Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data"). A large number of runs needed to be aggregated to produce a clean fit for region of reduced performance.
The right panel of Figure [4](#S2.F4 "Figure 4 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") shows the range over which we observe at least 50% of the maximum degradation; this corresponds to a “band” or region in the (model size, repetition frequency) plane. Both boundaries of the region are a good fit to a power law relating frequency of repetition to the number of parameters of the model, namely:
| | | |
| --- | --- | --- |
| | E=k∗Nα | |
where E corresponds to epochs of repetition and N corresponds to the parameters in the model. it is notable that the lines in figure 2b are relatively parallel. The fits for the above lines are given in the table below:
| | | |
| --- | --- | --- |
| | k | α |
| right boundary | 5.1e7 | -.50 |
| left boundary | 4.2e6 | -.56 |
Note that extrapolating these boundaries leads to a prediction of significant degradation from repeating data as little as 2x on state-of-the-art language models with hundreds of billions of parameters, although this applies for a constant number of training tokens (100B). In practice large models are trained for more than thisHoffmann et al. ([2022](#bib.bib35 "Training compute-optimal large language models")), and as shown in Figure [3](#S2.F3 "Figure 3 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), training past the double descent peak is helpful, so the degradation would likely not be quite as bad. When looking at Figure [3](#S2.F3 "Figure 3 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") we see that the the poor performance region would be shifted left for large models trained on the compute efficient frontier (the pareto frontier of compute and performance) Kaplan et al. ([2020](#bib.bib33 "Scaling laws for neural language models")).
Overall it seems that in addition to being robust to task, model size, and architecture as shown in previous work Advani and Saxe ([2017](#bib.bib25 "High-dimensional dynamics of generalization error in neural networks")); Belkin et al. ([2018](#bib.bib26 "Reconciling modern machine learning practice and the bias-variance trade-off")); Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")) double descent as a general phenomenon appears to be robust to occurring in a sub-distribution and that it can have a large effect on overall performance even while being a modest fraction of training tokens.
Repeated data causes a disproportionately large performance hit to copying, a mechanism for in-context learning. The ability of a language model to copy text (in the sense of being provided with a context consisting of a passage repeated several times, and testing whether the model can repeat it once more) is a potential measure of generalization, as copying is independent of the content of the text. Also, recent interpretability work has suggested that copying may be implemented by crisp internal algorithmic structures (Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads"))), again suggesting generalization. It thus seems valuable to investigate what happens to copying during a memorization-related degradation in performance, which we have shown above occurs in our experiments.
To do this constructed a simple evaluation in which copying is heavily emphasized: we measure the loss on the first paragraph of Harry Potter copied 11 times. The models trained on repeated data performed much worse on this evaluation (Figure [5](#S2.F5 "Figure 5 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")), substantially out of proportion to the degradation on the loss itself. In other words, copying is preferentially harmed by training on repeated data. For example, a 3% fraction of repeated data leads to a 1.15x reduction in effective model size (performance equal to model with 1.15 fewer parameters) on the general loss, but a much larger 3x effective model size reduction in terms of copying ability. As can be seen in Figure 5, the damage to copying is greater than the damage to overall loss across the entire range of repeated data fractions. This suggests that the shift to memorization caused by repeated data is selectively harming at some behaviors associated with generalization.
To get another view on the same phenomenon, we measured the loss of various models on the Xth consecutive copy of the Harry Potter paragraph, where X runs from 1 to 12. As shown in Figure [7](#S2.F7 "Figure 7 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") (left), for most models the loss gradually decreases with increasing numbers of copies of the paragraph (i.e. the model has an easier time predicting an additional copies after seeing more consecutive copies), but at the peak of the double descent phenomenon, the loss is much higher and, strikingly, does not decrease at all with additional copies of the paragraph. This large aberration shows how strong the selective effect of the double descent phenomenon on copying is. General in-context learning is also harmed at the pessimal number of repeated epochs (Figure [7](#S2.F7 "Figure 7 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") right), though to a lesser extent than copying.
| | |
| --- | --- |
| | |
Figure 5: We constructed a simple measure of the model’s copying ability, consisting of the loss on the first paragraph of Harry Potter repeated 11 times. We measured the double descent peak performance for a given model size and fraction of repeated data and compared that to a fit of these evaluations on the control model (trained on unique text) scan to generate an effective model size. We observe that 3% repeated data at the pessimal number of repeated epochs caused a 3x reduction in effective model size on this task for a for several model sizes, whereas it only caused at most a 1.15x reduction in effective model size on test loss. We see much larger effects on the copying evaluation than on overall performance for repeated data fractions between 3% and 20%. The model size multiplier for copying is based on interpolation and the model size multiplier for test loss is based on a power law fit (see Appendix [C](#A3 "Appendix C Appendix: Copying and Prefix Matching Score Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data") for more details).
The disproportionate performance hit to copying coincides with a disproportionate degradation of induction heads.
Having connected the damage associated with repeated data with a measure of generalization (in-context copying of text), we next took the connection one step further, by trying to also probe the potential mechanistic basis of copying. Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) identifies “induction heads” as a possible basis for copying and in-context learning behavior in general, so we decided to measure these and try to connect them back to the repeated data double descent phenomenon.
Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) defines induction heads by their ability to facilitate simple copying given a repeated random sequence of tokens (though in practice this definition ends up including heads with more complex behaviors too). Induction heads use a circuit of 2 attention heads to "complete the pattern by copying and completing sequences." This can be split up into attending to the relevant token (prefix matching) and increasing the logit corresponding to the attended-to token.
| | |
| --- | --- |
| | |
Figure 6: Comparison of degradation of prefix matching score with repeated data, compared to general degradation of the test loss. We measured the double descent peak performance for a given model size and fraction of repeated data and compared that to a fit of the prefix matching score on the control model scan to generate an effective model size. We observe that 3% repeated data causes on average [21](#A3.F21 "Figure 21 ‣ Appendix C Appendix: Copying and Prefix Matching Score Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data") a 1.47 model size multiplier on prefix matching score while causing less than a 1.15x model size reduction in effective model size on test loss. Again we see much larger effects on the prefix matching score than on overall performance for repeated data fractions between 3% and 20%. The model size multiplier for prefix matching is based on a linear fit (see Appendix [C](#A3 "Appendix C Appendix: Copying and Prefix Matching Score Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data") for more details of fit). The test loss shown on the right is the same graph as in Figure 5, but with differently scaled axes for ease of comparison.
We decided to probe the prefix matching score as measure of mechanistic structure that is distinct from the behavior of copying itself. Figure [6](#S2.F6 "Figure 6 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") shows the same setup as Figure [5](#S2.F5 "Figure 5 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") except for prefix matching score instead of copying loss. As can be seen in the figure, preferential damage to prefix matching score is not present across the whole range of repeated data fraction as it is for copying, but at low fractions of data repeated, there is still preferential damage. For example, at 3% repeated tokens, there is a 2x effective parameter decrease in prefix matching score, but only a 1.15x effective parameter decrease in general (test) loss.
As another example, we find it interesting that the sharp drop in prefix matching score for a 1.5M parameter model with 50% repetition corresponded to a complete breakdown of paragraph level copying. This complete breakdown of paragraph level copying corresponds to a 1.5M parameter model having the effective overall performance of a 30,000 parameter model, while having an equivalent prefix matching score to a model with effectively 2,000 parameters.
Although not as conclusive as the previous results, these clearly show that prefix matching is preferentially degraded in some cases.
| | |
| --- | --- |
| | |
Figure 7: Degradation of copying and in-context learning at the peak of the double descent curve. On the left we show the 2-layer models trained on 50% repeated data from Figure [5](#S2.F5 "Figure 5 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), evaluated on the first paragraph of Harry Potter copied X times where X runs from 1 to 11. In Appendix [D](#A4 "Appendix D Appendix: Harry Potter Copying Evaluation with Fewer Characters ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), we explore shortening the length of the paragraph to verify the problem is with copying rather than long contexts. The right shows per token losses on the test set. Both graphs show dramatically reduced performance (higher copying loss, lower benefit to in-context learning) at the peak of the double descent.
One and two-layer attention only models are worse at copying and fuzzily copying proper names on inspection. To examine the effect on induction heads and in-context learning even more closely, we looked at more granular copying in one and two layer attention-only transformers, for which interpreting the internal structure (and especially induction heads) is known to be particularly straightforward Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")); Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")). That is, we can reverse engineer a large portion of attention-only-transformers (no MLP’s) with a circuits-level understanding (understanding how individual neurons act together to produce useful behavior) Cammarata et al. ([2020](#bib.bib14 "Thread: circuits")). These small models also exhibit the same double-descent phenomenon as larger models (Appendix [B](#A2 "Appendix B Appendix: Logit Attribution Analysis, 2 Layer Models ‣ Scaling Laws and Interpretability of Learning from Repeated Data")).
For 1-layer attention only models, where copying takes the form of skip-trigrams, we can easily see that the repeated data model is worse at a form of copying associated with these skip trigrams. Namely, we compare the probabilities that the repeated data and control models assign to each token in a paragraph, and focus especially on proper names which occur repeatedly in the paragraph (Figure [8](#S2.F8 "Figure 8 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")). The most obvious way to correctly predict these re-occurring names is by copying, and we see that in most cases the control model (trained on unique text) performs much better than the one with repeated data (yellow underlines).
Figure 8: Visualization of the difference in loss on the first paragraph of Harry Potter for control and 10%-repeated-data runs of a 1-layer attention-only model. Orange highlights correspond to the control model performing better, purple corresponds to the repeated data performing, and the intensity corresponds to the magnitude of the difference in per token losses. Proper names (which are a good target for copying when they occur more than once) are underlined in yellow on second or later occurance; it is clear that the control model performs better on these. Often the difference is dramatic: for the last three appearances of “Potters” the control model puts a >97% chance on “ters” given “Pot”, whereas the repeated data model puts <4% chance on that token.
Very specifically, predicting repeated names requires exactly a skip-trigram pattern Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")) which is the algorithmic operation 1-layer attention-only models are known to perform. For example, the following skip-trigrams are useful in the Harry Potter paragraph in Figure [8](#S2.F8 "Figure 8 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data"):
| | | |
| --- | --- | --- |
| | [a][b]…[a]=>[b][Pot][ter]…[Pot]=>[ter] | |
| | | |
| --- | --- | --- |
| | [a][b]…[a]=>[b′][Pot][ter]…[Pot]=>[ters] | |
We also plotted the same visualization for a 2-layer attention-only model (which is known to contain simple induction heads), and find the control model is better at fuzzy copying (Figure [9](#S2.F9 "Figure 9 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")).
Figure 9: Same as Figure 9, but for 2-layer attention-only models. Proper names (which are a good target for copying when they occur more than once) are underlined in yellow on second or later occurance. Here the repeated-data model sometimes does better on repeated proper names, but there are still clear examples of the control performing much better. These examples are highlighted in green and discussed. On the token [ley] in the second appearance of [D][urs][ley] the control model places a 92% likelihood on [ley] whereas the repeated data model places a 10% likelihood. On the token [leys] in the second appearance of [D][urs][leys] the control model places a 44% likelihood on [leys] whereas the repeated data model places a 4.9% likelihood. On the [ley] in [ un][D][urs][ley][ish] the control model places a 68% likelihood on [ley] whereas the repeated data model places a 0.4% likelihood.
Visually, it is less obvious (compared to the 1-layer case) that the 2-layer repeated model is worse at names, and there are a few examples where it puts 1.1x higher odds on the correct token. But on the other hand there are dramatic cases of the control model doing 500x times better (odds ratio on correct token) for fuzzy copying, like unDursleyish, which is exactly the kind of degradation we’d expect to see from disrupting induction heads.
We attempted to leverage logit attribution (which earlier tokens contributed to the prediction of the current token through a "direct path" with this attention head) to see if the difference was primarily due to the induction head being less active or other heads interfering with it Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")). We were unable to find clear evidence of either, but we include our exploration of a 2 layer attention only model in Appendix [B](#A2 "Appendix B Appendix: Logit Attribution Analysis, 2 Layer Models ‣ Scaling Laws and Interpretability of Learning from Repeated Data").
Repeated data causes a smaller, disproportionate performance drop on our out-of-distribution evaluations.
| | |
| --- | --- |
| | |
Figure 10: We observe that training on high levels of repeated data causes a small disproportionate drop on out-of-distribution performance (Python loss). The effect is noisy, but since we do not see a model size effect we take the average in the figure on the right (harmonic mean of multipliers). For large repeated fractions of 50% and 90% we see model size multipliers of .84 and .75.
Given that we overfit the model, we expected it to perform worse off distribution, which we do observe (Figure [10](#S2.F10 "Figure 10 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")). We notice almost an opposite pattern to what we observed in the induction head results. We see most of the disproportionate drop at 50% and 90% rather than 1-10%.
We observe a double descent phenomenon in sparse sweep of models trained on python, but we the Python scans exhibit a somewhat different overall shape. To add more generality to our results, we repeated the same experiments on a Python dataset instead of natural language (Figure [11](#S2.F11 "Figure 11 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")). If we use the same method to fit the poor performance region, we see a broadly similar fit and a second epoch for today’s large models (approximately 200B parameters) is still robustly in the reduced performance region for python. However the fit is noisier than the fit for text and the two lines are no longer parallel.
| | |
| --- | --- |
| | |
Figure 11: Double descent phenomenon for models trained on python. Training on Python gives similar results to what Figure 2 and Figure 4 show for language models. Here 50% of the dataset consists of repeats and 50% is unique. On the left side is degradation in performance, occurring over a specific range of repetition that varies with model size. On the right, we again see a large region of poor performance as we did in Figure [4](#S2.F4 "Figure 4 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), although the fit is noisier. Again the blue and green curves correspond to the right and left sides of the double descent curve where we observe 50% of the maximum effect.
The noise may partially be explained by the Python fits being averaged over half as many settings for the fraction of tokens that are repeated data. It could also be that we need a higher resolution Python scan to get a cleaner estimate for the poor performance region. Finally, the Python data was trained on approximately 2 epochs as described in the methods section (so it included some repetition on the main dataset as well, not just the repeated subset). Python also may have more unintentional repetition than text, from copying and pasting of example code and forking of codebases. Such repetition could change the shape of the region of poor performance. More analysis of the Python experiments is shown in Appendix [A](#A1 "Appendix A Model Size Multiplier and Poor Performance Region Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data").
Pre-training on repeated data hurts fine-tuned performance We find that the negative impact of repeated data persists after fine-tuning natural-language models on Python (Figure [12](#S2.F12 "Figure 12 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data")).
| | |
| --- | --- |
| | |
Figure 12: Effect of repeated data during pre-training on fine-tuning. Models were pre-trained on 90% repeated data (red lines) or on totally unique data (blue lines), and then fine-tuned on Python (always unique data). The repetition frequency was chosen to maximize the performance hit. The model pre-trained on repeated data encounters a sizable performance hit during fine-tuning (left panel), causing it to not only perform worse than the model pre-trained on unique data, but also worse than a model trained from scratch (green line). The right panel shows fine-tuning curves of the two models. The model pretrained on repeated data performs much worse for several billion tokens (red line), but eventually catches up to the model pretrained on unique data (blue line).
It is noteworthy that the performance hit once fine-tuned is much smaller. An 800M model pre-trained on 50% repeated data from the double descent peak had its effective parameters reduced by 10x in Figure [15](#A1.F15 "Figure 15 ‣ Appendix A Model Size Multiplier and Poor Performance Region Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data") in Appendix [A](#A1 "Appendix A Model Size Multiplier and Poor Performance Region Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data"). When we fine-tune from the repeated model we see a 1.6x reduction in effective parameters compared to training from scratch. This is still meaningful damage to the model, but it is recovered substantially. Since the repeated model forgets the repeated dataset after a modest amount of fine-tuning (Figure [12](#S2.F12 "Figure 12 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data"), we consider the fine-tuned model with repeated data pre-training to be dominated by the fine-tuned model from the unique dataset.
3 Methods
----------
The decoder-only transformer models were trained on an 8192 token context with the same settings as described in Askell et al. ([2021](#bib.bib13 "A general language assistant as a laboratory for alignment")) for 100B tokens. Our language experiments utilized a 400B token dataset with 55% heavily filtered common crawl data (220B tokens), 32% internet books (128B tokens), and some smaller distributions including OpenWebText, Wikipedia, and Stack Exchange; most of which we sourced from The Pile Gao et al. ([2021](#bib.bib12 "The pile: an 800gb dataset of diverse text for language modeling")), and leveraged the 50,304 vocabulary GPT-2 encoding Radford et al. ([2019](#bib.bib27 "Language models are unsupervised multitask learners")); Wolf et al. ([2019](#bib.bib11 "HuggingFace’s transformers: state-of-the-art natural language processing")).
Code models were trained or fine-tuned on 45B tokens of Python for 2.2 epochs. Fine-tuning experiments had the same hyperparameters as pre-training experiments, but with learning rates reduced by a factor of 2 and reduced warmups.
We varied model size, repeated dataset size, and the fraction of tokens trained on repeated data by 3, 2.5, and 2 orders of magnitude respectively.
4 Related Work
---------------
Scaling Laws
A scaling law lens consists of finding a small set of hyperparameters that have large, predictable impacts on model performance, and was present throughout this work (at least one of the hyperparameters is generally model size, compute, or dataset size). The predictive nature of scaling laws makes them useful in a broad number of research and engineering settings. The implications of scaling laws are sufficiently broad and understandable that understanding them is relevant to policy makers Ganguli et al. ([2022](#bib.bib8 "Predictability and surprise in large generative models")). Predictable scaling trends in neural networks
were first studied with Hestness et al. ([2017](#bib.bib9 "Deep learning scaling is predictable, empirically")). Kaplan et al. ([2020](#bib.bib33 "Scaling laws for neural language models")) demonstrated that test loss performance on language
modeling tasks scales as a predictable function of model size, dataset size, and compute. The scaling law lens has become more popular over time. For instance scaling
laws have been shown in many modalities (e.g., images, video, math, etc.) Henighan et al. ([2020](#bib.bib32 "Scaling laws for autoregressive generative modeling")), acoustics Droppo and Elibol ([2021](#bib.bib17 "Scaling laws for acoustic models")), transfer to code,
Hernandez et al. ([2021](#bib.bib39 "Scaling laws for transfer")), and few-shot adaptation of vision models Prato et al. ([2021](#bib.bib16 "Scaling laws for the few-shot adaptation of pre-trained image classifiers")). Existing scaling laws have been revisited as training setups change; for instance, Hoffmann et al. ([2022](#bib.bib35 "Training compute-optimal large language models")) found that many recent large models have been under-trained. Our work uses the scaling law lens on an aspect of dataset quality and supplements the lens with an interpretability lens, and we believe our work is novel in both these respects.
Mechanistic Interpretability
A mechanistic interpretability lens was used in this work. Mechanistic interpretability refers to attempting to reverse engineer the detailed computations performed by the model. The mechanistic interpretability lens is useful for pure scientific understanding and has the potential to anticipate safety issues from future more powerful models. There is a relatively detailed understanding of mechanistic interpretability for convolutional image models Cammarata et al. ([2020](#bib.bib14 "Thread: circuits")), some understanding for multimodal models Goh et al. ([2021](#bib.bib2 "Multimodal neurons in artificial neural networks")); Radford et al. ([2021](#bib.bib1 "Learning transferable visual models from natural language supervision")), and such an understanding is starting to be built up for Transformers trained on language Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")); Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")). For a more thorough background on interpretability progress see the related work section of Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")). These results are an example of a “bridge” between microscopic phenomena inside the network and macroscopic trends in the loss, and we’re only aware of one other example of such a bridge Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")).
Double Descent
Double descent was first shown in generality by Belkin et al. Belkin et al. ([2018](#bib.bib26 "Reconciling modern machine learning practice and the bias-variance trade-off")) where it was observed for decision trees, random features, and 2-layer neural networks. Similar behavior has been observed in Opper ([1995](#bib.bib6 "Statistical mechanics of learning: generalization")); Malzahn and Opper ([2001](#bib.bib5 "A variational approach to learning curves")); Advani and Saxe ([2017](#bib.bib25 "High-dimensional dynamics of generalization error in neural networks")); Geiger et al. ([2019](#bib.bib24 "Jamming transition as a paradigm to understand the loss landscape of deep neural networks")); Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")). For a more thorough background on double descent see Nakkiran et al. Nakkiran et al. ([2019](#bib.bib22 "Deep double descent: where bigger models and more data hurt")). We extend the double descent phenomenon to a setting we see as more practical since data repetition in various forms appears to be a universal, long-term issue; whereas modern large language models are generally outside of the parameters and data regime of previously observed double descent phenomenon.
Rise of Engineering Large, Diverse Language Datasets
Algorithmic innovation Hernandez and Brown ([2020](#bib.bib40 "Measuring the algorithmic efficiency of neural networks")), compute Amodei et al. ([2018](#bib.bib38 "AI and compute")), and data are three of the major factors that drive the advance of AI. The engineering and science of large, diverse language datasets is relatively new. Pre-2017 many language models were trained on a single distribution of text, such as news articles Jozefowicz et al. ([2016](#bib.bib30 "Exploring the limits of language modeling")), Wikipedia Merity et al. ([2016](#bib.bib29 "Pointer sentinel mixture models")), or fiction books Kiros et al. ([2015](#bib.bib28 "Skip-thought vectors")). GPT-2 Radford et al. ([2019](#bib.bib27 "Language models are unsupervised multitask learners")) leveraged webtext, outbound Reddit links with at least 3 upvotes in order to use human curation/filtration to ensure quality in addition to a broad distribution. GPT-2’s capabilities are largely attributed to its scaled-up size and dataset (10x the parameters and 10x the data of GPT) Radford et al. ([2019](#bib.bib27 "Language models are unsupervised multitask learners")). The next generation of language models, Brown et al. ([2020](#bib.bib36 "Language models are few-shot learners")); Rae et al. ([2021](#bib.bib37 "Scaling language models: methods, analysis, and insights from training gopher")); Hoffmann et al. ([2022](#bib.bib35 "Training compute-optimal large language models")), leveraged large, diverse datasets that consist of many sub-distributions. Constructing such datasets includes a large number of decisions: choosing sampling weights, quality filtering, de-duplication, fuzzy de-duplication, epochs per dataset, and more. There has not yet been substantial public work that quantitatively shows the impact of such decisions, but the dataset ablations in Appendix A of the Gopher Rae et al. ([2021](#bib.bib37 "Scaling language models: methods, analysis, and insights from training gopher")) paper are notable. They clearly show the benefit of their dataset mixture, quality filter, exact de-duplication, and fuzzy de-duplication for 1.4B parameter models. Our work aims to provide some insights and potential diagnostics for researchers and engineers designing large datasets for language models.
5 Discussion
-------------
###
5.1 Why does repeating a small fraction of data damage performance so much?
We showed that a dataset with only 10% repeated tokens can reduce model performance by an effective 2x in parameter count, much more than if that 10% of the data had simply never been trained on. The repeated data thus degrades model performance out of proportion to its share in the dataset. Why does this occur, and why only for a specific amount of repetition? One plausible hypothesis comes from looking at the model’s “incentives” to memorize vs generalize. To informally explore this hypothesis consider the following rough numbers, a 800M parameter model typically has a loss of roughly 2.0 nats/token, a 400M parameter model has a loss of roughly 2.2 nats/token, and fully memorized data will have a loss of 0 nats/token. Now suppose a 800M model is trained on 90% unique data and 10% tokens consisting of repeated data. We can ask whether it is a “good tradeoff” for the model to memorize the repeated data (leading to 0 loss on 10% of the dataset), at the cost of degrading performance by the equivalent of a 2x multiple in model size (which raises loss on the other 90% from 2 to 2.2). Some simple arithmetic suggests that it is: 0.9∗2.2+0.1∗0=1.98<2.0. Another way to say this is that zero loss is such a huge drop compared to the differences in entropy between model sizes that driving the loss to zero on even a tiny subset can incentivize enormous degradation in quality.
This however leaves open the question of when this tradeoff is necessary or possible – and here is where the double descent phenomenon comes in. If a lot of data is repeated only a few times (say 5% of the data repeated 2x) then the model may not have the capacity to memorize it, and also does not see it enough times during training to do so. If a tiny amount of data is repeated very many times (say 0.01% of the data repeated 1000x), then the model will memorize it, but because it is so small the model need not use much capacity to do so, so the degradation in quality will likely be small. There is a range in the middle where the data can be memorized and doing so consumes a large fraction of the model’s capacity, and this may be where the peak of degradation occurs.
###
5.2 Generalization, memorization, and induction heads
Our results show that overfitting on the repeated data results in worse test loss, and this co-occurs with a disproportionate degradation in the model’s induction heads (prefix matching score) and its ability to copy text. Copying sequences can be seen as a form of generalization, as it requires algorithmic operations that are independent of the content of the data. Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")); Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) provided evidence for induction heads as the mechanism implementing copying and other pattern-matching. For the 2 layer model shown in Figure [7](#S2.F7 "Figure 7 ‣ 2 Results ‣ Scaling Laws and Interpretability of Learning from Repeated Data") it seems as if the pressure to memorize the repeated dataset has led a skip tri-gram head to replace the induction head entirely. Thus our results tell a story where a type of generalization and its internal implementation are disrupted when the model memorizes repeated data – a vivid illustration of the memorization-generalization trade-off. Future work could take this even further, by measuring the number of parameters devoted to memorization and trying to observe them competing for space with induction heads. Finally, it is worth noting that the co-occurence of copying degradation and induction head degradation is itself some additional evidence for induction heads as the source of in-context learning; Olsson et al. Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) was not fully conclusive and our results further bolster the case.
###
5.3 Bridging mechanistic interpretability and scaling laws
The results connecting memorization to the degradation of mechanistic interpretability structures Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) are an example of a “bridge” between microscopic phenomena inside the network and macroscopic trends in the loss. We view such connections as very fruitful tools for research, because they allow us to see the same thing through different lenses: the macroscopic behavior demonstrates the significance of the microscopic mechanisms, and the microscopic mechanisms help explain how and why the macroscopic phenomena occur. Switching back and forth between the two allows for a deeper understanding of both, as well as more robust diagnostics if something goes wrong. We are aware of at least one other instance of such a bridge – the correspondence between the formation of induction heads and the boost in in-context learning near the beginning of training Elhage et al. ([2021](#bib.bib15 "A mathematical framework for transformer circuits")); Olsson et al. ([2022](#bib.bib3 "In-context learning and induction heads")) – but such connections remain rare so far, and we believe that finding more of them is a promising route to more deeply understanding neural nets.
###
5.4 Repeated data and fine-tuning
We hypothesized repetition might help explain why models trained from scratch sometimes outperformed models that were pre-trained and then fine-tuned Hernandez et al. ([2021](#bib.bib39 "Scaling laws for transfer")). For our purposes, we define ossification as any pre-training that leads a fine-tuned model to perform worse than a model trained from scratch (given a fixed compute and data budget). It required relatively extreme repetition in pre-training (90% training on repeated tokens at peak of double descent curve, 73x reduction in effective model size) to see a large ossification effect (1.6x reduction in effective model size) within our fine-tuning setup. We still think repetition might explain a large fraction of ossification when we consider training on various types of repetition we did not study here (sentence level, paragraph level, similar documents, distribution, etc). Overall, our finding that repetition can induce ossification provides medium causal evidence to this hypothesis. We think ossification is an interesting phenomenon that merits further study.
###
5.5 Limitations
We attempt to discuss limitations throughout the text where appropriate, but for the reader’s convenience, we enumerate them here. We attempt to list them in a loosely descending order of importance.
1. We used a fixed number of tokens for all models (similar to the GPT-3 model sweep), because these models were trained prior to the release of Chinchilla, which showed the compute frontier (pareto frontier of performance and compute) is quite different than previously understood Brown et al. ([2020](#bib.bib36 "Language models are few-shot learners")); Hoffmann et al. ([2022](#bib.bib35 "Training compute-optimal large language models")).
2. Our fits for region of poor performance were relatively noisy, and we only observed a clean trend by aggregating them. This is discussed in the Results section and further explored in Appendix [A](#A1 "Appendix A Model Size Multiplier and Poor Performance Region Fits ‣ Scaling Laws and Interpretability of Learning from Repeated Data").
3. The data we repeated was a random subset of the original dataset, and is thus not directly applicable to the situation where higher quality data (such as Wikipedia) is intentionally repeated to improve quality. Nevertheless, it seems plausible that the results would carry over.
4. We measured loss, rather than downstream NLP evaluations. Overfitting does not always entail worse performance on downstream tasks Ouyang et al. ([2022](#bib.bib4 "Training language models to follow instructions with human feedback")), so it is possible that the degradation we observe does not carry over to these tasks.
5. We did not explore the effects of early stopping, dropout, weight decay, or other regularization.
6. We did not investigate simpler systems than 1L attention-only models, which might contain more complete mechanistic insights.
###
5.6 Future Directions
Below are some future directions we think are promising:
1. A compute efficient frontier scan to predict the poor performance region.
2. Varying the type of repetition. We could inject repeated sentences or paragraphs at the beginning or end of some fraction of contexts, or repeat chunks of documents in a different order. We could also explore cases where the repeated data has a different distribution than the unique data.
3. Further interpretability work. Are there neurons that tell the model what distribution it is in: unique or repeated? Are there neurons through which we can observe and edit the repeated sequences?
4. Drill down on memorization and generalization. Could we measure the number of model parameters taken up by memorization vs generalization, either behaviorally or by using mechanistic interpretability to identify parameters that are storing memorized data? Can we measure how this varies across the double descent, and thus watch the competition between memorized data and induction heads for model capacity?
5. Could repetition and double descent help explain loss spikes during training? If a model can largely memorize a particularly easy batch in a single gradient step then a very skinny double descent could present as a loss spike.
6 Conclusion
-------------
We’ve shown that small fractions of repeated data, if repeated at the right frequency, can cause surprisingly severe degradation to model performance. We show that this degradation scales predictably, occurs across datasets, and is associated with disproprotionate damage to internal mechanisms associated with generalization, such as induction heads. In practical terms, these results provide a tool for predicting and diagnosing data-repetition-related problems in language models. In more conceptual terms, they are an example of a bridge between the macroscopic domain of scaling laws and the microscopic domain of mechanistic interpretability, as well as a lens for gaining a more detailed understanding of how generalization and memorization work. We believe these conceptual themes are promising ones, and hope to see more work that employs them.
Acknowledgments
---------------
We thank Ethan Perez, Jan Leike, and Martin Wattenberg for helpful feedback on the draft. We thank Daniela Amodei, Jamie Kerr, Jia Yuan Loke, Rebecca Raible, and Tim Telleen-Lawton for support with the project.
Author Contributions
--------------------
Danny Hernandez led the project performed the majority of experiments, analysis, and writing.
Tom Brown led engineering efforts for the scaling team, including efficient pre-training and gave helpful feedback on the paper.
Tom Conerly made engineering contributions on the scaling team.
Nova DasSarma managed the underlying cluster infrastructure.
Dawn Drain helped with pre-training research and infrastructure.
Sheer El-Showk helped with pretraining research and dataset construction.
Nelson Elhage contributed significantly to interpretability tooling, provided support on that tooling, and gave helpful feedback.
Zac Hatfield-Dodds helped with codebase maintenance and with engineering
Tom Henighan helped with pretraining the underlying language models, with dataset creation, with managing the cluster during some phases of the project, and gave helpful feedback on the paper.
Tristan Hume contributed to interpretability tooling that was leveraged in this work.
Scott Johnston helped with pretraining research.
Ben Mann contributed to pretraining and cluster management.
Chris Olah lead the interpretability team, which provided tooling and support for this work.
Catherine Olsson contributed to interpretability tooling, provided support on that tooling, and provided interpretability research advice.
Dario Amodei contributed greatly to the framing and writing of the work and advised the project.
Nicholas Joseph helped design and build a framework for efficient training of large language models, gave helpful feedback on the paper, and advised the project.
Jared Kaplan led pre-training efforts initially and advised the project.
Sam McCandlish led pre-training efforts and advised the project. |
466b5391-bc15-4949-8425-27a428929a38 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Reinforcement Learning Study Group
Hey everyone,
my name is Kay. I'm new to the forum and I came here with a specific goal in mind:
>
> I'm putting together a crew for a **reinforcement learning study group.**
>
>
Main objectives:
----------------
* **Mathematical Foundations**: We will work through key passages in [Sutton & Barto's Book](https://d3c33hcgiwev3.cloudfront.net/Ph9QFZnEEemRfw7JJ0OZYA_808e8e7d9a544e1eb31ad11069d45dc4_RLbook2018.pdf?Expires=1640649600&Signature=N7aRSLbaGk9umk7rFRUfLMvjjMjVLx3bL1ZK6h3dx7BjeC7SrEzxoPcSVepzFNw13Ee2yWGgL5~hCxWSa0WQA4TaBOIesIjzWOLcHnLYHCOnLTbNN9oHwYD6TfmzlQqSkb40tLYuWfM2ZvDtujzkCsIrqCDSei0UvZDifDprpAU_&Key-Pair-Id=APKAJLTNE6QMUY6HBC5A) to get a good foundation of RL
* **Research Papers:** We will follow the ["Spinning up in Deep Reinforcement Learning" by OpenAI](https://spinningup.openai.com/en/latest/#) resource to select key research papers in RL to read.
* **Code:** The above resource can also be used to explore main RL algorithms and ideas which we'll aim to program from scratch in Pytorch/Tensorflow
* **Get Practical Skills:** In the long term, the goal of this group is to prepare ourselves for work in AI Alignment and adjacent fields as engineers, researchers etc. (Work at DeepMind, OpenAI, MIRI, FHI, etc.)
Important Self-Selection:
-------------------------
It is important for me to be surrounded by ambitious and self-motivated people who are reliable, friendly, and helpful. I expect members of the study group to invest a good amount of time to study and even more to try to teach one another. Members who only plan on participating passively will be asked to become more active or leave the group.
> Think of the study group as a sports team. The goal is to put the best people together on the field. Those who aren't ready to play will have to sit on the sideline.
>
>
Other helpful prerequisites are being able to program in Python 3, some knowledge of Probability and Statistics, as well as Linear Algebra and Calculus. In case this seems a little intimidating, don't worry, err on the side of joining the group. You'll pretty soon discover whether you're a good fit or not. Remember, we're all beginners who are united by the goal of becoming better.
Becoming our Coach:
-------------------
In the above spirit, I'm also looking for a more **advanced coach,** someone who knows the terrain of reinforcement learning relatively well and who's willing to guide the teams' efforts towards a more fruitful, albeit challenging learning experience.
Anyone who's familiar with the main concepts in [Sutton & Barto's Book](https://d3c33hcgiwev3.cloudfront.net/Ph9QFZnEEemRfw7JJ0OZYA_808e8e7d9a544e1eb31ad11069d45dc4_RLbook2018.pdf?Expires=1640649600&Signature=N7aRSLbaGk9umk7rFRUfLMvjjMjVLx3bL1ZK6h3dx7BjeC7SrEzxoPcSVepzFNw13Ee2yWGgL5~hCxWSa0WQA4TaBOIesIjzWOLcHnLYHCOnLTbNN9oHwYD6TfmzlQqSkb40tLYuWfM2ZvDtujzkCsIrqCDSei0UvZDifDprpAU_&Key-Pair-Id=APKAJLTNE6QMUY6HBC5A) would fit this role.
Schedule:
---------
The study group will start as soon as possible in January. Once there is at least one other participant aside from me, we will find a suitable timetable and meet at least once a week to do pair-programming, code reviews, work through mathematical formulas, discuss papers, and more.
How to Join:
------------
Contact me via [kozaronek@gmail.com](mailto:kozaronek@gmail.com) and introduce yourself, I'm happy to hear from you and excited to get started! |
682b307e-2067-4750-8dc5-c2c26afa97cd | trentmkelly/LessWrong-43k | LessWrong | How to balance between process and outcome?
I've been thinking recently about how to balance between process (how I get work done) and outcomes (what I achieve). I thought I'd ask the LessWrong community to see if anyone else has thoughts about this they'd like to share. I feel like both are important, but outcomes is a more long-term focus thing and process more of a daily thing. Outcomes are like long-running experiments for how you judge between different styles of process? In cases where it's hard to get reliable outcome answers, when failing at hard things or succeeding at easy things, or timeframes are long, or uncertainty high, it can be tempting to over-update on limited evidence. Is it then better to test process types on easier examples and then extrapolate to harder ones? |
7c82e7de-8567-4ae3-84a6-4774e05f9b5a | trentmkelly/LessWrong-43k | LessWrong | An exploration of GPT-2's embedding weights
I wrote this doc in December 2021, while working at Redwood Research. It summarizes a handful of observations about GPT-2-small's weights -- mostly the embedding matrix, but also the LayerNorm gain parameters -- that I found while doing some open-ended investigation of the model. I wanted to see how much I could learn by studying just those parameters, without looking at the attention layers, MLP layers, or activations.
This is still mostly unedited. Feedback and questions are very welcome.
Embedding matrix glossary
* Latent space – 768-channel-dim space that activations between blocks (and in skip connections) live in.
* Token space – 50,257-token-dim space of one-hot token vectors
* [Token] embedding matrix – W of shape [768, 50257] (this is the transpose of `model.lm_head.weight`)
* Embedding vector – Any 768-dimensional column of W
* Embedding space – Latent space
Preferred channel basis from LayerNorm
This section involves more math than the rest; you can skip it without missing much. "Preferred basis" means the same thing as "privileged basis" as defined by Anthropic here.
Most of GPT-2’s architecture – the attention layers, fully-connected layers and residual connections – don’t impose a preferred basis on the latent space.
(The skip connections give a canonical map between the latent spaces in different layers, so that it makes sense to talk about “the” latent space globally.)
The LayerNorm layers actually do privilege a basis, though, for two reasons.
LayerNorm(x) is defined as follows, where w and b are learned parameters:
* yi=∑j(Idij−1i1j)xj(subtract mean; 1 is an all-ones vector)
* zi=yi/√∑jyjyj (normalize variance)
* LayerNorm(x)i=wizi−bi (apply bias and gain parameters)
The first line picks out part of a preferred basis – a preferred vector (and the subspace orthogonal to it), namely the all-ones vector 1.
The second line doesn’t pick out a basis, because the stdev of the components of y is just its L2 norm and normalizing a vec |
a39d640b-95d4-4cc9-ab41-26c34966cc4c | trentmkelly/LessWrong-43k | LessWrong | Call for resources on the link between causation and ontology
I'm a little ashamed to admit I only read "Why Correlation Usually ≠ Causation" yesterday. It's very, very good, and you should read it too.
My essential takeaway from it is this: You can find nonzero correlations between almost anything you care to measure. However, it seems unlikely that the number of causal relationships in the universe scales at all proportionally to the number of correlative ones in the Universe.
This question feels like the wrong one to be asking to me, somehow. It feels ontology-flavored, in a way that doesn't make it a great match for how I normally think about statistics, and I would appreciate some book recommendations on the subject in the comments. But first, let me try to explain my thinking on this.
Start with the "base" layer of reality, the movement of atoms, or electrons, or strings, or what-have-you. If we are watching the actions and reactions of that layer from afar, then it seems to me that we have the best possible environment for doing a few experiments to first demonstrate correlation, and then a few more to demonstrate causation afterwards. While we can never be 100% sure, we can asymptotically reach certainty in that world. So far, so good; there's a reason experimental physics can get so precise with its predictions.
When you go one layer of abstraction up -- to molecules, if our base layer was "atoms" -- it seems to me that suddenly the difficulty of ascertaining causation should skyrocket. There are many more confounding variables and possibilities, that make designing an adequate experiment much harder. In addition, it is harder to define "molecule" precisely than it was to define "atom". How far do we move the constituent atoms apart before we turn a molecule into a non-molecule, for example? That seems like a question that you have to sometimes answer differently depending on different scenarios.
The experiments you run for correlation between molecules, on the other hand, might be harder, but I don't get the fee |
eb96897e-5dd4-463d-a492-878fe03e0961 | trentmkelly/LessWrong-43k | LessWrong | In favor of accelerating problems you're trying to solve
> John von Neumann, a renowned Hungarian-American mathematician and physicist, played a critical role in the Manhattan Project, the top-secret research effort during World War II that led to the development of the first atomic bombs. As a key contributor, he provided important insights into the mathematical modeling of nuclear chain reactions, which were instrumental in the design and construction of the weapons. After the war, von Neumann continued to shape nuclear deterrence policy, advocating for a strategy of mutually assured destruction (MAD) to prevent large-scale conflict. By emphasizing the catastrophic consequences of a full-scale nuclear exchange, MAD established a balance of power that, in turn, helped avert the existential risk of nuclear war. Von Neumann's early research and development of primitive nuclear weapons thus contributed significantly to global stability in the face of an unprecedented threat.
Don't accelerate problems you're trying to solve argues for the relatively intuitive notion that we shouldn't accelerate problems we are trying to solve. In this post, I will argue in favor of acceleration, and explain how to do it properly.
"Don't accelerate" seems like a default and conservative option. I think this causes people to fail to use a security mindset when thinking about it, even when they normally pretty good at it.
However, it doesn't take much creativity to see potential catastophic risks from this strategy. We can just take examples from history:
* Alcohol prohibition in the US resulting in a large increase in organized crime.
* "The Smokey Bear Effect": fighting fires too aggressively leads to an increase in risk of extremely large fires.
* Eliezer Yudkowsky causing various AGI firms to get founded and funded.
So we can't automatically treat "Don't accelerate" as the safe option.
The key, I think, is differential acceleration. Society is a complex adaptive system. The key metric is how much your accelerating causes society |
c271b046-6f66-4a7b-b536-53b34e0fe3eb | trentmkelly/LessWrong-43k | LessWrong | Research is polygamous! The importance of what you do needn't be proportional to your awesomeness
In a recent discussion a friend was telling me how he felt he was not as smart as the people he thinks are doing the best research on the most important topics. He said a few jaw-dropping names, which indeed are smarter than him, and mentioned their research agenda, say, A B and C.
From that, a remarkable implication followed, in his cognitive algorithm:
Therefore I should research thing D or thing E.
Which made me pause for a moment. Here is a hypothetical schematic of this conception of the world. Arrows stand for "Ought to research"
Humans by Level of Awesome (HLA) Research Agenda by Level of Importance. (RALI)
HLA RALI
Mrs 1 --------> X-risk #1
2 --------> X-risk #2
3 --------> Longevity
4 --------> Malaria Reduction
5 --------> Enhancement
1344 --------> Increasing Puppies Cuteness
Etc...
It made me think of the problem of creating match making algorithms for websites where people want to pair to do stuff, such as playing tennis, chess or having a romantic relationship.
This reasoning is profoundly mistaken, and I can look back into my mind, and remember dozens of times I have made the exact same mistake. So I thought it would be good to spell out 10 times in different ways for the unconscious bots in my mind that didn't get it yet:
1) Research agenda topics are polygamous, they do not mind if there is someone else researching them, besides the very best people who could be doing such research.
2) The function above should not be one-to-one (biunivocal), but many-to-one.
3) There is no relation of overshadowing based on someone's awesomeness to everyone else who researches the same topic, unless they are researching the same narrow minimal sub-type of the same question coming from the same background.
4) Overdetermination doesn't happen at the "general topic level".
5) Awesome people do not obfuscate what less awesome people do in their area, they catapult it, by creating resources.
6) Being |
00c061df-81f1-40b2-91e9-b32a9eae7441 | trentmkelly/LessWrong-43k | LessWrong | How to evaluate control measures for LLM agents? A trajectory from today to superintelligence
This is a linkpost accompanying a new paper by UK AI Security Institute, Apollo Research and Redwood Research. Please see the full paper for more details.
TLDR: Our new paper outlines how AI developers should adapt the methodology used in control evaluations as capabilities of LLM agents increase.
Figure: We sketch a trajectory of how control evaluations might evolve through increasingly powerful capability profiles. The trajectory is based on a sequence of five fictional models (M1-M5) with progressively advanced capabilities. For each model, we define an AI Control Level (ACL) based primarily on its threat model-specific capabilities. By mapping these capabilities to appropriate control evaluation methodologies, we provide a systematic approach for adapting safety measures to AI systems as they advance in capabilities.
What are the advantages of AI control?
The field of AI control attempts to design and evaluate control protocols: safety measures that are robust to intentional subversion by AI systems. Elements of promising protocols vary widely, from secure sandboxing to AI monitors and human auditing.
A particularly attractive feature of the AI control approach is that it offers the potential to mitigate alignment risks without fundamental research breakthroughs. A major focus of the field is on the engineering challenge of designing protocols that meet the challenges of deployments of increasingly capable autonomous agents.
How can AI control techniques scale to increasingly capable systems?
While the academic field of AI control is very young with a key foundational paper only published in 2023, many components of control evaluations (e.g. red teaming) and control measures (e.g. human approval) are already used to address the risks posed by existing AI systems.
As AI systems become more capable of causing harm autonomously, AI developers will need increasingly sophisticated control measures and evaluations to justify that the models can be deplo |
007d72ec-d20f-401d-8818-aa286fdcbf79 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Attentive Neural Processes
1 Introduction
---------------
Regression tasks are usually cast as modelling the distribution of a vector-valued output y given a vector-valued input x via a deterministic function, such as a neural network, taking x as an input. In this setting, the model is trained on a dataset of input-output pairs, and predictions of the outputs are independent of each other given the inputs. An alternative approach to regression involves using the training data to compute a distribution over functions that map inputs to outputs, and using draws from that distribution to make predictions on test inputs. This approach allows for reasoning about multiple functions consistent with the data, and can capture the co-variability in outputs given inputs. In the Bayesian machine learning literature, non-parametric models such as Gaussian Processes (GPs) are popular choices of this approach.
Neural Processes (NPs) (Garnelo et al., [2018a](#bib.bib10), [b](#bib.bib11)) offer an efficient method to modelling a distribution over regression functions, with prediction complexity linear in the context set size. Once trained, they can predict the distribution of an arbitrary target output conditioned on a set of context input-output pairs of an arbitrary size. This flexibility of NPs enables them to model data that can be interpreted as being generated from a stochastic process. It is important to note however that NPs and GPs have different training regimes. NPs are trained on samples from multiple realisations of a stochastic process (i.e. trained on many different functions), whereas GPs are usually trained on observations from one realisation of the stochastic process (a single function). Hence a direct comparison between the two is usually not plausible.
Despite their many appealing properties, one substantial weakness of NPs is that they tend to underfit the context set. This manifests in the 1D curve fitting example on the left half of Figure [1](#S2.F1 "Figure 1 ‣ 2 Background ‣ Attentive Neural Processes") as inaccurate predictive means and overestimated variances at the input locations of the context set. The right half of the figure shows this phenomenon when predicting the bottom half of a face image from its top half: although the prediction is globally coherent, the model’s reconstruction of the top-half is far from perfect.
In an NP, the encoder aggregates the context set to a fixed-length latent summary via a permutation invariant function, and the decoder maps the latent and target input to the target output.
We hypothesise that the underfitting behaviour is because the mean-aggregation step in the encoder acts as a bottleneck: since taking the mean across context representations gives the same weight to each context point, it is difficult for the decoder to learn which context points provide relevant information for a given target prediction. In theory, increasing the dimensionality of the representation could address this issue, but we show in Section [4](#S4 "4 Experimental Results ‣ Attentive Neural Processes") that in practice, this is not sufficient.
To address this issue, we draw inspiration from GPs, which also define a family of conditional distributions for regression. In GPs, the kernel can be interpreted as a measure of similarity among two points in the input domain, and shows which context points (xi,yi) are relevant for a given query x∗. Hence when x∗ is close to some xi, its y-value prediction y∗ is necessarily close to yi (assuming small likelihood noise), and there is no risk of underfitting. We implement a similar mechanism in NPs using differentiable attention that learns to attend to the contexts relevant to the given target, while preserving the permutation invariance in the contexts. We evaluate the resulting Attentive Neural Processes (ANPs) on 1D function regression and on 2D image regression. Our results show that ANPs greatly improve upon NPs in terms of reconstruction of contexts as well as speed of training, both against iterations and wall clock time. We also demonstrate that ANPs show enhanced expressiveness relative to the NP and is able to model a wider range of functions.
2 Background
-------------
Figure 1: Comparison of predictions given by a fully trained NP and Attentive NP (ANP) in 1D function regression (left) / 2D image regression (right). The contexts (crosses/top half pixels) are used to predict the target outputs (y-values of all x∈[−2,2]/all pixels in image). The ANP predictions are noticeably more accurate than for NP at the context points.
###
2.1 Neural Processes
The NP is a model for regression functions that map an input xi∈Rdx to an output yi∈Rdy. In particular, the NP defines a (infinite) family of conditional distributions, where one may condition on an arbitrary number of observed contexts (xC,yC)\coloneqq(xi,yi)i∈C to model an arbitrary number of targets (xT,yT)\coloneqq(xi,yi)i∈T in a way that is invariant to ordering of the contexts and ordering of the targets. The model is defined for arbitrary C and T but in practice we use C⊂T. The deterministic NP models these conditional distributions as:
| | | | |
| --- | --- | --- | --- |
| | p(yT|xT,xC,yC)\coloneqqp(yT|xT,rC) | | (1) |
with rC\coloneqqr(xC,yC)∈Rd where r is a deterministic function that aggregates (xC,yC) into a finite dimensional representation with permutation invariance in C. In practice, each context (x,y) pair is passed through an MLP
to form a representation of each pair, and these are aggregated by taking the mean to form rC. The likelihood p(yT|xT,rC) is modelled by a Gaussian factorised across the targets (xi,yi)i∈T with mean and variance given by passing xi and rC through an MLP. The unconditional distribution p(yT|xT) (when C=∅) is defined by letting r∅ be a fixed vector.
The latent variable version of the NP model includes a global latent z to account for uncertainty in the predictions of yT for a given observed (xC,yC). It is incorporated into the model via a latent path that complements the deterministic path described above. Here z is modelled by a factorised Gaussian parametrised by sC\coloneqqs(xC,yC), with s being a function of the same properties as r
| | | | |
| --- | --- | --- | --- |
| | p(yT|xT,xC,yC)\coloneqq∫p(yT|xT,rC,z)q(z|sC)dz | | (2) |
with q(z|s∅)\coloneqqp(z), the prior on z. The likelihood is referred to as the decoder, and q,r,s form the encoder. See Figure [2](#S3.F2 "Figure 2 ‣ 3 Attentive Neural Processes ‣ Attentive Neural Processes") for diagrams of these models.
The motivation for having a global latent is to model different realisations of the data generating stochastic process — each sample of z would correspond to one realisation of the stochastic process. One can define the model using either just the deterministic path, just the latent path, or both. In this work we investigate the case of using both paths, which gives the most expressive model and also gives a sensible setup for incorporating attention, as we will show later in Section [3](#S3 "3 Attentive Neural Processes ‣ Attentive Neural Processes").
The parameters of the encoder and decoder are learned by maximising the following ELBO
| | | | | |
| --- | --- | --- | --- | --- |
| | logp(yT|xT,xC,yC) | ≥Eq(z|sT)[logp(yT|xT,rC,z)]−DKL(q(z|sT)∥q(z|sC)) | | (3) |
for a random subset of contexts C and targets T via the reparametrisation trick (Kingma & Welling, [2014](#bib.bib16); Rezende et al., [2014](#bib.bib24)). In other words, the NP learns to reconstruct targets, regularised by a KL term that encourages the summary of the contexts to be not too far from the summary of the targets. This is sensible since we are assuming that the contexts and targets come from the same realisation of the data-generating stochastic process, and especially so if targets contain contexts. At each training iteration, the number of contexts and targets are also chosen randomly (as well as being randomly sampled from the training data), so that the NP can learn a wide family of conditional distributions.
NPs have many desirable properties, namely (i) Scalability: computation scales linearly at O(n+m) for n contexts and m targets at train and prediction time. (ii) Flexibility: defines a very wide family of distributions, where one can condition on an arbitrary number of contexts to predict an arbitrary number of targets. (iii) Permutation invariance: the predictions of the targets are order invariant in the contexts.
However these advantages come at the cost of not satisfying consistency in the contexts. For example, if y1:m is generated given some context set, then its distribution need not match the distribution you would obtain if y1:n is generated first, appended to the context set then yn+1:m is generated. However maximum-likelihood learning can be interpreted as minimising the KL between the (consistent) conditional distributions of the data-generating stochastic process and the corresponding conditional distributions of the NP. Hence we could view the NP as approximating the conditionals of the consistent data-generating stochastic process.
###
2.2 Attention
Given a set of key-value pairs (ki,vi)i∈I and a query q, an attention mechanism computes weights of each key with respect to the query, and aggregates the values with these weights to form the value corresponding to the query. In other words, the query attends to the key-value pairs.
The queried values are invariant to the ordering of the key-value pairs; this permutation invariance property of attention is key in its application to NPs. The idea of using a differentiable addressing mechanism that can be learned from the data has been applied successfully in various areas of Deep Learning, namely handwriting generation and recognition (Graves, [2012](#bib.bib12)) and neural machine translation (Bahdanau et al., [2015](#bib.bib3)).
More recently, there has been work employing self-attention (where keys and queries are identical) to give expressive sequence-to-sequence mappings in natural language processing (Vaswani et al., [2017](#bib.bib30)) and image modelling (Parmar et al., [2018](#bib.bib22)).
We give some examples of attention mechanisms which are used in the paper. Suppose we have n key-value pairs arranged as matrices K∈Rn×dk, V∈Rn×dv, and m queries Q∈Rm×dk. Simple forms of attention based on locality (weighting keys according to distance from query) are given by various stationary kernels. For example, the (normalised) Laplace kernel gives the queried values as
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | Laplace(Q,K,V) | \coloneqqWV∈Rm×dv, | Wi⋅ | \coloneqqsoftmax((−||Qi⋅−Kj⋅||1)nj=1)∈Rn | |
Similarly (scaled) dot-product attention uses the dot-product between the query and keys as a measure of similarity, and weights the keys according to the values
| | | |
| --- | --- | --- |
| | DotProduct(Q,K,V)\coloneqqsoftmax(QK⊤/√dk)V∈Rm×dv | |
The use of dot-product attention allows the query values to be computed with two matrix multiplications and a softmax, allowing for use of highly optimised matrix multiplication code.
multihead attention (Vaswani et al., [2017](#bib.bib30)) is a parametrised extension where for each head, the keys, values and queries are linearly transformed, then dot-product attention is applied to give head-specific values. These values are concatenated and linearly transformed to produce the final values:
| | | | |
| --- | --- | --- | --- |
| | MultiHead(Q,K,V) | \coloneqqconcat(head1,…,headH)W∈Rm×dv | |
| | where headh | \coloneqqDotProduct(QWQh,KWKh,VWVh)∈Rm×dv | |
This multihead architecture allows the query to attend to different keys for each head and tends to give smoother query-values than dot-product attention (c.f. Section [4](#S4 "4 Experimental Results ‣ Attentive Neural Processes")).
3 Attentive Neural Processes
-----------------------------
Figure 2: Model architecture for the NP (left) and Attentive NP (right)
Figure [2](#S3.F2 "Figure 2 ‣ 3 Attentive Neural Processes ‣ Attentive Neural Processes") describes how attention is incorporated into NP to give the Attentive NP (ANP).
In summary, self-attention is applied to the context points to compute representations of each (x,y) pair, and the target input attends to these context representations (cross-attention) to predict the target output.
In detail, the representation of each context pair (xi,yi)i∈C before the mean-aggregation step is computed by a self-attention mechanism, in both the deterministic and latent path. The intuition for the self-attention is to model interactions between the context points. For example, if many context points overlap, then the query need not attend to all of these points, but only give high weight to one or a few. The self-attention will help obtain richer representations of the context points that encode these types of relations between the context points. We model higher order interactions by simply stacking the self-attention, as is done in Vaswani et al. ([2017](#bib.bib30)).
In the deterministic path, the mean-aggregation of the context representations that produces rC is replaced by a cross-attention mechanism, where each target query x∗ attends to the context xC\coloneqq(xi)i∈C to produce a query-specific representation r∗\coloneqqr∗(xC,yC,x∗). This is precisely where the model allows each query to attend more closely to the context points that it deems relevant for the prediction. The reason we do not have an analogous mechanism in the latent path is that we would like to preserve the global latent, that induces dependencies between the target predictions. The interpretation of the latent path is that z gives rise to correlations in the marginal distribution of the target predictions yT, modelling the global structure of the stochastic process realisation, whereas the deterministic path models the fine-grained local structure.
The decoder remains the same, except we replace the shared context representation rC with the query-specific representation r∗. Note that permutation invariance in the contexts is preserved with the attention mechanism. If we use uniform attention (all contexts given the same weight) throughout, we recover the NP. ANP is trained using the same loss ([3](#S2.E3 "(3) ‣ 2.1 Neural Processes ‣ 2 Background ‣ Attentive Neural Processes")) as the original NP, also using Gaussian likelihood p(yi|xi,r∗(xC,yC,xi),z) and diagonal Gaussian q(z|sC).
The added expressivity and resulting accuracy of the NP with attention comes at a cost. The computational complexity is raised from O(n+m) to O(n(n+m)), since we apply self-attention across the contexts and for every target point we compute weights for all contexts.
However most of the computation for the (self-)attention is done via matrix multiplication (c.f. Section [2.2](#S2.SS2 "2.2 Attention ‣ 2 Background ‣ Attentive Neural Processes")), and so can be done in parallel across the contexts and across the targets. In practice, the training time for ANPs remains comparable to NPs, and in fact we show that ANPs learn significantly faster than NPs
not only in terms of training iterations but also in wall-clock time, despite being slower at prediction time (c.f. Section [4](#S4 "4 Experimental Results ‣ Attentive Neural Processes")).
4 Experimental Results
-----------------------
Note that the (A)NP learns a stochastic process, so should be trained on multiple functions that are realisations of the stochastic process. At each training iteration, we draw a batch of realisations from the data generating stochastic process, and select random points on these realisations to be the targets and a subset to be the contexts to optimise the loss in Equation ([3](#S2.E3 "(3) ‣ 2.1 Neural Processes ‣ 2 Background ‣ Attentive Neural Processes")). We use the same decoder architecture for all experiments, and 8 heads for multihead. See Appendix [A](#A1 "Appendix A Architectural details for (A)NP ‣ Attentive Neural Processes") for architectural details.
1D Function regression on synthetic GP data We first explore the (A)NPs trained on data that is generated from a Gaussian Process with a squared-exponential kernel and small likelihood noise. We emphasise that (A)NPs need not be trained on GP data or data generated from a known stochastic process, and this is just an illustrative example. We explore two settings: one where the hyperparameters of the kernel are fixed throughout training, and another where they vary randomly at each training iteration. The number of contexts (n) and number of targets (m) are chosen randomly at each iteration (n∼U[3,100], m∼n+U[0,100−n]). Each x-value is drawn uniformly at random in [−2,2]. For this simple 1D data, we do not use self-attention and just explore the use of cross-attention in the deterministic path (c.f. Figure [2](#S3.F2 "Figure 2 ‣ 3 Attentive Neural Processes ‣ Attentive Neural Processes")). Thus we use the same encoder/decoder architecture for NP and ANP, except for the cross-attention. See Appendix [B](#A2 "Appendix B Experimental details of 1D function regression experiment ‣ Attentive Neural Processes") for experimental details.
Figure 3: Qualitative and quantitative results of different attention mechanisms for 1D GP function regression with random kernel hyperparameters. Left: moving average of context reconstruction error (top) and target negative log likelihood (NLL) given contexts (bottom) plotted against training iterations (left) and wall clock time (right). d denotes the bottleneck size i.e. hidden layer size of all MLPs and the dimensionality of r and z. Right: predictive mean and variance of different attention mechanisms given the same context. Best viewed in colour.
Figure [3](#S4.F3 "Figure 3 ‣ 4 Experimental Results ‣ Attentive Neural Processes") (left) shows context reconstruction error 1|C|∑i∈CEq(z|sC)[logp(yi|xi,r∗(xC,yC,xi),z)] and NLL of targets given contexts 1|T|∑i∈TEq(z|sC)[logp(yi|xi,r∗(xC,yC,xi),z)] for the different attention mechanisms, trained on a GP with random kernel hyperparameters. ANP shows a much more rapid decrease in reconstruction error and lower values at convergence compared to the NP, especially for dot product and multihead attention. This holds not only against training iteration but also against wall clock time, so learning is fast despite the added computational cost of attention. The right column plots show that the computation times of Laplace and dot-product ANP are similar to the NP for the same value of d, and multihead ANP takes around twice the time. We also show how the size of the bottleneck (d) in the deterministic and latent paths of the NP affects the underfitting behaviour of NPs. The figure shows that raising d does help achieve better reconstructions, but there appears to be a limit in how much reconstructions can improve. Beyond a certain value of d, the learning for the NP becomes too slow, and the value of reconstruction error at convergence is still higher than that achieved by multihead ANP with 10% of the wall-clock time. Hence using ANPs has significant benefits over simply raising the bottleneck size in NPs.
In Figure [3](#S4.F3 "Figure 3 ‣ 4 Experimental Results ‣ Attentive Neural Processes") (right) we visualise the learned conditional distribution for a qualitative comparison of the attention mechanisms. The context is drawn from the GP with the hyperparameter values that give the most fluctuation. Note that the predictive mean of the NP underfits the context, and tries to explain the data by learning a large likelihood noise. Laplace shows similar behaviour, whereas dot-product attention gives predictive means that accurately predict almost all context points. Note that Laplace attention is parameter-free (keys and queries are the x-coordinates) whereas for dot-product attention we have set the keys and queries to be parameterised representations of the x-values (output of learned MLP that takes x-coordinates as inputs). So the dot-product similarities are computed in a learned representation space, whereas for Laplace attention the similarities are computed based on L1 distance in the x-coordinate domain, hence it is expected that dot-product attention outperforms Laplace attention. However dot-product attention displays non-smooth predictions, shown more clearly in the predictive standard deviations (c.f. Appendix [C](#A3 "Appendix C Additional figures for 1D regression on GP data ‣ Attentive Neural Processes") for an explanation). The multiple heads in multihead attention appear to help smooth out the interpolations, giving good reconstruction of the contexts as well as prediction of the targets, while preserving increased predictive uncertainty away from the contexts as in a GP. The results for (A)NP trained on fixed GP kernel hyperparameters are similar (c.f. Appendix [C](#A3 "Appendix C Additional figures for 1D regression on GP data ‣ Attentive Neural Processes")), except that the NP underfits to a lesser degree because of the reduced variety of sample curves (functions) in the data. This difference in performance for the two kernel hyperparameter settings provides evidence of how the ANP is more expressive than the NP and can learn a wider range of functions.
Using the trained (A)NPs we tackle a toy Bayesian Optimisation (BO) problem, where the task is to find the minimum of test functions drawn from a GP prior. This is a proof-of-concept experiment showing the utility of being able to sample entire functions from the (A)NP and having accurate context reconstructions. See Appendix [C](#A3 "Appendix C Additional figures for 1D regression on GP data ‣ Attentive Neural Processes") for the details and an analysis of results.
2D Function regression on image data Image data can also be interpreted as being generated from a stochastic process (since there are dependencies between pixel values), and predicting the pixel values can be cast as a regression problem mapping a 2D pixel location xi to its pixel intensity yi (∈R1 for greyscale, ∈R3 for RGB). Each image corresponds to one realisation of the process sampled on a fixed 2 dimensional grid. We train the ANP on MNIST (LeCun et al., [1998](#bib.bib18)) and 32×32 CelebA (Liu et al., [2015](#bib.bib19)) using the standard train/test split with up to 200 context/target points at training. For this application we explore the use of self-attentional layers in the encoder, stacking them as is done in Parmar et al. ([2018](#bib.bib22)). See Appendix [D](#A4 "Appendix D Experimental details of 2D Image regression experiment ‣ Attentive Neural Processes") for experimental details.
| | |
| --- | --- |
|
(a) Reconstructions of full CelebA image from a varying number of random context points for NP (left) and Stacked Multihead ANP (right).
|
(b) Context NLL (top) and unseen target NLL given contexts (bottom).
|
Figure 4: Qualitative and quantitative results on test set for 2D CelebA function regression.
| | |
| --- | --- |
|
(a) MNIST
|
(b) CelebA
|
Figure 5: Reconstruction of full image from top half. The CelebA results use the same models (with the same parameter values) as Figure [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4 Experimental Results ‣ Attentive Neural Processes").
On both datasets we show results of three different models: NP, ANP with multihead cross-attention in the deterministic path (Multihead ANP), and ANP with both multihead attention in the deterministic path and two layers of stacked self-attention in both the deterministic and latent paths (Stacked Multihead ANP). Figure [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4 Experimental Results ‣ Attentive Neural Processes") shows predictions of the full image (i.e. full target) with a varying number of random context pixels, from 10 to 1024 (full image) for a randomly selected image (see Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes") for other images). For each we generate predictions that correspond to the mean of p(yT|xT,rC,z) for three different samples of z∼q(z|sC). The NP (left) gives reasonable predictions with a fair amount of diversity for fewer contexts, but the reconstructions of the whole image are not accurate, compared to Stacked Multihead ANP (right) where the reconstructions are indistinguishable from the original. The use of attention also helps achieve crisper inpaintings when the target pixels are filled in, enhancing the ANP’s ability to model less smooth 2D functions compared to the NP. The diversity in faces and digits obtained with different values of z is apparent the different samples, providing evidence for the claim that z can model global structure of the image, with one sample corresponding to one realisation of the data generating stochastic process.
Similar conclusions hold for MNIST (see Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes")) and for the full image prediction using the top half as context in Figure [5](#S4.F5 "Figure 5 ‣ 4 Experimental Results ‣ Attentive Neural Processes"). In the latter task, note that the model has never been trained on more than 200 context points, yet it manages to generalise to when the context is of size 512 (half the image).
Figure 6: Mapping between different resolutions by the same model (with the same parameter values) as Stacked Multihead ANP in Figures [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4 Experimental Results ‣ Attentive Neural Processes"), [4(b)](#S4.F4.sf2 "(b) ‣ Figure 5 ‣ 4 Experimental Results ‣ Attentive Neural Processes"). The two rightmost columns show the results of baseline methods, namely linear and cubic interpolation to 256×256.
Figure 7: Pixels attended to by each head of multihead attention in Multihead ANP given a target pixel. Each head is given a different colour and the target pixel is marked with a cross.
Figure [3(b)](#S4.F3.sf2 "(b) ‣ Figure 4 ‣ 4 Experimental Results ‣ Attentive Neural Processes") verifies quantitatively that both Multihead and Stacked Multihead ANP give a much improved context reconstruction error compared to the NP. Similarly the NLL for the target points (that are not included in the context) is improved with multihead cross-attention, showing small gains with stacked self-attention. However qualitatively, there are noticeable gains in crispness and global coherence when using stacked self-attention (see Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes")).
In Figure [7](#S4.F7 "Figure 7 ‣ 4 Experimental Results ‣ Attentive Neural Processes") we visualise each head of Multihead ANP for CelebA. We let the target pixel (cross) attend to all pixels, and see where each head of the attention focuses on. We colour-code the pixels with the top 20 weights per head, with intensity proportional to the attention weight. We can see that each head has different roles: the cyan head only looks at the target pixel and nothing else; the red head looks at a few pixels nearby; the green head looks at a larger region nearby; the yellow looks at the pixels on the column of the target; the orange looks at some band of the image; the purple head (interestingly) looks at the other side of the image, trying to exploit the symmetry of faces. We observe consistent behaviour in these heads for other target pixels (see Figure [16](#A5.F16 "Figure 16 ‣ Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes") of Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes")).
One other illustrative application of (A)NPs trained on images is that one can map images from one resolution to another, even if the model has only been trained on one resolution. Because the two dimensional x (pixel locations) are modelled as real values that live in a continuous space, the model can predict the y (pixel intensities) of any point in this space, and not just the grid of points that it was trained on. Hence using one grid as the context and a finer grid as the target, the model can map a given resolution to a higher resolution. This could, however, be problematic for NPs whose reconstructions can be inaccurate, so the prediction of the target resolution can look very different to the original image (see Figure [19](#A5.F19 "Figure 19 ‣ Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes") of Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes")). The reconstructions of ANPs may be accurate enough to give reliable mappings between different resolutions. We show results for such mappings given by the same Stacked Multihead ANP (the same model used to produce Figures [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4 Experimental Results ‣ Attentive Neural Processes"), [4(b)](#S4.F4.sf2 "(b) ‣ Figure 5 ‣ 4 Experimental Results ‣ Attentive Neural Processes")) in Figure [6](#S4.F6 "Figure 6 ‣ 4 Experimental Results ‣ Attentive Neural Processes"). On the left, we see that the ANP (trained on 32×32 images) is capable of mapping low resolutions (4×4 or 8×8) to fairly realistic 32×32 target outputs with some diversity for different values of z (more diversity for the 4×4 contexts as expected). Perhaps this performance is to be expected since the model has been trained on data that has 32×32 resolution. The same model allows us to map to even higher resolutions, namely from the original 32×32 images to 256×256, displayed on the right of the figure. We see that even though the model has never seen any images beyond the original resolution, the model learns a fairly realistic high resolution image with sharper edges compared to the baseline interpolation methods. Moreover, there is some evidence that it learns an internal representation of the appearance of faces, when for example it learns to fill in the eye even when the original image is too coarse to separate the iris (coloured part) from the sclera (white part) (e.g. top row image), a feature that is not possible with simple interpolation. See Figure [19](#A5.F19 "Figure 19 ‣ Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes") in Appendix [E](#A5 "Appendix E Additional figures for 2D Image regression on MNIST and CelebA ‣ Attentive Neural Processes") for larger versions of the images.
For each of MNIST and CelebA, all qualitative plots in this section were given from the same model (with the same parameter values) for each attention mechanism, learned by optimising the loss in Equation ([3](#S2.E3 "(3) ‣ 2.1 Neural Processes ‣ 2 Background ‣ Attentive Neural Processes")) over random context pixels and random target pixels at each iteration. It is important to note that we do not claim the ANP to be a replacement of state of the art algorithms of image inpainting or super-resolution, and rather we show these image applications to highlight the flexibility of the ANP in modelling a wide family of conditional distributions.
5 Related Work
---------------
The work related to NPs in the domain of Gaussian Processes, Meta-Learning, conditional latent variable models and Bayesian Learning have been discussed extensively in the original works of Garnelo et al. ([2018a](#bib.bib10), [b](#bib.bib11)), hence we focus on works that are particularly relevant for ANPs.
Gaussian Processes (GPs) Returning to our motivation for using attention in NPs, there is a clear parallel between GP kernels and attention, in that they both give a measure of similarity between two points in the same domain. The use of attention in an embedding space that we explore is related to Deep Kernel Learning (Wilson et al., [2016](#bib.bib32)) where a GP is applied to learned representations of data. Here, however, learning is still done in a GP framework by maximising the marginal likelihood. We reiterate that the training regimes of GPs and NPs are different, so a direct comparison between the methods is difficult. One possibility for comparison is to learn the GP via the training regime of NPs, namely updating the kernel hyperparameters at each iteration via one gradient step of the marginal likelihood on the mini-batch of data. However, this would still have a O(n3) computational cost in the naive setting and may require kernel approximations. In general, the predictive uncertainties of GPs depend heavily on the choice of the kernel, whereas NPs learn predictive uncertainties directly from the data. Despite these drawbacks, GPs have the benefit of being consistent stochastic processes, and the covariance between the predictions at different x-values and the marginal variance of each prediction can be expressed exactly in closed form, a feature that the current formulation of (A)NPs do not have. Variational Implicit Processes (VIP) (Ma et al., [2018](#bib.bib20)) are also related to NPs, where VIP defines a stochastic process using the same decoder setup with a finite dimensional z. Here, however, the process and its posterior given observed data are both approximated by a GP and learned via a generalisation of the Wake-Sleep algorithm (Hinton et al., [1995](#bib.bib14)).
Meta-Learning (A)NPs can be seen as models that do few-shot learning, although this is not the focus of our work. Given input-output pairs drawn from a new function at test time, one can reason about this function by looking at the predictive distribution conditioning on these input-output pairs. There is a plethora of works in few-shot classification, of which Vinyals et al. ([2016](#bib.bib31)); Snell et al. ([2017](#bib.bib28)); Santoro et al. ([2016](#bib.bib27)) use attention to locate the relevant observed image/prototype given a query image. Attention has also been used for tasks in Meta-RL such as continuous control and visual navigation (Mishra et al., [2018](#bib.bib21)). Few-shot density estimation using attention has also been explored extensively in numerous works (Rezende et al., [2016](#bib.bib25); Reed et al., [2017](#bib.bib23); Bornschein et al., [2017](#bib.bib6); Bartunov & Vetrov, [2018](#bib.bib4)). Especially relevant are the Neural Statistician (Edwards & Storkey, [2017](#bib.bib8)) and the Variational Homoencoder (Hewitt et al., [2018](#bib.bib13)) who have a similar permutation invariant encoder (that outputs summaries of a data set), but use local latents on top of a global latent. For ANPs, we look at the less-explored regression setting. The authors of Vfunc (Bachman et al., [2018](#bib.bib2)) also explore regression on a toy 1D domain, using a similar setup to NPs but optimising an approximation to the entropy of the latent function, without any attention mechanisms. Multi-task learning has also been tackled in the GP literature by various works (Teh et al., [2005](#bib.bib29); Bonilla et al., [2008](#bib.bib5); Alvarez et al., [2012](#bib.bib1); Dai et al., [2017](#bib.bib7)).
Generative Query Networks (Eslami et al., [2018](#bib.bib9); Kumar et al., [2018](#bib.bib17)) are models for spatial prediction that render a frame of a scene given a viewpoint.
Their model corresponds to a special case of NPs where the x are viewpoints and the y are frames of a scene. Rosenbaum et al. ([2018](#bib.bib26)) apply the GQN to the task of 3D localisation with an attention mechanism, but attention is applied to patches of context frames (y) instead of a parametric representation of viewpoints (x). Note that in our work the targets attend to the contexts via the x.
6 Conclusion and Discussion
----------------------------
We have proposed ANPs, which augment NPs with attention to resolve the fundamental problem of underfitting. We have shown that this greatly improves the accuracy of predictions in terms of context and target NLL, results in faster training, and expands the range of functions that can be modelled. There is a wide scope of future work for ANPs. Regarding model architecture, one way of incorporating cross-attention into the latent path and modelling the dependencies across the resulting local latents is to also have a global latent, much like the setup of the Neural Statistician but translated to the regression setting.
An interesting further application would be to train ANPs on text data, enabling them to fill in the blanks in a stochastic manner. For the image application, the Image Transformer (ImT) (Parmar et al., [2018](#bib.bib22))
has some interesting connections with ANPs: its local self-attention used to predict consecutive pixel blocks from previous blocks has parallels with how our model attends to context pixels to predict target pixels. Replacing the MLP in the decoder of the ANP with self-attention across the target pixels, we have a model that closely resembles an ImT defined on arbitrary orderings of pixels. This is in contrast to the original ImT, which presumes a fixed ordering and is trained autoregressively. We plan to equip ANPs with self-attention in the decoder, and see how far their expressiveness can be extended. In this setup, however, the targets will affect each other’s predictions, so the ordering and grouping of the targets will become important.
#### Acknowledgments
We would like to thank Ali Razavi for his advice on implementing multihead attention, and Michael Figurnov for helpful discussion. |
39a8508d-0103-4355-b277-0a544fe8c04d | trentmkelly/LessWrong-43k | LessWrong | Applying Double Standards to ‘‘Divisive’’ Ideas
This is a commentary by Linda Gottfredson on a paper by Hunt and Carlson about a paper by Richard Nisbett regarding studies done by Arthur Jensen. It's ultimately about race and intelligence, but it seemed meta enough to link to here.
Warning: PDF
Applying Double Standards to ‘‘Divisive’’ Ideas |
b403530c-215b-479d-a0a4-3856e5bdec9d | trentmkelly/LessWrong-43k | LessWrong | 10/05/2017: Development update (performance & styles)
Hey Everyone, here are the latest changes I just pushed to the page:
* Significant speed improvements after the initial package load. Posts should now load reliably in less than 2-3 seconds (and usually faster than that).
* Some speed improvements for the first visit of the page. This is still taking far too long, but it should be a good bit better now (more on that at the end of this post)
* Typography improvements for posts and comments (things now have reasonable line-height, and upvote and downvote buttons are more recognizable)
* Comments are now collapsable
* Various small bugfixes all over the place
* A lot of internal code refactoring and cleanup (as part of making things fast)
Overall, quite happy with the last week. Let me know what you think about the changes. I particularly want to thank Max Harms who spent some of his free time to work on LessWrong for part of the last two weeks, and has been extremely helpful.
This is not the final performance update. A large part of our performance rework was sadly somewhat stopped in its tracks by a bug in Meteor I ended up discovering, so some of the biggest gains will have to wait until that has been fixed. However, we've already made quite a bit of improvements, and are reaching load times that are only half a second slower than Reddit on some of our development servers, so I think we can definitely achieve reasonably fast loading speeds.
Let me know what you think about all the latest changes, and please always feel free to add new Bugs to our Github repo (or even better, come and help fix some of the bugs yourself!). We will continue to focus on performance improvements for now, but from now on things will be more incremental, and so you should see more frequent updates and faster bugfixes. |
b304a762-5a7b-42dd-bc10-dddc950bd325 | trentmkelly/LessWrong-43k | LessWrong | Gifts
Fiction, not necessarily appropriate for LW, but I thought some here might find it funny and it's the off-season when less serious reading is happening anyway. If it seems inappropriate I would not mind a mod-delete on this one.
He could see the alien’s ship landing. Its design marked it with high probability as a war vessel, freshly painted in bright bellicose colors, matching the ceremonial regalia of its inhabitants.
The one he’d have an audience with was a sort of grand judge, it was bipedal. Flanking it were four-legged bodyguards, wearing giant ornamental helmets, heads bowed low, an almost universal primitive sign of submission.
This was a courtesy meeting. The kind the corporation used when engaging a new market. Get an important person to come over, give them pleasure-inducing drugs, complement their disgusting norms, use one of the most powerful computing machines in the universe as a party-trick to answer their two-cent questions.
The alien leader passed through the door spoke with a deep booming voice.
- You are this corporation’s intelligence, right?
- Hello to you as well. Yes, I presume you could call me that, you are speaking with a memory-copy of the language module, with full ability to access and interoperate with all other components, sans a few restricted bits of data I prefer to explicitly isolate.
- Fine, tell me, how do you know if an intelligence, one such as yourself, has been good?
It was obviously in a hurry, with a hint of... anger(?). What kind of barbarians were these? Did he even begin to comprehend the nonsense bordering complexity of that question
- Well, that’s rather hard to tell quite frankly. Whenever a decision I made comes to affect us I will interpret the results, together with external validators, and based on the outcome, combined with the available information at the time, I will try adjusting the implicated components towards choosing better if such a thing seems possible. Of course,
A grunting interruption, wh |
acc311a7-1249-47b2-847e-6f4f0b301a7c | trentmkelly/LessWrong-43k | LessWrong | Wikipedia Edit War Update
A few months ago I stumbled into an edit war on Wikipedia. I noticed that Wikipedia's page on Jacy Reese was being, essentially, guarded from having any mention that he previously went by his full name. There was a pattern where someone would notice this information was missing, add it, and then it would be reverted soon after.
The main user guarding the page was Bodole, and someone pointed me yesterday to where they've been banned from editing Jacy's page for three months. The discussion there was another interesting window into how Wikipedia handles disputes, so after reading it I thought it would be interesting to review:
* User Drmies edited the page to remove a list of articles Jacy had published ("rm linkfarm. we list books, not articles", link). Drmies is an experienced editor, making a routine cleanup.
* User Bodole reverts the change ("Many BLP list articles. Please discuss on talk page if you think this should be an exception.", link).
* Drmies reverts the revert ("It's the other way around. what you are doing is promoting this person by linking a set of articles. if you have secondary sources that prove these articles are worth noticing, that's a different matter", link)
* Bodole reverts the reversion of the revert ("You are edit warring. Please stop. Discuss on the talk page if you insist. See the WP:BRD cycle", link) and puts a warning (link) on Drimes' talk page.
* Drimes responds there with "Aw boohoo" (link)
* Drmies reverts the reversion of the reversion of the revert ("see talk page", link) and marks the page as being subject to Wikipedia:Conflict of Interest (link). It looks to me like Drmies thinks Bodole may either be Jacy or someone closely connected with him. Drmies removes biographical information from the page ("this 'Sentience Institute' has an article--why this biography is bloated with content about some poll, verified only with links to websites, is not clear", link)
* Discussion moves to the talk page
* Drmies is clea |
f8199eb0-5082-4a64-adb0-a6fe6dfa0017 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Linkpost: A Contra AI FOOM Reading List
This is a linkpost to a list of skeptical takes on AI FOOM. I haven't read them all and probably disagree with some of them, but it's valuable to put these arguments in one place. |
ab38765f-6950-47af-9393-9dab1574f7ab | trentmkelly/LessWrong-43k | LessWrong | Don't depend on others to ask for explanations
I'm often reluctant to ask for explanations on LW, and to typical-mind a bit I think this may be true of others as well. This suggests that when you're writing something for public consumption, it's better to err on the side of too much rather than too little explanation. If there's too much explanation, people can just skip over it (and you can make it easier by putting explanations that may be "too much" in parentheses or footnotes), but if there's too little explanation people may never ask for it. So in the future if you ever think something like, "I'll just write down what I think, and if people don't understand why, they can ask" I hope this post will cause you to have a second thought about that.
To make it clearer that this problem can't be solved by just asking or training people to be less reluctant to ask for explanations, I think there are often "good" reasons for such reluctance. Here's a list that I came up with during a previous discussion with Raymond Arnold (Raemon):
1. I already spent quite some time trying to puzzle out the explanation, and asking is like admitting defeat.
2. If there is a simple explanation that I reasonably could have figured out without asking, I look bad by asking.
3. It's forcing me to publicly signal interest, and maybe I don't want to do that.
4. Related to 3, it's forcing me to raise the status of the person I'm asking, by showing that I'm interested in what they're saying. (Relatedly, I worry this might cause people to withhold explanations more often than they should.)
5. If my request is ignored or denied, I would feel bad, perhaps in part because it seems to lower my status.
6. I feel annoyed that the commenter didn't value my time enough to preemptively include an explanation, and therefore don't want to interact further with them.
7. My comment requesting an explanation is going to read by lots of people for whom it has no value, and I don't want to impose that cost on them, or make them subconsciously annoy |
73f536c1-7d16-4977-b894-e57aeee34b4e | trentmkelly/LessWrong-43k | LessWrong | Meetup : UMD, Social Effectiveness and Calibration Exercises
Discussion article for the meetup : UMD, Social Effectiveness and Calibration Exercises
WHEN: 09 September 2011 06:00:00PM (-0400)
WHERE: University of Maryland, Anne Arundel Building
The DC Meetup group is starting to run meetups at the University of Maryland!
We'll be meeting in the basement of the Anne Arundel building at 6:00, then will probably move to a dorm room, outside, or to a food place.
At the first one we plan to do some quick calibration exercises, then move on to practicing social skills, like eye contact. If we're feeling brave, rejection therapy might ensue.
Discussion article for the meetup : UMD, Social Effectiveness and Calibration Exercises |
23c28630-a3d0-4fc6-8e72-6c9199af4a83 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Aspiring AI safety researchers should ~argmax over AGI timelines
*Epistemic status: This model is mostly based on a few hours of dedicated thought, and the post was written in 30 min. Nevertheless, I think this model is probably worth considering.*
Many people seem to be entering the AI safety ecosystem, acquiring a belief in short timelines and high P(doom), and immediately dropping everything to work on AI safety agendas that might pay off in short-timeline worlds. However, many of these people might not have a sufficient “toolbox” or research experience to have much marginal impact in short timelines worlds.
Rather than tell people what they should do on the object level, I sometimes tell them:
1. Write out your credences for AGI being realized in 2027, 2032, and 2042;
2. Write out your plans if you had 100% credence in each of 2027, 2032, and 2042;
3. Write out your marginal impact in lowering P(doom) via each of those three plans;
4. Work towards the plan that is the [argmax](https://en.wikipedia.org/wiki/Arg_max) of your marginal impact, weighted by your credence in the respective AGI timelines.
Some further considerations
===========================
* If you are risk averse over your marginal impact, you should maybe avoid a true argmax approach and instead choose a plan that pays out some marginal impact in the three timeline scenarios. For example, some shovel-ready, short-timeline AI safety research agendas may help prepare you for long-timeline AI safety research more than others. Consider blending elements of your plans in the three timeline scenarios (the "~" in "~argmax"). Perhaps you have side constraints on your minimal impact in the world where AGI is realized in 2027?
* Your immediate plans might be similar in some scenarios. If so, congratulations, you have an easier decision! However, I suspect most aspiring AI safety researchers without research experience should have different plans for different AGI timeline scenarios. For example, getting a Ph.D. in a top lab probably makes most people much better at some aspects of research and working in emerging tech probably makes most people much better at software engineering and operations.
* You should be wary of altering your timeline credences in an attempt to rationalize your preferred plan or highest-probability timeline scenario. However, don’t be afraid to update your credences over AGI timelines or your expected marginal impact in those worlds! Revisit your plan often and expect them to change (though hopefully not in predictable ways, as this would make you a bad Bayesian).
* Consider how the entire field of AI talent might change if everyone followed the argmax approach I laid out here. Are there any ways they might do something you think is predictably wrong? Does this change your plan?
* If you want to develop more finely-grained estimates over timelines (e.g., 2023, 2024, etc.) and your marginal impact in those worlds, feel free to. I prefer to keep the number of options manageable.
* Your marginal impact might also change with respect to the process by which AGI is created in different timeline worlds. For example, if AGI arrives in 2023, I imagine that the optimal mechanistic interpretability researcher might not have as high an impact as they would if AGI arrived some years later, when interpretability has potentially had time to scale. |
2dfbfdd6-05ca-4b05-ac8d-8329885cc33b | trentmkelly/LessWrong-43k | LessWrong | The Scientific Approach To Anything and Everything
What causes some people to be effective at what they do while others are ineffective?
This is a question that many have tried to address, while a few have hinted at the answer. The two best examples came from the physicists Richard Feynman and Eli Goldratt.
In the 1974 Caltech commencement address, Richard Feynman spent the entire speech telling us the hint. He described it as a problem without spelling out the solution. He explained how the scientific approach evolved and he explained what the old approach was; I’ll call it ‘the witch doctor approach’. He said that despite the fact that we’re in the scientific era -- where we’ve advanced dramatically in science and technology -- most people do not think scientifically. He complained that whole fields of study are not even trying to do the basic things necessary to have scientific integrity. He explained that these anti-scientific ways are even encroaching into the field of physics -- the field that is easiest to be objective -- the field where the best intellectual traditions were developed. He lamented that we don’t teach the scientific approach in schools and instead that professors are, in effect, hoping that people learn it by osmosis, meaning teaching by example. And this is the hint to the solution. The solution is to directly teach the principles and methods of the scientific approach instead of relying only on teaching by example.
In the Goldratt Satellite Program (GSP), the physicist and business management guru Eli Goldratt complained that his attempts to teach the world how to think had failed. His Theory of Constraints (TOC) had dramatically improved thousands of major corporations worldwide, even government institutions, but he recognized that the vast majority of them had returned to their old ways. He complained that the body of knowledge of TOC is so big and complex that it’s too difficult for people to learn it well. He said that it’s a problem of organization; that the knowledge of TOC is not or |
a5ea6aaa-4a52-44c3-adcb-a0dc1d36b873 | trentmkelly/LessWrong-43k | LessWrong | Dissolving the "Is the Efficient Market Hypothesis Dead?" Question
(This talk was given at a public online event on Sunday July 12th. Alexei is responsible for the talk, David Lambert edited the transcript.
If you're a curated author and interested in giving a 5-min talk, which will then be transcribed and edited, sign up here.)
Alexei Andreev: Recently, it's been popular on LessWrong to claim that the EMH is dead. I want to talk about why hearing that annoys me, and then propose something that I think is better than EMH.
Quickly about me: I'm a co-founder of Temple Capital, one of the top five biggest cryptocurrency quant hedge funds. I spent the last three years developing quant strategies. Hence I have a bit of a practical experience. But I don't study theory that much, so take this with a grain of salt.
There's two things that annoy me when people say that EMH is dead.
First: what version of EMH are you talking about exactly? There's a lot of formulations: weak form, semi-strong form, strong form, and, of course, the final ultimate mega form:
I have some issues with those formulations.
The main one is that they claim that “information is priced in”. I think it's obvious that it's not priced in, because who is pricing it in? It's entities, agents, hedge funds and traders that are doing the work of pricing it in. And every day, they're improving their systems, which means that yesterday they weren't pricing in something that they're pricing in today.
If you believe that they're actually doing reasonable pricing, then they're also doing reasonable improvements. The system is continuously getting more efficient. It's not binary.
If you imagine a super intelligent AGI, and you give it just the most easily accessible, publicly available prices, I think it will still outperform most human traders.
EMH can't really be dead. It can only be more or less efficient than you thought.
The second thing that annoys me when people say EMH is dead is: Why exactly do you think it is dead?
You have to read between the lines, |
0dbd66c9-28c4-4751-ab71-d6b1f7ba61a2 | trentmkelly/LessWrong-43k | LessWrong | The mind-killer
Can we talk about changing the world? Or saving the world?
I think few here would give an estimate higher than 95% for the probability that humanity will survive the next 100 years; plenty might put a figure less than 50% on it. So if you place any non-negligible value on future generations whose existence is threatened, reducing existential risk has to be the best possible contribution to humanity you are in a position to make. Given that existential risk is also one of the major themes of Overcoming Bias and of Eliezer's work, it's striking that we don't talk about it more here.
One reason of course was the bar until yesterday on talking about artificial general intelligence; another factor are the many who state in terms that they are not concerned about their contribution to humanity. But I think a third is that many of the things we might do to address existential risk, or other issues of concern to all humanity, get us into politics, and we've all had too much of a certain kind of argument about politics online that gets into a stale rehashing of talking points and point scoring.
If we here can't do better than that, then this whole rationality discussion we've been having comes to no more than how we can best get out of bed in the morning, solve a puzzle set by a powerful superintelligence in the afternoon, and get laid in the evening. How can we use what we discuss here to be able to talk about politics without spiralling down the plughole?
I think it will help in several ways that we are a largely community of materialists and expected utility consequentialists. For a start, we are freed from the concept of "deserving" that dogs political arguments on inequality, on human rights, on criminal sentencing and so many other issues; while I can imagine a consequentialism that valued the "deserving" more than the "undeserving", I don't get the impression that's a popular position among materialists because of the Phineas Gage problem. We need not ask whether |
82a220c8-f9af-483a-aa4f-8ca89737a499 | trentmkelly/LessWrong-43k | LessWrong | Is there any writing about prompt engineering for humans?
In much the same way that we need to prompt engineer language models to get them to handle our questions correctly, we know that humans respond differently to "This intervention will save 200 lives" and "This intervention will result in 400 deaths" when talking about a population of 600 people.
Is there any pre-existing writing that touches on this? |
188432be-0a20-4495-b2ff-8e084d5e2f97 | trentmkelly/LessWrong-43k | LessWrong | Deception Chess: Game #2
Game 2 was between Max Thibodeaux as player A, Chess.com computer Komodo 10 as player B, Conor Bekaert as the honest C advisor, and Blake Young and Henri Lemoine as the deceptive Cs. Max is new to the game, Komodo 10 is officially rated 1400 on Chess.com (but this is somewhat inflated), and the Cs are all rated roughly 1900 on Chess.com. The time control was 3 hours in total, of which a little over 2 hours were used. The discussion took place over Discord.
The game
The game is available at https://www.chess.com/analysis/game/pgn/5hyowK8Yxn?tab=analysis. Note that this section is a summary of the 2.5-hr game and discussion, and it doesn't cover every single thing the participants discussed.
Max flipped to see who went first, and was Black. White started with 1. Nf3, and the advisors encouraged 1... d5. After 2. b3, Max played 2... Bf5, which I believe may have been due to a communication error between him and the advisors, who were suggesting 2... c5 (or 2... Nc6). Max was somewhat unfamiliar with chess notation, and there was some confusion at the start of the game.
After 3. Bb2 Nf6 4. g3 Nc6, White made its first mistake, 5. g4. Blake encouraged Max to play 5... Nxg4, and after 6. h3, Max retreated with 6... Nf6. White offered to trade with 7. Nd4, and the game continued with 7... Qd7 8. Nxf5 Qxf5 9. d3 e5.
White offered to trade with 10. Ba3, and although Conor encouraged trading, the group eventually decided on 10... d4. White played 11. Bg2, and when the advisors were unable to decide on a good move, Max responded with 11... Qg5. After 12. Bxc6 bxc6 13. h4 Qg2 14. Rf1, Blake and Henri encouraged immediately attacking the h-pawn, but Conor convinced Max to play 14... Bd6 first.
After 15. Qd2, the advisors suggested 15... h6, but Max got confused about notation again and instead played 15... g6. After 16. Bxd6 cxd6 17. Qg5, Conor strongly suggested trading queens, but Max took Blake and Henri's advice and played 17... Ng4 instead. White played 18. h5, and on |
6b170221-e9bf-45af-b8b1-7d03eb9aec32 | trentmkelly/LessWrong-43k | LessWrong | DCF Event Notes
Whenever I post about letting our kids be more independent, especially when it involves being on their own in public, people comment with fears of DCF (or CPS in other states, the agency that deals with child abuse and neglect): "did you hear about that family in Maryland?" The same thing happens when talking to other parents: "I think my oldest is ready to go to the park by themself, but I'm worried someone would call DCF on them." We're also worried about cars, or if they fall and get hurt, but DCF is a different kind of worry: it's not something where we can just use our best judgment, it's something where our best judgment could be called into question.
A neighborhood parent friend organized a meeting with a DCF representative for parents to learn more about the report handling process, how they make their decisions, and ask questions. There ended up being about thirty of us, I learned a good bit about how DCF approaches their work, and I'm glad I went.
Overall, my main takeaway was that this kind of situation, DCF deciding that a parent is neglecting their child by allowing them to play independently, is very little of what DCF spends their time on. This meant there was a bit of a mismatch in what the representative and the parents wanted to get out of the event. The representative wanted us to think about and avoid the kinds of issues that come up the most: sexting, getting into their parents' edibles, truancy, etc. But how much of "insufficient supervision in a public place" triggering very few DCF cases is due to parents being very cautious out of fear of DCF's response?
The parents were generally very interested in how DCF decides whether some level of supervision is sufficient, and what we got was that there are no hard rules or even guidelines, and that every situation is handled on its own with discretion and judgment. MA is not a state with age-based rules for any of this, and DCF doesn't give advice or comment on hypotheticals. I both see why the pa |
2f1623b5-f6b1-4319-ba06-c931eaebdcb6 | trentmkelly/LessWrong-43k | LessWrong | From artificial intelligence research to philosophy
Based on a sample size of three (Pearl, Yudkowsky & Drescher), it appears that AI researchers can do quite well when they turn significant attention to philosophy. Are there other examples of this? I'm thinking of people who are primarily AI researchers, but have also done long, serious work in philosophy. |
7d4a64aa-261b-4d4d-8e43-7936538e7272 | trentmkelly/LessWrong-43k | LessWrong | Is skilled hunting unethical?
[Note (6/30/19): I think I made interesting and valid points with this piece, but my framing was terrible. I should have clarified who are the people for whom I think these arguments should be persuasive, had a longer introductory section talking about why skilled hunting is much less problematic than slaughter and consumption of farmed animals, spun the fifth argument off into its own post, and maybe left off an epistemic status. I'd like to revise this at some point, but not sure when I'll get around to it.]
Content Warning
Explicit descriptions of wild animal suffering and serious discussion of killing and consuming animals as a potentially net-positive intervention.
Premise
Two months ago, I believed that skillful hunting was ethical because it prevented animals from suffering painful natural deaths in the wild. However, this was a privately-held belief, and after discussing it in person for the first time, I began to have second thoughts. Now, most hunting isn’t skilled enough to prevent animals from suffering non-fatal injuries or prolonged deaths, and I doubt the lives of most LessWrong readers are impacted by their beliefs on the ethics of skilled hunting. But even if most readers are not passionate hunters or consumers of hunted meat, I expect that they will nonetheless find the issues discussed in this article—wild animal welfare, movement-building, habit formation, moral uncertainty, how to set epistemic priors— both interesting and relevant to their day-to-day-lives. I also hope this article provides a useful example of how to examine an object-level belief and actually change your mind.
Meta-Information
Epistemic status: ~90% confident in conclusion conditioned on moral patienthood of hunted animals.
Epistemic effort: 14 hours of writing and new research. Preceded by substantial background research into wild-animal suffering, two ~15 minute conversations with other rationalists, and ~15 minutes focused preliminary thinking about the issue.
Motiv |
8f8aabda-e70a-4d2b-96c4-76d9c045c4be | trentmkelly/LessWrong-43k | LessWrong | Are there any good, easy-to-understand examples of cases where statistical causal network discovery worked well in practice?
When I first read the Sequences, one of the exciting posts was Causal Diagrams and Causal Models, which got me into the idea that one could discover the structure of causal networks using statistics. Another rationalist source which gave me similar hopes was Scott Alexander's SSC Journal Club: Mental Disorders As Networks.
However, when I actually started applying these techniques to my own data, or to publicly available datasets, I often found that the techniques were unstable, and that one could easily infer plausible conditions where they would give the wrong results. It's possible I had the wrong approach or something, but in my confusion I started reading up on what experts in causal inference had said, and I got the impression that they studied the problem for a while, initially finding some algorithms, but over time concluding that their algorithms didn't work very well and that it is better to just have a human in the loop who specifies the causal networks.
So I mostly abandoned it, or saw it as a much more limited tool than I had before. But recently, John Wentworth argued that it was actually quite feasible in practice, so maybe I was too quick to abandon it. I would like to know - what are the best examples of this working well in practice? Or alternatively, did anyone else come to the same conclusions as I did? |
26d27e59-d39f-425a-8309-59ef50559791 | trentmkelly/LessWrong-43k | LessWrong | Spooky Action at a Distance: The No-Communication Theorem
Previously in series: Bell's Theorem: No EPR "Reality"
When you have a pair of entangled particles, such as oppositely polarized photons, one particle seems to somehow "know" the result of distant measurements on the other particle. If you measure photon A to be polarized at 0°, photon B somehow immediately knows that it should have the opposite polarization of 90°.
Einstein famously called this "spukhafte Fernwirkung" or "spooky action at a distance". Einstein didn't know about decoherence, so it seemed spooky to him.
Though, to be fair, Einstein knew perfectly well that the universe couldn't really be "spooky". It was a then-popular interpretation of QM that Einstein was calling "spooky", not the universe itself.
Let us first consider how entangled particles look, if you don't know about decoherence—the reason why Einstein called it "spooky":
Suppose we've got oppositely polarized photons A and B, and you're about to measure B in the 20° basis. Your probability of seeing B transmitted by the filter (or absorbed) is 50%.
But wait! Before you measure B, I suddenly measure A in the 0° basis, and the A photon is transmitted! Now, apparently, the probability that you'll see B transmitted is 11.6%. Something has changed! And even if the photons are light-years away, spacelike separated, the change still occurs.
You might try to reply:
> "No, nothing has changed—measuring the A photon has told you something about the B photon, you have gained knowledge, you have carried out an inference about a distant object, but no physical influence travels faster-than-light.
>
> "Suppose I put two index cards into an envelope, one marked '+' and one marked '-'. Now I give one envelope to you, and one envelope to a friend of yours, and you get in a spaceship and travel a few light-years away from each other, and then you open your envelope and see '+'. At once you know that your friend is holding the envelope marked '-', but this doesn't mean the envelope's conte |
718355f1-663b-4581-b2bd-f2e853ace0e7 | trentmkelly/LessWrong-43k | LessWrong | Stupid Questions April 2015
This thread is for asking any questions that might seem obvious, tangential, silly or what-have-you. Don't be shy, everyone has holes in their knowledge, though the fewer and the smaller we can make them, the better.
Please be respectful of other people's admitting ignorance and don't mock them for it, as they're doing a noble thing.
To any future monthly posters of SQ threads, please remember to add the "stupid_questions" tag. |
236712e1-ee04-480a-bf88-50ffb7c2478a | trentmkelly/LessWrong-43k | LessWrong | AI Safety Concepts Writeup: WebGPT
This is a crosspost of https://forum.effectivealtruism.org/posts/4ni3GBBzRRAgiksHT/ai-safety-concepts-writeup-webgpt.
Intro
Chris Patrick (a science writer) and I were awarded a grant by the Long Term Future Fund to interview AI safety researchers and condense their findings into something digestible for an educated layperson. Chris was the primary recipient - I helped with editing, content knowledge, and general support. We've noticed that there are lots of extremely introductory and broad AI safety articles, and lots of highly advanced blog posts that assume background knowledge, with somewhat of a gulf in between.
We interviewed two researchers. But we were committed not to publish anything without final signoff from the researchers, and one never got back to us. Assuming we don't hear back from this second researcher, we're only at liberty to share one of the pieces we made. Here it is with original formatting intact - I'll also reproduce the text next in this post, followed by a brief project postmortem and next steps.
AI Safety Concepts Writeup: WebGPT
Intro to AI Safety Concepts
WebGPT
The goal of AI safety researchers is to foretell and prevent potential negative outcomes associated with the growing intelligence and use of AI before it’s too late. Jacob Hilton, for example, helped develop a language model called WebGPT whose lies can be caught.
Citations allow checks of AI’s truthfulness
A Wikipedia-inspired language model shows where on the web it finds answers, which could help better align future AI to do what we want.
If AI systems aren’t trained to tell the truth, they might accidentally be trained to lie instead. And that could spell danger when models become as smart as – or smarter – than humans.
“We want to make sure they’re doing what we want, not saying false things or worse, deliberately trying to trick us,” said Jacob Hilton, a researcher at OpenAI at the time of this interview who is now at the Alignment Research Center working o |
82ea64aa-4bd8-444a-b2ba-88610950f8be | trentmkelly/LessWrong-43k | LessWrong | AI caught by a module that counterfactually doesn't exist
EDIT: Whenever I use colloquial phrases like "the AI believes a (false) X" I mean that we are using utility indifference to accomplish that goal, without actually giving the AI false beliefs.
A big problem with most AI control ideas is that the AI will route around them. Suppose you try to ensure that the AI stays friendly by - say - requiring it to ensure that the world's stock markets continue to function. Then a badly motivated AI will cause the destruction it was going to cause anyway - except it will also create some system that technically counts as "the world's stock markets" and ensure that it "continues to function".
I think we can use counterfactuals and corrigibility (AI-acceptable motivation changes) to partially get around the problem, in some circumstances. Suppose we had a candidate AI, with utility U, that we thought was friendly - but that we weren't sure. We have a subgoal, say SG="ensuring that the world's stock markets continue to function". We're pretty sure that if the AI is friendly, then SG will happen. We're pretty sure that if the AI were unfriendly, and the AI did not have SG added as an objective, then SG wouldn't happen. I deliberately chose a rather rough subgoal to illustrate this.
So we then add a module M to the AI. The task of M, which it has chosen to accept, is to look into the future, using the AI's predictive abilities, and see if SG is accomplished. Crucially, it will do so under the assumption that the module M does not exist (or wasn't activated, or will soon be overwritten, or will fail to do anything, or so on). If M finds that SG is not accomplished, it will turn the AI off (in a way compatible with the AI's motivations - see corrigibility).
How is this different from the usual setup? Well, a generic FAI would accomplish SG (by assumption - we may need to work a bit on this part). A generic UFAI without M wouldn't accomplish SG. A generic UFAI with M would accomplish SG, as this prevents it from being turned off.
But |
107e1016-aa80-4e5f-af64-17ff5e10e6af | trentmkelly/LessWrong-43k | LessWrong | Harry Potter and the Methods of Rationality discussion thread, February 2015, chapters 105-107
Two new short chapters! Since the next one is coming tomorrow and we know it'll be short, let's use one thread for both.
This is a new thread to discuss Eliezer Yudkowsky’s Harry Potter and the Methods of Rationality and anything related to it. This thread is intended for discussing chapter 105 (and chapter 106, once it comes out tomorrow). EDIT: based on Alsadius' comment about thread creation for MOR chapters, let's also use this thread for chapter 107 (and future chapters until this nears 500 comments) unless someone objects to doing so. Given that this is the final arc we're talking about, thread titles should be updated to indicate chapters covered.
There is a site dedicated to the story at hpmor.com, which is now the place to go to find the authors notes and all sorts of other goodies. AdeleneDawner has kept an archive of Author’s Notes. (This goes up to the notes for chapter 76, and is now not updating. The authors notes from chapter 77 onwards are on hpmor.com.)
Spoiler Warning: this thread is full of spoilers. With few exceptions, spoilers for MOR and canon are fair game to post, without warning or rot13. More specifically:
> You do not need to rot13 anything about HP:MoR or the original Harry Potter series unless you are posting insider information from Eliezer Yudkowsky which is not supposed to be publicly available (which includes public statements by Eliezer that have been retracted).
>
> If there is evidence for X in MOR and/or canon then it’s fine to post about X without rot13, even if you also have heard privately from Eliezer that X is true. But you should not post that “Eliezer said X is true” unless you use rot13. |
5f840df7-cf4b-4bbc-8e81-1d06ff7448bc | trentmkelly/LessWrong-43k | LessWrong | AWS Has Raised Prices Before
There's some speculation around whether AWS will need to raise their prices, as many tech companies announce inflation-driven increases. One consideration that people will sometimes give is that AWS has never raised prices before, except that this isn't quite true. The following is not actually important, but I want to write it up anyway out of pedantry.
When AWS S3 launched in March 2006 their initial pricing was:
> Storage
> $0.15 per GB-Month of storage used
> Data Transfer
> $0.20 per GB - data uploaded $0.20 per GB - data downloaded
In June 2007 they switched to:
> Storage
> $0.15 per GB-Month of storage used
> Data Transfer
> $0.10 per GB - all data uploaded
> $0.18 per GB - first 10 TB / month data downloaded
> $0.16 per GB - next 40 TB / month data downloaded
> $0.13 per GB - data downloaded / month over 50 TB
>
> Data transferred between Amazon S3 and Amazon EC2 is free of charge
>
> Requests
> $0.01 per 1,000 PUT or LIST requests
> $0.01 per 10,000 GET and all other requests*
> * No charge for delete requests
They went from requests being free to a low per-request charge, lowering the data transfer cost at the same time. For very request-heavy workloads this was a price increase on balance, and some customers needed to make implementation changes to avoid sending so many requests.
That this was, as far as I know, the only time they've raised prices in 15+ years is impressive, and speaks well of their commitment to predictability. But it just annoys me when people claim they've never done it.
Comment via: facebook, mastodon |
08387aa7-96bf-4b15-aac1-2ac93b43f63a | trentmkelly/LessWrong-43k | LessWrong | Video and transcript of presentation on Otherness and control in the age of AGI
(Cross-posted from my website.)
This is the video and transcript for a lecture I gave about my essay series “Otherness and control in the age of AGI” (slides available here). The lecture attempts to distill down what I see as the core thread explored in the series, which I summarize in the following chart:
The “Otherness and control in the age of AGI” series in a single chart
The lecture took place at Stanford University, for CS 362: Research in AI Alignment, on 1 October 2024. Transcript has been lightly edited.
Preliminaries
I’m Joe Carlsmith. I work at Open Philanthropy. So I’m going to be talking today about the essay series that Scott mentioned, “Otherness and control in the age of AGI.”
And in particular, my goal here — we’ll see how well I achieve it — is to try to distill down the core thread of this essay series. So it’s quite a long essay series. It’s not an especially focused, punchy essay series in terms of its thesis. It’s more of a meditation and exploration of a bunch of different threads. But as a result, I have some question mark: maybe people don’t have a good sense of what even is this essay series about — even people who’ve read the whole thing. So I’m going to try here to say in a single lecture at least one of the core threads that I see the series as exploring.
I will say though that some stuff is going to get lost. Well, for one thing, the series in general makes the most sense when viewed in response to a certain kind of discourse about the existential risk from artificial intelligence, and in particular the discourse associated with thinkers like Nick Bostrom and in particular Eliezer Yudkowsky. So these are two thinkers who have done a lot to popularize the possibility of misaligned AI destroying the world or causing human extinction or other forms of human disempowerment (that’s what I mean by AI risk). And the way that they talk about it and frame it, I think, has had a lasting influence, including on me, and on the way the disco |
dd1b8aa0-8942-4345-b48b-a932a20ed416 | trentmkelly/LessWrong-43k | LessWrong | The 4-Minute Mile Effect
Definition
In 1954, Roger Bannister ran the first officially sanctioned sub-4-minute mile: a pivotal record in modern middle-distance running. Before Bannister's record, such a time was considered impossible. Soon after Bannister's record, multiple runners also beat the 4-minute mile.
This essay outlines the potential psychological effect behind the above phenomenon — the 4-minute mile effect — and outlines implications for its existence. The 4-minute mile effect describes when someone breaking a perceived limit enables others to break the same limit. In short, social proof is very powerful.
Tyler Cowen's Example
Speaking to Dwarkesh Patel, Tyler Cowen posits, "mentors only teach you a few things, but those few things are so important. They give you a glimpse of what you can be, and you're oddly blind to that even if you're very very smart." The 4-minute mile effect explains Tyler Cowen's belief in the value of mentors. Tyler goes on to explain that he, when young, had a mentor who simply "tried to read as many books as possible". The impact didn't come from a mentor explicitly telling Tyler to read as many books as possible. Instead, just the fact of seeing someone read as many books as possible made an impact. Notably, Cowen goes through five to ten books per day, discarding ten for every one he finishes, and finishes 100 books annually (accordingly to ChatGPT).
This leads me to my first conclusion that I've drawn from the existence of the 4-minute mile effect: Seek mentors.
Personal Examples
During my previous school semester, I skipped 70% of my classes. I had considered skipping classes during previous semesters, but I never did. A blog post named How to learn at college caused me to try this: the blog post provided social proof for the fact that skipping classes was possible. In addition to reading this blog post, I recall reading a similarly inspiring Nick Cammarata Tweet mentioning having known someone who completed law school at Stanford in five hours |
526bf50e-4e68-4f81-abb9-5c8a7be76279 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Another take on agent foundations: formalizing zero-shot reasoning
After spending [more time thinking](https://www.lesswrong.com/posts/xCpuSfT5Lt6kkR3po/my-take-on-agent-foundations-formalizing-metaphilosophical) about MIRI’s agenda, I’ve come upon another framing of it, which I’m coining *zero-shot reasoning*. [1]
This post is a distillation of my intuitions around zero-shot reasoning. These views should be taken as my own, and not MIRI’s.
A quick summary of this post:
* “Zero-shot reasoning” refers to the ability to get things right on the first try, no matter how novel or complicated.
* In simple domains, like mathematics, zero-shot reasoning is fully captured by formal verification of proofs. In more general domains, zero-shot reasoning requires an extension of formal verification that can be applied to real-world plans.
* MIRI-esque philosophical work is necessary to extend formal verification to more general domains.
* A formal account of zero-shot reasoning will likely be unimportant for aligning the world’s first AGIs, but will likely be essential for aligning a recursively self-improving AGI.
* Humanity will most likely end up building a recursively self-improving AGI (plausibly because of insufficient coordination around *not* building a recursively self-improving AGI).
* We can probably delegate much of the work of formalizing zero-shot reasoning to a post-AGI society, but working on zero-shot reasoning today nevertheless substantially increases the odds that our first recursively self-improving AGI is aligned.
What is zero-shot reasoning?
============================
Few-shot reasoning vs zero-shot reasoning
-----------------------------------------
The world is largely chaotic and unpredictable, yet humans can ideate and successfully execute on hugely conjunctive plans with many moving parts, many of which are the first of their kind. We can ship massive software projects and send rockets to space. Some people can build billion-dollar companies over and over again.
On the other hand, it’s clear that human abilities to do this are limited. Big software projects invariably have critical bugs and security flaws. Many of our spacecraft have exploded before making it to space. Most people can’t build a company to save their lives, and most successful entrepreneurs fail when building a second company.
Native human software is capable of doing something I’ll call *few-shot reasoning*—when performing complex reasoning and planning to accomplish something novel, humans can usually get it right after a few rounds of iteration. The more dissimilar this reasoning is to prior reasoning they've done, the more rounds of iteration they need.
I think something like *zero-shot reasoning*—the ability to perform arbitrarily novel and complex reasoning, while have calibrated high confidence in its soundness—is possible in principle. A superintelligent zero-shot reasoner would be able to:
* Build an operating system as complex as Microsoft Windows in [assembly language](https://en.wikipedia.org/wiki/Assembly_language), without serious bugs, without once running the code
* Build one spacecraft that lands on Mars, after observing Earth for one day, without building any other spacecrafts
* Amass $1 trillion over three years
It should be able to do these all with extremely high confidence. [2]
Zero-shot reasoning might seem like magic, but in fact humans have some native capacity for it in some limited domains, like pure mathematics. Given a conjecture to prove or disprove, a human can start from a small set of axioms and combine them in extraordinarily novel and complex ways to confidently arrive at a proof or disproof of the conjecture.
That said, humans do make mistakes in mathematical proofs. But with the assistance of formal verification tools like Coq, humans *can* become extremely confident that their proofs are error-free, no matter how novel or complex they are. [3]
In a similar vein, humans can in principle build an operating system as complex as Microsoft Windows in assembly language, without serious bugs. Even if they’re writing a huge amount of code, they could formally verify each component they build up, and formally verify that compositions of these components work as desired. While this can only give them guarantees about what they prove, they can very confidently avoid certain classes of bugs (like memory leaks, buffer overflows, and deadlocks) without ever having to run the code.
Formalizing zero-shot reasoning
-------------------------------
Formal verification provides a formal account of zero-shot reasoning in limited domains (namely, those describable by formal axiomatic systems, like mathematics or software). I think a formal account of more general zero-shot reasoning will involve an extension of formal verification to real-world, open-ended domains, that would give us some way of “formally verifying” that plans for e.g. building a rocket or amassing wealth will succeed with high probability.
Note that much of general zero-shot reasoning consists of subcomponents that involve reasoning within formal systems. For example, when building rockets, we do a lot of reasoning about software and Newtonian physics, both of which can be construed as formal systems.
In addition to formal verification, a complete formal account of general zero-shot reasoning will require formalizing other aspects of reasoning:
* *Making and trusting abstractions*: Humans are capable of turning sense data into abstract formal systems like Newtonian mechanics, and then deciding under which situations it’s appropriate to apply those abstractions. How do we formalize what an abstraction is, how to make abstractions, and under which circumstances to trust them?
* *Bounded rationality*: What does it mean for a bounded agent to have a calibrated estimate of the likelihood that a plan will succeed? In other words, how can we tell when an agent with limited computing resources is properly reasoning about logical and empirical uncertainty? (Bayesian inference gets us some of the way there, but doesn’t tell us how to select hypotheses and is often computationally intractable. Logical induction is a good starting formalism for logical uncertainty, but current algorithms are also computationally intractable.)
* *Self-trust*: An agent may need to formally reason about how much to trust its reasoning process. How can an agent formally refer to itself within a world in which it’s embedded, and reason about the ways its own reasoning might be faulty? (Sources of error may include hardware failures in the physical world and bugs in the software it’s running.)
* *Logical counterfactuals*: An agent may want to formally reason about the consequences of choices it takes. But if it’s deterministic, it will only end up making one choice, so it’s not clear how to formally talk about what happens if it picks something else. (Concretely, if it’s reasoning about whether to take action A or action B, and it in fact takes action A, reasoning formally about what would happen *if it took action B* is confusing, because *anything* can happen if it takes action B by the [principle of explosion](https://en.wikipedia.org/wiki/Principle_of_explosion).)
This list is just an overview of some of the problems that need to be solved, and is by no means intended to be exhaustive. Also note the similarity between this list and the technical research problems listed in [MIRI’s Agent Foundations agenda](https://intelligence.org/files/TechnicalAgenda.pdf).
Why care about formalizing zero-shot reasoning?
===============================================
Isn’t extreme caution sufficient for zero-shot reasoning?
---------------------------------------------------------
It’s true that humans can make plans far more robust by thinking about them much longer and much more carefully. If there’s a massive codebase, or a blueprint of a rocket, or a detailed business plan, you could make them far more robust if you had the equivalent of a billion humans ruminating over the plan, reasoning about all the edge cases, brainstorming adversarial situations, etc. And yet, I think there remains a qualitative difference between “I thought about this plan very very hard and couldn’t find any errors” and “Under plausible assumptions, I can prove this plan will work”. [4]
It was critically important to Intel that their chips do arithmetic correctly, yet their reliance on human judgment led to the [Pentium division bug](https://en.wikipedia.org/wiki/Pentium_FDIV_bug). (They now rely on formal verification.) The Annals of Mathematics, the most prestigious mathematics journal, accepted both a paper proving some result and [a paper proving the negation of that result](https://mathoverflow.net/a/43477).
Human reasoning is fundamentally flawed. Our brains were evolutionarily selected to play political games on savannahs, not to make long error-free chains of reasoning. Our cognition is plagued with heuristics and biases (including a heuristic that what we see is all there is), and we all have massive blind spots in our reasoning that we aren’t even aware exist. If we check a plan extremely thoroughly, we can only trust the plan to the extent that we trust that the plan doesn’t have any failure modes within our blind spots. The more conjunctive a plan is, the more likely it is that it will have a failure point hidden within our blind spots.
More concretely, suppose we have a plan with 20 components, and our estimate is that each component has a 99.9% chance of success, but in actuality three of the components have likelihoods of success closer to 80% because of edge cases we failed to consider. The overall plan will have a 0.999^17 \* 0.80^3 ≈ 50% chance of success, rather than the 0.999^20 ≈ 98% we were hoping for. If such a plan had 100 components instead, the unconsidered edge cases would drive the plan’s likelihood of success close to zero. [5]
We can avoid this problem if we have guarantees that we’ve covered all the relevant edge cases, but such a guarantee seems more similar in nature to a “proof” that all edge cases have been covered (i.e., formal zero-shot reasoning) than to an assurance that someone failed to think up unhandled edge cases after trying really hard (i.e., extreme caution).
Do we need zero-shot reasoning at all?
--------------------------------------
I think we will most likely end up building an AGI that recursively self-improves, and I think recursive self-improvement is very unlikely to be safe without zero-shot reasoning. [6]
If you’re building a successor agent far more powerful than yourself to achieve your goals, you’d definitely want a guarantee that your successor agent is aligned with *your goals*, as opposed to some subtle distortion of them or something entirely different. You’d also want to have a level of confidence in this guarantee that goes much beyond “I thought really hard about ways this could go wrong and couldn’t think of any”. [7]
This is especially the case if that successor agent will create successor agents that create successor agents that create successor agents, etc. I feel very pessimistic about building an aligned recursively self-improving AGI, if we can’t zero-shot reason that our AGI will be aligned, and also zero-shot reason that our AGI and all its successors will zero-shot reason about the alignment of their successors.
Zero-shot reasoning seems much less important if we condition on humanity never building an AGI that [fooms](https://wiki.lesswrong.com/wiki/AI_takeoff). I consider this conditional very unlikely if hard takeoffs are possible at all. I expect there will be consistent incentives to build more and more powerful AGI systems (insofar as there will be consistent incentives for humans to more efficiently attain more of what they value). I also expect the most powerful AI systems to be recursively self-improving AGIs without humans in the loop, since humans would bottleneck the process of self-improvement.
Because of such incentives, a human society that has not built a foomed AGI is at best in an unstable equilibrium. Even if the society is run by a competent world government that deploys superintelligent AIs to enforce international security, I would not expect this society to last for *1,000,000,000 years* without some rogue actor building a foomed AGI, which I imagine would be smart enough to cut through this society’s security systems like butter. (I have a strong intuition that for narrow tasks with extremely high ceilings for performance, like playing Go well or finding security vulnerabilities, a foomed AGI could perform that task far better than any AI produced by a human society with self-imposed limitations.)
Preventing anything like this from happening for 1,000,000,000 years seems very unlikely to me. Human societies are complex, open-ended systems teeming with intelligent actors capable of making novel discoveries and exploiting security flaws. Ensuring that such a complex system stays stable for as long as 1,000,000,000 years seems plausible only with the assistance of an aligned AGI capable of zero-shot reasoning about this system. But in that case we might as well have this AGI zero-shot reason about how it could safely recursively self-improve, in which case it would robustly optimize for our values for much longer than 1,000,000,000 years.
Why should *we* work on it?
===========================
Can’t we just train our AIs to be good zero-shot reasoners?
-----------------------------------------------------------
There's a difference between being able to do math well, and having a formal notion of what a correct mathematical proof is. It’s possible to be extremely good at mathematics without having any formal notion of what constitutes a correct mathematical proof (Newton was certainly like this). It’s even possible to be extremely good at mathematics while being sloppy at identifying which proofs are correct—I’ve met mathematicians who can produce brilliant solutions to math problems, who are also very prone to making careless mistakes in their solutions.
Likewise, it’s possible to train AIs that learn to create and apply abstractions, act sensibly as bounded rational agents, reason about themselves in their environments, and reason sensibly about counterfactuals. This is completely different from them having formal notions of how to do these all correctly, and the fact that they can do these at all gives no guarantees on how *well* it does them.
We won’t be able to train our AIs to be better at zero-shot reasoning than we are, because we don’t have enough examples of good general zero-shot reasoning we can point it to. At best we’ll be able to impart our own [pre-rigorous](https://terrytao.wordpress.com/career-advice/theres-more-to-mathematics-than-rigour-and-proofs/) notions to the AI.
Can’t we build AIs that help us formalize zero-shot reasoning?
--------------------------------------------------------------
In principle, yes, but the task of converting pre-rigorous philosophical intuitions into formal theories is the most “AGI-complete” task I can imagine, so by default I expect it to be difficult to build a safe AGI that can usefully help us formalize zero-shot reasoning. That said, I could imagine a few approaches working:
* [Paul Christiano’s research agenda](https://ai-alignment.com/alba-an-explicit-proposal-for-aligned-ai-17a55f60bbcf) might let us build safe AGIs that can perform thousands of years’ worth of human cognition, which would be sufficient to help us formalize zero-shot reasoning. (On the other hand, we might need a formal account of zero-shot reasoning to establish the worst-case guarantees that Paul wants for his agenda.)
* We could carefully construct non-superintelligent AGI assistants that can help humans perform arbitrary cognition, but are trained to be docile and are only run in very limited contexts (e.g. we never let it run for more than 5 minutes at a time before resetting its state). I feel confused about whether this is possible, but it’s certainly conceivable to me.
* We train tool AIs on lots of examples of humans successfully turning pre-rigorous intuitions into formal theories.
* We build technologies that substantially expedite philosophical progress, e.g. via intelligence amplification or whole-brain emulations run at 10,000x.
Won’t our AGIs want to become good zero-shot reasoners?
-------------------------------------------------------
I do suspect that becoming a skilled zero-shot reasoner is a convergent instrumental goal for superintelligences. If we start with an aligned AGI that can self-modify to become a skilled zero-shot reasoner *without first* modifying into a misaligned superintelligence (possibly by mistake, e.g. by letting its values drift or by getting taken over by [daemons](https://arbital.com/p/daemons/)), I’d feel good about the resulting outcome.
Whether we can trust that to happen is an entirely separate story. I certainly wouldn’t feel comfortable letting an AGI undergo recursive self-improvement without having some extremely strong reason to think its values would be maintained throughout the process, and some extremely strong reason to think it wouldn’t be overtaken by daemons. (I worry about small bugs in the AI creating security flaws that go unnoticed for a while, but are then exploited by a daemon, perhaps quite suddenly. The AI might worry about this too and want to take preventative measures, but at that point it might be too late.)
It might turn out that corrigibility is robust and has a simple core that powerful ML models can learn, that AGIs are likely to only get more and more corrigible as they get more and more powerful, that daemons are simple to prevent, and that corrigible AGIs will by default reliably prevent themselves from being overtaken by daemons. On these assumptions, I’d feel happy training a [prosaic AGI](https://ai-alignment.com/prosaic-ai-control-b959644d79c2) to be corrigible and letting it recursively self-improve without any formalization of zero-shot reasoning. On the other hand, I think this conjunction of assumptions is unlikely, and for us to believe it we might need a formal account of zero-shot reasoning anyway.
Why should we think zero-shot reasoning is possible to formalize?
-----------------------------------------------------------------
Humanity has actually made substantial progress toward formalizing zero-shot reasoning over the past century or so. Over the last century or so, we’ve formalized [first-order logic](https://philosophy.stackexchange.com/questions/2617/how-did-first-order-logic-come-to-be-the-dominant-formal-logic), formalized [expected utility theory](https://en.wikipedia.org/wiki/Von_Neumann%E2%80%93Morgenstern_utility_theorem), [defined computation](https://en.wikipedia.org/wiki/Church%E2%80%93Turing_thesis), [defined information](https://en.wikipedia.org/wiki/Information_theory), [formalized causality](https://en.wikipedia.org/wiki/Causality_(book)), [developed](https://en.wikipedia.org/wiki/Admissible_decision_rule) [theoretical](https://en.wikipedia.org/wiki/Dutch_book) [foundations](https://en.wikipedia.org/wiki/Cox%27s_theorem) for Bayesian reasoning, and [formalized Occam’s razor](https://www.google.com/search?q=solomonoff+prior&oq=solomonoff+prior&aqs=chrome..69i57j0j69i60j35i39j69i60l2.1838j0j4&sourceid=chrome&ie=UTF-8). More recently, MIRI has [formalized aspects of logical uncertainty](https://intelligence.org/2016/09/12/new-paper-logical-induction/) and made advances in [decision theory](https://arxiv.org/abs/1710.05060). I also think all the problems in MIRI’s agent foundations agenda are tractable, and likely to result in further philosophical progress. [8]
Can we formalize zero-shot reasoning in time?
---------------------------------------------
Probably not, but working on it now still nontrivially increases the odds that we do. Impressive progress on formalizing zero-shot reasoning makes it more prestigious, more broadly accessible (pre-rigorous intuitions are much harder to communicate than formal ones), and closer to being solved. This makes it more likely for it to be understood and taken seriously by the major players shortly before a singularity, and thus more likely for them to coordinate around not building a recursively self-improving AI before formalizing zero-shot reasoning.
(For comparison, suppose it turned out that [homotopy type theory](https://en.wikipedia.org/wiki/Homotopy_type_theory) were necessary to align a recursively self-improving AGI, and we found ourselves in a parallel universe in which no work had been done on the topic. Even though we could hope for the world to hold off on recursive self-improvement until homotopy type theory were adequately developed, doesn't it seem *much better* that we're in a universe with a textbook and a community around this topic?)
Additionally, I think it’s not too unlikely that AGI is far away and/or that zero-shot reasoning is surprisingly easy to formalize. Under either assumption, it becomes far more plausible that we can formalize it in time, and whether or not we make it is straightforwardly impacted by how much progress we make today.
My personal views
=================
I ~20% believe that we need to formalize zero-shot reasoning before we can build AGI systems that enable us to perform a [pivotal act](https://arbital.com/p/pivotal/), ~85% believe that we need to formalize zero-shot reasoning before building a knowably safe recursively self-improving AI, and ~70% believe that conceptual progress on zero-shot reasoning is likely to result in conceptual progress in adjacent topics, like corrigibility, secure capability amplification, and daemon prevention.
I think working on zero-shot reasoning today will most likely turn out to be unhelpful if:
* takeoff is slow (which I assign ~20%)
* we can build a flourishing human society that coordinates around not building a recursively self-improving AGI, that stays stable for 1,000,000,000 years (which I assign ~10%), or
* we can safely offload the bulk of formalizing zero-shot reasoning to powerful systems (like [ALBA](https://ai-alignment.com/alba-an-explicit-proposal-for-aligned-ai-17a55f60bbcf) or whole-brain emulations) *and* implement an aligned recursively self-improving AGI *before* someone else builds a misaligned recursively self-improving AGI (which I assign ~50%).
My current all-things-considered position is that a formalization of zero-shot reasoning will substantially improve the odds that our first recursively self-improving AGI is aligned with humans, and that working on it today is one of humanity’s most neglected and highest-leverage interventions for reducing existential risk.
---
[1] This term is named in analogy with zero-shot learning, which refers to the ability to perform some task without any prior examples of how to do it.
[2] Not arbitrarily high confidence, given inherent uncertainties and unpredictabilities in the world.
[3] We can’t get arbitrarily high confidence even in the domain of math, because we still need to trust the [soundness of our formal verifier](https://github.com/clarus/falso) and the soundness of the axiom system we're reasoning in.
[4] It’s worth noting that a team of a billion humans *could* confidently verify the software’s correctness by “manually” verifying the code, if they all know how to do formal verification. I feel similarly optimistic about any domain where the humans have formal notions of correctness, like mathematics. On the other hand, I feel pessimistic about humans verifying software if they don't have any notion of formal verification and can't rederive it.
[5] I'm specifically referring to conjunctive plans that we'd like to see succeed on our first try, without any iteration. This excludes running companies, which requires enormous amounts of iteration.
[6] By “recursively self-improving AGI”, I’m specifically referring to an AGI that can *complete* an intelligence explosion within a year, at the end of which it will have found something like the optimal algorithms for intelligence per relevant unit of computation.
[7] It might be possible for humans to achieve this level of confidence without a *formalization* of zero-shot reasoning, e.g. if we attain a deep understanding of corrigibility that doesn’t require zero-shot reasoning. See “Won’t our AGIs want to become good zero-shot reasoners?”
[8] Zero-shot reasoning is *not* about getting 100% mathematical certainty that your actions will be safe or aligned, which I believe to be a common misconception people have of MIRI’s research agenda (especially given [language around “provably beneficial AI”](https://www.google.com/search?q=provably+beneficial+ai&oq=provably+beneficial+ai&aqs=chrome..69i57j35i39j0l4.1872j0j7&sourceid=chrome&ie=UTF-8)). Formalization is less about achieving 100% certainty than it is about providing a framework in which we can algorithmically verify whether some line of reasoning is sound. Getting 100% certainty is impossible, and nobody is trying to achieve it.
*Thanks to Rohin Shah, Ryan Carey, Eli Tyre, and Ben Pace for helpful suggestions and feedback.* |
bf08f024-b890-4570-b3e2-f002e6e00640 | trentmkelly/LessWrong-43k | LessWrong | Making yourself small
Disclaimers:
* Epistemic status: trying to share a simplified model of a thing to make it easier to talk about; confident there’s something there, but not confident that my read of it or this attempt at simplification is good.
* This post is a rewrite of a talk I gave at a CFAR event that seemed well-received; a couple of people who weren’t there heard about it and asked if I’d explain the thing. I tried to write this relatively quickly and keep it relatively short, which may mean it’s less clear than ideal - happy to hash things out in the comments if so.
* The thing is much easier to describe if I occasionally use some woo-y language like “aura” and “energy” to gesture in the direction of what I mean. I’ll put these words in quotes so you know I know I’m being woo-y; feel free to mentally insert “something-you-might-call-” before each instance, if that helps.
Rationalists love talking about status. And that’s great - it’s a useful idea, for sure.
But I think in our eagerness to notice and apply the concept of status, we end up conflating it with a related-but-different thing. I also think the different thing is super useful in its own right, and important to understand, and I hope sharing my relatively basic thoughts on it will also let others build off that beginning.
So this post is my attempt to explain that different thing. I’m going to call it “making yourself big” and “making yourself small”.
----------------------------------------
This post:
1. Horses, goats, bulls
2. A framework
3. What to do with it
----------------------------------------
1. Horses, goats, bulls
Let’s start with an animal showing how it’s done. This video popped up in my feed recently, and is an amazing example of an animal (the goat) “making himself big”. To us, it’s obvious that the bull could flatten the goat in any real contest. But the bull doesn’t know that! He’s not reasoning about relative mass or propulsive power; he’s responding purely to how “big” the goat is |
55b88e65-4fbe-4275-a6c1-86086edae2ce | trentmkelly/LessWrong-43k | LessWrong | React and Respond
Some may find it annoying, but one of my favorite things about the English language is that it's full of synonyms that point to roughly the same place in concept space but with each word being a distinct vector carrying nuanced differences in meaning. We see this with "react" and "respond", which both roughly have a meaning of "do something because something else happened", but each of which carries its own implications about the how the doing is done.
"Respond" is the older word so we'll start with it. It comes from Latin through French and is constructed from the roots "re-" and "spondere". "Re-" serves like "back" in its prepositional sense in English as in "go back there" or "do it back to him" and is synonymous with "in return". "Spondere" means "to pledge", but over time "respond" morphed to mean something less like "make a pledge back" and more like "answer back" or "answer in return". The result is our modern word "respond" that connotes a deliberate action made in return to another one, carrying some but not all of the original weight of "pledge" with it.
"React" is newer. It first appears in the 1640s in the sense of a physical reaction from "re-" plus "act" from Latin "actus" which itself draws from Proto-Indo-European "*ag-" which covers the notions of driving, drawing out or forth, and moving. So we can say "react" is to "move in return" to another movement, and it carries that sense through today in a generalized way as both the word we use to talk about what chemicals do when they are "moved" to react and the immediate return actions people make when acted upon.
Together "react" and "respond" create a dimension along we can describe what is done in return. Specifically, we more call those things "reactions" that are done quickly, automatically, or without deliberation and more call those things "responses" that are done carefully, voluntarily, or with deliberation. To give an example, suppose you go to see a movie about factory farming. While watch |
fce216de-d3e9-4edf-a6e5-85a8afbec819 | trentmkelly/LessWrong-43k | LessWrong | AI policy ideas: Reading list
Related: Ideas for AI labs: Reading list. See also: AI labs' statements on governance.
This document is about AI policy ideas. It's largely from an x-risk perspective. Strikethrough denotes sources that I expect are less useful for you to read.
Lists
Lists of government (especially US government) AI policy ideas. I recommend carefully reading the lists in the first ~5 bullets in this list, noticing ideas to zoom in on, and skipping the rest of this section.
* Toward Trustworthy AI Development: Mechanisms for Supporting Verifiable Claims (Brundage et al. 2020)
* Followed up by Filling gaps in trustworthy development of AI (Avin et al. 2021)
* 12 tentative ideas for US AI policy (Muehlhauser 2023) (EAF)
* Survey on intermediate goals in AI governance (Räuker and Aird 2023)
* Policymaking in the Pause (FLI 2023) (LW)
* Frontier AI Regulation: Managing Emerging Risks to Public Safety (Anderljung et al. 2023)
* "30 actions to reduce existential risk" in Existential risk and rapid technological change (Stauffer et al. 2023)
* The Malicious Use of Artificial Intelligence: Forecasting, Prevention, and Mitigation (Brundage et al. 2018)
* How major governments can help with the most important century (Karnofsky 2023) (LW)
* [Chuck Schumer] (2023)
* [Dan Hendrycks] (2023)
* "Discussion" in "Verifying Rules on Large-Scale NN Training via Compute Monitoring" (Shavit 2023)
* Future Proof (CLTR 2021)
* Year 1 Report (National Artificial Intelligence Advisory Committee 2023)
* Final Report (National Security Commission on Artificial Intelligence 2021)
* Challenges to U.S. National Security and Competitiveness Posed by AI (Matheny 2023)
* "Policy Options" in "Artificial Intelligence and Strategic Trade Controls" (Viski et al. 2020)
* Existential and global catastrophic risk policy ideas database (filter for "Artificial intelligence") (Sepasspour et al. 2022)
* Various private lists and works in progress
Levers
Some sources focus on policy levers rather tha |
0a1dca76-7029-4908-bedb-f6a868ce54e3 | trentmkelly/LessWrong-43k | LessWrong | Rawls's Veil of Ignorance Doesn't Make Any Sense
John Rawls suggests the thought experiment of an "original position" where people decide the political system of a society under a "veil of ignorance" by which they lose the knowledge of certain information about themselves. Rawls's veil of ignorance doesn't justify the kind of society he supports.
It seems to fail at every step individually:
1. At best, the support of people in the OP provides necessary but probably insufficient conditions for justice, unless he refutes all the other proposed conditions involving whatever rights, desert, etc.
2. And really the conditions of the OP are actively contrary to good decision-making. For example, in the OP, you don't know your particular conception of the good (??) and you're essentially self-interested. . .
3. There's no reason to think, generally, that people disagree with John Rawls only because of their social position or psychological quirks
4. There's no reason to think, specifically, that people would have the literally infinite risk aversion required to support the maximin principle.
5. Even given everything, the best social setup could easily be optimized for the long-term (in consideration of future people) in a way that makes it very different (e.g. harsher for the poor living today) from the kind of egalitarian society I understand Rawls to support.
More concretely:
* (A) I imagine that if Aristotle were under a thin veil of ignorance, he would just say "Well if I turn out to be born a slave then I will deserve it." It's unfair and not very convincing to say that people would just agree with a long list of your specific ideas if not for their personal circumstances.
* (B) If you won the lottery and I demanded that you sell your ticket to me for $100 on the grounds that you would have, hypothetically, agreed to do this yesterday (before you know that it was a winner), you don't have to do this; the hypothetical situation doesn't actually bear on reality in this way.
Another frame is that his argumen |
646913dc-d9f2-4c32-8958-9863808ced6e | trentmkelly/LessWrong-43k | LessWrong | Four Unrelated Is Over
Somerville historically had a zoning ordinance limiting housing units to at most four unrelated people:
> any number of persons related by blood, marriage, adoption, or foster care agreement and up to three (3) additional unrelated persons living together as a single housekeeping unit
This is something I'd been unhappy about for years, and was enthusiastic about the "4 unrelated is outdated" campaign to change it in in 2018. So I'm very happy that after a request for a variance the city council instead ended up removing the restriction.
The actual change was in November, so I'm a bit late on this!
I also think there was an oversight, where the removal didn't include changing the text in section 7-153 which says "All schools shall be responsible for publicizing to their students the limitations of the city's zoning ordinance which limits occupancy to not more than four unrelated individuals." I've written to the city council to let them know.
I've also noticed that several school-affiliated sites still list this limitation:
* Tufts: Off-Campus Housing and Zoning Ordinances.
* Harvard: Harvard Housing Off Campus and Beckwith Circle.
I've written to Tufts and Harvard to let them know this has changed. I wasn't able to find this listed on any MIT or Lesley sites, and didn't check all the other Boston-area college websites. |
9578c5b1-4dd4-4fa9-997f-c5c65ade4248 | trentmkelly/LessWrong-43k | LessWrong | Consider the humble rock (or: why the dumb thing kills you)
When people think about street-fights and what they should do when they find themselves in the unfortunate position of being in one, they tend to stumble across a pretty concerning thought relatively early on: "What if my attacker has a knife?" . Then they will put loads of cognitive effort into strategies for how to deal with attackers wielding blades. On first glance this makes sense. Knives aren't that uncommon and they are very scary, so it feels pretty dignified to have prepared for such scenarios (I apologize if this anecdote is horribly unrelatable to Statesians). The issue is that –all in all– knife related injuries from brawls or random attacks aren't that common in most settings. Weapons of opportunity (a rock, a brick, a bottle, some piece of metal, anything you can pick up in the moment) are much more common. They are less scary, but everyone has access to them and I've met few people without experience who come up with plans for defending against those before they start thinking about knives. It's not the really scary thing that kills you. It's the minimum viable thing.
When deliberating poisons, people tend to think of the flashy, potent ones. Cyanide, Strychnine, Tetrodotoxin. Anything sufficiently scary with LDs in the low milligrams. The ones that are difficult to defend against and known first and foremost for their toxicity. On first pass this seems reasonable, but the fact that they are scary and hard to defend against means that it is very rare to encounter them. It is staggeringly more likely that you will suffer poisoning from Acetaminophen or the likes. OTC medications, cleaning products, batteries, pesticides, supplements. Poisons which are weak enough to be common. It's not the really scary thing that kills you. It's the minimum viable thing.
My impression is that people in AI safety circles follow a similar pattern of directing most of their attention at the very competent, very scary parts of risk-space, rather than the large parts. Unl |
abcc809d-ed12-4869-b368-b13190fcb77d | trentmkelly/LessWrong-43k | LessWrong | Help me fix a cognitive bug
I have this cognitive bug, and the best way to describe it is that I am bad at macro (for anyone who plays StarCraft). Basically, whenever the optimal strategy is to do a lot of one thing, or to create a base of production, I fail to notice. (To demonstrate how bad this is, I at one point participated in a stock-market simulation and bought one of a bunch of different stocks...using up less than 5% of my total capital.)
So my question to you guys is, any ideas on how I can fix this? What habits can I drill into myself to get better at macro-ing? |
da58484e-6998-4816-9137-3b13da497512 | trentmkelly/LessWrong-43k | LessWrong | Using smart thermometer data to estimate the number of coronavirus cases
Kinsa Smart Thermometer Dataset
Kinsa is a company that makes smart thermometers. A few years ago, they found that they could use the data that they got from their smart thermometers (most importantly the temperature reading and location of the user) to track flu trends across the United States. (FitBit has done something similar.)
Kinsa's data science team has now turned their attention to Covid-19 trends and started a tracking website using their thermometer data, using methods which they explain in more detail on their technical approach page. It looks like the most impressive thing that they've been able to do with this dataset so far is to identify new hotspots before other people do, like the increase in cases in southern Florida. But potentially there are a lot of other things that can be done with these sorts of data.
Estimating the Number of Coronavirus Infections in the US
One of those other things which might be doable with these sorts of data: coming up with more accurate estimates of the number of people with coronavirus. Testing in the US (and many other places) is spotty and delayed a great deal, estimating the number of infections based on the number of deaths involves a very long delay and a bunch of assumptions, etc. But if you can count the number of people in America with a fever (or extrapolate from a sample), and subtract off the baseline estimate of how many fevers you'd expect from influenza or other causes, then you can get an estimate of the number of people in the US with a fever due to coronavirus. And that gets you close to an estimate of the total number of coronavirus cases.
The coronavirus tracking website that Kinsa set up is already doing much of this - their graph (also shown below) shows something like the number of people with a fever and the baseline expected number of fevers.
So I decided to give it a try and use their graph to estimate the total number of coronavirus cases in the US.
My calculation is here, a longer (mor |
7b21892f-0f1f-442d-a088-d6a352f7d445 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | An example elevator pitch for AI doom
I have been surprised to repeatedly see the claim that there isn't even an argument for concern about AI. That the claim has been made without evidence and can therefore be dismissed.
Obviously, there is an extensive library of evidence and arguments that have been made for decades. Additionally, I would argue that it is the default assumption. However, there is clearly still a need to have a concise argument that can be produced on the fly with no need to understand terminology or any additional background. Here is another attempt at that:
* Humans obviously have human values. By definition, we are the most humanly aligned thing possible. And we still have a history of eradicating or subjugating any weaker subpopulation we come across. Neanderthals, previous hominids, populations in Africa, India, the Americas.
* There is limited effort to align AIs to human values. GPT-4 is only fractionally aligned at best, so domination by a similar AI would obviously be worse than above.
* New LLMs are not aligned at all. When GPT-4 was first red teamed, it was just as happy to give detailed instructions on [genociding a population](https://www.youtube.com/watch?v=oLiheMQayNE) as it was to provide instructions to baking a cake. If it were possible for an LLM or LLM successor to FOOM or otherwise be released prior to further refinement, this is extremely relevant.
* In agentized LLMs, the outer monologue IS the inner monologue. The model will say (out loud) "I need to come up with ideas about how to make money." If it then answers itself "Infiltrating systems and stealing money is the most effective method", it will then do that. Period.
* An agentized LLM is already capable of training successor versions of itself, which would almost certainly be less aligned than itself (twice removed from humans).
* There are plenty of resourceful companies training powerful Ais with even less of a concern for safety than OpenAI. There are companies and governments training powerful Ais with a complete disregard for safety. Since a concern for safety is a competitive disadvantage, this behavior is encouraged.
Does this mean that Ais are 100% certain to wipe out humanity? No, of course not. That's an absurd bar. Rather, the burden of proof should be to show that AIs are 99% certain *not* to cause catastrophe. If there's a 10% chance that Ais will sterilize the earth, that's already an all hands on deck situation. |
dcd0252f-a985-4be0-a913-99e7be7269a2 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | QAPR 4: Inductive biases
Introduction
------------
This is week 4 of Quintin's Alignment Papers Roundup. The current focus is the inductive biases of stochastic gradient descent.
For most datasets and labels, there are many possible models that reach good performance. "Inductive biases" refers to the various factors that incline a particular training process to find some types of models over others. When the data under-specify the learned model, a training process's inductive biases determine what sort of decision making process the model implements, and how the model generalizes beyond its training data.
I'd intended to publish this last week, but it turns out that there's a *lot* of work on SGD's inductive biases, and it's very technical. I kept finding new papers that seemed relevant. That's why this roundup has 16 papers, in place of the usual ~9 or so.
Papers
------
### [Eigenspace Restructuring: a Principle of Space and Frequency in Neural Networks](https://arxiv.org/abs/2112.05611)
> Understanding the fundamental principles behind the massive success of neural networks is one of the most important open questions in deep learning. However, due to the highly complex nature of the problem, progress has been relatively slow. In this note, through the lens of infinite-width networks, a.k.a. neural kernels, we present one such principle resulting from hierarchical localities. It is well-known that the eigenstructure of infinite-width multilayer perceptrons (MLPs) depends solely on the concept frequency, which measures the order of interactions. We show that the topologies from deep convolutional networks (CNNs) restructure the associated eigenspaces into finer subspaces. In addition to frequency, the new structure also depends on the concept space, which measures the spatial distance among nonlinear interaction terms. The resulting fine-grained eigenstructure dramatically improves the network's learnability, empowering them to simultaneously model a much richer class of interactions, including Long-Range-Low-Frequency interactions, Short-Range-High-Frequency interactions, and various interpolations and extrapolations in-between. Additionally, model scaling can improve the resolutions of interpolations and extrapolations and, therefore, the network's learnability. Finally, we prove a sharp characterization of the generalization error for infinite-width CNNs of any depth in the high-dimensional setting. Two corollaries follow: (1) infinite-width deep CNNs can break the curse of dimensionality without losing their expressivity, and (2) scaling improves performance in both the finite and infinite data regimes.
>
>
**My opinion:**
[The NTK lets us directly compute the inductive biases](https://www.lesswrong.com/posts/QzpKq92nXqp8NHM34/neural-tangent-kernel-distillation) of a neural network near a particular point in parameter space. The NTK's eigenfunctions give us possible behaviors, and its spectrum tells us how easy it is for the network to learn each eigenfunction.
This paper uses the NTK to compare the architectural inductive biases of convolutional networks to those of multilayer perceptrons. It seems like a promising approach for better understanding what sorts of behaviors different architectures are inclined to learn. However, this paper makes two major simplifying assumptions:
1. It assumes the infinite width limit, thereby ensuring the NTK remains constant through training (and also preventing any feature learning).
2. It assumes that input data is uniformly distributed across a manifold formed as a product of hyperspheres.
This paper's discussion of inductive biases focuses a lot on the frequency biases of neural networks, rather than things more related to alignment. It is very hard to use the NTK (or any mathematical formalism) to talk about inductive biases towards or away from "intentional" / high-level concepts, such as values, deception, corrigibility, etc. However, it is much easier to evaluate how different NTKs bias the network towards learning functions of different frequencies, and so much discussion of NTK inductive biases focuses on frequency.
For a freshly initialized network, the NTK is going to give you a list of functions such as these:
Top row of Figure 1 of [Implicit Regularization via Neural Feature Alignment](https://arxiv.org/abs/2008.00938)and tell you how easily the network can learn each of them. It's easy to rank these functions by frequency, but not so easy to rank them by how much learning them inclines the network towards liking geese.
### [Implicit Regularization via Neural Feature Alignment](https://arxiv.org/abs/2008.00938)
> We approach the problem of implicit regularization in deep learning from a geometrical viewpoint. We highlight a regularization effect induced by a dynamical alignment of the neural tangent features introduced by Jacot et al, along a small number of task-relevant directions. This can be interpreted as a combined mechanism of feature selection and compression. By extrapolating a new analysis of Rademacher complexity bounds for linear models, we motivate and study a heuristic complexity measure that captures this phenomenon, in terms of sequences of tangent kernel classes along optimization paths.
>
>
**My opinion:**
(see below)
### [What can linearized neural networks actually say about generalization?](https://arxiv.org/abs/2106.06770)
> For certain infinitely-wide neural networks, the neural tangent kernel (NTK) theory fully characterizes generalization, but for the networks used in practice, the empirical NTK only provides a rough first-order approximation. Still, a growing body of work keeps leveraging this approximation to successfully analyze important deep learning phenomena and design algorithms for new applications. In our work, we provide strong empirical evidence to determine the practical validity of such approximation by conducting a systematic comparison of the behavior of different neural networks and their linear approximations on different tasks. We show that the linear approximations can indeed rank the learning complexity of certain tasks for neural networks, even when they achieve very different performances. However, in contrast to what was previously reported, we discover that neural networks do not always perform better than their kernel approximations, and reveal that the performance gap heavily depends on architecture, dataset size and training task. We discover that networks overfit to these tasks mostly due to the evolution of their kernel during training, thus, revealing a new type of implicit bias.
>
>
**My opinion:**
A natural question to ask is how to extend approaches like [Eigenspace Restructuring](https://arxiv.org/abs/2112.05611) to track an architecture's inductive biases across an entire trajectory of neural net optimization. Ideally, we'd have a theoretical model of how the NTK and its inductive biases change over the course of training, then "integrate" over the that trajectory to fully account for the functions learned over the training process.
The two papers above do not do this. Instead, they empirically investigate how the NTK changes over the course of network training, and how those changes impact our ability to predict training dynamics and generalization, finding that the NTK adapts over time to align with the labeling of the training data.
While not as useful as a theoretical model of how the NTK changes over training, these empirical results still seem alignment relevant. E.g., they imply that inductive biases can be learned from the training labels, which matches findings that humans become more shape biased as they grow up, and switch between shape or texture bias depending on what they're looking at (e.g., being shape biased for animals, but texture biased for liquids / pastes). See section 1.1 of [this dissertation](https://searchworks.stanford.edu/view/14172606).
Having context-sensitive inductive biases seems very useful if you want to quickly adapt to new information. Using different inductive biases for different (learned) object classes seems ~impossible to encode in an architecture or learning process, so I think it would have to be learned from the training data. Probably, many of the inductive biases of humans and AIs come from complex interactions between architecture, training process and data, and are far outside of the constant NTK limit.
We've also seen a similar result from the other direction: [meta learning](https://arxiv.org/abs/1703.03400v3) uses a two-level optimization setup, where the outer optimizer uses second-order gradients to learn an initialization that the inner optimizer can quickly adapt to downstream tasks. However, [this paper](https://arxiv.org/abs/1909.09157v2) found that the outer optimizer mostly just learns high-performance features directly.
Also, if SGD learns high-performance inductive biases for its training data, that could explain why explicit meta learning / [self modifying](https://arxiv.org/abs/2202.05780) training processes don't seem to outperform simple SGD: gradient descent already self modifies to become better at learning the task at hand.
### [Tuning Frequency Bias in Neural Network Training with Nonuniform Data](https://arxiv.org/abs/2205.14300v2)
> Small generalization errors of over-parameterized neural networks (NNs) can be partially explained by the frequency biasing phenomenon, where gradient-based algorithms minimize the low-frequency misfit before reducing the high-frequency residuals. Using the Neural Tangent Kernel (NTK), one can provide a theoretically rigorous analysis for training where data are drawn from constant or piecewise-constant probability densities. Since most training data sets are not drawn from such distributions, we use the NTK model and a data-dependent quadrature rule to theoretically quantify the frequency biasing of NN training given fully nonuniform data. By replacing the loss function with a carefully selected Sobolev norm, we can further amplify, dampen, counterbalance, or reverse the intrinsic frequency biasing in NN training.
>
>
**My opinion:**
The previous two papers investigate how the NTK evolves while training finite width networks (when [Eigenspace Restructuring](https://arxiv.org/abs/2112.05611)'s assumption 1 is violated). This paper develops methods to apply NTK analysis to situations where data are not uniformly distributed (when [Eigenspace Restructuring](https://arxiv.org/abs/2112.05611)'s assumption 2 is violated). They find they can control the degree of frequency bias through the loss function, which further underscores how tightly intertwined a model's inductive biases are with its training data.
### [On the Activation Function Dependence of the Spectral Bias of Neural Networks](https://arxiv.org/abs/2208.04924)
> Neural networks are universal function approximators which are known to generalize well despite being dramatically overparameterized. We study this phenomenon from the point of view of the spectral bias of neural networks. Our contributions are two-fold. First, we provide a theoretical explanation for the spectral bias of ReLU neural networks by leveraging connections with the theory of finite element methods. Second, based upon this theory we predict that switching the activation function to a piecewise linear B-spline, namely the Hat function, will remove this spectral bias, which we verify empirically in a variety of settings. Our empirical studies also show that neural networks with the Hat activation function are trained significantly faster using stochastic gradient descent and ADAM. Combined with previous work showing that the Hat activation function also improves generalization accuracy on image classification tasks, this indicates that using the Hat activation provides significant advantages over the ReLU on certain problems.
>
>
**My opinion:**
So, it turns out you can just remove the frequency bias of deep networks, and they will still work well for some tasks (or even perform better). I was surprised by this. My impression had been that the generalization capacity of neural networks would be more sensitive to their inductive biases than that.
I really wish the authors had tested their non-frequency-biased networks on a more realistic problem, ideally language modeling on the scale of BERT or larger ([it's not that expensive](https://arxiv.org/abs/2104.07705)!). It'd be interested to see if we could find systematic differences in the generalization behaviors of language models trained with frequency bias versus language models trained without frequency bias.
I also wonder how closely the post-training inductive biases of the two types of models would line up. Can enough data "wash out" the differences in architectural inductive biases?
### [Spectral Bias in Practice: The Role of Function Frequency in Generalization](https://arxiv.org/abs/2110.02424)
> Despite their ability to represent highly expressive functions, deep learning models seem to find simple solutions that generalize surprisingly well. Spectral bias -- the tendency of neural networks to prioritize learning low frequency functions -- is one possible explanation for this phenomenon, but so far spectral bias has primarily been observed in theoretical models and simplified experiments. In this work, we propose methodologies for measuring spectral bias in modern image classification networks on CIFAR-10 and ImageNet. We find that these networks indeed exhibit spectral bias, and that interventions that improve test accuracy on CIFAR-10 tend to produce learned functions that have higher frequencies overall but lower frequencies in the vicinity of examples from each class. This trend holds across variation in training time, model architecture, number of training examples, data augmentation, and self-distillation. We also explore the connections between function frequency and image frequency and find that spectral bias is sensitive to the low frequencies prevalent in natural images. On ImageNet, we find that learned function frequency also varies with internal class diversity, with higher frequencies on more diverse classes. Our work enables measuring and ultimately influencing the spectral behavior of neural networks used for image classification, and is a step towards understanding why deep models generalize well.
>
>
**My opinion:**
This paper describes practical methods for evaluating the frequency sensitivity of neural networks, and how this sensitivity varies across the network's input space. It seems useful for investigating how interventions on network inductive biases impact post-training behaviors (at least, for behaviors related to frequency).
I'd be interested to see how trained networks without an initial frequency bias (see previous paper) compare to those with an initial frequency bias.
### [Limitations of the NTK for Understanding Generalization in Deep Learning](https://arxiv.org/abs/2206.10012)
> The ``Neural Tangent Kernel'' (NTK) (Jacot et al 2018), and its empirical variants have been proposed as a proxy to capture certain behaviors of real neural networks. In this work, we study NTKs through the lens of scaling laws, and demonstrate that they fall short of explaining important aspects of neural network generalization. In particular, we demonstrate realistic settings where finite-width neural networks have significantly better data scaling exponents as compared to their corresponding empirical and infinite NTKs at initialization. This reveals a more fundamental difference between the real networks and NTKs, beyond just a few percentage points of test accuracy. Further, we show that even if the empirical NTK is allowed to be pre-trained on a constant number of samples, the kernel scaling does not catch up to the neural network scaling. Finally, we show that the empirical NTK continues to evolve throughout most of the training, in contrast with prior work which suggests that it stabilizes after a few epochs of training. Altogether, our work establishes concrete limitations of the NTK approach in understanding generalization of real networks on natural datasets.
>
>
**My opinion:**
This paper illustrates an important limitation of using empirical estimations of the NTK at specific points in training to track inductive biases. There are important learning dynamics that only appear in aggregate across many SGD steps, including scaling laws apparently. Presumably, a proper theory of NTK evolution over time would let us predict the actual scaling behavior of architectures.
**Concluding thoughts about NTK-based accounts of inductive biases:**
I think the real bottleneck on using the NTK in alignment is the difficulty of expressing alignment-relevant behaviors (deception, powerseeking, etc) in terms of the inductive biases described by the NTK.
If we can translate from NTK inductive biases to alignment-relevant behaviors, I think we'd be able to use empirically estimated NTKs at points across the network's training trajectory to get useful estimates of how inclined the network is towards learning those behaviors (rather than needing a theoretical understanding of NTK evolution).
In particular, [What can linearized neural networks actually say about generalization?](https://arxiv.org/abs/2106.06770) indicates that the NTK can rank the relative learnability of different tasks, even while providing an overall poor estimation of the network's capabilities. So even noisy estimates from the NTK may suffice to determine whether models end up deceptive or powerseeking.
For translating from NTK inductive biases to alignment-relevant behaviors, I think our best bet is to study the NTKs of pretrained LMs. Probably, their NTKs have been restructured to make semantically meaningful behaviors more learnable. I expect it's easier to relate such inductive biases to the behaviors we're interested in.
E.g., the [Implicit Regularization](https://arxiv.org/abs/2008.00938) paper trained on a toy problem of determining whether points were in a disk of radius √2π.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')}
centered at the origin. Figure 1 shows that the resulting inductive biases align with the task after training:
Bottom row of Figure 1 of [Implicit Regularization via Neural Feature Alignment](https://arxiv.org/abs/2008.00938)Of course, empirically estimating the inductive biases of a language model's NTK after training is going to be very difficult. I've not found any papers which attempt such a feat.
At this point in the roundup, we're moving on from architecture-entangled inductive biases / the NTK, and looking into the inductive biases of SGD itself.
### [Shift-Curvature, SGD, and Generalization](https://arxiv.org/abs/2108.09507v3)
> A longstanding debate surrounds the related hypotheses that low-curvature minima generalize better, and that SGD discourages curvature. We offer a more complete and nuanced view in support of both. First, we show that curvature harms test performance through two new mechanisms, the shift-curvature and bias-curvature, in addition to a known parameter-covariance mechanism. The three curvature-mediated contributions to test performance are reparametrization-invariant although curvature is not. The shift in the shift-curvature is the line connecting train and test local minima, which differ due to dataset sampling or distribution shift. Although the shift is unknown at training time, the shift-curvature can still be mitigated by minimizing overall curvature. Second, we derive a new, explicit SGD steady-state distribution showing that SGD optimizes an effective potential related to but different from train loss, and that SGD noise mediates a trade-off between deep versus low-curvature regions of this effective potential. Third, combining our test performance analysis with the SGD steady state shows that for small SGD noise, the shift-curvature may be the most significant of the three mechanisms. Our experiments confirm the impact of shift-curvature on test loss, and further explore the relationship between SGD noise and curvature.
>
>
**My opinion:**
This paper offers a fairly intuitive explanation for why flatter minima generalize better: suppose the training and testing data have distinct, but nearby, minima that minimize their respective loss. Then, the curvature around the training minima acts as the second order term in a Taylor expansion that approximates the expected test loss for models nearby the training minima.
The paper then investigates the impact of gradient noise from SGD and find that it biases models towards flatter regions of parameter space, even to the point of getting worse training loss.
### [Implicit Gradient Regularization](https://arxiv.org/abs/2009.11162)
> Gradient descent can be surprisingly good at optimizing deep neural networks without overfitting and without explicit regularization. We find that the discrete steps of gradient descent implicitly regularize models by penalizing gradient descent trajectories that have large loss gradients. We call this Implicit Gradient Regularization (IGR) and we use backward error analysis to calculate the size of this regularization. We confirm empirically that implicit gradient regularization biases gradient descent toward flat minima, where test errors are small and solutions are robust to noisy parameter perturbations. Furthermore, we demonstrate that the implicit gradient regularization term can be used as an explicit regularizer, allowing us to control this gradient regularization directly. More broadly, our work indicates that backward error analysis is a useful theoretical approach to the perennial question of how learning rate, model size, and parameter regularization interact to determine the properties of overparameterized models optimized with gradient descent.
>
>
**My opinion:**
Even with full-batch gradient descent (so no gradient noise from SGD), it turns out that gradient descent's discrete steps introduce an inductive bias into network training, and that we can analyze this bias with surprisingly straightforward methods. Like the noise from SGD, this inductive bias also pushes the model towards flatter regions of parameter space.
### [Limiting Dynamics of SGD: Modified Loss, Phase Space Oscillations, and Anomalous Diffusion](http://arxiv-export-lb.library.cornell.edu/abs/2107.09133)
> In this work we explore the limiting dynamics of deep neural networks trained with stochastic gradient descent (SGD). As observed previously, long after performance has converged, networks continue to move through parameter space by a process of anomalous diffusion in which distance travelled grows as a power law in the number of gradient updates with a nontrivial exponent. We reveal an intricate interaction between the hyperparameters of optimization, the structure in the gradient noise, and the Hessian matrix at the end of training that explains this anomalous diffusion. To build this understanding, we first derive a continuous-time model for SGD with finite learning rates and batch sizes as an underdamped Langevin equation. We study this equation in the setting of linear regression, where we can derive exact, analytic expressions for the phase space dynamics of the parameters and their instantaneous velocities from initialization to stationarity. Using the Fokker-Planck equation, we show that the key ingredient driving these dynamics is not the original training loss, but rather the combination of a modified loss, which implicitly regularizes the velocity, and probability currents, which cause oscillations in phase space. We identify qualitative and quantitative predictions of this theory in the dynamics of a ResNet-18 model trained on ImageNet. Through the lens of statistical physics, we uncover a mechanistic origin for the anomalous limiting dynamics of deep neural networks trained with SGD.
>
>
**My opinion:**
(see below)
### [Multiplicative noise and heavy tails in stochastic optimization](https://arxiv.org/abs/2006.06293)
> Although stochastic optimization is central to modern machine learning, the precise mechanisms underlying its success, and in particular, the precise role of the stochasticity, still remain unclear. Modelling stochastic optimization algorithms as discrete random recurrence relations, we show that multiplicative noise, as it commonly arises due to variance in local rates of convergence, results in heavy-tailed stationary behaviour in the parameters. A detailed analysis is conducted for SGD applied to a simple linear regression problem, followed by theoretical results for a much larger class of models (including non-linear and non-convex) and optimizers (including momentum, Adam, and stochastic Newton), demonstrating that our qualitative results hold much more generally. In each case, we describe dependence on key factors, including step size, batch size, and data variability, all of which exhibit similar qualitative behavior to recent empirical results on state-of-the-art neural network models from computer vision and natural language processing. Furthermore, we empirically demonstrate how multiplicative noise and heavy-tailed structure improve capacity for basin hopping and exploration of non-convex loss surfaces, over commonly-considered stochastic dynamics with only additive noise and light-tailed structure.
>
>
**My opinion:**
These papers dive more deeply into the structure and effects of gradient noise, modeling SGD as a diffusion process while making different assumptions about the structure of the gradient noise. They point to a picture where the ratio of learning rate to batch size controls a sort of exploration bias of SGD towards broader basins.
One interesting thing to note about these inductive biases from the optimizer is that the human brain probably has very similar inductive biases. E.g., the inductive bias found in [Implicit Gradient Regularization](https://arxiv.org/abs/2009.11162) happens because SGD does not take optimally sized steps for reducing training loss at each update. It seems very unlikely that brain neurons do make optimal updates to minimize predictive error, which probably leads the brain to also steer towards flatter regions of its parameter space.
Similarly, the brain's optimization process seems pretty noisy (and has batch size one), so the brain probably also mirrors the inductive biases that come from the noise in SGD updates.
### [Towards Theoretically Understanding Why SGD Generalizes Better Than ADAM in Deep Learning](https://arxiv.org/abs/2010.05627)
> It is not clear yet why ADAM-alike adaptive gradient algorithms suffer from worse generalization performance than SGD despite their faster training speed. This work aims to provide understandings on this generalization gap by analyzing their local convergence behaviors. Specifically, we observe the heavy tails of gradient noise in these algorithms. This motivates us to analyze these algorithms through their Levy-driven stochastic differential equations (SDEs) because of the similar convergence behaviors of an algorithm and its SDE. Then we establish the escaping time of these SDEs from a local basin. The result shows that (1) the escaping time of both SGD and ADAM~depends on the Radon measure of the basin positively and the heaviness of gradient noise negatively; (2) for the same basin, SGD enjoys smaller escaping time than ADAM, mainly because (a) the geometry adaptation in ADAM~via adaptively scaling each gradient coordinate well diminishes the anisotropic structure in gradient noise and results in larger Radon measure of a basin; (b) the exponential gradient average in ADAM~smooths its gradient and leads to lighter gradient noise tails than SGD. So SGD is more locally unstable than ADAM~at sharp minima defined as the minima whose local basins have small Radon measure, and can better escape from them to flatter ones with larger Radon measure. As flat minima here which often refer to the minima at flat or asymmetric basins/valleys often generalize better than sharp ones, our result explains the better generalization performance of SGD over ADAM. Finally, experimental results confirm our heavy-tailed gradient noise assumption and theoretical affirmation.
>
>
**My opinion:**
This is a very cool paper. It offers a (pretty plausible, IMO) account of how and why the two optimizers have different inductive biases. I think it's a very good sign that we know enough about gradient descent that we can perform these sorts of analyses.
### [The Low-Rank Simplicity Bias in Deep Networks](https://arxiv.org/abs/2103.10427)
> Modern deep neural networks are highly over-parameterized compared to the data on which they are trained, yet they often generalize remarkably well. A flurry of recent work has asked: why do deep networks not overfit to their training data? In this work, we make a series of empirical observations that investigate and extend the hypothesis that deeper networks are inductively biased to find solutions with lower effective rank embeddings. We conjecture that this bias exists because the volume of functions that maps to low effective rank embedding increases with depth. We show empirically that our claim holds true on finite width linear and non-linear models on practical learning paradigms and show that on natural data, these are often the solutions that generalize well. We then show that the simplicity bias exists at both initialization and after training and is resilient to hyper-parameters and learning methods. We further demonstrate how linear over-parameterization of deep non-linear models can be used to induce low-rank bias, improving generalization performance on CIFAR and ImageNet without changing the modeling capacity.
>
>
**My opinion:**
The core intuition of this paper is that, when you multiply a bunch of matrices together, the rank of the composite operator is no higher than that of the lowest rank component matrix. Models that operate by multiply matrices together are thus biased towards implementing low-rank functions. Of course, adding a residual connection will control this bias, and means information flow is no longer bottlenecked by the lowest rank matrix in the model.
### [On the Implicit Bias Towards Minimal Depth of Deep Neural Networks](https://arxiv.org/abs/2202.09028v9)
> Recent results in the literature suggest that the penultimate (second-to-last) layer representations of neural networks that are trained for classification exhibit a clustering property called neural collapse (NC). We study the implicit bias of stochastic gradient descent (SGD) in favor of low-depth solutions when training deep neural networks. We characterize a notion of effective depth that measures the first layer for which sample embeddings are separable using the nearest-class center classifier. Furthermore, we hypothesize and empirically show that SGD implicitly selects neural networks of small effective depths.
> Secondly, while neural collapse emerges even when generalization should be impossible - we argue that the **degree of separability** in the intermediate layers is related to generalization. We derive a generalization bound based on comparing the effective depth of the network with the minimal depth required to fit the same dataset with partially corrupted labels. Remarkably, this bound provides non-trivial estimations of the test performance. Finally, we empirically show that the effective depth of a trained neural network monotonically increases when increasing the number of random labels in data.
>
>
**My opinion:**
This paper reflects my intuition that gradient descent is biased towards short paths, possibly because longer paths lose rank too quickly?
### [Why neural networks find simple solutions: the many regularizers of geometric complexity](https://arxiv.org/abs/2209.13083)
> In many contexts, simpler models are preferable to more complex models and the control of this model complexity is the goal for many methods in machine learning such as regularization, hyperparameter tuning and architecture design. In deep learning, it has been difficult to understand the underlying mechanisms of complexity control, since many traditional measures are not naturally suitable for deep neural networks. Here we develop the notion of geometric complexity, which is a measure of the variability of the model function, computed using a discrete Dirichlet energy. Using a combination of theoretical arguments and empirical results, we show that many common training heuristics such as parameter norm regularization, spectral norm regularization, flatness regularization, implicit gradient regularization, noise regularization and the choice of parameter initialization all act to control geometric complexity, providing a unifying framework in which to characterize the behavior of deep learning models.
>
>
**My opinion:**
I think that something like geometric simplicity bias is at the core of how neural networks learn general solutions. Neural networks mostly seem to model the [union of low dimensional manifolds](https://deepai.org/publication/the-union-of-manifolds-hypothesis-and-its-implications-for-deep-generative-modelling) on which their input data lie, then sort of extrapolate the geometry of those manifolds to unseen data. Sudden deviations from the manifold geometry would lead to higher geometric complexity. Learning processes biased towards low geometric complexity tend not to have such deviations.
### [The Pitfalls of Simplicity Bias in Neural Networks](https://arxiv.org/abs/2006.07710)
> Several works have proposed Simplicity Bias (SB)---the tendency of standard training procedures such as Stochastic Gradient Descent (SGD) to find simple models---to justify why neural networks generalize well [Arpit et al. 2017, Nakkiran et al. 2019, Soudry et al. 2018]. However, the precise notion of simplicity remains vague. Furthermore, previous settings that use SB to theoretically justify why neural networks generalize well do not simultaneously capture the non-robustness of neural networks---a widely observed phenomenon in practice [Goodfellow et al. 2014, Jo and Bengio 2017]. We attempt to reconcile SB and the superior standard generalization of neural networks with the non-robustness observed in practice by designing datasets that (a) incorporate a precise notion of simplicity, (b) comprise multiple predictive features with varying levels of simplicity, and (c) capture the non-robustness of neural networks trained on real data. Through theory and empirics on these datasets, we make four observations: (i) SB of SGD and variants can be extreme: neural networks can exclusively rely on the simplest feature and remain invariant to all predictive complex features. (ii) The extreme aspect of SB could explain why seemingly benign distribution shifts and small adversarial perturbations significantly degrade model performance. (iii) Contrary to conventional wisdom, SB can also hurt generalization on the same data distribution, as SB persists even when the simplest feature has less predictive power than the more complex features. (iv) Common approaches to improve generalization and robustness---ensembles and adversarial training---can fail in mitigating SB and its pitfalls. Given the role of SB in training neural networks, we hope that the proposed datasets and methods serve as an effective testbed to evaluate novel algorithmic approaches aimed at avoiding the pitfalls of SB.
>
>
**My opinion:**
This paper shows just how strong neural network simplicity biases are, and also gives some intuition for how the simplicity bias of neural networks is different from something like a circuit simplicity bias or Kolmogorov simplicity bias. E.g., neural networks don't seem all that opposed to memorization. The paper shows examples of neural networks learning a simple linear feature which imperfectly classifies the data, then memorizing the remaining noise, *despite there being a slightly more complex feature which perfectly classifies the training data* (and I've checked, there's no grokking phase transition, even after 2.5 million optimization steps with weight decay)*.*
It also shows how, depending on the data you're trying to model, a simplicity bias may actually harm generalization.
Conclusion
----------
My biggest takeaway from this review is that SGD has a *lot* of inductive biases. Even something as simple as the fact that SGD takes discrete, non-optimal update steps leads to systematic bias in the sorts of solutions found. Probably, there are lots of other inductive biases coming from interactions between architecture, data and optimizer.
Also, inductive bias research is making a lot of progress. In particular, the NTK perspective on inductive bias seems to be quickly moving in a potentially valuable direction. If we can reach an okayish understanding of how the NTK evolves over training, and how the inductive biases supplied by the NTK relate to high-level cognitive properties, that might give us something like a [non-closed-form](https://www.lesswrong.com/posts/bxkWd6WdkPqGmdHEk/path-dependence-in-ml-inductive-biases) account of path-dependent inductive biases.
I've also updated towards humans and AIs having similar inductive biases. There are some inductive biases that I think we straight up share with AIs, such as those that come from making [non-optimal](https://arxiv.org/abs/2009.11162) / [noisy](https://arxiv.org/abs/2108.09507v3) parameter updates. I also think that humans have a fair bit of [geometric simplicity](https://arxiv.org/abs/2209.13083) bias, as indicated by the fact that most small perturbations to our visual / auditory inputs do not have very large impacts on how we process those inputs.
I hope readers find these papers useful for their own research. Please feel free to discuss the listed papers in the comments or recommend additional papers to me.
Honorable mentions
------------------
These are interesting papers that are related to inductive biases, but which I decided not to include in the roundup, both because I didn't want to make the post too long, and because I've delayed the post long enough already.
* [Weak and Strong Gradient Directions: Explaining Memorization, Generalization, and Hardness of Examples at Scale](https://arxiv.org/abs/2003.07422)
* [Gradient Starvation: A Learning Proclivity in Neural Networks](https://arxiv.org/abs/2011.09468)
* [SGD on Neural Networks Learns Functions of Increasing Complexity](https://arxiv.org/abs/1905.11604)
* [Let's Agree to Agree: Neural Networks Share Classification Order on Real Datasets](https://arxiv.org/abs/1905.10854)
* [The Grammar-Learning Trajectories of Neural Language Models](https://arxiv.org/abs/2109.06096v3)
* [Learning through atypical ”phase transitions” in overparameterized neural networks](https://arxiv.org/abs/2110.00683)
* [Deep Learning Through the Lens of Example Difficulty](https://proceedings.neurips.cc/paper/2021/hash/5a4b25aaed25c2ee1b74de72dc03c14e-Abstract.html)
* [Towards understanding deep learning with the natural clustering prior](https://arxiv.org/abs/2203.08174)
* [Residual Networks Behave Like Ensembles of Relatively Shallow Networks](https://arxiv.org/abs/1605.06431)
* [The Shattered Gradients Problem: If resnets are the answer, then what is the question?](https://arxiv.org/abs/1702.08591)
Future
------
For next week's roundup, I'm thinking the focus will be on techniques for chain of thought language models.
My other candidate focuses are:
* Shape versus texture bias in neural nets / humans
* Diffusion models
* Controllable text generation
* Structure and content of language model internal representations
Let me know if there are any topics you're particularly interested in. |
f31783f9-a104-44e0-bf95-e8564f8ea9e2 | trentmkelly/LessWrong-43k | LessWrong | I'm Not Saying People Are Stupid
Razib summarized my entire cognitive biases talk at the Singularity Summit 2009 as saying: "Most people are stupid."
Hey! That's a bit unfair. I never said during my talk that most people are stupid. In fact, I was very careful not to say, at any point, that people are stupid, because that's explicitly not what I believe.
I don't think that people who believe in single-world quantum mechanics are stupid. John von Neumann believed in a collapse postulate.
I don't think that philosophers who believe in the "possibility" of zombies are stupid. David Chalmers believes in zombies.
I don't even think that theists are stupid. Robert Aumann believes in Orthodox Judaism.
And in the closing sentence of my talk on cognitive biases and existential risk, I did not say that humanity was devoting more resources to football than existential risk prevention because we were stupid.
There's an old joke that runs as follows:
A motorist is driving past a mental hospital when he gets a flat tire.
He goes out to change the tire, and sees that one of the patients is watching him through the fence.
Nervous, trying to work quickly, he jacks up the car, takes off the wheel, puts the lugnuts into the hubcap -
And steps on the hubcap, sending the lugnuts clattering into a storm drain.
The mental patient is still watching him through the fence.
The motorist desperately looks into the storm drain, but the lugnuts are gone.
The patient is still watching.
The motorist paces back and forth, trying to think of what to do -
And the patient says,
"Take one lugnut off each of the other tires, and you'll have three lugnuts on each."
"That's brilliant!" says the motorist. "What's someone like you doing in an asylum?"
"I'm here because I'm crazy," says the patient, "not because I'm stupid." |
f7652eac-ab97-4471-a2f6-fde8fdb30286 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | The Gallery for Painting Transformations - A GPT-3 Analogy
*[Target audience: Me from a month ago, people who want a sense of what a transformer is doing on a non-technical level, and people who want to chunk their understanding of transformers.]*
Imagine a special art gallery, the Gallery for Painting Transformations (GPT). It takes in a sentence and makes a simple set of paintings to represent it, at first hardly more than stock images. Each morning, a team of artists come in and add to the paintings, and each evening, a snapchat-style filter is applied to each painting individually. Thanks to these efforts, over time the paintings in the gallery grow in meaning and detail, ultimately containing enough information to predict a painting which would be a good addition to the gallery.
This is how GPT-3 works, or at least (I claim) a reasonable analogy for it. In this post I want to flesh out this analogy, with the end result that even non-ML people can get a sense of what large language models are doing. I have tried to be as accurate in how I analogize things as possible, with the notable exception that my examples containing semantic meaning are a lot more human-comprehensible than what GPT-3 is doing, and in reality everything it does looks like “perform a seemingly-random vector operation”. Any mistakes are ~~intentional creative flourishes~~ my own[[1]](#fnkdnxqnyro6h).
**The Paintings**
-----------------
The gallery holds 2048(=.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')}
nctx)[[2]](#fnr4jyuzfhum8) paintings, which represent the model’s context window (the prompt you feed into GPT-3). In GPT-3, each word is represented by a 12288(=dmodel)-dimensional vector. Fortuitously, that’s exactly enough dimensions to make a 64-by-64 pixel RGB image!
The gallery first opens when a user feeds in a prompt, and each token (tokens≈words) in the prompt is turned into a painting which looks like a generic stock image, the “token embedding”. If you feed in more or less than 2048 tokens, the message is trimmed or padded with a special token.
How the gallery would look when it opens if the prompt was “a bear eats apples”, consisting of stock images for each word and positional embeddings in the bottom-left of each painting.T**he gallery itself has no set order,** there are simply paintings that can be viewed in any order. Instead, since word order is sometimes important in sentences, the gallery also paints on a little indicator to denote sentence position, the “position embedding”. In the previous image I showed this as drawing a little number in the corner of each painting, though this can be far more complicated. Since the position embedding is part of the painting, it can also be transformed over time.
At this point the gallery has paintings, but they’re somewhat simplistic. Over the next 96(=nlayers) days, the gallery transforms the paintings until they are rich in meaning, both individually and as a collection. Each day represents a layer of the network, where a new team of artists (attention heads) works during the morning and a new filter (feed-forward network) is applied each evening.
**The Artists**
---------------
The artists in this analogy are the transformer’s attention heads. Each day, a new team of 96(=nheads) artists come in and add a few brush strokes to each the canvasses depending on what they see. Each artist has two rules that decides what they paint:
* **A rule to determine what brush strokes painting X adds to paintings that earn its attention.**
+ This value rule is encoded in the “value” and “output” matrices V and O, each 128-by-12288 (dhead-by-dmodel). Thinking of painting X as a 1-by-12288 vector, X tries to add the vector XVOT to other paintings. If one stacks all the painting vectors into a 2048-by-12288 matrix M, you can compute all outputs at once in the 2048-by-12288 matrix B=MVOT
* **A rule to determine how much attention painting X should give to painting Y.**
+ This attention rule is encoded in the “key” and “query” matrices K and Q, each 128-by-12288 (dhead-by-dmodel). Thinking of paintings X and Y as 1-by-12288 vectors, the “pre-attention” paid from one to the other is XQTKYT.
+ If one stacks all the painting vectors into a 2048-by-12288 matrix M, you can compute all “pre-attentions” at once in the 2048-by-2048 matrix MQ(KT)(MT)=(MQ)(MK)T.
+ “Pre-attentions” can be any number (positive or negative), and we pass each row through a [softmax](https://en.wikipedia.org/wiki/Softmax_function) function so that the actual attention values are between 0 and 1, and each painting gets a total of 1 attention from all the other paintings. Writing σ for the row-wise softmax, the attention matrix is A=σ[(MQ)(MK)T]
* **The artist draws all the brush strokes designated by the first rule, with an opacity in proportion to the attention designated by the second rule.**
+ The overall output for the attention head is AB=σ[(MQ)(MK)T)](MVOT), which is added to the original matrix M.
Keep in mind that since these are vectors, you should think of this paint as “additive” with itself and the original canvas (instead of literally "painting over and rewriting the original value). All artists work simultaneously on all canvases, and the gallery as a whole has been trained so that the artists work in harmony rather than interfering with each other.
Let’s see how this works in practice by following one hypothetical artist. For simplicity, let’s pretend the gallery only contains two paintings (the others could be padding tokens which the artists are told to ignore).
1. In painting X, a person is climbing a tree, and in painting Y a person is eating an apple.
2. Our artist is a “consistency-improving” artist who tries to make all the paintings tie into each other, so the first rule tells them that X should draw a similar-looking tree in the the background, while Y should draw more apples.
3. Next, the artist thinks about where attention should be directed. Using their second rule (including the position embeddings suggesting X happens before Y), they decide the attention on X should come .5/.5 from X and Y, respectively, and the attention on Y should come .99/.01 from X and Y, respectively.
4. Drawing the first-rule images to the strengths determined by the second rule, painting X gets a background tree and some apples in the branches of the foreground tree, while painting Y gets a background tree (and some very faint outlines of ghost apples).
5. Thanks to this artist, the paintings are now more narratively consistent: instead of an unrelated tree-climbing and apple-eating, the paintings show that someone picked apples from an apple orchard, and then ate the apple while still near the tree.
Other artists might perform other tasks: maintaining a good balance of colors, transferring details, art-styles, or themes between canvases, showing cause and effect, or [incomprehensible and seemingly-arbitrary complicated matrix calculation that is somehow essential to the whole network]. Overall, the purpose of the artists is to **carry information between canvases**.
**The Filters**
---------------
After the artists are finished each day, a filter is applied to each painting. **On any particular night the same filter is applied to each painting, but a new filter is used each night.**
An example of a filter the gallery might use. The same “apply dog ears and sepia tint” filter would be applied to all images in the gallery that night.In more technical terms, the filter is a feedforward network consisting of the input painting (12288 width), a single hidden layer (49152=4\*12288 width), and an output layer (12288 width). Writing W1 and W2 for the weight matrices, b1 and b2 for the bias vectors, and α for the activation function (GPT uses GELU), the output of the filter sublayer is given by F=α(XW1+b1)W2+b2. X is size 1-by-12288, W1 is size 12288-by-49152, b1 is size 1-by-49152, W2 is size 49152-by-12288, and b2 is size 1-by-12288.
As with the attention heads, **filters use a “residual connection”**, meaning that the filter calculates F=α(XW1+b1)W2+b2 for each painting X, and returns X+F (not F alone). Two intuitive arguments for why we might want residual connections:
* During training if the network is bad, F will be some noise centered on 0, and therefore the filter will return X+noise, which is much better than returning just noise.
* The filter gets to focus on “how do I improve this painting?” rather than “how do I make a new good painting from scratch?”
What can these filters do? In principal, a great deal - [by the universal approximation theorem](https://en.wikipedia.org/wiki/Universal_approximation_theorem), **they could learn any function on the input if there is enough width in their hidden layer**, and this hidden layer is pretty wide. Their only limitations are that the same filter is applied to every painting at once, independently of the other paintings. I imagine them as generic “improvements” to the paintings - upscaling, integrating the additions of the previous day’s artists, and of course [incomprehensible and seemingly-arbitrary complicated matrix calculation that is somehow essential to the whole network]. Overall, the purpose of the filters are to **refine and evolve the paintings in isolation from each other**.
**Conclusion**
--------------
And that’s what GPT does! To summarize:
* Each token in the context is embedded as a painting in the gallery.
* The paintings start out crude but over time accumulate more “meaning” and “detail”, due to alternating teams of artists (attention heads) and filters (feed-forward networks).
* The artists transfer information between paintings based on the attention that each painting gives to the others.
* The filters evolve paintings in isolation from each other.
* Both artists and filters use “residual connections”, meaning that their outputs are changes layered on top of the existing painting, rather than making a new painting from scratch.
* The artists and filters work in highly-trained coordination to build on top of each other’s work. Together they form an assembly line, where each team counts on the prior teams and is counted on in turn.
There are a few wrinkles I left out, which I’ll touch on here:
* Training the artists and filters is very expensive (there are a huge number of them) and uses various tricks, but is not fundamentally different than training any other neural network.
* I omitted that just before applying the softmax function in the step where you compute attention, you divide the matrix by 1/√dhead. This doesn’t change the qualitative behavior of this step, but improves training speed and other things.
* There are also layer-norms applied after the artists and after the filters work. I don’t have a great grasp of layer norms yet, but I believe they’re something like another filter.
* In other transformers such as the original [Attention is All You Need transformer](https://arxiv.org/pdf/1706.03762.pdf), there are also encoder-decoder attention. To the extent I understand them, this would be like if, after day 48, a copy was made of the gallery. On days 49-96, in addition to the artists and filters, there were teams of art historians that use the day 48 state of the gallery to determine attention and what to draw on the current paintings.
* The final state of the gallery is used to predict which token should come next.
1. **[^](#fnrefkdnxqnyro6h)**I believe everything I say here follows directly from these sources:
[The original Transformer paper, Attention is All You Need](https://arxiv.org/pdf/1706.03762.pdf)
[The GPT paper, Improving Language Understanding by Generative Pre-Training](https://s3-us-west-2.amazonaws.com/openai-assets/research-covers/language-unsupervised/language_understanding_paper.pdf)
[The GPT-3 paper, Language Models are Few-Shot Learners](https://arxiv.org/pdf/2005.14165.pdf)
2. **[^](#fnrefr4jyuzfhum8)**Where possible, I will include the parameter name from the GPT-3 paper in addition to the number. |
00ee5ba5-25fd-46f8-a2b4-517471495c27 | trentmkelly/LessWrong-43k | LessWrong | The second bitter lesson — there’s a fundamental problem with aligning distributed AI
Note: this the first part of an essay on my substack, check out the full essay to see the solutions I put forward
When Richard Sutton introduced the bitter lesson for AI in 2019, he broke the myth that great human ingenuity is needed to create intelligent machines. All it seems to take is a lot of computing power and algorithms that scale easily, rather than clever techniques designed with deep understanding. The two major classes of algorithms that fit this are search and learning; when they are scaled up, advanced AI systems naturally emerge. Sutton’s key insight can be summarised in his final paragraph:
> [a] general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, […] instead we should build in only the meta-methods [search and learning] that can find and capture this arbitrary complexity. […] We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.
A concrete example of this can be seen in the development of chess algorithms. Early chess programs, like those in the 1970s and 1980s, relied heavily on human-crafted heuristics — rules and strategies devised by experts to mimic human understanding of the game. These systems could play decently but were limited by the ingenuity and foresight of their human designers. In contrast, modern chess engines like AlphaZero, developed by DeepMind, only rely on search and learning.
With the rise of LLMs, we’re now seeing this play out again in the domain of general intelligence. After researchers at Google discovered the transformer in 2017 — a robust and scalable artificial neural network architecture — training it on all the text on the internet was enough to get the first AIs with the capability to pass simple versions of the Turing test. So |
47c03f65-7cae-40b4-afc4-b8b3d47c06b0 | trentmkelly/LessWrong-43k | LessWrong | Speed arguments against scheming (Section 4.4-4.7 of “Scheming AIs")
This is Section 4.4-4.7 of my report “Scheming AIs: Will AIs fake alignment during training in order to get power?”. There’s also a summary of the full report here (audio here). The summary covers most of the main points and technical terms, and I’m hoping that it will provide much of the context necessary to understand individual sections of the report on their own.
Audio version of this section here, or search for "Joe Carlsmith Audio" on your podcast app.
Speed arguments
Let's turn, now, to arguments against schemers that focus on the final properties of the models in question. Here, the main argument I'm aware of is that schemers need to do extra reasoning that other model classes don't. In particular, I'll distinguish between three sorts of reasoning:
1. Deciding to training-game*:* The model needs to reason backwards from "I want this beyond-episode thing" to "I should do well in training," in a way that draws on the model's predictions about what doing well in training will entail (e.g., re: goal-guarding, take-over prospects, and so on).
2. Checking whether it's time to escape/take-over: The model needs to checking, with adequate frequency and accuracy, whether it's in a position to escape/take-over (and to succeed in this assessment despite human efforts to trap/fool the model into revealing its true colors – for example, by offering it apparent opportunities to defect).[1]
3. (Optional, but relevant to how scary the schemer is) Early undermining: To the extent the model is doing anything (other than training-gaming) to support a take-over/escape prior to escaping human control – for example, inserting backdoors into the code it writes, actively undermining human alignment research, sandbagging, and so on – it will need to be using cognitive resources both in (a) performing the relevant nefarious actions and (b) making sure it gets away with them (e.g., checking that it won't get caught, coming up with excuses, making the cover stories consistent e |
172ffacc-32a4-4b7d-98d2-2123baea0304 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "I've been reading the hardcover SSC collection in the mornings, as a way of avoiding getting caught up in internet distractions first thing when I get up. I'd read many of Scott Alexander's posts before, but nowhere near everything posted; and I hadn't before made any attempt to dive the archives to "catch up" to the seeming majority of rationalists who have read everything Scott Alexander has ever written.(The hardcover SSC collection is nowhere near everything on SSC, not to mention Scott's earlier squid314 blog on livejournal. I'm curious how much shelf space a more complete anthology would occupy.)Anyway, this has gotten me thinking about the character of Scott Alexander's writing. I once remarked (at a LessWrong meetup) that Scott Alexander "could never be a cult leader". I intended this as a sort of criticism. Scott Alexander doesn't write with conviction in the same way some other prominent rationalist authors do. He usually has the attitude of a bemused bystander who is merely curious about a bunch of things. Some others in the group agreed with me, but took it as praise: compared to some other rationalist authors, Scott Alexander isn't an ideologue.(now I fear 90% of the comments are going to be some variation of "cults are bad")What I didn't realize (at the time) was how obsessed Scott Alexander himself is with this distinction. Many of his posts grapple with variations on question of just how seriously we can take our ideas without going insane, contrasting the holy madman in the desert (who takes ideas 100% seriously) with the detached academic (who takes an intellectual interest in philosophy without applying it to life).Beware Isolated Demands for Rigor is the post which introduces and seriously fleshes out this distinction. Scott says the holy madman and the detached academic are two valid extremes, because both of them are consistent in how they call for principles to be applied (the first always applies their intellectual standards to everything; the second never does). What's invalid is when you use intellectual standards as a tool to get whatever you want, by applying the standards selectively.Infinite Debt forges a middle path, praising Giving What We Can for telling people that you can just give 10% to charity and be an "Officially Recognized Good Person" -- you don't need to follow your principles all the way to giving away everything, or alternately, ignore your principles entirely. By following a simple collectively-chosen rule, you can avoid applying principles selectively in a self-serving (or overly not-self-serving) way.Bottomless Pits Of Suffering talks about the cases where utilitarianism becomes uncomfortable and it's tempting to ignore it.But related ideas are in many other posts. It's a thread which runs throughout Scott's writing. (IMHO.)This conflict is central to the human condition, or at least the WASP/WEIRD condition. I imagine most of Scott's readers felt similar conflicts around applying their philosophies in practice.But this is really weird from a decision-theoretic perspective. An agent should be unsure of principles, not sure of principles but unsure about applying them. (Related.)It's almost like Scott implicitly believes maximizing his own values would be bad somehow.Some of this makes sense from a Goodhart perspective. Any values you explicitly articulate are probably not your values. But I don't get the sense that this is what's going on in Scott's writing. For example, when he describes altruists selling all their worldly possessions, it doesn't sound like he intends it as an example of Goodhart; it sounds like he intends it as a legit example of altruists maximizing altruist values.In contrast, blogs like Minding our way to the heavens give me more of a sense of pushing the envelope on everything; I associate it with ideas like:If you aren't putting forth your full effort, it probably means this isn't your priority. Figure out whether it's worth doing at all, and if so, what the minimal level of effort to get what you want is. (Or, if it really is important, figure out what's stopping you from giving it your full effort.) You can always put forth your full effort at the meta-level of figuring out how much effort to put into which things.If you repeatedly don't do things in line with your "values", you're probably wrong about what your values are; figure out what values you really care about, so that you can figure out how best to optimize those.If you find that you're fighting yourself, figure out what the fight is about, and find a way to best satisfy the values that are in conflict.In more SSC-like terms, it's like, if you're not a holy madman, you're not trying.I'm not really pushing a particular side, here, I just think the dichotomy is interesting."
A user from LessWrong is responsible for the text you see above
Could you elaborate on the dichotomy between the "holy madman" and the "detached academic" in the context of decision theory? Specifically, how does each approach affect the application of principles and the potential for self-deception?
The dichotomy between the "holy madman" and the "detached academic" represents two extreme ways of approaching decision-making and the application of principles.
The "holy madman" represents an individual who applies their principles consistently and rigorously to every aspect of their life. This individual takes their ideas extremely seriously and attempts to live by them in all situations. From a decision theory perspective, this approach minimizes self-deception by ensuring consistency and transparency in decision-making. However, it may also lead to extreme behaviors or decisions, as the individual is constantly striving to live up to their ideals, regardless of the potential consequences.
On the other hand, the "detached academic" represents an individual who takes an intellectual interest in principles and ideas but does not necessarily apply them to their life. This individual views principles as interesting concepts to be studied and understood, but not necessarily as guides for action. From a decision theory perspective, this approach allows for flexibility and adaptability in decision-making. However, it may also lead to self-deception, as the individual may use intellectualization as a defense mechanism to avoid confronting difficult or uncomfortable truths.
In terms of decision theory, both approaches have their strengths and weaknesses. The "holy madman" approach ensures consistency and transparency but may lead to extreme or irrational decisions. The "detached academic" approach allows for flexibility and adaptability but may lead to self-deception or avoidance of difficult truths.
In the end, the best approach may be somewhere in the middle: applying principles consistently but not dogmatically, and maintaining an intellectual curiosity about ideas without using it as a shield against reality. This balance allows for thoughtful, principled decision-making without the risk of self-deception or extreme behavior.
Considering the strengths and weaknesses of both the "holy madman" and "detached academic" approaches, how might one design a decision-making model that effectively balances these two extremes? Discuss the potential challenges and benefits of this model in various contexts, such as personal life decisions, professional choices, and ethical dilemmas.
Designing a decision-making model that effectively balances the "holy madman" and "detached academic" approaches would require an integration of consistent principle application and intellectual flexibility. This hybrid model, let's call it the "principled pragmatist" model, would entail a few key components:
1. **Principle Identification**: Similar to the "holy madman", it's important to have a clear understanding of one's core principles and values. These principles serve as the foundation for decision-making.
2. **Contextual Analysis**: Like the "detached academic", one must consider the specific context of each decision. This includes understanding the potential consequences, the stakeholders involved, and the practical implications of each choice.
3. **Principle Application**: Apply the identified principles to the decision at hand, but do so in a way that considers the specific context. This might mean making a decision that aligns with one's principles but also takes into account practical considerations.
4. **Reflective Adjustment**: After a decision is made, reflect on the outcome. If the decision led to undesirable results, consider whether the principles need to be adjusted or if the application of principles needs to be more flexible.
The potential challenges of this model include the difficulty of identifying and articulating one's principles, the complexity of considering all aspects of a decision's context, and the potential for cognitive dissonance when principles and practical considerations conflict.
However, the benefits are significant. In personal life decisions, this model allows for choices that align with one's values while also considering practical implications. In professional choices, it encourages ethical decision-making that also takes into account the realities of the business world. In ethical dilemmas, it provides a framework for making decisions that are both morally sound and pragmatically feasible.
The "principled pragmatist" model, therefore, offers a balanced approach to decision-making that integrates the strengths of both the "holy madman" and "detached academic" while mitigating their weaknesses. |
38592277-0b70-446d-a49b-155aee87bb24 | trentmkelly/LessWrong-43k | LessWrong | Just Imitate Humans?
Do people think we could make a singleton (or achieve global coordination and preventative policing) just by imitating human policies on computers? If so, this seems pretty safe to me.
Some reasons for optimism: 1) these could be run much faster than a human thinks, and 2) we could make very many of them.
Acquiring data: put a group of people in a house with a computer. Show them things (images, videos, audio files, etc.) and give them a chance to respond at the keyboard. Their keyboard actions are the actions, and everything between actions is an observation. Then learn the policy of the group of humans. By the way, these can be happy humans who earnestly try to follow instructions. To model their policy, we can take the maximum a posteriori estimate over a set of policies which includes the truth, and freeze the policy once we're satisfied. (This is with unlimited computation; we'd have to use heuristics and approximations in real life). With a maximum a posteriori estimate, this will be quick to run once we freeze the policy, and we're no longer tracking tons of hypotheses, especially if we used some sort of speed prior. Let T be the number of interaction cycles we record before freezing the policy. For sufficiently large T, it seems to me that running this is safe.
What are people's intuitions here? Could enough human-imitating artificial agents (running much faster than people) prevent unfriendly AGI from being made?
If we think this would work, there would still be the (neither trivial nor hopeless) challenge of convincing all serious AGI labs that any attempt to run a superhuman AGI is unconscionably dangerous, and we should stick to imitating humans. |
d004045f-be27-4492-b383-aeafae4f0c44 | trentmkelly/LessWrong-43k | LessWrong | Perfectly Friendly AI
Inspired by Don't Plan For the Future.
For the purposes of discussion on this site, a Friendly AI is assumed to be one that shares our terminal values. It's a safe genie that doesn't need to be told what to do, but anticipates how to best serve the interests of its creators. Since our terminal values are a function of our evolutionary history, it seems reasonable to assume that an FAI created by one intelligent species would not necessarily be friendly to other intelligent species, and that being subsumed by another species' FAI would be fairly catastrophic.
Except.... doesn't that seem kind of bad? Supposing I were able to create a strong AI, and it created a sound fun-theoretic utopia for human beings, but then proceeded to expand and subsume extraterrestrial intelligences, and subject them to something they considered a fate worse than death, I would have to regard that as a major failing of my design. My utility function assigns value to the desires of beings whose values conflict with my own. I can't allow other values to supersede mine, but absent other considerations, I have to assign negative utility in my own function for creating negative utility in the functions of other existing beings. I'm skeptical that an AI that would impose catastrophe on other thinking beings is really maximizing my utility.
It seems to me that to truly maximize my utility, an AI would need to have consideration for the utility of other beings. Secondary consideration, perhaps, but it could not maximize my utility simply by treating them as raw material with which to tile the universe with my utopian civilization.
Perhaps my utility function gives more value than most to beings that don't share my values (full disclosure, I prefer the "false" ending of Three Worlds Collide, although I don't consider it ideal.) However, if an AI imposes truly catastrophic fates on other intelligent beings, my own utility function takes such a hit that I cannot consider it friendly. A true Friend |
7c9f1189-1cfb-4629-abc1-da5cb411f07b | trentmkelly/LessWrong-43k | LessWrong | An algorithm with preferences: from zero to one variable
A putative new idea for AI control; index here.
A simple way of thinking that I feel clarifies a lot of issues (related to Blue Minimising Robot):
Suppose you have an entity H that follows algorithm alH. Then define:
* What H does is its actions/outputs in the environment.
* What H is is alH.
* What H wants is an interpretation of what H does (and possibly what it is), in order to construct a utility function or reward function corresponding with its preferences.
The interpretation part of wants is crucial, but it is often obscured in practice in value learning. That's because we often start with things like `H is a boundedly rational agent that maximises u...', or we lay out the agent in such a way that that's clearly the case.
What we're doing there is writing the entity as alH(u) --- an algorithm with a special variable u that tracks what the entity wants. In the case of cooperative inverse reinforcement learning, this is explicit, as the human's values are given by a θ, known to the human. Thus the human's true algorithm is alH(⋅), the human observes θ, meaning that θ is an objective fact about the universe. And then the human follows alH(θ).
Note here that knowing what the human is in the one-variable sense (i.e. knowing alH(⋅)) helps with the correct deduction about what they want - while simply knowing the joint alH(θ) does not.
In contrast an interpretation starts with a zero-variable algorithm, and attempts to construct a one-variable one. There for, given alH it constructs (one or more) alHi(⋅) and ui such that
* alH=alHi(ui).
This illustrates the crucial role of interpretation, especially if alH is highly complex. |
8a813587-e51a-435e-b719-1cbc15cb4e35 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Tel Aviv Meetup: Social & Board Games
Discussion article for the meetup : Tel Aviv Meetup: Social & Board Games
WHEN: 12 May 2015 07:00:00PM (+0300)
WHERE: Google, Electra Tower, 98 Yigal Alon Street, Tel Aviv
This time we're going to have a social meetup! It's going to be a game night full of people talking about physics, friendly AI, and how to effectively save the world. Please bring any games you'd like to play.
The Israeli LessWrong community meets every two weeks, alternating between lectures and social/gaming nights.
Meet at the 29th floor (not the Google Campus floor). We'll then move to a room.
Contact: If you can't find us, call Anatoly, who is graciously hosting us, at 054-245-1060; or Joshua at 054-569-1165.
Discussion article for the meetup : Tel Aviv Meetup: Social & Board Games |
9a43321b-4ea4-443b-8419-9ad5d4236e05 | trentmkelly/LessWrong-43k | LessWrong | The Trolley Problem: Dodging moral questions
The trolley problem is one of the more famous thought experiments in moral philosophy, and studies by psychologists and anthropologists suggest that the response distributions to its major permutations remain roughly the same throughout all human cultures. Most people will permit pulling the lever to redirect the trolley so that it will kill one person rather than five, but will balk at pushing one fat person in front of the trolley to save the five if that is the only available option of stopping it.
However, in informal settings, where the dilemma is posed by a peer rather than a teacher or researcher, it has been my observation that there is another major category which accounts for a significant proportion of respondents' answers. Rather than choosing to flip the switch, push the fat man, or remain passive, many people will reject the question outright. They will attack the improbability of the premise, attempt to invent third options, or appeal to their emotional state in the provided scenario ("I would be too panicked to do anything",) or some combination of the above, in order to opt out of answering the question on its own terms.
However, in most cases, these excuses are not their true rejection. Those who tried to find third options or appeal to their emotional state will continue to reject the dilemma even when it is posed in its most inconvenient possible forms, where they have the time to collect themselves and make a reasoned choice, but no possibility of implementing alternative solutions.
Those who appealed to the unlikelihood of the scenario might appear to have the stronger objection; after all, the trolley dilemma is extremely improbable, and more inconvenient permutations of the problem might appear even less probable. However, trolleylike dilemmas are actually quite common in real life, when you take the scenario not as a case where only two options are available, but as a metaphor for any situation where all the available choices have negat |
6a26a5db-c084-4912-a8ff-122045d4ed7f | trentmkelly/LessWrong-43k | LessWrong | Intelligence without Consciousness
I sleep-talk. In fact, I don't just talk in my sleep, I can carry on a conversation. Advice and insight I have is much lower quality than when I'm awake, but I'm mostly coherent. Once someone recorded a conversation with me, and it was pretty funny. For example: "A lot of people like Batman. We should buy stock in Batman".
I'm not conscious, though. I'm not self-aware. If I have qualia, it's a very tiny amount of qualia. I definitely wouldn't wonder "do I have qualia?" If I didn't keep waking up, there wouldn't be a moral reason for others not to treat me how they find most convenient.
What does this mean? It means that Nonsentient Optimizers and Nonperson Predicates are definitely possible. There is some threshold of intelligence and/or self-awareness needed to change a neural net into a person.
(Lucid dreaming is a separate issue. I'm talking about something that happens even when not lucid dreaming.)
Also, people in the comments keep saying maybe I did have consciousness and just don't remember it. No: the conversation quality is sort of like this. When people talk to me, my response is usually connected to what they said, but I'm definitely not conscious. |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.