id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
207780f8-6a80-450c-b93c-83b03e22dbb0 | trentmkelly/LessWrong-43k | LessWrong | Motivating Abstraction-First Decision Theory
Let’s start with a prototypical embedded decision theory problem: we run two instances of the same “agent” from the same source code and in the same environment, and the outcome depends on the choices of both agents; both agents have full knowledge of the setup.
Roughly speaking, Functional decision theory and its predecessors argue that each instance of the agent should act as if it’s choosing for both.
I suspect there’s an entirely different path to a similar conclusion - a path which is motivated not prescriptively, but descriptively.
When we say that some executing code is an “agent”, that’s an abstraction. We’re abstracting away the low-level model structure (conditionals, loops, function calls, arithmetic, data structures, etc) into a high-level model. The high-level model says “this variable is set to a value which maximizes <blah>”, without saying how that value is calculated. It’s an agent abstraction: an abstraction in which the high-level model is some kind of maximizer.
Key question: when is this abstraction valid?
Abstraction = Information at a Distance talks about what it means for an abstraction to be valid. At the lowest level, validity means that the low-level and high-level models return the same answers to some class of queries. But the linked post reduces this to a simpler definition, more directly applicable to the sort of abstractions we use in real life: variables in a high-level model should contain all the information in corresponding low-level variables which is relevant “far away”. We abstract far-away stars to point masses, because the exact distribution of the mass and momentum within the star is (usually) not relevant from far away.
Or, to put it differently: an abstraction is “valid” when components of the low-level model are independent given the values of high-level variables. The roiling of plasmas within far-apart stars is (approximately) independent given the total mass, momentum, and center-of-mass position of each star. As |
bfa2e2af-dcd1-45bf-9390-e251b66f7850 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization
1 Introduction
---------------
Reinforcement learning can be used to acquire complex behaviors from high-level specifications.
However, defining a cost function that can be optimized effectively and encodes the correct task can be challenging in practice, and techniques like cost shaping are often used to solve complex real-world problems (Ng et al., [1999](#bib.bib18)).
Inverse optimal control (IOC) or inverse reinforcement learning (IRL) provide an avenue for addressing this challenge by learning a cost function directly from expert demonstrations, e.g. Ng et al. ([2000](#bib.bib19)); Abbeel & Ng ([2004](#bib.bib1)); Ziebart et al. ([2008](#bib.bib31)).
However, designing an effective IOC algorithm for learning from demonstration is difficult for two reasons. First, IOC is fundamentally underdefined in that many costs induce the same behavior.
Most practical algorithms therefore require carefully designed features to impose structure on the learned cost. Second, many standard IRL and IOC methods require solving the forward problem (finding an optimal policy given the current cost) in the inner loop of an iterative cost optimization.
This makes them difficult to apply to complex, high-dimensional systems, where the forward problem is itself exceedingly difficult, particularly real-world robotic systems with unknown dynamics.
DdemoDsampDtrajcθcθθ←argminθ LIOC
Figure 1: Right: Guided cost learning uses policy optimization to adaptively sample trajectories for estimating the IOC partition function. Bottom left: PR2 learning to gently place a dish in a plate rack.
To address the challenge of representation, we propose to use expressive, nonlinear function approximators, such as neural networks, to represent the cost. This reduces the engineering burden required to deploy IOC methods, and makes it possible to learn cost functions for which expert intuition is insufficient for designing good features. Such expressive function approximators, however, can learn complex cost functions that lack the structure typically imposed by hand-designed features. To mitigate this challenge, we propose two regularization techniques for IOC, one which is general and one which is specific to episodic domains, such as to robotic manipulation skills.
In order to learn cost functions for real-world robotic tasks, our method must be able to handle unknown dynamics and high-dimensional systems. To that end, we propose a cost learning algorithm based on policy optimization with local linear models, building on prior work in reinforcement learning (Levine & Abbeel, [2014](#bib.bib12)). In this approach, as illustrated in Figure [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), the cost function is learned in the inner loop of a policy search procedure, using samples collected for policy improvement to also update the cost function. The cost learning method itself is a nonlinear generalization of maximum entropy IOC (Ziebart et al., [2008](#bib.bib31)), with samples used to approximate the partition function. In contrast to previous work that optimizes the policy in the inner loop of cost learning, our approach instead updates the cost in the inner loop of policy search, making it practical and efficient. One of the benefits of this approach is that we can couple learning the cost with learning the policy for that cost. For tasks that are too complex to acquire a good global cost function from a small number of demonstrations, our method can still recover effective behaviors by running our policy learning method and retaining the learned policy. We elaborate on this further in Section [4.4](#S4.SS4 "4.4 Learning Costs and Controllers ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
The main contribution of our work is an algorithm that learns nonlinear cost functions from user demonstrations, at the same time as learning a policy to perform the task. Since the policy optimization “guides” the cost toward good regions of the space, we call this method guided cost learning. Unlike prior methods, our algorithm can handle complex, nonlinear cost function representations and high-dimensional unknown dynamics, and can be used on real physical systems with a modest number of samples. Our evaluation demonstrates the performance of our method on a set of simulated benchmark tasks, showing that it outperforms previous methods. We also evaluate our method on two real-world tasks learned directly from human demonstrations. These tasks require using torque control and vision to perform a variety of robotic manipulation behaviors, without any hand-specified cost features.
2 Related Work
---------------
One of the basic challenges in inverse optimal control (IOC), also known as inverse reinforcement learning (IRL), is that finding a cost or reward under which a set of demonstrations is near-optimal is underdefined. Many different costs can explain a given set of demonstrations. Prior work has tackled this issue using maximum margin formulations (Abbeel & Ng, [2004](#bib.bib1); Ratliff et al., [2006](#bib.bib22)), as well as probabilistic models that explain suboptimal behavior as noise (Ramachandran & Amir, [2007](#bib.bib21); Ziebart et al., [2008](#bib.bib31)). We take the latter approach in this work, building on the maximum entropy IOC model (Ziebart, [2010](#bib.bib30)). Although the probabilistic model mitigates some of the difficulties with IOC, there is still a great deal of ambiguity, and an important component of many prior methods is the inclusion of detailed features created using domain knowledge, which can be linearly combined into a cost, including:
indicators for successful completion of the task for a robotic ball-in-cup game (Boularias et al., [2011](#bib.bib5)),
learning table tennis with features that include distance of the ball to the opponent’s elbow (Muelling et al., [2014](#bib.bib17)),
providing the goal position as a known constraint for robotic grasping (Doerr et al., [2015](#bib.bib7)),
and learning highway driving with indicators for collision and driving on the grass (Abbeel & Ng, [2004](#bib.bib1)). While these features allow for the user to impose structure on the cost, they substantially increase the engineering burden. Several methods have proposed to learn nonlinear costs using Gaussian processes (Levine et al., [2011](#bib.bib14)) and boosting (Ratliff et al., [2007](#bib.bib23), [2009](#bib.bib24)), but even these methods generally operate on features rather than raw states. We instead use rich, expressive function approximators, in the form of neural networks, to learn cost functions directly on raw state representations. While neural network cost representations have previously been proposed in the literature (Wulfmeier et al., [2015](#bib.bib29)), they have only been applied to small, synthetic domains. Previous work has also suggested simple regularization methods for cost functions, based on minimizing ℓ1 or ℓ2 norms of the parameter vector (Ziebart, [2010](#bib.bib30); Kalakrishnan et al., [2013](#bib.bib11)) or by using unlabeled trajectories (Audiffren et al., [2015](#bib.bib3)). When using expressive function approximators in complex real-world tasks, we must design substantially more powerful regularization techniques to mitigate the underspecified nature of the problem, which we introduce in Section [5](#S5 "5 Representation and Regularization ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
Another challenge in IOC is that, in order to determine the quality of a given cost function, we must solve some variant of the forward control problem to obtain the corresponding policy, and then compare this policy to the demonstrated actions.
Most early IRL algorithms required solving an MDP in the inner loop of an iterative optimization (Ng et al., [2000](#bib.bib19); Abbeel & Ng, [2004](#bib.bib1); Ziebart et al., [2008](#bib.bib31)).
This requires perfect knowledge of the system dynamics and access to an efficient offline solver, neither of which is available in, for instance, complex robotic control domains.
Several works have proposed to relax this requirement, for example by learning a value function instead of a cost (Todorov, [2006](#bib.bib26)), solving an approximate local control problem (Levine & Koltun, [2012](#bib.bib13); Dragan & Srinivasa, [2012](#bib.bib8)), generating a discrete graph of states (Byravan et al., [2015](#bib.bib6); Monfort et al., [2015](#bib.bib16)), or only obtaining an optimal trajectory rather than policy (Ratliff et al., [2006](#bib.bib22), [2009](#bib.bib24)).
However, these methods still require knowledge of the system dynamics.
Given the size and complexity of the problems addressed in this work, solving the optimal control problem even approximately in the inner loop of the cost optimization is impractical. We show that good cost functions can be learned by instead learning the *cost* in the inner loop of a *policy optimization*. Our inverse optimal control algorithm is most closely related to other previous sample-based methods based on the principle of maximum entropy, including relative entropy IRL (Boularias et al., [2011](#bib.bib5)) and path integral IRL (Kalakrishnan et al., [2013](#bib.bib11)), which can also handle unknown dynamics. However, unlike these prior methods, we adapt the sampling distribution using policy optimization. We demonstrate in our experiments that this adaptation is crucial for obtaining good results on complex robotic platforms, particularly when using complex, nonlinear cost functions.
To summarize, our proposed method is the first to combine several desirable features into a single, effective algorithm: it can handle unknown dynamics, which is crucial for real-world robotic tasks, it can deal with high-dimensional, complex systems, as in the case of real torque-controlled robotic arms, and it can learn complex, expressive cost functions, such as multilayer neural networks, which removes the requirement for meticulous hand-engineering of cost features. While some prior methods have shown good results with unknown dynamics on real robots (Boularias et al., [2011](#bib.bib5); Kalakrishnan et al., [2013](#bib.bib11)) and some have proposed using nonlinear cost functions (Ratliff et al., [2006](#bib.bib22), [2009](#bib.bib24); Levine et al., [2011](#bib.bib14)), to our knowledge no prior method has been demonstrated that can provide all of these benefits in the context of complex real-world tasks.
3 Preliminaries and Overview
-----------------------------
We build on the probabilistic maximum entropy inverse optimal control framework (Ziebart et al., [2008](#bib.bib31)). The demonstrated behavior is assumed to be the result of an expert acting stochastically and near-optimally with respect to an unknown cost function. Specifically, the model assumes that the expert samples the demonstrated trajectories {τi} from the distribution
| | | | |
| --- | --- | --- | --- |
| | p(τ)=1Zexp(−cθ(τ)), | | (1) |
where τ={x1,u1,…,xT,uT} is a trajectory sample, cθ(τ)=∑tcθ(xt,ut) is an unknown cost function parameterized by θ, and xt and ut are the state and action at time step t. Under this model, the expert is most likely to act optimally, and can generate suboptimal trajectories with a probability that decreases exponentially as the trajectories become more costly. The partition function Z is difficult to compute for large or continuous domains, and presents the main computational challenge in maximum entropy IOC. The first applications of this model computed Z exactly with dynamic programming (Ziebart et al., [2008](#bib.bib31)). However, this is only practical in small, discrete domains. More recent methods have proposed to estimate Z by using the Laplace approximation (Levine & Koltun, [2012](#bib.bib13)), value function approximation (Huang & Kitani, [2014](#bib.bib10)), and samples (Boularias et al., [2011](#bib.bib5)). As discussed in Section [4.1](#S4.SS1 "4.1 Sample-Based Inverse Optimal Control ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), we take the sample-based approach in this work, because it allows us to perform inverse optimal control without a known model of the system dynamics. This is especially important in robotic manipulation domains, where the robot might interact with a variety of objects with unknown physical properties.
To represent the cost function cθ(xt,ut), IOC or IRL methods typically use a linear combination of hand-crafted features, given by cθ(ut,ut)=θTf(ut,xt) (Abbeel & Ng, [2004](#bib.bib1)). This representation is difficult to apply to more complex domains, especially when the cost must be computed from raw sensory input. In this work, we explore the use of high-dimensional, expressive function approximators for representing cθ(xt,ut). As we discuss in Section [6](#S6 "6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), we use neural networks that operate directly on the robot’s state, though other parameterizations could also be used with our method. Complex representations are generally considered to be poorly suited for IOC, since learning costs that associate the right element of the state with the goal of the task is already quite difficult even with simple linear representations. However, as we discuss in our evaluation, we found that such representations could be learned effectively by adaptively generating samples as part of a policy optimization procedure, as discussed in Section [4.2](#S4.SS2 "4.2 Adaptive Sampling via Policy Optimization ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
4 Guided Cost Learning
-----------------------
In this section, we describe the guided cost learning algorithm, which combines sample-based maximum entropy IOC with forward reinforcement learning using time-varying linear models. The central idea behind this method is to adapt the sampling distribution to match the maximum entropy cost distribution p(τ)=1Zexp(−cθ(τ)), by directly optimizing a trajectory distribution with respect to the current cost cθ(τ) using a sample-efficient reinforcement learning algorithm. Samples generated on the physical system are used both to improve the policy and more accurately estimate the partition function Z. In this way, the reinforcement learning step acts to “guide” the sampling distribution toward regions where the samples are more useful for estimating the partition function. We will first describe how the IOC objective in Equation ([1](#S3.E1 "(1) ‣ 3 Preliminaries and Overview ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization")) can be estimated with samples, and then describe how reinforcement learning can adapt the sampling distribution.
###
4.1 Sample-Based Inverse Optimal Control
In the sample-based approach to maximum entropy IOC, the partition function Z=∫exp(−cθ(τ))dτ is estimated with samples from a background distribution q(τ). Prior sample-based IOC methods use a linear representation of the cost function, which simplifies the corresponding cost learning problem (Boularias et al., [2011](#bib.bib5); Kalakrishnan et al., [2013](#bib.bib11)). In this section, we instead derive a sample-based approximation for the IOC objective for a general nonlinear parameterization of the cost function. The negative log-likelihood corresponding to the IOC model in Equation ([1](#S3.E1 "(1) ‣ 3 Preliminaries and Overview ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization")) is given by:
| | | | |
| --- | --- | --- | --- |
| | LIOC(θ) | =1N∑τi∈Ddemocθ(τi)+logZ | |
| | | ≈1N∑τi∈Ddemocθ(τi)+log1M∑τj∈Dsampexp(−cθ(τj))q(τj), | |
where Ddemo denotes the set of N demonstrated trajectories, Dsamp the set of M background samples, and q denotes the background distribution from which trajectories τj were sampled. Prior methods have chosen q to be uniform (Boularias et al., [2011](#bib.bib5)) or to lie in the vicinity of the demonstrations (Kalakrishnan et al., [2013](#bib.bib11)).
To compute the gradients of this objective with respect to the cost parameters θ, let wj=exp(−cθ(τj))q(τj) and Z=∑jwj. The gradient is then given by:
| | | |
| --- | --- | --- |
| | dLIOCdθ=1N∑τi∈Ddemodcθdθ(τi)−1Z∑τj∈Dsampwjdcθdθ(τj) | |
When the cost is represented by a neural network or some other function approximator, this gradient can be computed efficiently by backpropagating −wjZ for each trajectory τj∈Dsamp and 1N for each trajectory τi∈Ddemo.
###
4.2 Adaptive Sampling via Policy Optimization
The choice of background sample distribution q(τ) for estimating the objective LIOC is critical for successfully applying the sample-based IOC algorithm. The optimal importance sampling distribution for estimating the partition function ∫exp(−cθ(τ))dτ is q(τ)∝|exp(−cθ(τ))|=exp(−cθ(τ)).
Designing a single background distribution q(τ) is therefore quite difficult when the cost cθ is unknown. Instead, we can adaptively refine q(τ) to generate more samples in those regions of the trajectory space that are good according to the current cost function cθ(τ). To this end, we interleave the IOC optimization, which attempts to find the cost function that maximizes the likelihood of the demonstrations, with a policy optimization procedure, which improves the trajectory distribution q(τ) with respect to the current cost.
Since one of the main advantages of the sample-based IOC approach is the ability to handle unknown dynamics, we must also choose a policy optimization procedure that can handle unknown dynamics. To this end, we adapt the method presented by Levine & Abbeel ([2014](#bib.bib12)), which performs policy optimization under unknown dynamics by iteratively fitting time-varying linear dynamics to samples from the current trajectory distribution q(τ), updating the trajectory distribution using a modified LQR backward pass, and generating more samples for the next iteration. The trajectory distributions generated by this method are Gaussian, and each iteration of the policy optimization procedure satisfies a KL-divergence constraint of the form DKL(q(τ)∥^q(τ))≤ϵ, which prevents the policy from changing too rapidly (Bagnell & Schneider, [2003](#bib.bib4); Peters et al., [2010](#bib.bib20); Rawlik & Vijayakumar, [2013](#bib.bib25)).
This has the additional benefit of not overfitting to poor initial estimates of the cost function. With a small modification, we can use this algorithm to optimize a maximum entropy version of the objective, given by minqEq[cθ(τ)]−H(τ), as discussed in prior work (Levine & Abbeel, [2014](#bib.bib12)). This variant of the algorithm allows us to recover the trajectory distribution q(τ)∝exp(−cθ(τ)) at convergence (Ziebart, [2010](#bib.bib30)), a good distribution for sampling. For completeness, this policy optimization procedure is summarized in Appendix [A](#A1 "Appendix A Policy Optimization under Unknown Dynamics ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
1: Initialize qk(τ) as either a random initial controller or from demonstrations
2: for iteration i=1 to I do
3: Generate samples Dtraj from qk(τ)
4: Append samples: Dsamp←Dsamp∪Dtraj
5: Use Dsamp to update cost cθ using Algorithm [2](#alg2 "Algorithm 2 ‣ 4.3 Cost Optimization and Importance Weights ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization")
6: Update qk(τ) using Dtraj and the method from (Levine & Abbeel, [2014](#bib.bib12)) to obtain qk+1(τ)
7: end for
8: return optimized cost parameters θ and trajectory distribution q(τ)
Algorithm 1 Guided cost learning
Our sample-based IOC algorithm with adaptive sampling is summarized in Algorithm [1](#alg1 "Algorithm 1 ‣ 4.2 Adaptive Sampling via Policy Optimization ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"). We call this method guided cost learning because policy optimization is used to guide sampling toward regions with lower cost. The algorithm consists of taking successive policy optimization steps, each of which generates samples Dtraj from the latest trajectory distribution qk(τ). After sampling, the cost function is updated using all samples collected thus far for the purpose of policy optimization. No additional background samples are required for this method. This procedure returns both a learned cost function cθ(xt,ut) and a trajectory distribution q(τ), which corresponds to a time-varying linear-Gaussian controller q(ut|xt). This controller can be used to execute the learned behavior.
###
4.3 Cost Optimization and Importance Weights
1: for iteration k=1 to K do
2: Sample demonstration batch ^Ddemo⊂Ddemo
3: Sample background batch ^Dsamp⊂Dsamp
4: Append demonstration batch to background batch: ^Dsamp←^Ddemo∪^Dsamp
5: Estimate dLIOCdθ(θ) using ^Ddemo and ^Dsamp
6: Update parameters θ using gradient dLIOCdθ(θ)
7: end for
8: return optimized cost parameters θ
Algorithm 2 Nonlinear IOC with stochastic gradients
The IOC objective can be optimized using standard nonlinear optimization methods and the gradient dLIOCdθ. Stochastic gradient methods are often preferred for high-dimensional function approximators, such as the neural networks. Such methods are straightforward to apply to objectives that factorize over the training samples, but the partition function does not factorize trivially in this way. Nonetheless, we found that our objective could still be optimized with stochastic gradient methods by sampling a subset of the demonstrations and background samples at each iteration. When the number of samples in the batch is small, we found it necessary to add the sampled demonstrations to the background sample set as well; without adding the demonstrations to the sample set, the objective can become unbounded and frequently does in practice.
The stochastic optimization procedure is presented in Algorithm [2](#alg2 "Algorithm 2 ‣ 4.3 Cost Optimization and Importance Weights ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), and is straightforward to implement with most neural network libraries based on backpropagation.
Estimating the partition function requires us to use importance sampling. Although prior work has suggested dropping the importance weights (Kalakrishnan et al., [2013](#bib.bib11); Aghasadeghi & Bretl, [2011](#bib.bib2)), we show in Appendix [B](#A2 "Appendix B Consistency Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization") that this produces an inconsistent likelihood estimate and fails to recover good cost functions. Since our samples are drawn from multiple distributions, we compute a fusion distribution to evaluate the importance weights.
Specifically, if we have samples from k distributions q1(τ),…,qκ(τ), we can construct a consistent estimator of the expectation of a function f(τ) under a uniform distribution as E[f(τ)]≈1M∑τj11k∑κqκ(τj)f(τj).
Accordingly, the importance weights are zj=[1k∑κqκ(τj)]−1, and the objective is now:
| | | | |
| --- | --- | --- | --- |
| | LIOC(θ)= | 1N∑τi∈Ddemocθ(τi)+log1M∑τj∈Dsampzjexp(−cθ(τj)) | |
The distributions qκ underlying background samples are obtained from the controller at iteration k. We must also append the demonstrations to the samples in Algorithm [2](#alg2 "Algorithm 2 ‣ 4.3 Cost Optimization and Importance Weights ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), yet the distribution that generated the demonstrations is unknown. To estimate it, we assume the demonstrations come from a single Gaussian trajectory distribution and compute its empirical mean and variance. We found this approximation sufficiently accurate for estimating the importance weights of the demonstrations, as shown in Appendix [B](#A2 "Appendix B Consistency Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
###
4.4 Learning Costs and Controllers
In contrast to many previous IOC and IRL methods, our approach can be used to learn a cost while simultaneously optimizing the policy for a new instance of the task not in the demos, such as a new position of a target cup for a pouring task, as shown in our experiments. Since the algorithm produces both a cost function cθ(xt,ut) and a controller q(ut|xt) that optimizes this cost on the new task instance, we can directly use this controller to execute the desired behavior. In this way, the method actually learns a policy from demonstration, using the additional knowledge that the demonstrations are near-optimal under some unknown cost function, similar to recent work on IOC by direct loss minimization (Doerr et al., [2015](#bib.bib7)). The learned cost function cθ(xt,ut) can often also be used to optimize new policies for new instances of the task without additional cost learning. However, we found that on the most challenging tasks we tested, running policy learning with IOC in the loop for each new task instance typically succeeded more frequently than running IOC once and reusing the learned cost. We hypothesize that this is because training the policy on a new instance of the task provides the algorithm with additional information about task variation, thus producing a better cost function and reducing overfitting. The intuition behind this hypothesis is that the demonstrations only cover a small portion of the degrees of variation in the task. Observing samples from a new task instance provides the algorithm with a better idea of the particular factors that distinguish successful task executions from failures.
5 Representation and Regularization
------------------------------------
We parametrize our cost functions as neural networks, expanding their expressive power and enabling IOC to be applied to the state of a robotic system directly, without hand-designed features.
Our experiments in Section [6.2](#S6.SS2 "6.2 Real-World Robotic Control ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization") confirm that an affine cost function is not expressive enough to learn some behaviors.
Neural network parametrizations are particularly useful for learning visual representations on raw image pixels.
In our experiments, we make use of the unsupervised visual feature learning method developed by Finn et al. ([2016](#bib.bib9)) to learn cost functions that depend on visual input.
Learning cost functions on raw pixels is an interesting direction for future work, which we discuss in Section [7](#S7 "7 Discussion and Future Work ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
While the expressive power of nonlinear cost functions provide a range of benefits, they introduce significant model complexity to an already underspecified IOC objective.
To mitigate this challenge, we propose two regularization methods for IOC.
Prior methods regularize the IOC objective by penalizing the ℓ1 or ℓ2 norm of the cost parameters θ (Ziebart, [2010](#bib.bib30); Kalakrishnan et al., [2013](#bib.bib11)).
For high-dimensional nonlinear cost functions, this regularizer is often insufficient, since different entries in the parameter vector can have drastically different effects on the cost.
We use two regularization terms. The first term encourages the cost of demo and sample trajectories to change locally at a constant rate (lcr), by penalizing the second time derivative:
| | | |
| --- | --- | --- |
| | glcr(τ)=∑xt∈τ[(cθ(xt+1)−cθ(xt))−(cθ(xt)−cθ(xt−1))]2 | |
This term reduces high-frequency variation that is often symptomatic of overfitting, making the cost easier to reoptimize. Although sharp changes in the cost slope are sometimes preferred, we found that temporally slow-changing costs were able to adequately capture all of the behaviors in our experiments.
The second regularizer is more specific to one-shot episodic tasks, and it encourages the cost of the states of a demo trajectory to decrease strictly monotonically in time using a squared hinge loss:
| | | |
| --- | --- | --- |
| | gmono(τ)=∑xt∈τ[max(0,cθ(xt)−cθ(xt−1)−1)]2 | |
The rationale behind this regularizer is that, for tasks that essentially consist of reaching a target condition or state, the demonstrations typically make monotonic progress toward the goal on some (potentially nonlinear) manifold. While this assumption does not always hold perfectly, we again found that this type of regularizer improved performance on the tasks in our evaluation. We show a detailed comparison with regard to both regularizers in Appendix [E](#A5 "Appendix E Regularization Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
6 Experimental Evaluation
--------------------------
We evaluated our sampling-based IOC algorithm on a set of robotic control tasks, both in simulation and on a real robotic platform. Each of the experiments involve complex second order dynamics with force or torque control and no manually designed cost function features, with the raw state provided as input to the learned cost function.
We also tested the consistency of our algorithm on a toy point mass example for which the ground truth distribution is known. These experiments, discussed fully in Appendix [B](#A2 "Appendix B Consistency Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), show that using a maximum entropy version of the policy optimization objective (see Section [4.2](#S4.SS2 "4.2 Adaptive Sampling via Policy Optimization ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization")) and using importance weights are both necessary for recovering the true distribution.
###
6.1 Simulated Comparisons
In this section, we provide simulated comparisons between guided cost learning and prior sample-based methods. We focus on task performance and sample complexity, and also perform comparisons across two different sampling distribution initializations and regularizations (in Appendix [E](#A5 "Appendix E Regularization Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization")).
green: goalred: obstaclesinitialstategoalstate
Figure 2: Comparison to prior work on simulated 2D navigation, reaching, and peg insertion tasks. Reported performance is averaged over 4 runs of IOC on 4 different initial conditions . For peg insertion, the depth of the hole is 0.1m, marked as a dashed line. Distances larger than this amount failed to insert the peg.
To compare guided cost learning to prior methods, we ran experiments on three simulated tasks of varying difficulty, all using the MuJoCo physics simulator (Todorov et al., [2012](#bib.bib27)). The first task is 2D navigation around obstacles, modeled on the task by Levine & Koltun ([2012](#bib.bib13)). This task has simple, linear dynamics and a low-dimensional state space, but a complex cost function, which we visualize in Figure [2](#S6.F2 "Figure 2 ‣ 6.1 Simulated Comparisons ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"). The second task involves a 3-link arm reaching towards a goal location in 2D, in the presence of physical obstacles. The third, most challenging, task is 3D peg insertion with a 7 DOF arm. This task is significantly more difficult than tasks evaluated in prior IOC work as it involves complex contact dynamics between the peg and the table and high-dimensional, continuous state and action spaces. The arm is controlled by selecting torques at the joint motors at 20 Hz. More details on the experimental setup are provided in Appendix [D](#A4 "Appendix D Detailed Description of Task Setup ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
In addition to the expert demonstrations, prior methods require a set of “suboptimal” samples for estimating the partition function. We obtain these samples in one of two ways: by using a baseline random controller that randomly explores around the initial state (random), and by fitting a linear-Gaussian controller to the demonstrations (demo). The latter initialization typically produces a motion that tracks the average demonstration with variance proportional to the variation between demonstrated motions.
Between 20 and 32 demonstrations were generated from a policy learned using the method of Levine & Abbeel ([2014](#bib.bib12)), with a ground truth cost function determined by the agent’s pose relative to the goal. We found that for the more precise peg insertion task, a relatively complex ground truth cost function was needed to afford the necessary degree of precision. We used a cost function of the form wd2+vlog(d2+α), where d is the distance between the two tips of the peg and their target positions, and v and α are constants. Note that the affine cost is incapable of exactly representing this function. We generated demonstration trajectories under several different starting conditions. For 2D navigation, we varied the initial position of the agent, and for peg insertion, we varied the position of the peg hole. We then evaluated the performance of our method and prior sample-based methods (Kalakrishnan et al., [2013](#bib.bib11); Boularias et al., [2011](#bib.bib5)) on each task from four arbitrarily-chosen test states. We chose these prior methods because, to our knowledge, they are the only methods which can handle unknown dynamics.
We used a neural network cost function with two hidden layers with 24–52 units and rectifying nonlinearities of the form max(z,0) followed by linear connections to a set of features yt, which had a size of 20 for the 2D navigation task and 100 for the other two tasks. The cost is then given by
| | | | |
| --- | --- | --- | --- |
| | | | (2) |
with a fixed torque weight wu and the parameters consisting of A, b, and the network weights. These cost functions range from about 3,000 parameters for the 2D navigation task to 16,000 parameters for peg insertion. For further details, see Appendix [C](#A3 "Appendix C Neural Network Parametrization and Initialization ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"). Although the prior methods learn only linear cost functions, we can extend them to the nonlinear setting following the derivation in Section [4.1](#S4.SS1 "4.1 Sample-Based Inverse Optimal Control ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization").
Figure [2](#S6.F2 "Figure 2 ‣ 6.1 Simulated Comparisons ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization") illustrates the tasks and shows results for each method after different numbers of samples from the test condition. In our method, five samples were used at each iteration of policy optimization, while for the prior methods, the number of samples corresponds to the number of “suboptimal” samples provided for cost learning. For the prior methods, additional samples were used to optimize the learned cost. The results indicate that our method is generally capable of learning tasks that are more complex than the prior methods, and is able to effectively handle complex, high-dimensional neural network cost functions. In particular, adding more samples for the prior methods generally does not improve their performance, because all of the samples are drawn from the same distribution.
###
6.2 Real-World Robotic Control
We also evaluated our method on a set of real robotic manipulation tasks using the PR2 robot, with comparisons to relative entropy IRL, which we found to be the better of the two prior methods in our simulated experiments. We chose two robotic manipulation tasks which involve complex dynamics and interactions with delicate objects, for which it is challenging to write down a cost function by hand.
For all methods, we used a two-layer neural network cost parametrization and the regularization objective described in Section [5](#S5 "5 Representation and Regularization ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), and compared to an affine cost function on one task to evaluate the importance of non-linear cost representations. The affine cost followed the form of equation [2](#S6.E2 "(2) ‣ 6.1 Simulated Comparisons ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization") but with yt equal to the input xt.111Note that a cost function that is quadratic in the state is linear in the coefficients of the monomials, and therefore corresponds to a linear parameterization. For both tasks, between 25 and 30 human demonstrations were provided via kinesthetic teaching, and each IOC algorithm was initialized by automatically fitting a controller to the demonstrations that roughly tracked the average trajectory. Full details on both tasks are in Appendix [D](#A4 "Appendix D Detailed Description of Task Setup ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), and summaries are below.
human demoinitial posefinal pose
Figure 3: Dish placement and pouring tasks. The robot learned to place the plate gently into the correct slot, and to pour almonds, localizing the target cup using unsupervised visual features. A video of the learned controllers can be found at <http://rll.berkeley.edu/gcl>
In the first task, illustrated in Figure [3](#S6.F3 "Figure 3 ‣ 6.2 Real-World Robotic Control ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), the robot must gently place a grasped plate into a specific slot of dish rack. The state space consists of the joint angles, the pose of the gripper relative to the target pose, and the time derivatives of each; the actions correspond to torques on the robot’s motors; and the input to the cost function is the pose and velocity of the gripper relative to the target position. Note that we do not provide the robot with an existing trajectory tracking controller or any manually-designed policy representation beyond linear-Gaussian controllers, in contrast to prior methods that use trajectory following (Kalakrishnan et al., [2013](#bib.bib11)) or dynamic movement primitives with features (Boularias et al., [2011](#bib.bib5)). Our attempt to design a hand-crafted cost function for inserting the plate into the dish rack produced a fast but overly aggressive behavior that cracked one of the plates during learning.
The second task, also shown in Figure [3](#S6.F3 "Figure 3 ‣ 6.2 Real-World Robotic Control ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), consisted of pouring almonds from one cup to another. In order to succeed, the robot must keep the cup upright until reaching the target cup, then rotate the cup so that the almonds are poured. Instead of including the position of the target cup in the state space, we train autoencoder features from camera images captured from the demonstrations and add a pruned feature point representation and its time derivative to the state, as proposed by Finn et al. ([2016](#bib.bib9)). The input to the cost function includes these visual features, as well as the pose and velocity of the gripper. Note that the position of the target cup can only be obtained from the visual features, so the algorithm must learn to use them in the cost function in order to succeed at the task.
| | | | |
| --- | --- | --- | --- |
| *dish* (NN) | RelEnt IRL | GCL q(ut|xt) | GCL reopt. |
| success rate | 0% | 100% | 100% |
| # samples | 100 | 90 | 90 |
| *pouring* (NN) | RelEnt IRL | GCL q(ut|xt) | GCL reopt. |
| success rate | 10% | 84.7% | 34% |
| # samples | 150,150 | 75,130 | 75,130 |
| *pouring* (affine) | RelEnt IRL | GCL q(ut|xt) | GCL reopt. |
| success rate | 0% | 0% | – |
| # samples | 150 | 120 | – |
Table 1: Performance of guided cost learning (GCL) and relative entropy (RelEnt) IRL on placing a dish into a rack and pouring almonds into a cup.
Sample counts are for IOC, omitting those for optimizing the learned cost.
An affine cost is insufficient for representing the pouring task, thus motivating using a neural network cost (NN).
The pouring task with a neural network cost is evaluated for two positions of the target cup; average performance is reported.
The results, presented in Table [1](#S6.T1 "Table 1 ‣ 6.2 Real-World Robotic Control ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), show that our algorithm successfully learned both tasks. The prior relative entropy IRL algorithm could not acquire a suitable cost function, due to the complexity of this domain. On the pouring task, where we also evaluated a simpler affine cost function, we found that only the neural network representation could recover a successful behavior, illustrating the need for rich and expressive function approximators when learning cost functions directly on raw state representations.222We did attempt to learn costs directly on image pixels, but found that the problem was too underdetermined to succeed. Better image-specific regularization is likely required for this.
The results in Table [1](#S6.T1 "Table 1 ‣ 6.2 Real-World Robotic Control ‣ 6 Experimental Evaluation ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization") also evaluate the generalizability of the cost function learned by our method and prior work. On the dish rack task, we can use the learned cost to optimize new policies for different target dish positions successfully, while the prior method does not produce a generalizable cost function. On the harder pouring task, we found that the learned cost succeeded less often on new positions. However, as discussed in Section [4.4](#S4.SS4 "4.4 Learning Costs and Controllers ‣ 4 Guided Cost Learning ‣ Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization"), our method produces both a policy q(ut|xt) and a cost function cθ when trained on a novel instance of the task, and although the learned cost functions for this task were worse, the learned policy succeeded on the test positions when optimized with IOC in the inner loop using our algorithm. This indicates an interesting property of our approach: although the learned cost function is local in nature due to the choice of sampling distribution, the learned policy tends to succeed even when the cost function is too local to produce good results in very different situations. An interesting avenue for future work is to further explore the implications of this property, and to improve the generalizability of the learned cost by successively training policies on different novel instances of the task until enough global training data is available to produce a cost function that is a good fit to the demonstrations in previously unseen parts of the state space.
7 Discussion and Future Work
-----------------------------
We presented an inverse optimal control algorithm that can learn complex, nonlinear cost representations, such as neural networks, and can be applied to high-dimensional systems with unknown dynamics. Our algorithm uses a sample-based approximation of the maximum entropy IOC objective, with samples generated from a policy learning algorithm based on local linear models (Levine & Abbeel, [2014](#bib.bib12)). To our knowledge, this approach is the first to combine the benefits of sample-based IOC under unknown dynamics with nonlinear cost representations that directly use the raw state of the system, without the need for manual feature engineering. This allows us to apply our method to a variety of real-world robotic manipulation tasks. Our evaluation demonstrates that our method outperforms prior IOC algorithms on a set of simulated benchmarks, and achieves good results on several real-world tasks.
Our evaluation shows that our approach can learn good cost functions for a variety of simulated tasks.
For complex robotic motion skills, the learned cost functions tend to explain the demonstrations only locally. This makes them difficult to reoptimize from scratch for new conditions. It should be noted that this challenge is not unique to our method. In our comparisons, no prior sample-based method was able to learn good global costs for these tasks. However, since our method interleaves cost optimization with policy learning, it still recovers successful policies for these tasks. For this reason, we can still learn from demonstration simply by retaining the learned policy, and discarding the cost function. This allows us to tackle substantially more challenging tasks that involve direct torque control of real robotic systems with feedback from vision.
To incorporate vision into our experiments, we used unsupervised learning to acquire a vision-based state representation, following prior work (Finn et al., [2016](#bib.bib9)). An exciting avenue for future work is to extend our approach to learn cost functions directly from natural images. The principal challenge for this extension is to avoid overfitting when using substantially larger and more expressive networks. Our current regularization techniques mitigate overfitting to a high degree, but visual inputs tend to vary dramatically between demonstrations and on-policy samples, particularly when the demonstrations are provided by a human via kinesthetic teaching. One promising avenue for mitigating these challenges is to introduce regularization methods developed for domain adaptation in computer vision (Tzeng et al., [2015](#bib.bib28)), to encode the prior knowledge that demonstrations have similar visual features to samples.
Acknowledgements
----------------
This research was funded in part by ONR through a Young Investigator Program
award, the Army Research Office through the MAST program, and an NSF
fellowship. We thank Anca Dragan for thoughtful discussions. |
d4f6b0e7-36f8-4874-96c5-779d27f8765d | trentmkelly/LessWrong-43k | LessWrong | Background material for reading Judgment Under Uncertainty?
After seeing it constantly referenced in the Sequences and elsewhere, I've picked up Kahneman and Tversky's book/collection of papers Judgment Under Uncertainty: Heuristics and Biases. I was wondering if anyone here who's read it or knows the subject would recommend any prefatory material so that it makes more sense/is more meaningful.
Personal background: [Personal information deleted] From that and from this site, I'm passingly familiar with e.g. the representativeness heuristic and Bayesian probability, but I've never had to use it much in any academic setting.
Any advice before I delve into it? |
5f06fa72-a1b8-4356-917b-e207f7117c08 | trentmkelly/LessWrong-43k | LessWrong | Writing feedback requested: activists should pursue a positive Singularity
I managed to turn an essay assignment into an opportunity to write about the Singularity, and I thought I'd turn to LW for feedback on the paper. The paper is about Thomas Pogge, a German philosopher who works on institutional efforts to end poverty and is a pledger for Giving What We Can.
I offer a basic argument that he and other poverty activists should work on creating a positive Singularity, sampling liberally from well-known Less Wrong arguments. It's more academic than I would prefer, and it includes some loose talk of 'duties' (which bothers me), but for its goals, these things shouldn't be a huge problem. But maybe they are - I want to know that too.
I've already turned the assignment in, but when I make a better version, I'll send the paper to Pogge himself. I'd like to see if I can successfully introduce him to these ideas. My one conversation with him indicates that he would be open to actually changing his mind. He's clearly thought deeply about how to do good, and may simply have not been exposed to the idea of the Singularity yet.
I want feedback on all aspects of the paper - style, argumentation, clarity. Be as constructively cruel as I know only you can.
If anyone's up for it, fee free to add feedback using Track Changes and email me a copy - mjcurzi[at]wustl.edu. I obviously welcome comments on the thread as well.
You can read the paper here in various formats.
Upvotes for all. Thank you! |
141bae9f-047a-48e1-aaa8-606c5fbe8d17 | trentmkelly/LessWrong-43k | LessWrong | Unattributed Good
Cross-posted from 10-year Horizon.
If a good deed is done and no one is around to see it, does it make a sound?
In 1939, Nicholas Winton found himself on a volunteer mission in pre-war Prague. Recognizing the darkening position of the Jews, he helped organize train transports that saved 669 children from concentration camps. Today, many of us have heard this fascinating story of good character. But the story was only uncovered in 1988 by the BBC, nearly 50 years later. Sir Nicholas is said to have never talked about it himself. To most of us, those uncredited decades are powerful and puzzling.
If a good deed is done and no one is around to see it, does it make a sound?
> "I used to believe that a legacy was solely based on your achievements."
>
> Dayana Sabatin
On an everyday scale, this happens to all of us. It’s the goodwill gesture that goes unnoticed in a relationship. The overlooked patience. The brushed-off compliment.
From the trivial to the heroic, there is a yawning mismatch between what’s done and what’s appreciated. It makes a glaring contrast with the visible measures of appreciation today: namely, social media likes. Who gets thousands of likes? What for? And who goes through life helping others, perhaps by simply being a little nicer than strictly necessary, yet stays invisible?
> As with economic and other resources, the judgement of others is very unevenly distributed. Some are rich with recognition, applause, goodwill, trust, reputation and others are starved of a good word.
>
> Ziyad Marar - Judged
Unattributed good is the delta between merit and credit. What an enormous amount of misallocated energy it represents! What an amount of inner tension in people! Especially considering that both merit and popularity are so highly prized today. Imagine a social network where the most popular people were the ones who did the most good. What would that feed look like? Very different from Instagram. We can bridge that. We don’t even have |
9021664b-fbf5-4fb6-a042-4047ce5a2eae | trentmkelly/LessWrong-43k | LessWrong | In which I fantasize about drugs
We operate like this: the "overseer process" tells the brain, using blunt instruments like chemicals, that we need to find something to eat, somewhere to sleep or someone to mate with. Then the brain follows orders. Unfortunately the orders we receive from the "overseer" are often wrong, even though they were right in the ancestral environment. It seems the easiest way to improve humans isn't to augment their brains - it's to send them better orders, e.g. using drugs. Here's a list of fantasy brain-affecting drugs that I would find useful, even though they don't seem to do anything complicated except affecting "overseer" chemistry:
1) A drug against unrequited love, aka "infatuation" or 'limerence".
2) A drug that makes you become restless and want to exercise.
3) A drug that puts you in the state of random creativity that you normally experience just before falling asleep.
4) A drug that puts you in the optimal PUA "state".
5) A drug that boosts your feeling of curiosity. Must be great for doing math or science.
Anything else? |
4a4bef51-3610-4b3d-83a5-d37d3c484966 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Voluntary safety commitments provide an escape from over-regulation in AI development
Abstract
--------
With the introduction of Artificial Intelligence (AI) and related technologies in our daily lives, fear and anxiety about their misuse as well as the hidden biases in their creation have led to a demand for regulation to address such issues. Yet blindly regulating an innovation process that is not well understood, may stifle this process and reduce benefits that society may gain from the generated technology, even under the best intentions. In this paper, starting from a baseline model that captures the fundamental dynamics of a race for domain supremacy using AI technology, we demonstrate how socially unwanted outcomes may be produced when sanctioning is applied unconditionally to risk-taking, i.e. potentially unsafe, behaviours. As an alternative to resolve the detrimental effect of over-regulation, we propose a voluntary commitment approach wherein technologists have the freedom of choice between independently pursuing their course of actions or establishing binding agreements to act safely, with sanctioning of those that do not abide to what they pledged. Overall, this work reveals for the first time how voluntary commitments, with sanctions either by peers or an institution, leads to socially beneficial outcomes in all scenarios envisageable in a short-term race towards domain supremacy through AI technology. These results are directly relevant for the design of governance and regulatory policies that aim to ensure an ethical and responsible AI technology development process.
Keywords: Evolutionary Game Theory, AI development race, Commitments, Incentives, Safety.
1 Introduction
---------------
With the rapid advancement of AI and related technologies, there has been significant fear and anxiety about their potential misuse as well as the social and ethical consequences that may result from biases within the design of such systems (Tzachor et al.,, [2020](#bib.bib45), Bostrom,, [2014](#bib.bib5), Stix and Maas,, [2020](#bib.bib42)).
While expectations associated with these advanced technologies increase and monetary profits stimulate rapid deployment, there is a serious risk for taking unethical or risky short cuts to enter a market first with the next innovation, ignoring safety checks and ethical development procedures.
As different disagreeable examples have emerged (Coeckelbergh,, [2020](#bib.bib12)), governments and regulating bodies have been catching up by debating new forms and frameworks for regulating this technology (Baum,, [2017](#bib.bib4), Cave and ÓhÉigeartaigh,, [2018](#bib.bib8), Taddeo and Floridi,, [2018](#bib.bib43)); notably, the recent EU White Paper on AI.
Such debates have produced proposals for mechanisms on how to avoid, mediate, or regulate the development and deployment of AI (Baum,, [2017](#bib.bib4), Cave and ÓhÉigeartaigh,, [2018](#bib.bib8), Geist,, [2016](#bib.bib19), Shulman and Armstrong,, [2009](#bib.bib40), Han et al.,, [2019](#bib.bib24), Vinuesa et al.,, [2020](#bib.bib47), Nemitz,, [2018](#bib.bib35), Taddeo and Floridi,, [2018](#bib.bib43), Askell et al.,, [2019](#bib.bib2), O’Keefe et al.,, [2020](#bib.bib38)). Essentially, regulatory measures such as restrictions and incentives are proposed to limit harmful and risky practices in order to promote beneficial designs (Baum,, [2017](#bib.bib4)). Examples of such approaches (Baum,, [2017](#bib.bib4)) include financially supporting the research into beneficial AI (McGinnis,, [2010](#bib.bib34)) and making AI companies pay
fines when found liable for the consequences of harmful AI (Gurney,, [2013](#bib.bib20)).
Although these regulatory measures may provide solutions for particular scenarios, one needs to ensure that they do not overshoot their targets since over-regulation could stifle innovation, potentially hindering investments into the development of novel innovations as they become too risky endeavors (Hadfield,, [2017](#bib.bib21), Lee,, [2018](#bib.bib32)).
Worries have been expressed by different organisations/societies that too strict policies may unnecessarily affect the benefits and societal advances that novel AI technologies may have to offer (EDRI,, [2021](#bib.bib18)).
Regulations affect moreover big and small tech companies differently: A highly regulated domain makes it more difficult for small new start-ups, introducing an inequality and dominance of the market by a few big players (Lee,, [2018](#bib.bib32)). It has been emphasised that neither over-regulation nor a laissez-faire approach suffices when aiming to regulate AI technologies (Dawson et al.,, [2019](#bib.bib15)). In order to find a balanced answer, one clearly needs to have first an understanding of how a competitive development dynamic may work.
Starting from a baseline game theoretical model, referred to as the DSAIR model (Han et al.,, [2020](#bib.bib28)) (a model of domain supremacy through an AI race), which defines the process through which multiple stake-holders aim for market supremacy, we demonstrate first that unconditional sanctioning will negatively influence social welfare in certain conditions of a short-term race towards domain supremacy through AI technology. Afterwards, we examine the DSAIR model with an alternative mechanism for resolving the issue that can be shown to lead to less detrimental effects. Our approach is to allow technologists or race participants to voluntarily commit themselves to safe innovation procedures, signaling to others their intentions. Specifically, this bottom-up, binding agreement (or commitment) is established for those who want to take a safe choice, with sanctioning applied to violators of such an agreement.
#### Previous Developments.
As was shown in (Han et al.,, [2020](#bib.bib28)) participants either follow safety precautions (the SAFE option) or ignore them (the UNSAFE option) in each step of the development process of the DSAIR model. The main assumption in the model was that it requires more time and more effort to comply with the precautionary requirements, making the SAFE option not only costlier, but also slower compared to the UNSAFE option. Accordingly, it was assumed that in playing SAFE, participants must pay a cost c>0𝑐0c>0italic\_c > 0, whereas playing UNSAFE costs nothing. Furthermore, whenever playing UNSAFE, the development speed differs and is s>1𝑠1s>1italic\_s > 1 whereas in playing SAFE the speed is simply normalised to 1.
Decisions to act SAFE or UNSAFE in AI development are repeated until one or more teams attain the designated objective, which can be translated into having completed W𝑊Witalic\_W development steps, on average (Han et al.,, [2020](#bib.bib28)). As a result, they earn a large benefit or prize B𝐵Bitalic\_B (e.g. windfall profits (O’Keefe et al.,, [2020](#bib.bib38))), equally shared among those reaching the target at the same time. A development disaster or a setback may however come to occur with some probability, which is presumed to increase with the number of times that safety requirements were ignored by the winning team(s) at each step.
Whenever a disaster of this kind occurs, all the benefits of a risk-taking participant are lost. This risk probability is denoted by prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT (see the Models and Methods section for more details).
Figure 1: Behavioural regions (zones) as identified in (Han et al.,, [2020](#bib.bib28)) when the time to reach domain supremacy is short. Region II: Inside the plots, the two solid lines delimit the boundaries wherein the collective prefers safety compliant behavior, yet unsafe development is individually preferred. Regions III and I exhibit where unsafe (respectively, safe) development is both the individually and collectively preferred outcome
It was observed in the DSAIR model that, in case the time-scale to reach the target is short, so that over the whole of the development process the average of the accumulated benefit is much smaller than the final benefit B𝐵Bitalic\_B, only for a certain window of parameter settings societal interest conflicts with the individual ones: In that region, individual unsafe behaviour dominates, despite that safe development would lead to a larger collective outcome or social welfare (cf. region II in Figure [1](#S1.F1 "Figure 1 ‣ Previous Developments. ‣ 1 Introduction ‣ Voluntary safety commitments provide an escape from over-regulation in AI development")).
From the regulatory perspective, it is only region II that thus requires governance in order to promote or enforce safe actions in order to avoid any disaster that may occur during the technology development race.
A peer punishment mechanism against unsafe behaviour was proposed in order to mediate the behavior in that region (without affecting the desirable safe outcome in region I) (Han et al.,, [2021](#bib.bib25)). It however may lead to a reduction of the societal welfare that can be obtained in region III (see Figure [1](#S1.F1 "Figure 1 ‣ Previous Developments. ‣ 1 Introduction ‣ Voluntary safety commitments provide an escape from over-regulation in AI development")) where the desired unsafe (risk-taking/innovative) behaviour becomes significantly reduced whenever punishment is not very costly to the punisher while strongly affecting the punished. The problem with applying system-wide sanctioning for any risk-taking behavior is that it does not take into account the region wherein the AI innovation is taking place, as was visualised by Figure [1](#S1.F1 "Figure 1 ‣ Previous Developments. ‣ 1 Introduction ‣ Voluntary safety commitments provide an escape from over-regulation in AI development"). That is, if it is the case that a development race falls in region III (low risk and innovation is collectively beneficial and preferred), unsafe behaviour should not be punished as it is beneficial for the overall social welfare. To enact such a region-dependent targeted punishment scheme, one would require the ability to estimate exactly the risk level associated with each AI development scenario, i.e. knowing beforehand the risk as well as speed of development. Clearly that may not be easy due to the lack of data on how such an AI innovation dynamics works. This is especially true in the early stages of the development and adoption of many new technologies, which has become known as the so-called double-bind problem (Collingridge,, [1980](#bib.bib13)): the impact of technologies including AI ones cannot be predicted until it becomes a reality, and controlling or changing it is difficult or no longer possible at that point.
#### Objectives
In this article an alternative solution is proposed that circumvents the need of being able to estimate correctly the speed of development and risk in order to appropriately regulate UNSAFE behavior: race participants can choose whether or not to establish a bilateral commitment to act safely, which also exposes them to a sanction in case they do not uphold their commitment. While, on the one had, the development teams are allowed to work in an UNSAFE manner without repercussions if they do not commit; a prior agreement allows, on the other hand, the safety compliant participants to identify easily unsafe ones while having also the capacity to punish the dishonest ones (who might also be sanctioned by an external party such as a regulating institution).
Our results reveal that this freedom of choice, if enabled through a prior bilateral commitment to SAFE actions, can, on the one hand, avoid over-regulating unsafe behaviour and, on the other, improve significantly the desired safety outcome in the dilemma zone when compared to punishment alone. Interestingly, this type of binding pledges has been argued to be of relevance within other types of global conflicts and dilemmas, such as environmental governance (Barrett,, [2003](#bib.bib3), Cherry and McEvoy,, [2013](#bib.bib11)).
2 Models and Methods
---------------------
We first recall the DSAIR model with pairwise interactions then extend it with the option to bilaterally commit to acting safely and associated punishment of violations of such commitments.
The Evolutionary Game Theory (EGT) methods being used to analyse the models will then be described.
###
2.1 Summary of the DSAIR model and prior results
The DSAIR model (Han et al.,, [2020](#bib.bib28)) was originally defined as a two-player game repeated with a certain probability, consisting thus on average of W𝑊Witalic\_W rounds111An N-player version of this game was discussed also in (Han et al.,, [2020](#bib.bib28)). Yet in order to keep things easy to access, we focus here on the two-player scenario..
At each round of development, players gather benefits arising from their intermediate AI developments, subject to whether or not they chose to act UNSAFE or SAFE. Presuming some fixed benefit, b𝑏bitalic\_b, resulting from the AI market, the teams will share this gain proportionately to their development speed.
Accordingly, at every round of the race one can write a payoff matrix denoted by ΠΠ\Piroman\_Π with respect to row players i𝑖iitalic\_i, whose entries are denoted by ΠijsubscriptΠ𝑖𝑗\Pi\_{ij}roman\_Π start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT (j𝑗jitalic\_j corresponding to some column), as shown
| | | | |
| --- | --- | --- | --- |
| | Π=𝑆𝐴𝐹𝐸𝑈𝑁𝑆𝐴𝐹𝐸𝑆𝐴𝐹𝐸( −c+b2−c+bs+1) 𝑈𝑁𝑆𝐴𝐹𝐸sbs+1b2.\Pi=\bordermatrix{~{}&\textit{SAFE}&\textit{UNSAFE}\cr\textit{SAFE}&-c+\frac{b}{2}&-c+\frac{b}{s+1}\cr\textit{UNSAFE}&\frac{sb}{s+1}&\frac{b}{2}\cr}.roman\_Π = start\_ROW start\_CELL end\_CELL start\_CELL end\_CELL start\_CELL SAFE end\_CELL start\_CELL UNSAFE end\_CELL start\_CELL end\_CELL end\_ROW start\_ROW start\_CELL SAFE end\_CELL start\_CELL italic\_( end\_CELL start\_CELL - italic\_c + divide start\_ARG italic\_b end\_ARG start\_ARG 2 end\_ARG end\_CELL start\_CELL - italic\_c + divide start\_ARG italic\_b end\_ARG start\_ARG italic\_s + 1 end\_ARG end\_CELL start\_CELL italic\_) end\_CELL end\_ROW start\_ROW start\_CELL UNSAFE end\_CELL start\_CELL end\_CELL start\_CELL divide start\_ARG italic\_s italic\_b end\_ARG start\_ARG italic\_s + 1 end\_ARG end\_CELL start\_CELL divide start\_ARG italic\_b end\_ARG start\_ARG 2 end\_ARG end\_CELL start\_CELL end\_CELL end\_ROW . | | (1) |
The payoff matrix can be explained as follows. Firstly, wherever there is an interaction between two players selecting the SAFE action, each shall pay a cost c𝑐citalic\_c and the resulting benefit b𝑏bitalic\_b is shared. Differently, whenever interaction is between two players selecting the UNSAFE action, they shall share benefit b𝑏bitalic\_b without having had to pay cost c𝑐citalic\_c.
Whenever an UNSAFE choice is matched with a SAFE one, the SAFE choice necessitates a cost c𝑐citalic\_c and receives a (smaller) part b/(s+1)𝑏𝑠1b/(s+1)italic\_b / ( italic\_s + 1 ) of b𝑏bitalic\_b, whereas the UNSAFE choice collects a larger sb/(s+1)𝑠𝑏𝑠1sb/(s+1)italic\_s italic\_b / ( italic\_s + 1 ) whilst not ever having had to pay c𝑐citalic\_c. Note that ΠΠ\Piroman\_Π is a simplification of the matrix defined in (Han et al.,, [2020](#bib.bib28)) for, in the current time-scale, it was shown that the parameters as defined here sufficiently explain the obtained results.
We analyse the evolutionary outcomes of this game in a well-mixed and finite population consisting of Z𝑍Zitalic\_Z players. Given the choices each player can make and the fact that these choices need to be repeated for W𝑊Witalic\_W round, each adopts player one of the two following strategies (Han et al.,, [2020](#bib.bib28)):
* •
AS: complies every time with safety precautions, acting thus SAFE in every round.
* •
AU: complies not once with safety precautions, acting thus UNSAFE in every round.
The averaged payoffs for AS vs AU are expressed by payoff matrix
| | | | |
| --- | --- | --- | --- |
| | 𝐴𝑆𝐴𝑈𝐴𝑆( B2W+Π11Π12) 𝐴𝑈p(sBW+Π21)p(sB2W+Π22),\bordermatrix{~{}&\textit{AS}&\textit{AU}\cr\textit{AS}&\frac{B}{2W}+\Pi\_{11}&\Pi\_{12}\cr\textit{AU}&p\left(\frac{sB}{W}+\Pi\_{21}\right)&p\left(\frac{sB}{2W}+\Pi\_{22}\right)\cr},start\_ROW start\_CELL end\_CELL start\_CELL end\_CELL start\_CELL AS end\_CELL start\_CELL AU end\_CELL start\_CELL end\_CELL end\_ROW start\_ROW start\_CELL AS end\_CELL start\_CELL italic\_( end\_CELL start\_CELL divide start\_ARG italic\_B end\_ARG start\_ARG 2 italic\_W end\_ARG + roman\_Π start\_POSTSUBSCRIPT 11 end\_POSTSUBSCRIPT end\_CELL start\_CELL roman\_Π start\_POSTSUBSCRIPT 12 end\_POSTSUBSCRIPT end\_CELL start\_CELL italic\_) end\_CELL end\_ROW start\_ROW start\_CELL AU end\_CELL start\_CELL end\_CELL start\_CELL italic\_p ( divide start\_ARG italic\_s italic\_B end\_ARG start\_ARG italic\_W end\_ARG + roman\_Π start\_POSTSUBSCRIPT 21 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL italic\_p ( divide start\_ARG italic\_s italic\_B end\_ARG start\_ARG 2 italic\_W end\_ARG + roman\_Π start\_POSTSUBSCRIPT 22 end\_POSTSUBSCRIPT ) end\_CELL start\_CELL end\_CELL end\_ROW , | | (2) |
wherein, for presentation purposes alone, let us denote p=1−pr𝑝1subscript𝑝𝑟p=1-p\_{r}italic\_p = 1 - italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT (note that prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT was explained in the Introduction section).
As has been shown in (Han et al.,, [2020](#bib.bib28)), by contemplating where AU is risk-dominant against AS (cf. Methods below), then three distinct regions are identifiable within the parameter space s𝑠sitalic\_s-prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT (cf. Figure [1](#S1.F1 "Figure 1 ‣ Previous Developments. ‣ 1 Introduction ‣ Voluntary safety commitments provide an escape from over-regulation in AI development")): (I) if pr>1−13ssubscript𝑝𝑟113𝑠p\_{r}>1-\frac{1}{3s}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT > 1 - divide start\_ARG 1 end\_ARG start\_ARG 3 italic\_s end\_ARG, AU is risk-dominated by AS: safety compliance affords both the collectively preferred outcome and the one evolution selects; (II) if 1−13s>pr>1−1s113𝑠subscript𝑝𝑟11𝑠1-\frac{1}{3s}>p\_{r}>1-\frac{1}{s}1 - divide start\_ARG 1 end\_ARG start\_ARG 3 italic\_s end\_ARG > italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT > 1 - divide start\_ARG 1 end\_ARG start\_ARG italic\_s end\_ARG: even though safety compliance is the more desirable strategy for ensuring the highest collective outcome, the social learning dynamics leads the population to the state within which safety precautions have been mostly ignored; (III) if pr<1−1ssubscript𝑝𝑟11𝑠p\_{r}<1-\frac{1}{s}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT < 1 - divide start\_ARG 1 end\_ARG start\_ARG italic\_s end\_ARG (AU is risk-dominant against AS), then unsafe development is both collectively preferred and selected by the social learning dynamics.
So it is important to remember for the rest of the paper that UNSAFE actions are preferred and established in zone III, SAFE actions are preferred and established in zone I and a conflict exists between the individual and the collective in zone II as the former prefers UNSAFE actions and the latter SAFE actions.
Figure 2: Behavioural dominance in different zones for varying prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT, in presence of prior commitments (top row: panels a, b) and comparison of its overall unsafe behaviour against when there are no commitments (bottom row: panels c, d). The black dotted lines in panels a and b indicate the total unsafe frequency (i.e. the sum of AU-in and AU-out frequencies). The desired collective behaviour is indicated for each zone (i.e. unsafe in zone III and safe in zones I and III). We show results for two important scenarios: when efficient punishment can be made for a small cost (left column: sα=0.3subscript𝑠𝛼0.3s\_{\alpha}=0.3italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 0.3, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1) and when punishment is not highly efficient (right column:, sα=1subscript𝑠𝛼1s\_{\alpha}=1italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 1, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1).
Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100.
### Introducing commitment strategies
We now extend the DSAIR model with strategies that can bilaterally commit to safety compliant behavior: Before an interaction, participants can commit to play SAFE in each round. The commitment stands when all parties agree.
The players can refuse to commit, preferring to proceed without being pushed into the safe direction and being able to take risks. Those that committed but later select the UNSAFE action are potentially subject to sanctioning. Two sanctioning scenarios are considered here: (a) peer punishment (PP), which is performed by the co-player who kept her side of the deal, and (b) institutional punishment (IP), which is performed by a third-party that is not actively participating in the race for supremacy in some domain through AI (e.g. the European Union or United Nations). Each player has the freedom of behavioural choice whereby those who do not commit will not be punished when playing UNSAFE in the DSAIR model. We call the latter behavior ”honest” unsafe behavior whereas those that act unsafely after committing are referred to as ”dishonest”.
Sanctioning an opponent who played UNSAFE in a previous round consists in imposing a reduction sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT on the opponent’s speed (Han et al.,, [2021](#bib.bib25)). In case of PP, the punishing player also incurs a reduction sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT on her own speed.
Committing may also be costly for all that do as they give up other choices. A commitment cost ϵitalic-ϵ\epsilonitalic\_ϵ (per round) is thus introduced for those performing this pre-play action.
Figure 3: Transitions and stationary distributions when agreement is present (top row) against when it is absent (bottom row), for three regions. For clarity, only stronger transitions (than the ones in the opposite directions) are shown; no transition either way means neutral. The choice of sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT values were chosen to illustrate the main difference between with vs without commitment scenarios, in the three zones. Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100.
With the possibility of joining or not a commitment to behave safely and sanctioning dishonest unsafe behaviour, one can now define the possible strategies. AS and AU (as defined above) can either commit to safe actions, and furthermore, when involved in a commitment, decide whether to punish a dishonest co-player. If no commitment can be made, the player will select the UNSAFE action222We also consider the version where these players unconditionally play SAFE regardless of the commitment. The safety outcomes in all regions are similar, just the strategies’ dominance is slightly different. Results are provided in Appendix (see Figures A3 and A4).. The choices listed before lead to five strategies:
1. 1.
AS-in: willing to commit, plays SAFE when commitment is in place and UNSAFE otherwise, but does not use (costly) punishment. This strategy can be considered a second-order free-rider on the punishment effort of others;
2. 2.
AS-out: does not commit, but always selects the SAFE action;
3. 3.
AU-in: claiming to commit to the SAFE action, but always plays UNSAFE in the interaction. This strategy makes a commitment, trying to exploit safe players who only want to interact under an agreement;
4. 4.
AU-out: does not commit and plays UNSAFE in the interaction. This strategy wants to freely take risk or innovate without any repercussions otherwise imposed by a commitment to follow SAFE actions;
5. 5.
PS: willing to commit and plays SAFE when the other player also commits; plays UNSAFE otherwise, and also punishes an UNSAFE action of a co-player that committed with her. This strategy is only present in the case of PP, and is not present in the IP sanctioning model.
###
2.2 Evolutionary Dynamics of Finite Populations
Herein are adopted the methods of EGT for finite populations (Sigmund,, [2010](#bib.bib41), Traulsen et al.,, [2006](#bib.bib44), Hindersin et al.,, [2019](#bib.bib30)), whether in the analytical or numerical results obtained here. In such settings, the payoffs of players stand for their social *success* or *fitness*, and the evolutionary dynamics is shaped by social learning, in accordance to which the players that are most successful will tend more often to be copied by other players. The so-called pairwise rule of comparison is utilised to model social learning (Traulsen et al.,, [2006](#bib.bib44)), which ensures a player A𝐴Aitalic\_A with fitness fAsubscript𝑓𝐴f\_{A}italic\_f start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT resorts to adopt the strategy of player B𝐵Bitalic\_B with fitness fBsubscript𝑓𝐵f\_{B}italic\_f start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT with a probability established by the Fermi function,
PA,B=(1+e−β(fB−fA))−1subscript𝑃𝐴𝐵superscript1superscript𝑒𝛽subscript𝑓𝐵subscript𝑓𝐴1P\_{A,B}=\left(1+e^{-\beta(f\_{B}-f\_{A})}\right)^{-1}italic\_P start\_POSTSUBSCRIPT italic\_A , italic\_B end\_POSTSUBSCRIPT = ( 1 + italic\_e start\_POSTSUPERSCRIPT - italic\_β ( italic\_f start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT - italic\_f start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT ) end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT,
where the intensity of selection is conveniently described by β𝛽\betaitalic\_β. In a population wherein several strategies are in co-presence, their long-term frequency can be computed simply by calculating the stationary distribution of a Markov chain the states of which represent each strategy. Absent behavioural exploration or mutations, the end states of evolution are inevitably monomorphic. Meaning whenever
such a state is reached, escape by imitation is impossible. Hence, we presume further that, given some mutation probability, each agent may freely explore its behavioural space (which consists of the two actions, UNSAFE and SAFE, in our case), by randomly adopting an action as a result of mutation. At the limit of a small probability of mutating, the population is comprised of at most one of two strategies at any time. Accordingly, the social dynamics is describable utilising a Markov Chain, in which each state represents a monomorphic population whose transition probabilities to another state are expressed by the fixation probability of one single mutant (Imhof et al.,, [2005](#bib.bib31), Nowak et al.,, [2004](#bib.bib36)). The Markov Chain’s stationary distribution depicts the average time the whole population spends at each monomorphic end state (see the examples in Figure [3](#S2.F3 "Figure 3 ‣ Introducing commitment strategies ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development") for illustration).
Let πX,Ysubscript𝜋𝑋𝑌\pi\_{X,Y}italic\_π start\_POSTSUBSCRIPT italic\_X , italic\_Y end\_POSTSUBSCRIPT denote the payoff some strategist X𝑋Xitalic\_X gathers from a pairwise interaction with some strategist Y𝑌Yitalic\_Y (as defined in the payoff matrix). Assume there exist two strategies at most in the population, for example k𝑘kitalic\_k agents using strategy A (0≤k≤Z0𝑘𝑍0\leq k\leq Z0 ≤ italic\_k ≤ italic\_Z) and (Z−k)𝑍𝑘(Z-k)( italic\_Z - italic\_k ) agents using instead strategy B. Hence, the (average) payoff of agents using A and B can be formulated, respectively, as
| | | | |
| --- | --- | --- | --- |
| | ΠA(k)=(k−1)πA,A+(Z−k)πA,BZ−1,ΠB(k)=kπB,A+(Z−k−1)πB,BZ−1.formulae-sequencesubscriptΠ𝐴𝑘𝑘1subscript𝜋𝐴𝐴𝑍𝑘subscript𝜋𝐴𝐵𝑍1subscriptΠ𝐵𝑘𝑘subscript𝜋𝐵𝐴𝑍𝑘1subscript𝜋𝐵𝐵𝑍1\begin{split}\Pi\_{A}(k)&=\frac{(k-1)\pi\_{A,A}+(Z-k)\pi\_{A,B}}{Z-1},\\
\Pi\_{B}(k)&=\frac{k\pi\_{B,A}+(Z-k-1)\pi\_{B,B}}{Z-1}.\end{split}start\_ROW start\_CELL roman\_Π start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT ( italic\_k ) end\_CELL start\_CELL = divide start\_ARG ( italic\_k - 1 ) italic\_π start\_POSTSUBSCRIPT italic\_A , italic\_A end\_POSTSUBSCRIPT + ( italic\_Z - italic\_k ) italic\_π start\_POSTSUBSCRIPT italic\_A , italic\_B end\_POSTSUBSCRIPT end\_ARG start\_ARG italic\_Z - 1 end\_ARG , end\_CELL end\_ROW start\_ROW start\_CELL roman\_Π start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT ( italic\_k ) end\_CELL start\_CELL = divide start\_ARG italic\_k italic\_π start\_POSTSUBSCRIPT italic\_B , italic\_A end\_POSTSUBSCRIPT + ( italic\_Z - italic\_k - 1 ) italic\_π start\_POSTSUBSCRIPT italic\_B , italic\_B end\_POSTSUBSCRIPT end\_ARG start\_ARG italic\_Z - 1 end\_ARG . end\_CELL end\_ROW | | (3) |
As a result, at each step in time, the probability of changing by ±plus-or-minus\pm±1 of the number of k𝑘kitalic\_k agents using strategy A is specified as (Traulsen et al.,, [2006](#bib.bib44))
| | | | |
| --- | --- | --- | --- |
| | T±(k)=Z−kZkZ[1+e∓β[ΠA(k)−ΠB(k)]]−1.superscript𝑇plus-or-minus𝑘𝑍𝑘𝑍𝑘𝑍superscriptdelimited-[]1superscript𝑒minus-or-plus𝛽delimited-[]subscriptΠ𝐴𝑘subscriptΠ𝐵𝑘1T^{\pm}(k)=\frac{Z-k}{Z}\frac{k}{Z}\left[1+e^{\mp\beta[\Pi\_{A}(k)-\Pi\_{B}(k)]}\right]^{-1}.italic\_T start\_POSTSUPERSCRIPT ± end\_POSTSUPERSCRIPT ( italic\_k ) = divide start\_ARG italic\_Z - italic\_k end\_ARG start\_ARG italic\_Z end\_ARG divide start\_ARG italic\_k end\_ARG start\_ARG italic\_Z end\_ARG [ 1 + italic\_e start\_POSTSUPERSCRIPT ∓ italic\_β [ roman\_Π start\_POSTSUBSCRIPT italic\_A end\_POSTSUBSCRIPT ( italic\_k ) - roman\_Π start\_POSTSUBSCRIPT italic\_B end\_POSTSUBSCRIPT ( italic\_k ) ] end\_POSTSUPERSCRIPT ] start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT . | | (4) |
The fixation probability of a single mutant adopting strategy A,in a population (Z−1)𝑍1(Z-1)( italic\_Z - 1 ) of agents adopting B, is defined by (Traulsen et al.,, [2006](#bib.bib44), Nowak et al.,, [2004](#bib.bib36))
| | | | |
| --- | --- | --- | --- |
| | ρB,A=(1+∑i=1Z−1∏j=1iT−(j)T+(j))−1.subscript𝜌𝐵𝐴superscript1superscriptsubscript𝑖1𝑍1superscriptsubscriptproduct𝑗1𝑖superscript𝑇𝑗superscript𝑇𝑗1\rho\_{B,A}=\left(1+\sum\_{i=1}^{Z-1}\prod\_{j=1}^{i}\frac{T^{-}(j)}{T^{+}(j)}\right)^{-1}.italic\_ρ start\_POSTSUBSCRIPT italic\_B , italic\_A end\_POSTSUBSCRIPT = ( 1 + ∑ start\_POSTSUBSCRIPT italic\_i = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_Z - 1 end\_POSTSUPERSCRIPT ∏ start\_POSTSUBSCRIPT italic\_j = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_i end\_POSTSUPERSCRIPT divide start\_ARG italic\_T start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT ( italic\_j ) end\_ARG start\_ARG italic\_T start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT ( italic\_j ) end\_ARG ) start\_POSTSUPERSCRIPT - 1 end\_POSTSUPERSCRIPT . | | (5) |
When considering a set {1,…,s}1…𝑠\{1,...,s\}{ 1 , … , italic\_s } of different strategies, such probabilities of fixation define the Markov Chain transition matrix M={Tij}i,j=1s𝑀superscriptsubscriptsubscript𝑇𝑖𝑗𝑖𝑗
1𝑠M=\{T\_{ij}\}\_{i,j=1}^{s}italic\_M = { italic\_T start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT } start\_POSTSUBSCRIPT italic\_i , italic\_j = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_s end\_POSTSUPERSCRIPT, with Tij,j≠i=ρji/(s−1)subscript𝑇𝑖𝑗𝑗
𝑖subscript𝜌𝑗𝑖𝑠1T\_{ij,j\neq i}=\rho\_{ji}/(s-1)italic\_T start\_POSTSUBSCRIPT italic\_i italic\_j , italic\_j ≠ italic\_i end\_POSTSUBSCRIPT = italic\_ρ start\_POSTSUBSCRIPT italic\_j italic\_i end\_POSTSUBSCRIPT / ( italic\_s - 1 ) and Tii=1−∑j=1,j≠isTijsubscript𝑇𝑖𝑖1subscriptsuperscript𝑠formulae-sequence𝑗1𝑗𝑖subscript𝑇𝑖𝑗T\_{ii}=1-\sum^{s}\_{j=1,j\neq i}T\_{ij}italic\_T start\_POSTSUBSCRIPT italic\_i italic\_i end\_POSTSUBSCRIPT = 1 - ∑ start\_POSTSUPERSCRIPT italic\_s end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_j = 1 , italic\_j ≠ italic\_i end\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i italic\_j end\_POSTSUBSCRIPT. The normalized eigenvector of the transposed matrix of M𝑀Mitalic\_M associated with eigenvalue 1 produces the above defined stationary distribution (Imhof et al.,, [2005](#bib.bib31)), depicting the relative time the population stays adopting each of the strategies.
###
2.3 Risk-dominance
A major standpoint of comparison of two strategies A and B is in which direction the transition is more probable or stronger, the one of some B mutant fixating in a population of agents that employ A, ρA,Bsubscript𝜌𝐴𝐵\rho\_{A,B}italic\_ρ start\_POSTSUBSCRIPT italic\_A , italic\_B end\_POSTSUBSCRIPT, or that of a mutant A fixating in the population of agents that employ B, ρB,Asubscript𝜌𝐵𝐴\rho\_{B,A}italic\_ρ start\_POSTSUBSCRIPT italic\_B , italic\_A end\_POSTSUBSCRIPT. At the limit, for a large enough population size (i.e. a large Z𝑍Zitalic\_Z), the condition simplifies to (Sigmund,, [2010](#bib.bib41))
| | | | |
| --- | --- | --- | --- |
| | πA,A+πA,B>πB,A+πB,B.subscript𝜋𝐴𝐴subscript𝜋𝐴𝐵subscript𝜋𝐵𝐴subscript𝜋𝐵𝐵\pi\_{A,A}+\pi\_{A,B}>\pi\_{B,A}+\pi\_{B,B}.italic\_π start\_POSTSUBSCRIPT italic\_A , italic\_A end\_POSTSUBSCRIPT + italic\_π start\_POSTSUBSCRIPT italic\_A , italic\_B end\_POSTSUBSCRIPT > italic\_π start\_POSTSUBSCRIPT italic\_B , italic\_A end\_POSTSUBSCRIPT + italic\_π start\_POSTSUBSCRIPT italic\_B , italic\_B end\_POSTSUBSCRIPT . | | (6) |
Figure 4: Frequency of unsafe in all regions for varying sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT, when commitment is absent (top row) vs when it is in use (bottom row). Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100.
3 Results
----------
We first focus on analysing the system of self-enforcing bilateral commitment (PP) then showing the results of bilateral commitment complemented with institutional enforcement (IP).
###
3.1 Self-enforcing bilateral commitments
In Figure [2](#S2.F2 "Figure 2 ‣ 2.1 Summary of the DSAIR model and prior results ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development") (top row), we show the stationary distributions of the five strategies for varying prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT across the three zones I, II and III.
We show results for two important scenarios: when efficient punishment can be made for a small cost (left column: sα=0.3subscript𝑠𝛼0.3s\_{\alpha}=0.3italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 0.3, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1) and when punishment is not highly efficient (right column:, sα=1subscript𝑠𝛼1s\_{\alpha}=1italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 1, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1). Punishment is considered efficient when its effect (sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT) on the development speed of the player performing the punishment is significantly smaller than the effect (sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT) it has on the player undergoing the punishment.
In both cases, AU-out dominates when prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT is small (zone III), PS dominates when it is intermediate (first part of zone II), while AS-out dominates when it is large (part of zone II and zone I). It is important to notice that these results reflect the most desirable outcomes for all three zones: risky innovation in III and safety compliance in I and II (see the black line in the bottom row of Figure [2](#S2.F2 "Figure 2 ‣ 2.1 Summary of the DSAIR model and prior results ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development")). Moreover, a more efficient punishment leads to better safety outcome in zone II (see panels c and d). It is worth noting that inefficient punishment may result in the presence of AU-in strategists, in other words cheaters. It is thus important in this self-organised commitment solution that participants have an effective mechanism to punish commitment violators. One solution could be directly affecting the player’s publicly available reputation.
To clarify the benefit of voluntary bilateral commitment, Figure [2](#S2.F2 "Figure 2 ‣ 2.1 Summary of the DSAIR model and prior results ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development") (bottom row) compares the overall frequency of unsafe/risk-taking when commitment is possible (black line) vs when it is absent (red line). In the absence of bilateral commitments,
over-regulation occurs, i.e. safety compliance is abundant but not desired for a large part of zone III when efficient punishment can be carried out for a small cost (panel c), while a safety dilemma occurs (i.e. safety compliance is desired but infrequent) for a large part of zone II when sanctioning is not highly efficient.
These observations can be better understood by examining
the transitions and stationary distributions in Figure [3](#S2.F3 "Figure 3 ‣ Introducing commitment strategies ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development").
In region III (first column, low risk, where risky innovation is desired), low-cost highly-efficient punishment can lead to significant reduction of innovation since PS dominates AU in absence of bilateral commitments (panel d). Addition of AU-out provides an escape as this strategy is not punished since it never commits to safe course of actions, thereby dominating PS (panel a). In region II (dilemma zone, safety is wanted but needs to be enforced), when punishment is not highly efficient, there is a large amount of unwanted AU in the population (panel e). As AU-out is dominated by PS, since PS plays also UNSAFE when not having a commitment partner, it leads to a lower frequency of unsafe behaviour (comparing panels b and e). In region III, addition AU-out does not change the desired outcome of the dominance of safety compliance.
In Figure [4](#S2.F4 "Figure 4 ‣ 2.3 Risk-dominance ‣ 2 Models and Methods ‣ Voluntary safety commitments provide an escape from over-regulation in AI development"), we show that these remarkable observations regarding voluntary bilateral commitment are robust for different regimes of punishment effectiveness. In particular, when comparing the unsafe frequency with commitments against when it is absent for varying sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT.
In region III, over-regulation occurs when there is only punishment whenever sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT is sufficiently small and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT is large (purple area, top row, first column), but when an agreement is in place that is not the case (bottom row, first column). In region II with the lower range of prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT (second column, pr=0.6subscript𝑝𝑟0.6p\_{r}=0.6italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT = 0.6), unsafe frequency is lower in the latter in for most sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT. In the higher range of region II and region III (third and fourth columns, with pr=0.75subscript𝑝𝑟0.75p\_{r}=0.75italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT = 0.75 and pr=0.9subscript𝑝𝑟0.9p\_{r}=0.9italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT = 0.9), desired safe behaviour is dominant in both cases.
It’s noteworthy that all results are robust also for other regimes of weaker or stronger selection intensities, i.e. β=0.1𝛽0.1\beta=0.1italic\_β = 0.1 or β=10𝛽10\beta=10italic\_β = 10 (see e.g. Figures [A1](#S5.F1 "Figure A1 ‣ 5.2 Other scenarios of modelling ‣ 5 Appendix ‣ Voluntary safety commitments provide an escape from over-regulation in AI development") and A2 in Appendix).
###
3.2 Institutionally governed bilateral commitments
When assuming that it is an institution that governs the dynamics of the race to supremacy through AI in some domain, punishing unsafe behaviour if a bilateral commitment is violated. Similarly to PP, IP reduces the development speed of a participant that selected the UNSAFE action by sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT.
On the one hand, when no bilateral commitments can be made, the population consists of two strategies, AS and AU, where AU’s speed will be s−sβ𝑠subscript𝑠𝛽s-s\_{\beta}italic\_s - italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT when being punished (by the institution). On the other hand, when voluntary commitments are possible, there will only be the four strategies, AS-in, AU-in, AS-out, and AU-out. There is no PS strategy as there is no peer sanctioning, which implies also no parameter sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT given that the institution is not part of the population333Future work might look at how to minimise the cost and efforts from the institution while ensuring a desired level of safety compliance (Han and Tran-Thanh,, [2018](#bib.bib29), Han et al.,, [2018](#bib.bib22), Chen et al.,, [2015](#bib.bib9), Wang et al.,, [2019](#bib.bib48), Couto et al.,, [2020](#bib.bib14)).
Concretely, the institution can only sanction AU-in when both players committed. In that case, its speed will be reduced, from s𝑠sitalic\_s to s−sβ𝑠subscript𝑠𝛽s-s\_{\beta}italic\_s - italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT. AU-out as before will be free of any sanction.
In Figure [5](#S3.F5 "Figure 5 ‣ 3.2 Institutionally governed bilateral commitments ‣ 3 Results ‣ Voluntary safety commitments provide an escape from over-regulation in AI development"), we compare the effect of institutional sanctioning with and without the option of having prior voluntary bilateral commitments between the two players. The outcomes are similar to the case of PP discussed in the previous section. In particular, one can observe that the presence of commitments can significantly improve safety outcome in region II without being detrimental to the desired behaviour and thus social outcome in the other two regions. Without commitments, institutional punishment leads again to over-regulation of unsafe/innovation in region III, which, as was mentioned before, is not a desired outcome. Moreover, with a smaller sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT (compare top and bottom rows of the figure), the outcomes remain similar, with a lower level of over-regulation but weaker results for region II.
Figure 5: Institutional punishment, with (left column) or without a safety agreement (right column). We report the frequency of unsafe behaviour (UNSAFE) in a population of AU and AS in the latter case and a population of AS-in, AU-in, AS-out, AU-out, in the former case. An institutional punishment reduces the speed of AU in the latter and AU-in (when an agreement is formed) in the former, by sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1 (top row) and by sβ=0.5subscript𝑠𝛽0.5s\_{\beta}=0.5italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 0.5 (bottom row). The lines within the figures are as in Figure 1. Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100.
In short, we observe that freedom of choice between bilateral commitments to safety regulations (which can be self-enforcing or institutionally enforced) and unilateral risky/unsafe endeavours provides an efficient solution to ensure desirable outcomes in all scenarios of the DSAIR model. First, it alleviates the problem of over-regulation when the risk is low (which happens when punishment can be very efficient), while at the same time it reduces the frequency of unsafe behaviour in the dilemma zone (region II). Furthermore, the desirable safety compliance outcomes in region I remain unaffected.
Discussion and conclusions
--------------------------
This paper proposes and analyses, using a multi-agent and population dynamics modelling approach, a novel solution that promotes desirable safety compliance in a technology development race, while at the same time avoiding stifling beneficial innovation due to over-regulation. We base our study on a previously proposed EGT model (Han et al.,, [2020](#bib.bib28)) that describes the dynamics of a competition between safety and risk-taking (unsafe) behaviours, within an (alleged) race for supremacy through AI in a marketable domain.
We show that, by allowing race participants the freedom of choosing to enter or not in bilateral commitments to act safely and avoid risks, accepting thus to be sanctioned in case of misbehavior, high levels of the most beneficial behaviour as a whole, are achieved in all regions of the parameter space.
This system of voluntary bilateral commitments, either depending on sanctioning actions of peers or by an institution, provides a mechanism to overcome the problems associated with over-regulation, which might occur whenever risk taking behaviour, which may be perceived as unsafe, is unconditionally penalised without taking into account the true risk level (in relation to the cost and benefit of the safety compliance and of the opting for risky behaviour), as shown in (Han et al.,, [2021](#bib.bib25)).
On the one hand, allowing for participants to explicitly commit to safety precautions in their development process ensures higher levels of safety compliance when that is desired, because then unsafe/risk-taking strategies become either clearly identified (AU-out) and suitably coped with, or punished through the binding commitment (AU-in). On the other hand, the commitment option enables the (honest) risk-taking behaviour (namely, AU-out) to prevail whenever it is collectively preferred, since then it is not punished, as a result of not joining the commitment in the first place.
There have been several theoretical modelling studies based on EGT, showing the benefit of establishing prior commitments or agreements for promoting the evolution of certain positive behaviours, such as cooperation and coordination in a population of self-regarding agents, see e.g. (Han et al.,, [2013](#bib.bib27), Sasaki et al.,, [2015](#bib.bib39), [Han et al., 2017a,](#bib.bib23) , Ogbo et al.,, [2021](#bib.bib37)). Behavioral experiments with human subjects have also been performed showing the promoting role of commitments for cooperative behaviours, see e.g. (Chen and Komorita,, [1994](#bib.bib10), Cherry and McEvoy,, [2013](#bib.bib11)). In all such contexts, it is well-defined what is the correct (positive) behaviour most beneficial collectively and thus the one that should be promoted.
In our case that no longer holds, since whether or not a behaviour is the most beneficial collectively depends on the behavioural region in which the race occurs at the time. This region is not well defined, especially when data is not abundantly available. For example, we might need to wait for a technology to be audited and even largely adopted by users to know the level of risk associated with a particular technology (but then it might be too late already to provide pertinent regulations). Such level of uncertainty can also lead to other externalities, such as the emergence of strongly polarized positions (Domingos et al.,, [2020](#bib.bib17)), an effect not yet characterized by our model.
Interestingly, we show here that establishing the possibility of a voluntary bilateral commitments is as well highly effective in promoting collectively desired/preferred behaviour in all cases.
Commitment to cooperative actions in social dilemmas work because they allow cooperators to avoid free-riding strategies, for the case of an AI race, because they permit players the freedom to follow their preferred course of development without being punished (as long as they are being honest). One can also see it as a particular form of binding signal or pledge. The results of this work are in line with Hadfield’s arguments (Hadfield,, [2017](#bib.bib21)) in the general context of technology globalisation; that is, legal contracts and institutions should allow freedom (via coordination) instead of exacting it from individuals in a population. For the sake of clarity, here we illustrate this idea with a pairwise race model; however, as shown in (Han et al.,, [2020](#bib.bib28)), this model can be easily extended to N-player interactions, with a concomitant increase in complexity, yet leading to the same qualitative messages. Future work may explicitly consider N-player commitments and more complex communication and bottom-up dynamics typical of this type of agreements, see e.g. ([Han et al., 2017a,](#bib.bib23) , [Han et al., 2017b,](#bib.bib26) , Vasconcelos et al.,, [2013](#bib.bib46)).
Last but not least, in (O’Keefe et al.,, [2020](#bib.bib38)) the authors propose the so-called Windfall Clause for mediating the tension in AI competition. It is proposed there to arrange prior commitments or agreements from the race participants (e.g., companies, governments) for sharing a large part of the windfall benefit (i.e. the large prize B𝐵Bitalic\_B in our model) with society. As discussed in that paper, it requires significant legal infrastructures to be put in place to enable such a mechanism. Our approach does not so require because legal contracts can be used to ensure compliance with prior safety agreements.
It is noteworthy that, although we focus in this paper on an AI development race, the model proposed can be more generally applicable to other kinds of long-term situations of competition, such as the development of technological innovation and its racing for patents in which there is a significant advantage (i.e. a large B𝐵Bitalic\_B) to be gotten in being one of the first ever to reach an important target (Denicolò and Franzoni,, [2010](#bib.bib16), Campart and Pfister,, [2014](#bib.bib7), Lemley,, [2012](#bib.bib33)).
Other important domains involve pharmaceutical and vaccines development race, where companies might attempt to cut a few corners by not following strictly the safe clinical trial protocols, with the view to be the first ones to develop and put on the market some pharmaceutical product, in order to reap the highest possible share of benefit in the market (Abbott et al.,, [2009](#bib.bib1), Burrell and Kelly,, [2020](#bib.bib6)). However, certain aspects or factors of the current model might need to be revised.
For example, the large profit B𝐵Bitalic\_B for a vaccine development race would be more difficult to achieve than in the case of AI, since a developed vaccine needs to be approved by suitable authorities before users have trust in it and use it.
Vaccine developers only can generate the profit B𝐵Bitalic\_B if they could show evidence of the safety of the vaccine and that they have closely followed all the safety procedures.
In case of AI, the winning developers can simply deploy the technologies and get the profit B (in most cases). One can expand the current model to capture also the vaccine race (and also generalised the current AI race model) by for example having a new parameter to capture when B𝐵Bitalic\_B can be generated for the race winners, which depends on the nature of the race (e.g. vaccine vs AI) and also the frequency of safety compliance in the past. our future works will address these issues.
4 Acknowledgements
-------------------
T.A.H., L.M.P., T.L. and T.C., are supported by Future of Life Institute grant RFP2-154. L.M.P. acknowledges support from FCT/MEC NOVA LINCS PEst UID/CEC/04516/2019. F.C.S. acknowledges support from FCT Portugal (grants PTDC/EEI-SII/5081/2014, PTDC/MAT/STA/3358/2014). T.L. acknowledges support by the FuturICT2.0 (www.futurict2.eu) project funded by the FLAG-ERA JCT 2016. T.A.H. is also supported by a Leverhulme Research Fellowship (RF-2020-603/9).
5 Appendix
-----------
###
5.1 Effect of larger intensity of selection
See Figures A1 and A2.
###
5.2 Other scenarios of modelling
See Figures A3 and A4.
Figure A1: Same as Figure 2, with other values of β𝛽\betaitalic\_β.
Figure A2: Frequency of unsafe in all regions for varying sαsubscript𝑠𝛼s\_{\alpha}italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and sβsubscript𝑠𝛽s\_{\beta}italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT, when safety agreement is in use, for different values of intensities of selection β𝛽\betaitalic\_β. Similar observations as in the main text. Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, Z=100𝑍100Z=100italic\_Z = 100.
Figure A3: When AU-in always plays SAFE in an interaction. Behavioural dominance of in different zones for varying prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT, in presence of prior commitment (top row: panels a, b) and comparison of its overall unsafe behaviour against when commitment is absent (bottom row: panels c, d). We show results for two important scenarios: when efficient punishment can be made for a small cost (left column: sα=0.3subscript𝑠𝛼0.3s\_{\alpha}=0.3italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 0.3, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1) and when punishment is not highly efficient (right column:, sα=1subscript𝑠𝛼1s\_{\alpha}=1italic\_s start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = 1, sβ=1subscript𝑠𝛽1s\_{\beta}=1italic\_s start\_POSTSUBSCRIPT italic\_β end\_POSTSUBSCRIPT = 1).
In both cases, AU-out dominates when prsubscript𝑝𝑟p\_{r}italic\_p start\_POSTSUBSCRIPT italic\_r end\_POSTSUBSCRIPT is small (zone III), PS dominates when it is intermediate (first part of zone II), while PS and AU-in together dominate when it is large (part of zone II and zone I). That results in the fact that when a commitment is present desirable outcomes are achieved in all three zones: unsafe/innovation in III and safe in I and II, see bottom row. In absence of a commitment,
over-regulation occurs (i.e. safe is abundant but not desired) for a large part of zone I when efficient punishment can be done for a small cost (panel c), while safety dilemma occurs (i.e. safe is desired but infrequent) for a large part of zone II when punishment is not highly efficient. Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100.
Figure A4: When AU-in always play SAFE in an interaction. Transitions and stationary distributions when a commitment is present (top row) against when it is absent (bottom row), for three regions (Only stronger transitions are shown; no transition either way means neutral). In region III (first column, low risk, where UNSAFE or innovation is wanted), low-cost highly efficient punishment can lead to significant reduction of innovation since PS dominates AU in absence of a commitment (panel d). Addition of AU-out provides an escape as this strategy is not punished, thereby dominating PS (panel a). In region II (dilemma zone, safety is wanted), when punishment is not highly efficient, there is still some amount of unwanted AU in the system. As AU-out is dominated by PS since PS plays UNSAFE in absence of an commitment, it leads to lower unsafe behaviour. In region III, addition AU-out does not change the desired outcome of safety. Parameters: b=4𝑏4b=4italic\_b = 4, c=1𝑐1c=1italic\_c = 1, s=1.5𝑠1.5s=1.5italic\_s = 1.5, W=100𝑊100W=100italic\_W = 100, B=104𝐵superscript104B=10^{4}italic\_B = 10 start\_POSTSUPERSCRIPT 4 end\_POSTSUPERSCRIPT, β=1𝛽1\beta=1italic\_β = 1, Z=100𝑍100Z=100italic\_Z = 100. |
713a9a50-044f-48f2-93e2-7227c5b4f7c6 | trentmkelly/LessWrong-43k | LessWrong | Formalization is a rationality technique
We are interested in developing practical techniques of rationality. One practical technique, used widely and successfully in science and technology is formalization, transforming a less-formal argument into a more-formal one. Despite its successes, formalization isn't trivial to learn, and schools rarely try to teach general techniques of thinking and deciding. Instead, schools generally only teach domain-specific reasoning. We end up with graduates who can apply formalization skillfully inside of specific domains (e.g. electrical engineering or biology), but fail to apply, or misapply, their skills to other domains (e.g. politics or religion).
A side excursion, to be used as an example:
> Is it true that a real-world decision can be perfectly justified via a mathematical proof? No.
>
> Here is one reason: Any real-world "proof", whether it is a publication in a math journal or the output of a computer proof-checker, might contain a mistake. Human mathematicians routinely make mistakes in publications. The computer running the proof-checker is physical, and might be struck by a cosmic ray. Even if the computer were perfect, the proof-checker might have a bug in it.
>
> Here is another reason: Even if you had a proof known to be mistake free, mathematical proof always reasons from assumptions to conclusions. To apply the proof to the real world, you must make an informal argument that the assumptions apply in this real-world situation, and a second informal argument that the conclusion of the proof corresponds to the real-world decision. These first and last links in the chain will still be there, no matter how rigorous and trusted the proof is. The problem with duplicating real-world spheres with Banach-Tarski isn't that there's a mistake in the proof, it's that the real-world spheres don't fit the assumptions.
If you were fitting this argument into your beliefs, you might produce a number, a "gut" estimate of how likely this informal argument is wrong. Can we |
2924e50c-de7f-4a73-b5ab-6987ba4287e1 | trentmkelly/LessWrong-43k | LessWrong | What's wrong with being dumb?
I like to make fun of my dog sometimes for being dumb. Affectionately. I'll say things like "That's not for you dumb dumbs!" when he tries to eat food that will clearly end up giving him a stomach ache and making him nauseous.
A few years ago I did this in front of my girlfriend's mom. She laughed and told me to stop making fun of him. Perhaps due to my instinct to be disagreeable, my response was "What's wrong with being dumb?".
My dog doesn't care about intelligence. He just cares about cuddles, love, being part of the pack, food, naps, and some good sniffs. And I for one endorse those values. So I'll ask the question again: what's wrong with being dumb?
Well, intelligence is a useful characteristic. As a society, by rewarding intelligence with higher social status, we incentivize things like scientific discoveries and new technologies. And by punishing a lack of intelligence — dumbness — we prevent people from injecting bleach into their veins. Or at least reduce the number of times that it happens.
But that's looking at things through a macroscopic, society-oriented lens. What if we zoom in and, to stereotype, look at a dumb farmer named Frank?
Frank is dumb. You can reward him for being smart all you want, but that's just not gonna get him to take any community college classes. He is content with his life. It works for him. He enjoys his job, loves his family, owns his home, and just isn't interested in change.
He does do some dumb things that are instrumentally harmful though. For example, he doesn't believe in medicine and hasn't seen a doctor in 30 years. Seeing a doctor would, let's assume for the hypothetical, extend his life by 10 years. But in practice, you're just not gonna change his mind. So why badger him about it? Why piss him off and frustrate yourself along the way? And similarly, he's not gonna be winning any Nobel Prizes or inventing new technologies, so why push for those community college classes?
We have a confusing situation here. If |
577058c5-acd4-43ae-82aa-73c2bc0f9243 | trentmkelly/LessWrong-43k | LessWrong | [link] Cybathlon (an Olympics for bionic athletes)
News: http://www.theverge.com/2014/3/26/5551180/switzerland-to-host-the-first-cybathlon-an-olympics-for-bionic
I have been dreaming about the day the Paralympics would become more popular than the Olympics because of advance technologies that would help handicapped human cyborgs to perform better than non-handicapped non-cyborg/biological humans. But now, it seems Cybathlon will take the lead in this area. I'm looking forward to watch this awesome competition! |
3b140157-7afb-434c-8d33-0e2a1d234a67 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington, D.C.: Intro to Confirmation Bias or "Encore" Singing Game Meetup
Discussion article for the meetup : Washington, D.C.: Intro to Confirmation Bias or "Encore" Singing Game Meetup
WHEN: 10 August 2014 03:00:00PM (-0400)
WHERE: National Portrait Gallery
Assuming that the forecast of inclement weather[1] holds, we will meet at the Kogod Courtyard of the National Portrait Gallery (8th and F Sts or 8th and G Sts NW, go straight past the information desk from either entrance) for a brief introduction to the idea of confirmation bias. We will congregate in the courtyard between 3:00 and 3:30 p.m., and the introduction to confirmation bias should run from 3:30 to about 4:00. The remainder of the meetup will be left open for unstructured discussion and socializing.
If we discover that the weather is likely to be fair, an announcement will be made Saturday night or Sunday morning via the Less Wrong DC Google Group - lesswrong-dc@googlegroups.com - moving the meetup to the Navy Memorial (south on 8th St from the Portrait Gallery, between D St and Pennsylvania Ave) to play the singing game "Encore". Again, we plan to congregate between 3:00 and 3:30 p.m. In this case, the game will start at 3:30 and likely conclude between 5 and 6.
[1] At the time of this announcement, the Area Forecast Discussion issued by the National Weather Service announced "unsettled [i.e. likely to produce rain] conditions this weekend into early next week".
Discussion article for the meetup : Washington, D.C.: Intro to Confirmation Bias or "Encore" Singing Game Meetup |
f33e7894-5513-41bd-94b4-ddceda8d2de3 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Conditions under which misaligned subagents can (not) arise in classifiers
**Core claim**: *Misaligned subagents are very unlikely to arise in a classification algorithm unless that algorithm is directly or indirectly (e.g. in a subtask) modeling interactions through time at a significant level of complexity.*
**Definition 1**: Agent - a function from inputs and internal state (or memory) to an output / action and new internal state. Note that this includes things that would not usually be considered as "agents" - e.g. plants or bacteria. Also note that not all "agents" of this type have consistent (or even coherent) "goals"
This definition of agent might be considered too broad; the reason I have decided to use it is that I believe it covers basically everything that could be dangerous - if an AI is not an agent under this definition, then I think it is extremely likely that this AI would be safe.
**Definition 2**: A function that was selected by an optimization procedure has a misaligned subagent if it spawns a subprocess that is an agent whose "goals" are different from (and potentially in conflict with) the optimization criteria.
Example: Consider an optimization process that selects for functions that accurately predict human actions, and assume that this optimization process finds a function that does this prediction by creating an extremely accurate simulations of humans. These simulations would be misaligned subagents, since humans are agents and the goals of the simulations would likely be very different from "predict human actions accurately".
For brevity, let us abbreviate classifiers with misaligned subagents as **CWMS**. Note that I might use classifier a bit more broadly than the strict definition - for example, I may call certain more general question-answering machines "classifiers". I do not believe this directly affects the general argument.
**Claim 1**: An agent will "perform" "better" than a memoryless function given the same sequence of inputs only if (almost) every input is highly correlated with the previous input. To phrase this in a different way, having a "memory" only helps if your "memory" gives good evidence for either what is going to happen next or what to do next.
**Claim / Assumption 2**: For most optimization processes powerful enough to find a classifier that performs well, we have:
.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-ex-box-test {position: absolute; overflow: hidden; width: 1px; height: 60ex}
.mjx-line-box-test {display: table!important}
.mjx-line-box-test span {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')}
P(Optimization process finds a CWMS)≈Density of CWMS that "perform well"Density of all classifiers that "perform well"
There might be certain optimization processes that lean more towards CWMS (or toward classifiers without misaligned subagents), but I think this is a reasonable base assumption given our current level of information.
**Claim 3**: For a certain task, if most good classifiers for that task are CWMS, then there exists one or more subtasks that involve processing highly correlated inputs, and doing some of these subtasks very well is important for being a good classifier. This follows from Claims 1 and 2.
Conversely, if such key subtasks do not exist, most good classifiers will not be CWMS
**Intuition 4**: For most tasks which have key subtasks of the type mentioned in Claim 3, those subtasks are very likely to involve some sort of modeling how things change / interact through time (example: understanding the contents of a video or the meaning of a paragraph in a book require this type of modeling).
**Intuition 5**: There are a lot of interesting, difficult tasks where modeling how things change through time is not key for solving the task.
**Claim / Intuition 6**: Optimizers for the tasks mentioned in Intuition 5 have very low probability of finding a CWMS. This follows from Claims 2 and 3 and Intuition 4.
Thanks to Scott Garrabrant and Evan Hubinger for discussing some the ideas in this post with me. |
533dd69b-699d-4063-86e4-66292ac32f17 | trentmkelly/LessWrong-43k | LessWrong | Debt, Submission, and a Technical Definition of Drama
Two public answers to private question in which I employ a technical definition of drama.
Origin Debt
Contractual debts can only exist within some sort of civil law wherein commitments can be publicly and unambiguously evaluated.
In the absence of civil law institutions, dramatic strategies posturing around owing and being owed parasitize norms around debts, until only the dramatic versions remain.
A debt too large to be repaid, with no resolution process (e.g. bankruptcy or formal debt slavery), effectively reduces the debtor's creditworthiness within that system to nothing. If a hobo somehow owes a billion dollars, it makes no sense for me to draw up an enforceable contract to lend him $10, since the prior creditor has first claim on any money he might get to repay me. I might lend him the cash anyway informally if our informal credit system is decoupled from the one under which he owes a billion. It makes more sense to think of parents (in practice usually fathers) owning their children in some systems but not in others. If I want to lend to someone within a system where they're owned by someone else, I need to investigate their owner's credit as well.
I can also be said to "owe" my parents in the sense of causal attribution.
The decision for each individual is in what sorts of systems, potential or actual, we wish to maintain creditworthiness. Seems to be like right now we barely even have language, so other sorts of debts are mostly dramatic or customary, as are property claims.
Violence by Submission
There are object-level situations where for reasons external to the social graph, one person is in charge of another. E.g. some kinds of advice, or a military officer commanding soldiers. Then sometimes people play out dramatic scenes in which one person plays the role of being in charge of another.
Someone can invoke a dramatic scene or influence/control how it plays out, regardless of what their role within the scene is. So there's no reason you can't agg |
ff566aec-55f0-44fd-9bf7-d9f48e9f09ae | trentmkelly/LessWrong-43k | LessWrong | Stages of Survival
This post outlines a fake framework for thinking about how we might navigate the future. I found it useful to my thinking, hopefully you will find it useful as well.
Whether or not we build AGI, one of five things must result.
1. Dune solution: The new world permanently lacks AGIs.
2. Human control: Humans retain control over AGIs indefinitely, avoid catastrophe.
3. Human-induced catastrophe: Humans retain control, cause catastrophe.
4. Aligned AGIs in control: AGIs do take control, but it is not catastrophic.
5. AGI-induced catastrophe: AGIs do take control, it is catastrophic.
There are disagreements in classifying many scenarios, both in terms of the source of catastrophe (#3 vs. #5) and in whether the scenario is catastrophic (#2 vs. #3, #4 vs. #5), as well as how to think about #1 in terms of whether it is inherently catastrophic, and whether if so it would count as being due to loss of control.
These disagreements are an underappreciated source of cruxes and disagreement over the probability of doom, including the conditional probability of doom.
At any given time, we could, as a simplification, describe the current world state as having:
1. Some degree of AGI capabilities progress (a-h) and some distribution of meaningful potential capabilities (i-iii).
1. No base model exists that could result in an AGI without a new training run.
2. At least one base model exists that could result in an AGI with sufficient scaffolding work, or some other process with radically lower compute than required to create the base model. We don’t yet know how to do this, and may not know that it is possible.
1. There is one such model. Access is highly restricted.
2. There is more than one such model. Access is meaningfully restricted.
3. Access to such models is not meaningfully restricted.
3. At least one base model exists that could result in an AGI with sufficient scaffolding work. We know how to do this.
1. There is one such |
beb24b6d-c848-4376-abbf-8c6c210ec1c6 | trentmkelly/LessWrong-43k | LessWrong | Covid 5/20: The Great Unmasking
The CDC has lifted its mask mandate for vaccinated people. In one fell swoop, it is suddenly fine for anyone fully vaccinated to go maskless anywhere, without social distancing, at any time, so long as regulations and rules permit it. This retains the madness of ignoring partial vaccination entirely, and the madness of making children as young as two (!) wear masks around groups of fully vaccinated adults, but it is certainly a huge step towards sanity. It also implies that the rules and regulations that remain need to be changed, and many of them are indeed being changed.
Reactions were mixed.
There was much rejoicing. Many cried tiers of joy. Others pointed out that this was a damn good reason to get vaccinated.
There was also much confusion and opposition. Very Serious People expressed panic and horror. Others expressed reasonable concern that this was premature. This would destroy mask mandates, causing maskless people to go around maskless. People would go around acting like the pandemic was over before there was full ‘herd immunity.’
The opposition has some legitimate points. The control system could still have some fight in it. We have no way to verify verification status that anyone is willing to use. Vaccination checks are mostly going to be either non-existent or on the honor system, and America does not highly value honor. Norms that are ignored by half the people are easy for the other half to choose to ignore as well. At a minimum, this will change norms in ways that slow down the decline in cases. If everyone were to use this as a reason to go fully ‘back to normal’ at current vaccination rates cases would presumably start rising again at least for a while, and it’s not clear we could do much of anything to reverse course if that happened beyond pushing harder to get more vaccinations.
A lot of people also pointed out the contradictions between this new policy and the old one, with the change from ‘take precautions that wouldn’t even make sense i |
635ac76f-06c7-4bf9-8efa-235c9df222cd | trentmkelly/LessWrong-43k | LessWrong | Conservation of Expected Evidence and Random Sampling in Anthropics
This is the second post in my series on Anthropics. The previous one is Anthropical Motte and Bailey in two versions of Sleeping Beauty. The next one is Antropical probabilities are fully explained by difference in possible outcomes.
Introduction
Since the first time I've heard about anthropics something felt off. Be it updating on awakening in Sleeping Beauty, accepting high probability of doom in the Doomsday Argument, or the premise of Grabby Aliens; following SSA or SIA, the whole pattern of reasoning seemed wrong to me.
That it's cheating. That there is something obviously unlawful going on. That it doesn't look at all like the way cognition engines produce map-territory correspondence.
It was hard to point out what exactly had been wrong, though. The discourse seemed to be focused around either accepting SIA or SSA, both of which are obviously absurd and wrong in some cases. Or discarding all the anthropic reasoning altogether - a position towards which I'm quite sympathetic, but which also seemed to be an overcorrection, as it tends to discard some sound reasoning.
It took me some time to formalize this feeling of wrongness into an actual rule that can be applied to anthropical problems, separating correct reasoning from wrong reasoning. In this post I want to explore this principle and its edge cases, and show how using it can make anthropics add up to normality.
The metaphysics of existence vs having a blue jacket
Let's consider the Blue Jacket Experiment (BJE):
> A fair coin is tossed. On Tails two people will be created, wearing blue jackets. On Heads two people will also be created but only one, randomly chosen, will have a blue jacket. You are created and notice that you are wearing a blue jacket. What's the probability that the coin landed Heads?
Here the situation is quite obvious. As having a blue jacket is twice as likely when the coin is Tails, we can do a simple Bayesian update:
P(Heads|BlueJacket)=P(BlueJacket|Heads)P(Heads)/P(BlueJac |
5a5f829f-8939-442c-88cb-7b8f12f0718d | trentmkelly/LessWrong-43k | LessWrong | Idea for LessWrong: Video Tutoring
Update 7/9/17: I propose that Learners individually reach out to Teachers, and set up meetings. It seems like the most practical way of getting started, but I am not sure and am definitely open to other ideas. Other notes:
* There seems to be agreement that the best way to do this is individualized guidance, rather than lectures and curriculums. Eg. the Teacher "debugging" the Learner. Assuming that approach, it is probably best for the amount of Learners in a session to be small.
* Consider that it may make sense for you to act as a Teacher, even if you don't have a super strong grasp of the topic. For example, I know a decent amount about computer science, but don't have a super strong grasp of it. Still, I believe it would be valuable for me to teach computer science to others. I can definitely offer value to people with no CS background. And for people who do have a CS background, there could be value in us taking turns teaching/learning, and debugging each other.
* We may not be perfect at this in the beginning, but let's dive in and see what we can do! I think it'd be a good idea to comment on this post with what did/didn't work for you, so we as a group could learn and improve.
* I pinned http://lesswrong.com/r/discussion/lw/p69/idea_for_lesswrong_video_tutoring/ to #productivity on the LessWrongers Slack group.
Update 6/28/17: With 14 people currently interested, it does seem that there's enough to get started. However, I'd like to give it a bit more time and see how much overall interest we get.
Idea: we coordinate to teach each other things via video chat.
* We (mostly) all like learning. Whether it be for fun, curiosity, a stepping stone towards our goals.
* My intuition is that there's a lot of us who also enjoy teaching. I do, personally.
* Enjoyment aside, teaching is a good way of solidifying ones knowledge.
* Perhaps there would be positive unintended consequences. Eg. socially.
* Why video? a) I assume that medium is better for educatio |
abf0b74d-f4cc-404a-9179-b804d929c005 | trentmkelly/LessWrong-43k | LessWrong | Seeking paid help for SPARC logistics
This August CFAR is running the Summer Program on Applied Rationality & Cognition, an academic high school summer program, and we are looking for a part-time, temporary logistics manager for the program. If you are interested in the role and live near Berkeley, CA, please fill out the short form here with additional information.
SPARC will be held at UC Berkeley from August 5 - August 17. The work is full time or more from August 3 - August 17, and about 10 hours per week between the start of the job and August 3. The participants are very talented high school math students.
If you are interested in the work CFAR is doing or in SPARC in particular, this will be a great opportunity to see the work first hand, help out, and meet many of the people involved.
Some examples of logistical tasks that you would help with:
* Ensure that facilities are prepared, help arrange housing, determine the schedule and coordinate with visiting instructors, and coordinate with UC Berkeley conference services
* Talk with students and parents in advance of the program, answer questions, give directions, etc.
* Acquire supplies for the program, print and distribute course materials and schedules
* Arrange liability insurance, medical histories, and releases.
SPARC volunteers and instructors will be available to help with some tasks, especially during the program itself, but you would be the main staff responsible for logistics. We strongly prefer applicants who have experience organizing events and ensuring that they run smoothly. Applicants should be local and able to easily get to the UC Berkeley campus. (They should also be happy to interact with 16-18 year old students and their parents.)
|
9efc0b0e-ea54-4843-996f-f75c29d7b57d | trentmkelly/LessWrong-43k | LessWrong | Variably compressibly studies are fun
I like studying calculus, physics, and English etymology.
Calculus includes problems like ∫dxarctanx. arctan is an elementary function, so you may be expected to have its integral memorised. I'd have to find it otherwise. By parts: ∫dxarctanx=xarctanx−∫dx(xarctan′x). I remember arctan's derivative: 1x2+1. Short of that, I could rederive it in terms of trigonometric identities and the chain rule. Then ∫dxarctanx=xarctanx−∫dxxx2+1.
I can easily enough express 1x2+1 as a power series, based on that of 11−x, and the numerator x just multiplies each term. Integrate term-by-term and get a new power series reminiscent of that for the logarithm. From that, I could guess-and-check to get the exact logarithm — or I do it precisely from the knowledge that (ln(f))′=f′f, recognising x as roughly the derivative of x2+1.
In physics, you can derive orbital velocity in terms of centripetal acceleration (a=v2r), and hesitantly get √GMr. You can also derive escape velocity in terms of kinetic (12mv2) and potential (GMmr) energies, and hesitantly get √2GMr.
If you recall approximate values for each for any one planet, like Earth's 8 km/s and 11 km/s, respectively, you can notice that they differ by a factor around 1.4. The small factor suggests the similar formulae are valid. If you recall √2≈1.414, it's clear that those formulae line up: both right or both wrong.
With enough English etymology knowledge, you can look at "acromegaly", recognise "acro-" to mean "tip", "-mega-" to mean "huge", and "-y" to mean "state", and infer that "acromegaly" is "huge extremities condition", with vague or absent context.
Somewhat less so, I can enjoy organic chemistry and 汉字. They repeat each other to enable easy memory, in ways described earlier. From the properties of one fatty acid, we can infer many properties of others, as their difference is the length of an alkane carbon chain, which only react by combustion. From the component 青, learned from the right of the common 情, we can infer a pro |
24ced15e-b12a-4bb2-828b-7404139ab7e8 | trentmkelly/LessWrong-43k | LessWrong | Questions about YIMBY
If you want to change the world, start by giving people free books. Or at least, this is the approach of the Adam Smith Institute, a free-market think tank in the UK. At an event of theirs, they gave me a copy of John Myers’ little book (or long paper) YIMBY: How to end the housing crisis, boost the economy and win more votes. Weirdly, the book starts with specific solutions to the housing crisis, and then only in the middle talks about why this is a problem you should care about.
You probably already know this story: NIMBYs (Not in My Back Yard) want to preserve the value of their homes, so they lobby the government to restrict supply through zoning laws, historic preservation rules, and widely distributed veto power. Hence, when demand goes up, prices increase. YIMBYism (Yes in My Back Yard) is a recent movement of people who want to allow more building to solve this (the author is a cofounder of London YIMBY).
I feel that whether you view NIMBYism as a problem of excessive government intervention or civil society is dependent on your pre-existing biases. I take a central claim of libertarianism to be that the government is overrated, and civil society is underrated. But if we had been trying to show the opposite, we could use the fact that civil society obstructs building so much to argue that the government needs to socialise more housing.
This gets at my general confusion about decentralisation arguments. There is a well-known thesis that decentralisation is good because it allows for the synthesis of information from many different sources, in a way that couldn’t plausibly be done by central authorities. The problem of NIMBYism is a problem of decentralised town councils, part of the centralised government, being lobbied by decentralised homeowners, to obstruct the ability of decentralised individuals to pay centralised developers to build new buildings. For whom is this a win – the centralisers or the decentralisers?
There’s a paper from Hsieh and Moretti |
d68e9a59-3ad3-4c7a-befd-485f01a1aa3f | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] Religion's Claim to be Non-Disprovable
Today's post, Religion's Claim to be Non-Disprovable, was originally published on 04 August 2007. A summary (taken from the LW wiki):
> Religions used to claim authority in all domains, including biology, cosmology, and history. Only recently have religions attempted to be non-disprovable by confining themselves to ethical claims. But the ethical claims in scripture ought to be even more obviously wrong than the other claims, making the idea of non-overlapping magisteria a Big Lie.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, in which we're 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 Belief as Attire, 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. |
eec18434-d70a-429e-994a-4401ecd10e2e | trentmkelly/LessWrong-43k | LessWrong | -
|
f7bebe6f-696f-461f-b0d8-0113f63b849e | trentmkelly/LessWrong-43k | LessWrong | Li and Vitanyi's bad scholarship
In An Introduction to Kolmogorov Complexity and Its Applications 4th edition Li and Vitanyi claim that Solomonoff Induction solves the problem of induction. In Section 5.1.4 they write:
> The philosopher D. Hume (1711–1776) argued that true induction is impossible because we can reach conclusions only by using known data and methods. Therefore, the conclusion is logically already contained in the start configuration. Consequently, the only form of induction possible is deduction. Philosophers have tried to find a way out of this deterministic conundrum by appealing to probabilistic reasoning such as using Bayes’s rule. One problem with this is where the prior probability one uses has to come from. Unsatisfactory solutions have been proposed by philosophers such as R. Carnap (1891–1970) and K.R. Popper.
What Hume actually wrote about induction was (Section IV, Part II, pp. 16-17):
> Should it be said that, from a number of uniform experiments, we infer a connexion between the sensible qualities and the secret powers; this, I must confess, seems the same difficulty, couched in different terms. The question still recurs, on what process of argument this inference is founded? Where is the medium, the interposing ideas, which join propositions so very wide of each other? It is confessed that the colour, consistence, and other sensible qualities of bread appear not, of themselves, to have any connexion with the secret powers of nourishment and support. For otherwise we could infer these secret powers from the first appearance of these sensible qualities, without the aid of experience; contrary to the sentiment of all philosophers, and contrary to plain matter of fact. Here, then, is our natural state of ignorance with regard to the powers and influence of all objects. How is this remedied by experience? It only shows us a number of uniform effects, resulting from certain objects, and teaches us that those particular objects, at that particular time, were endowed with |
805cc354-2b6e-411a-a735-c03f44b8b29b | StampyAI/alignment-research-dataset/arbital | Arbital | Codomain vs image
A function $f : X \to Y$ with [domain](https://arbital.com/p/3js) $X$ and [codomain](https://arbital.com/p/3lg) $Y$ cannot necessarily produce all values in $Y$. For example, consider the function that takes a [real number](https://arbital.com/p/real_number) and squares it. We could say that this function has codomain $\mathbb R$, because it produces a real number, even though its image is the set of non-negative real numbers.
If $f$ maps $X$ to a subset $I$ of $Y$, why not just say that its codomain is $I$ and do away with the distinction between codomain and image? There are at least two cases where the distinction is useful.
1. It is useful to distinguish codomain from image in cases where the _type of thing_ that the function produces is easy to specify, but the _specific values_ are hard to calculate. Consider, for instance, the [Ackermann function](https://arbital.com/p/43x). It definitely produces numbers (i.e. its codomain is $\mathbb N$), but it's not easy to figure out what numbers (exactly) it produces. When talking about what sort of outputs to expect from the Ackermann function, the natural answer is "a number," rather than "either 1 through 7, 9, 11, 13, 29, 61, 125, 65533, $2^{65536} − 3$, or some other ungodly large numbers that I lack the resources to calculate." For this purpose, the codomain can be thought of as a natural/simple boundary drawn around a complex image.
2. Cases where we're considering a set of all functions that map into a certain set. For example, if we're discussing all possible functions from the [natural numbers](https://arbital.com/p/natural_number) to the set $\{0, 1\},$ we want to be able to talk about "functions with codomain $\{0, 1\}$" even if some of those functions are things like "always return 0" (which have an image smaller than $\{0, 1\}$). For this purpose, the codomain can be thought of as a context-dependent tool that's useful when considering (or making use of) a bunch of different functions that are all producing objects of the same type. |
1d20e7c2-ecb7-4a6e-8ef0-7ded99e374f4 | trentmkelly/LessWrong-43k | LessWrong | Should we write more about social life?
Social life is a thorny subject. Writings on that topic can be noxious, for a variety of reasons. Yet we live in a social world, and instrumental rationality ought to make us better at navigating it. Furthermore, if status competition is a war of all against all, and "all warfare is based on deception," as Sun Tzu says, then we should expect that this is one of the areas where rationality might be especially helpful. On this website, people post self-help advice through a rationalist lens all the time. Yet I rarely see rationalist self-help advice on topics such as how to have a thriving social life, get along with your parents, strike up a conversation with a stranger, build a blog audience, or field a disagreement without lying, submitting, or looking like a jerk.
I'd consider the following to be virtues of posts about social life:
1. Striving for generality as much as (but no more than) is possible.
2. Avoiding moral judgments and political resentments.
3. Balancing a basically positive and constructive attitude with a healthy skepticism towards platitudes.
Do you agree that this forum might benefit from more posts on social life? What would you be most interested in writing or reading about? What do you tend to find most frustrating when reading people's takes on social life, here or elsewhere? |
fd3d10cd-5aa6-4e8a-b127-fdd1f657b00f | trentmkelly/LessWrong-43k | LessWrong | Meetup : Rationality Reading Group (62-66)
Discussion article for the meetup : Rationality Reading Group (62-66)
WHEN: 13 July 2015 06:30:00PM (-0700)
WHERE: Paul G. Allen Center (185 Stevens Way, Seattle, WA) Room 503
Reading group for Yudkowsky's "Rationality: AI to Zombies", which is basically an organized and updated version of the Sequences from LW (see http://wiki.lesswrong.com/wiki/Sequences).
The group meets to discuss the topics in the book, how to apply and benefit from them, and related topics in areas like cognitive biases, applied rationality, and effective altruism. You can get a copy of the book here: https://intelligence.org/rationality-ai-zombies/
The reading list for this week is five topics from the "Politics and Rationality" section of "How To Actually Change Your Mind". They are (actually 62-66, LW's auto-formatting is screwing it up):
1. Reversed Stupidity Is Not Intelligence
2. Argument Screens Off Authority
3. Hug the Query
4. Rationality and the English Language
5. Human Evil and Muddled Thinking
We previously covered the "Map and territory" sequence a few months ago, but please don't feel a need to have read everything up to this point to participate in the group.
Event is also on Facebook: https://www.facebook.com/events/1641756126111169/
We're meeting on the 5th floor. If you show up and the door into the room is locked, knock and look around for us elsewhere on the fifth floor if nobody answers. If the doors to the building are locked, try the other ones and don't believe the little red lights; try anyway. If the doors are, in fact, locked, we'll try to have somebody to let people in.
There's usually snacks at the meetup, though feel free to bring something. We usually get dinner afterward, around 9PM or so.
Discussion article for the meetup : Rationality Reading Group (62-66) |
bb532c43-0d84-4ef5-97c9-abb1b17aab32 | trentmkelly/LessWrong-43k | LessWrong | Variations of the Garrabrant-inductor
|
265d35f9-aefb-44d5-8248-4ceb12a9dc26 | trentmkelly/LessWrong-43k | LessWrong | What should I study to understand real-world economics? (I.e., To gain financial literacy)
I have taken a basic economics course and read some academic economic stuff here and there, but these have been viscerally “theoretical” in nature. I want to understand real-world concepts like value chains, stocks, venture capital, IPO, corporations (I can’t quite say I know what a corporation IS.) ... . The best resource I have found to date is http://stratechery.com . I am wondering if more efficient learning material exists. I don’t want advice/prediction books (on management or investing). I simply want to understand the basic stuff about these concepts. Financial literacy, I would say. I don’t know if there is a name for this ”field of study.” |
f4c3c371-04b9-4163-a814-f3d401e9cee7 | trentmkelly/LessWrong-43k | LessWrong | Adversarial vs Collaborative Contexts
You can often make pretty clear predictions about what sort of effect something would have in the world as it exists today, but when you start to take into account how the world might change in response it gets very difficult. One tool here is thinking about whether you mostly expect collaborative or adversarial effects.
In a collaborative context other people respond to your change by building on it, and it typically has a larger effect than you'd naively expect. For example, say you create an icon to communicate an idea in your program, other people start using it, and it becomes the standard representation for the concept. While an A/B test might have been only mildly positive, or even negative, after the world adapted to the new icon it would have become very clearly the right symbol to use.
In an adversarial context other people respond to your change by countering it, and it typically has a smaller effect. You figure out a way to detect bot traffic, but the bot operators are very motivated to shift their behavior to undo your work.
Some contexts are also neutral, where both collaborative and adversarial effects are minimal or nonexistent. A baby toy design doesn't get better or worse with the generations.
Some examples:
* Diseases are adversarial: antibiotics lose resistance, new variants evade immunity. While something looks good in initial testing, that advantage will erode over time. An interesting exception here is cancer, which (almost always) has to evolve anew in each patient, and so cancer treatments don't become less effective over time (in new patients).
* Adopting standards is collaborative: each computer with USB increased demand for devices that spoke USB and vice-versa. Someone thinking "let's not bother with USB, it's not that widely supported yet" would have been ignoring collaborative effects.
* Same with building on a platform: the more iPhones sold the more people wanted apps, and the more apps there were the more people wanted i |
1495d516-5595-4646-a816-4ff18e19eb10 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post278
In this post, I’ll make a technical point that comes up when thinking about risks from scheming AIs from a control perspective . In brief: Consider a deployment of an AI in a setting where it’s going to be given a sequence of tasks, and you’re worried about a safety failure that can happen suddenly. Every time the AI tries to attack, one of three things happens: we win, we lose, or the deployment continues (and the AI probably eventually attempts to attack again). So there’s two importantly different notions of attacks “failing”: we can catch the attack (in which case we win), or the attack can fail silently , either because it’s defective or because it’s defused by safety techniques, such that the attack doesn’t end up causing an unacceptable outcome but the AI also doesn’t get caught (in which case the deployment continues). When reasoning about how safe a protocol is, it’s crucial to distinguish between these two ways an attack can fail. And when designing protocols, it’s often worth thinking separately about the techniques that make the AI lose and the techniques that let the deployment continue without a win or loss. This point has been made in passing in many places (our control paper , our toy models post , Charlie Griffin’s control paper , Towards Safety Cases For AI Scheming , various upcoming papers) but I wanted to have a single place that makes the point clearly. Thanks to Alex Mallen, Josh Clymer, Joe Benton, and Justis Mills for comments, and to Monte MacDiarmid for providing experimental data that I analyze. Here’s a more precise description of the scenario: You’re deploying your AI to do a large number of discrete tasks (e.g. implementing many features across the codebase inside a company, or doing many ML research projects). Let’s assume that each task will be completed fairly quickly compared to the timescale of the whole deployment: for example, you’re trying to analyze the safety of deploying this AI for six months, and the tasks will take an average of an hour; each task lasts about 1/4000 of the whole deployment. The system will be doing many tasks in parallel. You’re worried about a concentrated failure (a.k.a. “high-stakes failure”): in absence of countermeasures, the AI would be able to cause an unacceptable outcome just by behaving badly on a small number of tasks. Examples of concentrated failures include: Causing a rogue deployment (see examples here ). Sabotaging an ML research task that is particularly crucial, where a single false result suffices for the lab to make a catastrophically bad decision. (I don’t think most ML research tasks are crucial enough that the AIs could cause an unacceptable outcome by just sabotaging a small number of experiments.) Leaking sensitive information (e.g. describing key algorithmic secrets in a document that will be shared with people who aren’t cleared to know those secrets; one extreme example is the model leaking algorithmic secrets to someone outside the AI lab, which it is able to do if it has access to both algorithmic secrets and the internet). Our first control paper (and various follow-up papers and related posts ) analyzes a setting with these properties. So you deploy the AI with some control measures. In many cases, the optimal strategy for the AI is to normally act aligned, and on very rare occasions, try to cause the unacceptable outcome—we’ll call those attempts “attacks”. Execute/replace/audit protocols For concreteness, let’s talk about a particularly important simple class of control protocols, which I’ll call execute/replace/audit protocols . An execute/audit/replace protocol works as follows: We have an agent loop that repeatedly generates inputs to a LLM and then calls an inference service on those inputs, getting a proposed action in response (see here for more on what I’m imagining). Then we produce an estimate of how dangerous or suspicious this action is. There are many techniques you could use for this: asking trusted or untrusted models to evaluate the danger of proposed actions , using probes on the policy trained on off-policy examples or produced other ways , or using interpretability techniques like sparse autoencoders to detect deceptive intent. All of these methods take a proposed action (together with some surrounding context) and produce a score describing how dangerous and/or suspicious the action is. I’ll call that the monitor and call its response the monitor score . Then, based on the monitor score, we need to make two decisions. Do we execute the action or replace it? If the action looks fine, we want to execute it. If the action looks dangerous, we don’t want to execute it. So we want to replace it with a new action. In many current systems, this is implemented by replacing the model’s response with a refusal. In the context of an agent, there are many different options for how to replace the action (which will then be monitored itself): Resample from the original, untrusted policy. Defer to trusted: get an action from a weaker, trusted model . Defer to trusted edit: get a trusted model to edit the action to remove parts that seemed suspicious. Abort the trajectory, reporting that you failed to solve the task. This is similar to refusals in current language models, except it aborts the whole task instead of just the one message. Block the trajectory, waiting until a human audit. Shut down the whole deployment. I personally think that this is so costly that it’s very unlikely you’d want to do this without many people from the AI company first carefully investigating the action. Do we want to audit the action? Auditing is when you review the action to determine whether it’s evidence that your model is intentionally trying to cause problems. The simplest strategy here is to put the action in a queue to eventually be reviewed by a human, who can escalate the investigation if they’re suspicious (perhaps using techniques described here ); this escalation might end with you concluding that the model is a misaligned power-seeker. So in general we want to get two scores from our monitor, which we use to make the two choices of whether to replace and whether to audit. You could use the same monitor with different thresholds for these two choices, which is how we handled it in our original control paper , but it’s probably better to make the decisions separately. The two defensive actions of auditing and replacing have quite different safety properties. Auditing is probably drastically more expensive, so you’ll probably do it much less. In particular, if you need to use human auditors and you’re able to replace actions by resampling, it’s totally affordable to resample your model’s actions 10% of the time, but the proportion of the model actions you can afford to have a human look at is drastically lower than that. E.g. OpenAI obviously can’t afford to have a human look at even a tiny fraction of the (very roughly estimated [1] ) 100 trillion tokens that their API handles per year; with a staff of 1000 fast readers, they could read 0.1% of it. One consequence is that you should generally replace whenever you audit. False negatives are catastrophic when choosing whether to replace, but not when choosing whether to audit. If you execute a catastrophically dangerous action, you might immediately lose. Whereas if you choose not to audit a catastrophically dangerous action, you’re just missing out on the opportunity to get evidence that the model is scheming. Execute/replace/audit protocols lead to win/continue/lose scenarios Let’s reason through what happens when the AI attacks. First, let’s think about a simple version. In this version, auditing is perfect: the auditing is able to incriminate the AI so convincingly that it can shut down after observing any attack. And replacing is perfect: every attack is neutralized by replacing (which might not be true in reality, e.g. for a replace strategy like trusted editing). So the ways the deployment can end here are: The attack is above the audit threshold, and we audit the attack; we then win. The attack is below the replace threshold, and also succeeds; we lose. (Note that attacks can fail. For example, in our code backdooring paper , a large fraction of attempted backdoors didn’t actually fit the requirements of a backdoor and so didn’t lead to an AI victory. One of the reasons that trusted monitoring is safer than upfront auditing in that paper is that trusted monitoring forces the attack policy to submit backdoors that, as a result of needing to be sneaky, are less likely to work.) To calculate safety here (as described in Toy models of AI control for concentrated catastrophe prevention ), you can make the following assumptions: The model rarely diverges from the behavior of an aligned model, attacking occasionally (but often enough that the deployment will end early—this is usually a good assumption, because the model loses if the deployment completes, as described here ). Successful attacks immediately cause the scheming AI to win. Then the odds of a blue team victory is the ratio between the probability of the following two events occurring on a particular step: The AI developer wins because they audited an attempted attack and then decided to shut down based on the result of the audit. The scheming AI wins because a successful attack was executed (rather than replaced). For example, if on the average attack attempt, there’s a 10% chance that the attack will be executed and be successful, and there’s a 20% chance that the attack will be audited and then the deployment will be shut down, then the overall P(blue team victory) is ⅔. It’s easy to modify this rule of thumb to relax some of those assumptions. Example: analyzing the safety of defer-to-trusted using the probes from “Simple probes can catch sleeper agents” In Simple probes can catch sleeper agents , Monte MacDiarmid makes a linear probe monitor with a simple contrast pair technique on the model from Sleeper Agents . He reports the following ROC curve , graphing true positive rate (TPR) against false positive rate (FPR): An obvious question is: if this was the actual ROC curve that we got with the actual monitor we were using on an actual scheming model, would we be safe? Let’s pretend we’re using a perfect auditor that we can afford to use 0.1% of the time. We need to use the monitor as a classifier with a 0.1% FPR in order to keep our auditor fully utilized. According to the data that Monte generously gave me, at FPR=0.001, TPR=0.26. So the probability of an attack getting audited is 26%. Now, let’s assume we’re also using this probe’s output for a replace strategy with a 5% replacement rate. At FPR=0.05, FNR=0.03. So when the model attacks, the chance that the attack is executed is 3%. (Let’s assume that every attack succeeds.) Thus when the model attacks, there’s a 26% chance it gets audited, a 3% chance the attack gets executed, and the rest of the probability is that the attack is replaced. So there’s a 0.26/(0.26 + 0.03) = 90% chance that the model is caught before an attack is executed, and in expectation the model will have a few attacks replaced before it either wins or loses. Some things to observe here: If we don't use a replace protocol, the safety is atrocious: with a 0% replace rate, there’s a 74% chance that the model successfully attacks (that is, an attack is executed before an attack is audited and caught). The safety is very sensitive to the audit rate. A lot of people intuitively think that monitoring (for auditing) needs a super high TPR to be effective (probably because of a cached intuition about intelligent adversaries being able to break unrobust measurements), but as we see here even a fairly low TPR of 26% gets you a 10x risk reduction in the simplified setting, when paired with replacement. So whether this probe is good enough for safety crucially depends on the details of the protocol in which the probe is being used as a monitor. ^ I’m going off the $510 million annualized API revenue mentioned here , and arbitrarily using the GPT-4o price of $5 per million input tokens, which gets you 100 trillion tokens. |
56677402-4c5f-4149-9225-b460baddce57 | trentmkelly/LessWrong-43k | LessWrong | Formalising decision theory is hard
In this post, I clarify how far we are from a complete solution to decision theory, and the way in which high-level philosophy relates to the mathematical formalism. I’ve personally been confused about this in the past, and I think it could be useful to people who casually follows the field. I also link to some less well-publicized approaches.
The first disagreement you might encounter when reading about alignment-related decision theory is the disagreement between Causal Decision Theory (CDT), Evidential Decision Theory (EDT), and different logical decision theories emerging from MIRI and lesswrong, such as Functional Decision Theory (FDT) and Updateless Decision Theory (UDT). This is characterized by disagreements on how to act in problems such as Newcomb’s problem, smoking lesion and the prisoner's dilemma. MIRI’s paper on FDT represents this debate from MIRI’s perspective, and, as exemplified by the philosopher who refereed that paper, academic philosophy is far from having settled on how to act in these problems.
I’m quite confident that the FDT-paper gets those problems right, and as such, I used to be pretty happy with the state of decision theory. Sure, the FDT-paper mentions logical counterfactuals as a problem, and sure, the paper only talks about a few toy problems, but the rest is just formalism, right?
As it turns out, there are a few caveats to this:
1. CDT, EDT, FDT, and UDT are high-level clusters of ways to go about decision theory. They have multiple attempted formalisms, and it’s unclear to what extent different formalisms recommend the same things. For FDT and UDT in particular, it’s unclear whether any one attempted formalism (e.g. the graphical models in the FDT paper) will be successful. This is because:
2. Logical counterfactuals is a really difficult problem, and it’s unclear whether there exists a natural solution. Moreover, any non-natural, arbitrary details in potential solutions are problematic, since some formalisms require everyb |
8ad3cfb2-a879-4588-8684-21f452469dd4 | trentmkelly/LessWrong-43k | LessWrong | Social Capital - Does it Matter?
Within the Walls
There's plenty of funny business. My dad was enthusiastically arguing that we should invest in this little known bank offering a high return. Your money doubles every month! Who told you that, I asked? The dentist. He now lives in a big house in Dedinje and has a huge painting of Petar Lubarda on the wall.
The next day he was fixing the VW Beetle - not his. It belonged to a friend who lived in Germany and occasionally needed to be chauffeured from and to the airport. The thing was rotten in places, but the engine started every time and kept up with other traffic by constantly being revved out of its wits. Now it was lying in bits on the street, oil in slicks and drops and dad's feet protruding from under the body.
My mum loved socialising. She took pride in her taste in the arts and sought company of writers, painters and actors. Her old school friends would often come for drinks in the evening and most of them looked like they came out of 70s French movies.
My great grandmother was visiting us again. She spent days sitting on a chair, narrating same old stories. This time It was about wisdom - how she systematically refused to follow doctor's advice and evaded the surgical knife. "He warned me, I've got 6 months max if I didn't have it removed. He died of a stroke 15 years ago. What a pity, he was a young man..."
Uncle and aunt lived with us - 7 people altogether. Me and my sister in one room, parents in the adjacent one, grandma in the living room and uncle and aunt in the second living room. People walked in bathrobes, queued in front of the only bathroom, smoked and chatted. There was always something cooking. When the phone rang - it was a fight who'd get to it first.
The dining room was the meeting place. My dad never felt comfortable joining in, he didn't smoke or drink coffee. Being socially awkward - he preferred staying in his room and reading motoring magazines that he was subscribed to. They'd arrive every month |
2fcd00d6-34cc-4885-9139-ce6870085d3f | trentmkelly/LessWrong-43k | LessWrong | Math appendix for: "Why you must maximize expected utility"
This is a mathematical appendix to my post "Why you must maximize expected utility", giving precise statements and proofs of some results about von Neumann-Morgenstern utility theory without the Axiom of Continuity. I wish I had the time to make this post more easily readable, giving more intuition; the ideas are rather straight-forward and I hope they won't get lost in the line noise!
The work here is my own (though closely based on the standard proof of the VNM theorem), but I don't expect the results to be new.
*
I represent preference relations as total preorders on a simplex ; define , , and in the obvious ways (e.g., iff both and , and iff but not ). Write for the 'th unit vector in .
In the following, I will always assume that satisfies the independence axiom: that is, for all and , we have if and only if . Note that the analogous statement with weak preferences follows from this: holds iff , which by independence is equivalent to , which is just .
Lemma 1 (more of a good thing is always better). If and , then .
Proof. Let . Then, and . Thus, the result follows from independence applied to , , , and .
Lemma 2. If and , then there is a unique such that for and for .
Proof. Let be the supremum of all such that (note that by assumption, this condition holds for ). Suppose that . Then there is an such that . By Lemma 1, we have , and the first assertion follows.
Suppose now that . Then by definition of , we do not have , which means that we have , which was the second assertion.
Finally, uniqueness is obvious, because if both and satisfied the condition, we would have .
Definition 3. is much better than , notation or , if there are neighbourhoods of and of (in the relative topology of ) such that we have for all and . (In other words, the graph of is the interior of the graph of .) Write or when ( is not much better than ), and ( is about as good as ) when both and .
Theorem 4 (existence of a utility function). There is a such that for all ,
Unl |
0f544928-37f0-40b6-b3f3-67e0c9865344 | trentmkelly/LessWrong-43k | LessWrong | Indifferent vs false-friendly AIs
A putative new idea for AI control; index here.
For anyone but an extreme total utilitarian, there is a great difference between AIs that would eliminate everyone as a side effect of focusing on their own goals (indifferent AIs) and AIs that would effectively eliminate everyone through a bad instantiation of human-friendly values (false-friendly AIs). Examples of indifferent AIs are things like paperclip maximisers, examples of false-friendly AIs are "keep humans safe" AIs who entomb everyone in bunkers, lobotomised and on medical drips.
The difference is apparent when you consider multiple AIs and negotiations between them. Imagine you have a large class of AIs, and that they are all indifferent (IAIs), except for one (which you can't identify) which is friendly (FAI). And you now let them negotiate a compromise between themselves. Then, for many possible compromises, we will end up with most of the universe getting optimised for whatever goals the AIs set themselves, while a small portion (maybe just a single galaxy's resources) would get dedicated to making human lives incredibly happy and meaningful.
But if there is a false-friendly AI (FFAI) in the mix, things can go very wrong. That is because those happy and meaningful lives are a net negative to the FFAI. These humans are running dangers - possibly physical, possibly psychological - that lobotomisation and bunkers (or their digital equivalents) could protect against. Unlike the IAIs, which would only complain about the loss of resources to the FAI, the FFAI finds the FAI's actions positively harmful (and possibly vice versa), making compromises much harder to reach.
And the compromises reached might be bad ones. For instance, what if the FAI and FFAI agree on "half-lobotomised humans" or something like that? You might ask why the FAI would agree to that, but there's a great difference to an AI that would be friendly on its own, and one that would choose only friendly compromises with a powerful other AI |
95b72d79-26e0-4131-bc4c-1de786dedb91 | trentmkelly/LessWrong-43k | LessWrong | Will issues are quite nearly skill issues
This post contains content that might be regarded as spoilers for the free will problem
I cannot jump off a cliff. I do not know any cliffs near me, but that's not the reason I care about, so we'll counterfactually skip over that. The more interesting reason I can't jump off a cliff is that I don't want to. When we confidently know that I don't want to jump off a cliff, and that no one in the near future would coerce or manipulate me to, we can conclude a negligibly tiny probability for my jumping off a cliff, or, in practice, that I will not do that. Once we accept that "could" expresses an option's plausibility, we're saying that "it cannot happen that I'd jump off a cliff", and, equivalently, "I couldn't jump off a cliff".
There is a straightforward sequence of muscle commands from my brain to my legs that would lead me to jump. But the existence of such motor outputs is not what we typically mean by "ability". If there was a cliff near me, but I didn't know about it (probably true anyway), that pattern of leg movements would just as much exist, but I'd be unable (in the conventional sense, not just "my" "weird" sense) to go to it and jump.
Not wanting to do something inhibits the person from doing it just as effectively as an explicit "inability". You can't just object "but free will!" (and be correct), both sith I just anticipated it, and so you wouldn't want to (unless you're trying to be annoying), and sith human minds are not mystically unpredictable blobs of chaos. People's minds have a theoretically-measurable internal state that constrains what they do.
If someone (honestly, after thorough thought, and not opposing their established habit) resolves to not do something (nontrivial and easily avoided), they will tend to not do it, unless they are affected by (sufficiently convincing) new input. A failure to accurately constrain someone's future actions is a matter of your limited knowledge about what they think or will encounter, not their inscrutable f |
eb1bf3fa-8f5b-4753-bb45-afe9b0aaf325 | trentmkelly/LessWrong-43k | LessWrong | Aubrey de Grey has responded to his IAMA - Now with Transcript!
Watch the video response here: http://www.youtube.com/watch?v=-tsI_28O3Ws
This was posted here on lesswrong a while ago, but they recently uploaded a new version of the video and I took the liberty of typing up a transcript.
The video is fairly long, about 25 minutes. But it's incredibly engaging and I highly recommend watching it. For those who prefer text (because it's faster or because you are a computer), you can read the transcript in this google doc, or below in the comments. Enjoy! |
d895ef5a-e2b2-4957-aa8d-9c07bd52cb12 | trentmkelly/LessWrong-43k | LessWrong | Back to Basics: Truth is Unitary
It was a dark and stormy night.
The prospect held the front of his cloak tight to his chest. He stumbled, fell over into the mud, and picked himself back up. Shivering, he slammed his body against the front doors of the Temple and collapsed under its awning.
He picked himself up and slammed his fists against the double ironwood doors. He couldn't hear his own knocks above the gale. He banged harder, then with all his strength.
"Hello! Is anyone in there? Does anyone still tend the Fire?" he implored.
There was no answer.
The Temple's stone walls were built to last, but rotting plywood covered the apertures that once framed stained glass. The prospect slumped down again, leaning his back against the ironwood. He listened to the pitter-patter of rain on overgrowth. It wasn't a bad place to think.
The trouble was, he didn't want to think. Not right now. Thinking creates depression. Action cures it. The prospect put his stiff hands in his pockets. His fingers traced the delicate forms of a disposable lighter bought on the darkweb and a short cheap aluminum-wrapped wax candle.
He considered lighting the candle under the Temple's awning. But that felt pathetic. If the Temple was abandoned then he should at least do it at the altar. The acolyte eyed the plywood. Surely he could punch through it and climb in that way. He left the shelter of the awning and tapped on the former window. His taps left fingerprints in the myceliation.
The ironwood doors opened. A young girl poked her head out. The prospect shouted in surprised and fell into the mud.
"What are you doing out there in the mud?" the girl asked.
"Choosing to dunk myself in the mud wasn't exactly an explicit rational choice," said the prospect while shaking himself off.
"Well come inside. Hypothermia impairs one's ability to make rational decisions," said the girl. She poked her head back inside the Temple and closed the door behind herself to keep out the rain.
The prospect looked at the door. He noticed |
0499ddcd-6918-4ee2-81d3-19d10d04a240 | StampyAI/alignment-research-dataset/special_docs | Other | The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents
1 THE SUPERINTELLIGENT WILL: MOTIVA TION AND
INSTRUMENT AL RATIONALITY IN ADVANCED
ARTIFICIAL AGENTS
(2012) Nick Bostrom
Future of Humanit y Institute
Facult y of Philosophy & Oxford Martin School
Oxford University
www.nickbostrom.com
[Minds and Machines, Vol. 22, Iss. 2, May 2012] [translation: Portuguese]
ABSTRACT
This paper discusses the relation between intelligence an d motivation in artificial agen ts,
developing and briefly arguing for tw o theses. The first, the orthogonality thesis , hol ds (with some
caveats) that intelligence and final goals (purposes) are orthogonal axes along which possible
artificial intellects can freely vary—more or less any level of intelligence could be combined with
more or less any final goal. The secon d, the instrumental conver gence thesis , holds that as long as
they possess a sufficient level of intelligence, agents having any of a wide range of final goals
will pursue simil ar intermediary goals becaus e the y have instrumental reasons to do so. In
combination , the two theses help us underst and the possibl e range of behavior of
superintelligent agents, and they point to some potential dangers i n building such an agent.
KEYWO RDS: superintelligence, artificial intelligence, AI, goal, instrumental reason , intelligent
agent
1. The orthogonali ty of motivati on and intelligence
1.1 Avoiding anthropomorphism
If we imagine a space in which all possible minds can be represented, we must imagine all human
minds as constituting a small and fairl y tight cluster wit hin that space. The personality
differences between Hannah Arendt and Benny Hill might seem vast to us, but this is because
the scale bar in our intuitive judgment is calibrat ed on the existing human distribution. In the
wider space of all logical possibilities , thes e tw o personalities are close neighb ors. I n terms of
neural architecture, at least, Ms. Arendt and Mr. Hill are nearly identical. Imagine their brains
laying si de by side in quiet repose. The differences would appear minor an d you would quite
readily recogni ze them as two of a kind; y ou mi ght even be unable to tell which brain was
whose. If you studied the morphology of the two brains more closely under a microsco pe, the
2
impression of fundamental similarity would only be strengthened: you would then see the same
lamellar organization of the cortex, made up of the same types of neuron, soaking in the same
bath of neurotransmitter molecules.1
It is well k nown th at naïve observers often anthropomorphize the capabilities of simple r
insensate systems. We might say, for example, “This vending machine is taking a long time to
think about my hot chocolate.” This might lead one either to underestimate the cognitive
complexity of capabilities which come naturally to human being s, such as motor control and
sensory perception, or, alternatively, to ascribe significant degrees of mindfulness and
intelligence to very dumb systems, such as chatterbo xes like Weizenbaum’s ELIZ A
(Weizenbaum 1976) . In a similar manner, there is a common tendency to anthropomorphize the
motivations of intelligent systems in which there is really no ground for expecting human -like
drives and passions (“My car really didn’t want to start this morning”) . Eliezer Yudkowsky
gives a nice illustration of this phenomenon :
Back in the era of pulp science fiction, magazine covers occasionally depicted a sentient
monstrous alien —colloquially known as a bug -eyed monster (BEM) —carrying off an
attractive human fem ale in a torn dress. It would seem the artist believed that a non -
humanoid alien, with a wholly different evolutionary history, woul d sexually desire
human females … Probably the artist did not ask whether a giant bug perceives human
females as attractiv e. Rather, a human female in a torn dress is sexy—inherently so, as an
intrinsic property. They who made this mistake did not think about the insectoid’s
mind: they focused on the woman’s torn dress. If the dress were not torn, the woman
would be less s exy; the BEM does not enter into it . (Yudkowsky 2008)
An artificial intelligence can be far less human -like in its motivations than a space alien .
The extraterrestrial (let us assume) is a biological creature who has arisen through a process of
evolution and may theref ore b e expected to have the kinds of motivation typical of evolved
creatures . For example, it would not be hugely surprising to find that some random intelligent
alien would have motives related to the attaining or avoiding of food, air, temperature, ener gy
expenditure, the threat or occurrence of bodily injury, disease, predators, reproduction, or
protection of offspring. A member of an in telligent social species might also have motivations
related t o coope ration and competition: like us, it might show i n-group loyalty, a resentment of
free-riders, perhaps even a concern with r eputation and appearance.
By contrast, an artificial mind need not care intrinsically about any of those things, not
even to the slightest degree. One can easily conceive of an artificial intelligence whose sole
fundamental goal is to count the grains of sand on Boracay , or to calculate decimal places of pi
indefinitely, or to maximize the total number of paperclips in its future lightcone. In fact, it
would be easier to create an AI with simple goals like these, than to build one that has a human -
like set of values and dispositions.
1 This is of course not to deny that differences that appear small visually can be functionally profound.
3
1.2 The orthogonality thesis
For our purposes, “intelligence” will be roughly taken to correspond to the capacity for
instrumental reasoning (more on this later) . Intelligent search for instrumentally optimal plans
and policies can be performed in the service of any goal. Intelligence and motivation can in this
sense be thought of as a pair of orthogonal axes on a graph whose points represent intelligent
agents of different paired specifications. Each point in the graph represents a logically possible
artificial agent , modulo some weak constraints —for instance, it might be impossi ble for a very
unintelligent system to have very complex motivations , since complex motivations would place
significant demands on memory. Furthermore, in order for an agent to “have” a set of
motivations , this set may need to be functionally integrated w ith the agent’s decision -processes,
which again would place demands on processing power and perhaps on intelligence. For minds
that can modify themselves, there may also be dynamical constra ints; for instance, an intelligent
mind with a n urgent desire to be stupid might not remain intelligent for very long. But these
qualifications should not obscure the main idea , which we can express as follows:
The Orthogonality Thesis
Intelligence and final goals are orthogonal axes along which possible agents can fr eely
vary . In other words, more or less any level of intelligence could in principle be
combined with more or less any final goal.
A comparison may be made here with the Humean theory of motivation . David Hume
thought that beliefs alone (say, about wha t is a good thing to do) cannot motivate action: some
desire is required.2 This would support the orthogonality thesis by undercutting one possible
objection to it, namely, that sufficient intelligence might entail the acquisition of certain beliefs,
and that these beliefs would necessaril y produce certain motivations. Not so, according to David
Hume: belief and motive are separate.
Although the orthogonality thesis can draw su pport from the Humean theory of
motivation, it does not presuppose it. In particular, one need not maintain that beliefs alone can
never motivate action. It would suffice to a ssume, for example, that an agent —be it ever so
intelligent —can be motivated to pursue any course of action if the agent happens to have certain
standing desires of some su fficient, overriding strength. Another way in which the
orthogonality thesis could be true even if the Humean theory of motivation is false is if
arbitrarily high intelligence does not entail the acquisition of any such beliefs as are (putatively)
motivating on their own . A third way in which it might be possible for the orthogonality thesis
to be true even if the Humean theory were false is if it is possible to build a cognitive system (or
more neutrally, an “optimization process”) with arbitrarily high intell igence but with
constitution so alien as to contain no clear functional analogues to what in humans we call
2 For some recent attempts to defend the Humean theory of motivation, see Smith (1987 ), Le wis (1988), and
Sinhababu (2009) .
4
“beliefs” and “desires” . This would be the case if such a system could be constructed in a way
that would make it motivated to pursue any given final goal.
The o rthogonality thesis , as formulated here, makes a claim about the relationship
between motivation and intelligence , rather than between motivat ion and rationality (or
motivation and reason ). This is because some philosophers use the word “rationality” to connote
a “normatively thicker ” concept than we seek to connote here with the word “intelligence”. For
instance , in Reasons and Persons Derek Parfit argues that certain ba sic preferences would be
irrational, such as that of an otherwise normal agent who has “Future -Tuesday -Indifference” :
A certain hedonist cares greatly about the quality of his future experiences. With one
exception, he cares equally about all the parts of his future. The exception is that he has
Future -Tuesday -Indifference. Throughout every Tuesday he cares in the normal way
about what is happening to him. But he never cares about possible pains or pleasures on
a future Tu esday... This indifference is a bare fact. When he is planning his future, it is
simply true that he always prefers the prospect of great suffering on a Tuesday to the
mildest pain on any other day. (Parfit 1984)3
Thus, the agent is now indifferent to his own future suffering if and only if it occurs on a future
Tuesday. For our purposes, we need take no stand on whether Parfit is right that this is
irrational, so long as we grant that it is not necessarily unintelligent . By “intelligence ” here we
mean something lik e instrumental rationality —skill at prediction, planning, and means -ends
reasoning in general . Parfit’s imaginary Future -Tuesday -Indifferent agent could have
impeccable instrumental rationality, and therefore great intelligence, even if he falls short on
some kind of sensitivity to “objective reason ” that might be required of a fully rational agent.
Consequently, this kind of example does not undermine the orthogonality t hesis.
In a similar vein, even if there are objective moral facts that any fully rat ional agent
would comprehend, and even if these moral facts are somehow intrinsically motivating (such
that anybody who fully comprehends them is necessarily motivated to act in accord ance with
them) this need not undermine the orthogonality thesis. The t hesis could still be true if an agent
could have impeccable instrumental rationality even whilst lacking some other faculty
constitutive of rationality proper, or some faculty required for the full comprehension of the
objective moral facts . (An agent cou ld also be extremely intelligent, even superintelligent ,
without having full instrumental rational ity in every domain. )
One reason for focusing on intelligence, that is, on instrumental rationality, is that this is
the most relevant concept if we are trying to figure out what different kinds of systems would
do. Normative questions , such as whether their behavior would count as being prudentially
rational or morally justifiable , can be important in various ways. However, such questions
should not bli nd us to the possibility of cognitive systems that fail to satisfy substantial
3 See also Parfit (2011) .
5
normative criteria but which are nevertheless very powerful and able to exert strong influence
on the world.4
1.3 Predicting superintelligence motivation and behavior
The orth ogonality thesis implies that synthetic minds can have utterly non -anthropomorphic
goals—goals as bizarre by our lights as sand -grain -counting or paperclip -maximizing. T his
hold s even (indeed especially ) for artificial agents that are extremely intelligen t or
superintelligent. Yet it does not follow from the orthogonality thesis that it is impossible to make
predictions about what parti cular agents will do. P redictability is important if one seeks to
design a system to achieve particular outcomes, and th e issue becomes more important the more
powerful the artif icial agent in question is. S uperintelligent agent s could be extremely powerful,
so it is important to develop a way of analyzing and predicting their behavior. Yet despite the
independence of intelligence and final goals implied by the orthogonality thesis, the problem of
predicting an agent’s behavior need not be intractable —not even with regard to hypothetical
superintelligent agents whose cognitive complexity and performance characteristics might
render them in certain respects opaque to human analysis.
There are at least three directions from which one can approach the problem of
predicting superintelligen t motivation :
(1) Predictability through design competence . If we can suppose that the designers of a
superintellig ent agent can successfully engineer the goal system of the agent so that it
stably pursues a particular goal set by the programmers, then one prediction we can
make is that the agent will pursue that g oal. The more intelligent t he agent is, the greater
the cognitive resourcefulness it will have to pursue that goal. So even before an agent
has been created we might be able to predict something about its behavior , if we know
something about who will build it and what goals they wi ll want it to have.
(2) Predictability th rough inheritance . If a digital intelligence is created directly from a human
template (as would be the case in a high -fidelity whole brain emulation), then the digital
intelligence might inherit the motivation s of the human template .5 The agent might
retain some of these motivations even if its cognitive capacities are subsequently
enhanced to make it superintell igent. This kind of inference requires caution . The
agent’s goals and values could easily become corrupted in the uploading process or
during its subse quent operation and enhancement, depending on how the procedure is
implemented.
4 The orthogonality thesis implies that most any combination of final goal and intelligence level is logically
possible; it does not imply that it would be practically easy to endow a superintelligent agent with some
arbitrary or human -respecting final goal —even if we knew how to construct the intelligence part. For
some preliminary notes on the value -loading problem, see , e.g., Dewey (2011 ) and Yudkowsky (2011) .
5 See Sandberg & Bostrom (2008 ).
6
(3) Predictability through convergent instrumental reasons . Even without detailed knowledge of
an agent’s final goals, we m ay be able to infer something about its more immediate
objectives by considering the instrumental reasons that would arise for any of a wide
range of possible final goals in a wide range of situations. This way of predicting
becomes more useful the greate r the intelligence of the agent , because a more intelligent
agent is more likely to recognize the true instrumental reasons for its actions, and so act
in ways that make it more likely to achieve its goals .
The next section explores this third way of pre dictability and develops an “instrumental
convergence thesis” which complement s the orthogo nality thesis .
2. Instrumental convergence
According to the orthogonality thesis, artificial intelligent agents may have an enormous range
of possible final goals . Ne vertheless, according to what we may term the “instrumental
convergence ” thesis, there are some instrumental goals likely to be pursued by almost any
intelligent agent , because there are some objectives that are useful intermediaries to the
achievement of almost any final goal. We can formulate this thesis as follows:
The Instrumental Convergence Thesis
Several instrumental values can be identified which are convergent in the sense that their
attainment would increase the chances of the agent’s goal being realized for a wide range
of final goals and a wide range of situations, implying that these instrumental values are
likely to be pursued by many intelligent agents.
In the following we will consider several categories where s uch convergent instrumental
values may be found.6 The likelihood that an agent will recognize the instrumental values it
conf ronts increases (ceteris paribus) with the agent’s intelligence . We will therefore focus mainly
on the case of a hypot hetical superintelligent agent whose instrumental reasoning capacities far
6 Stephen Omohundro has written two pioneering pape rs on t his topic (Omohundro 2008 a, 2008b) .
Omohundro argues that all advanced AI systems are likely to exhibit a number of “basic drives ”, by which
he means “ tendencies which will be present unless explicitly countera cted.” The term “AI drive” has the
advantage of being short and evocative, but it has the disadvantage of suggesting that the instrumental
goals to which it refers influence the AI’s decision -making in the same way as psychological drives
influence human decision -making, i.e. via a kind of phenomenological tug on our ego which our willpower
may occasionall y succeed in resisting. That connotation is unhelpful. One would not normally say that a
typical human being has a “drive” to fill out their tax return, even though filing taxes may be a fairly
convergent instrumental goal for humans in contemporary soci eties (a goal whose realization averts
trouble that would prevent us from realizing many of our final goals). Our treatment here also differs from
that of Omohundro in some other more substantial ways, although the underlying idea is the same. (See
also Chalmers (2010) and Omohundro (2012) .
7 exceed those of any human. We will also comment on how the instrumental convergen ce thesis
applies t o the case of human beings, as this gives us occasion to elaborate some essential
qualifications concernin g how the instrumental convergence thesis should b e interpreted and
applied. Whe re there are convergent instrumental values, w e may be able t o predict some
aspects of a superintel ligence’s behavior even if we know virtually nothing about that
superintelligence’s final goals.
2.1 Self-preservation
Suppose that an agent has some final goal that extends some way into the future. There are
many scenarios in which the agent, if it is still around in the future, is then able to perform
actions that increase the probability of achieving the goal . This creates an instrumental reason
for the agent to try to b e around i n the future—to hel p achieve its present future-oriented goal.
Agents with human-li ke motivational structures often seem to place som e final val ue on
their own survival . This is not a necessary feature of artificial agen ts: som e may b e designed to
place no final value whatever on thei r own survival. Nevertheless, even agents that do not care
intrinsicall y about their own survival would, under a fairly wide range of conditions, care
instrumentally to some degree about their own survival i n order to accomplish the final goals
they do value.
2.2 Goal-conten t integrity
An age nt is more likel y to act in the future to maximize the realization of its present final goals if
it still has thos e goals in the future. This gives the agent a present instrumental reason to prevent
alterations of its final goals. (This argument applies onl y to fi nal goals. In order t o attain its final
goals, an intelligent agen t will of course routinely want to chan ge its subgoals in light of new
information and insight.)
Goal-content integrit y for final goals is in a sense even mo re fundamental than survival
as a convergent instrumental motivation. Among humans , the opposite may seem to be the case,
but that is because survival is usually part of our final goals . For software agents, which can
easily switch bodies or create exact duplicates of themselves, preservation of self as a particular
implementation or a particular physical object nee d not be an important instrumental value.
Advanced software agents might als o be able to swap memories, download skills, an d radically
modif y their cognitive architecture and personalities. A population of such agents might operate
more like a “functional soup” than a society composed of distinct semi-permanent persons.7 For
some purposes, processes i n such a system mi ght b e bett er individuated as teleological threads ,
based on their final values, rath er than on the basis of bodies , personalities , memories, or
abilities. In such scenarios, goal-continuity might be said to constitute a key aspect of survival.
Even so, there are situations in which an agent may intentionally chan ge its ow n final
goals. Such situations can arise when any of the following factors is significant:
7 See Chislenko ( 1997) .
8
Social signaling . When others can perceive an agent’s goals and use that information to
infer instrumentally relevant dispositions or other correlated attributes, it can be in the
agent’s interest to modify its goals to make whatever desired impression. For example,
an age nt might miss out on beneficial deals if potential partners cannot trust it to fulfill
its side of the bargain. In order to make credible commitments, an agent might therefore
wish to adopt as a final goal the honoring of its earlier commitments, and to a llow others
to verify that it has indeed adopted this goal. Agents that could flexibly and
transparently modify their own goals could use this ability to enforce deals among one
another.8
Social preferences . Others may also have preferences about an agen t’s goals. The agent
could then have r eason to modify its goal s, either to satisfy or to frustrate those
preferences.
Preferences concerning own goal content . An agent might have some final goal concerned
with the agent’s own goal content. For example, the a gent might have a final goal to
become the type of agent that is motivated by certain values, such as compassion .
Storage costs . If the cost of storing or processing some part of an agent’s utility function is
large compared to the chance that a situ ation will arise in which applying that part of the
utility function will make a difference, then the agent has an instrumental reason to
simplify its goal content, and it may trash that part of the utility function.9 10
We h uman s often seem happy to let our final goals and values drift . This might often be
because we do not know precisely what they are . We obviously want our beliefs about o ur final
goals and values to be able to change in light of continuing self -discovery or changing self -
presentat ion needs. However, there are cases in which we willingly change the goals and values
themselves, n ot just our beliefs or interpretations of them . For example, somebody deciding to
have a child might predict that they will come to value the child for its own sake, even though at
the time of the decision they may not particularly value their future child or even like children in
general.
8 See also Shulman (2010) .
9 An agent might also change its goal representation if it changes its ontology, in order to transpose its old
representation into the new ontology. Cf. de Blanc (2011) .
10 Another type of factor that mig ht make an evidential decision theorist undertake various actions, including
changing its final goals, is the evidential import of deciding to do so. For example, an agent that follows
evidential decision theory might believe that there exist other agents like it in the universe, and that its own
actions will provide some evidence about how those other agents will act. The agent might therefore
choose to adopt a final goal that is altruistic towards those other evidentially -linked agents, on grounds
that this will give the agent evidence that those other agents will have chosen to act in like manner. An
equivalent outcome might be obtained, however, without changing one’s final goals, by choosing in each
instant to act as if one had those final goals.
9
Humans are complicated, and many factors might be at play in a situation like this.11 For
instance, one might have a fi nal value that involves becoming the kind of person who cares
about some other individual for his or her own sake (here one places a final value on having a
certain final value ). Alternatively, one might have a final value that involves having certain
experiences and occupying a certain soci al role ; and becoming a parent —and undergoing an
associated goal s hift—might be a necessary part of that. Human goals can also have inconsistent
content, goal content; and so some people might want to modify some of th eir final goals to
reduce the inconsistencies.
2.3 Cognitive enhancement
Improvements in rationality and intelligence will tend to improve an agent’s decision -making,
making the agent more likely to achieve her final goals . One would therefore expect co gnitive
enhancement to emerge as an instrumental goal for many types of intelligent agent. For similar
reasons, agents will tend to instrumentally value many kinds of information.12
Not all kinds of rationality, intelligence, and knowledge need be instrum entally useful in
the attainment of an agent’s final goals. “Dutch book arguments” can be used to show that an
agent whose credence function does not obey the rules of probability theory is susceptible to
“money pump” procedures, in which a savvy bookie a rranges a set of bets, each of which
appears favorable according to the agent’s beliefs, but which in combination are guaranteed to
result in a loss to the agent, and a corresponding gain for the bookie. However, this fact fails to
provide any strong gene ral instrumental reasons to seek to iron out all probabilistic incoherency .
Agents who do not expect to encounter savvy bookies, or who adopt a general policy against
betting, do not stand to lose much from having some incoherent beliefs —and they may gain
important benefits of the types mentioned: reduced cognitive effort, social signaling, etc. There
is no general reason to expect an agent to seek instrumentally useless forms of cognitive
enhancement , as an agent might not value knowledge and understandi ng for their own sakes.
Which cognitive abilities are instrumentally useful depends both on the agent’s f inal
goals and its situation . An agent that has access to reliable expert advice may have little need for
its own intelligence and knowledge , and it may therefore be indifferent to these resources . If
intelligen ce and knowledge come at a cost, such as time and effort expended in acquisition, or in
increased sto rage or processing requirements, then an agent might prefer less knowledge and
11 An extensive psychological literature explores adaptive preference fo rmation. See, e.g., Forgas et al.
(2009 ).
12 In formal models, the value of information is quantified a s the difference between the expected value
realized by optimal decisions made with that information and the expected value realized by optimal
decision s made without it. (See, e.g., Russell & Norvig 2010 .) It follows that the value of information is
never negative. It also follows that any information you know will never affect any decision you will ever
make has zero value for you. However, this kind of m odel assumes several idealizations which are often
invalid in the real world —such as that knowledge has no final value (meaning that knowledge has only
instrumental value and is not valuable for its own sake), and that agents are not transparent to other
agents.
10
less intellig ence.13 The same can hold if the agent has final goals that involve bei ng ignorant of
certain facts: likewise if an agent faces incentives arising from strategic commitments, signaling,
or social preferences, as noted above.14
Each of these countervailing reasons often comes into play for human beings . Much
information is irrelevant to our goals; we can often rely on others’ skill and expertise; acquiring
knowledge takes time and effort; we might intrinsically value certain kinds of ignorance; and we
opera te in an environment in which the ability to make strategic commitments, socially signal,
and satisfy other people’s direct preferences over our own epistemic states , is often important to
us than simple cognitive gains .
There are special situations in wh ich cognitive enhancement may result in an enormous
increase in an agent’s ability to achieve its final goals —in particular, if the agent’s final goals are
fairly unbounded and the agent is in a position to become the first superintelligence and thereby
potentially obtain a decisive advantage enabling the agent to shape the future of Earth -
originating life and accessible cosmic resources according to its preferences . At least in this
special case, a rational intelligent agent would place a very high instru mental value on cognitive
enhancement.
2.4 Technolog ical perfection
An agent may often have instrumental reasons to seek better technology, which at its simplest
means seeking more efficient ways of transforming some given set of inputs into valued outputs.
Thus , a software agent might place an instrumental value on more efficient algorithms that
enable its mental functions to run faster on given hardware. Similarly, agents whose goals
require some form of physical construction might instrumentally value improved engineering
technology which enable s them to create a wider range of structures more quickly and reliably ,
using fewer or cheaper materials and less energy. Of course , there is a tradeoff: the p otential
benefits of better technology mus t be weighed against its costs, including not only the cost of
obtaining the technology but also the costs of learning how to use it, integrating it with other
technologies already in us e, and so forth.
Proponents of some new technology, confident in its superiority to existing alternatives ,
are often dismayed when other people do not share their enthusiasm, but peoples’ resistance to
novel and nominally superior technology need not be b ased on ignorance or irrationality . A
technology’s valence or normative character depends not only on the context in which it is
deployed, but also the vantage point from which its impacts are evaluated: what is a boon from
one person’s perspective can be a liability from another’s. Thus, although mechanized looms
increased the economic efficie ncy of textile production, the Luddite handloom weavers who
13 This strategy is exemplified by the sea squirt larva, which swims about until it finds a suitable rock, to
which it then permanently affixes itself. Cemented in place, the larva has less need for complex
information processing, whence it proceeds to digest part of its own brain (its cerebral ganglion).
Academics can sometimes observe a similar phenomenon in colleagues who are granted tenure.
14 Cf. Bostrom (2012 ).
11
anticipated that the innovation would render their artisan skills obsolete may have had good
instrumenta l reason s to oppose it. The point here is that if “technolog ical perfection” is to name
a widely convergent instrumental goal for intelligent agents, then the term must be understood
in a spec ial sense—technology must be construed as embedded in a particu lar social context, and
its costs and benefits must be evaluated with reference to some specified agents’ final values.
It seems that a superintelligent singleton—a superintelligent agent that faces no
significant intelligent rivals or opposition , and is thus in a position to determ ine global policy
unilaterally —would have instrumental reason to perfect the technologies that would make it
better able to shape the world according to its preferred designs.15 This would probably include
space colonization tec hnology, such as von Neumann probes —automatic, self -mending and self -
replicating spaceships that can extend its reach beyond the Solar System. Molecular
nanotechnology, or some alternative still more capable physical manufacturing technology , also
seem s potentially very useful in the service of an extremely wide range of final goals.16
2.5 Resource acquisition
Finally, resource acquisition is a nother common emergent instrumental goal, for muc h the same
reasons as technological perfection: both technology and resources facilitate physical
construction projects.
Human beings tend to seek to acquire resources sufficient to meet their basic biological
needs. But people usually seek to acquire resources far beyond this minim um level. In doing so,
they may b e partially driven by lesser physical desiderata, such as in creased comfort and
convenience. A great deal of resource accumulation is motivated by social concerns —gaining
status, mates, friends and influence, through wealth accumulation and conspicuous
consumption. Perhaps l ess commonly, some people seek additional resources to achieve
altruistic or expensive non -social aims.
15 Cf. Bostrom (2006 ).
16 One could reverse the question and look instead at possible reasons for a superintelligent singleton not to
develop some technolo gical capabilities. These include: (a) The singleton foreseeing that it will have no
use of some technolo gical capability ; (b) The development c ost being too large relative to its anticipated
utility . This would be the case if, for instance , the technology will never be suitable for achieving any of the
singleton’s ends, or if the singleton has a very high discount rate that s trongly discourages investment; (c)
The singleton having some final value that requires abstention from particular avenues of technology
development; (d) If the singleton is not certain it will remain stable, it might prefer to refrain from
developing tech nologies that could threaten its internal stability or that would make the consequences of
dissolution worse (e.g., a world government may not wish to develop technologies that would facilitate
rebellion, even if they had some good uses, nor develop t echno logies for the easy production of weapons of
mass destruction which could wreak havoc i f the world government were to dissolve); (e) Similarly, the
singleton might have made some kind of binding strategic commitment not to develop some technology, a
commit ment that remains operative even if it would now be convenient to develop it. (Note, however, that
some current reasons for technology -development would not apply to a singleton: e.g., reasons arising from
unwanted arms races.)
12
On the basis of such observations it might be tempting to suppose that a
superintelligence not facing a competitive social world wo uld see no instrumental reason to
accumulate resources beyond some modest level, for instance whatever computational resources
needed to run its mind along with some virtual reality. Yet such a supposition would be entirely
unwarranted. First, the value of resources depends on the uses to which they can be put, which
in turn depends on the available technology. With mature technology, basi c resources such as
time, space, and matter , and other forms of free energy , could be processed to serve almost any
goal. For instance, such basic resources could be converted into life. Increased computational
resources could be used to run the super intelligence at a greater speed and for a longer duration,
or to create additional physical or simulated (virtual) lives and civilizations. Extra physical
resources could also be used to create backup systems or perimeter defenses, enhancing security.
Such projects could easily c onsume far more than one planet’s worth of resources.
Furthermore, the cost of acquiring additional extraterrestrial resources will decline
radically as the technology matures . Once von Neumann probes can be built, a large portion of
the observable universe (assuming it is uninhabited by intelligent lif e) could be gradually
colonized —for the one -off cost of building and launching a single successful self -reproducing
probe. This low cost of celestial resource acquisition would mean that such expansion could be
worthwhile even if the value of the additional resources gained were somewhat marginal. For
example, even if a superintelligence cared non -instrumentally only about what happens within
some pa rticular small volume of space, such as the space occupied by its origina l home planet, it
would still have in strumental reasons to harvest the resources of the cosmos beyond. It could
use those surplus resources to build computers to calculate more optimal way s of using
resources within the small spatial region of primary concern. It could also use the extra
resources to build ever -more robust defenses to safeguard the privileged real estate. Since the
cost of acquiring additional resources would keep declining, this process of optimizing and
increasing safeguards might well continue indefinitely even if it were subject to steeply declining
returns.17 18
17 Suppose that an agent dis counts resources obtained in the future at an exponential rate, and that because
of the light speed limitation the agent can only increase its resource e ndowment at a polynomial rate.
Would this mean that there will be some time after which the agent woul d not find it worthwhile to
continue acquisitive expansion? No, because although the present value of the resources obtained at future
times would asymptote to zero the further into the future we look, so would the present cost of obtaining
them. The pre sent cost of sending out one more von Neumann probe a 100 million years from now
(possibly using some resource acquired some short time earlier) would be diminished by the same
discount factor that would diminish the present value of the future resources t he extra probe would
acquire (modulo a constant factor).
18 Even an agent that has an apparently very limited final goal, such as “to make 32 paperclips”, could
pursue unlimited resource acquisition if there were no relevant cost to the agent of doing so. For example,
even after an expected -utility -maximizing agent had built 32 paperclips, it could use some extra resources
to verify that it had indeed successfully built 32 paperclips meeting all the specifications (and, if necessary,
to take corrective acti on). After it had done so, it could run another batch of tests to make doubly sure that
no mistake had been made. And then it could run another test, and another. The benefits of subsequent
tests would be subject to steeply diminishing returns; however, so long as there were no alternative action
13
Thus, there is an extremely wide range of possible final goals a superintelligent singleton
could have that would generate the instrumental goal of unlimited resource acquisition . The
likely manifestation of thi s would be the superintelligence’s initiation of a colonization process
that would expand in all directions using von Neumann probes . This would roughly result in a
sphere of expanding infrastructure centered on the originating planet and growing in radiu s at
some fraction of the speed of light; and the colonization of the universe would continue in this
manner until the accelerating speed of cosmic expansion ( a consequence of the positive
cosmological constant) makes further material acquisition physicall y impossible as remoter
regions drift permanently out of reach .19 By contrast, agents lacking the technology required for
inexpensive resource acquisition , or for the conversion of generic physical resources into useful
infrastructure , may often find it not cost-effective to invest any present resources in increa sing
their material endowment. The same may hold for agents operating in competition with other
agents of similar powers. For instance, if competing agents have already secured accessible
cosmic resources, a late -starting agent may have no colonization opportunities. The conver gent
instrumental reasons for superintelligences un certain of the non-existence of other powerful
superintelligent agents are complicated by strategic considerations in wa ys that we do not
currently fully comprehend but which may constitute important qualifications to the examples
of convergent instrumental reasons we have looked at here.20
It should be emphasized that the existence of convergent instrumental reasons, even if
they apply to and are recognized by a particular agent, does not imply that the agent’s behavior
is easily predictable. An agent might well think of ways of pursuing the relevant instrumental
values that do not readily occur to us. This is especially true for a superintelligence, which could
devi se extremely clever but counterintuitive plans to realize its goals, possibly even exploiting
as-yet undiscovered physical phenomena. What is predictable is that the convergent
with a higher expected utility, the agent would keep testing and re -testing (and keep acquiring more
resources to enable these tests).
19 While the volume reached by colonization probes at a given time might be roughly spherical and
expanding with a rate proportional to the square of time elapsed since the first probe was launched (~t2),
the amount of resources contained within this volume will follow a less regular growth pattern, since the
distribution of resou rces is inhomogeneous and varies over several scales. Initially, the growth rate might
be ~t2 as the home planet is colonized; then the growth rate might become spiky as nearby planets and
solar systems are colonized; then, as the roughly disc -shaped volu me of the Milky Way gets filled out, the
growth rate might even out, to be approximately proportional to t; then the growth rate might again
become spiky as nearby galaxies are colonized; then the growth rate might again approximate ~t2 as
expansion procee ds on a scale over which the distribution of galaxies is roughly homogeneous; then
another period of spiky growth followed by smooth ~t2 growth as galactic superclusters are colonized; until
ultimately the growth rate starts a final decline, eventually rea ching zero as the expansion speed of the
universe accelerates to such an extent as to make further colonization impossible.
20 The simulation argument may be of particular importance in this context. A superintelligent agent may
assign a significant probab ility to hypotheses according to which it lives in a computer simulation and its
percept sequence is generated by another superintelligence, and this might various generate convergent
instrumental reasons depending on the agent’s guesses about what types o f simulations it is most likely to
be in. Cf. Bostrom (2003 ).
14
instrumental values would be pur sued and used to realize the agent’s final goals, not the specific
actions that the agent would take to achieve this.
Conclusions
The orthogonality thesis suggests that we cannot blithely assume that a superintelligence will
necessarily share any of the f inal values stereotypically associated with wisdom and in tellectual
development in humans —scientific curiosity, benevolent concern for others, spiritual
enlightenment and contemplation , renunciation of material acquisitiveness , a taste for refined
culture or for the simple pleasures in life , humility and selflessness, and so forth. It might be
possible through deliberate effort to construct a superintelligence that values such things , or to
build one that values human welfare, moral goodness, or any other complex purpose that its
designers might want it to ser ve. But it is no less possible—and probably technically easie r—to
build a superintelligence that places final value on nothing but calculating the decimals of pi .
The instrumental convergence thesis suggests that we cannot blithely assume that a
superintelligence with the final goal of calculating the decimals of pi (or making paperclips, or
counting grains of sand) would limit its activities in such a way as to not materially infringe on
human intere sts. An agent with such a final goal would have a convergent instrumental reason ,
in many situations, to acquire an unlimited amount of physical resources and, if possible, to
eliminate potential threats to itself and its goal system .21 It might be possib le to set up a situation
in which the optimal way for the agent to pursue these instrumental values (and thereby its final
goals) is by promoting human welfare, acting morally, or serving some beneficial purpose as
intended by its creators. However, if and when such an agent finds itself in a different situation,
one in which it expects a greater number of decimals of pi to be calculated if it destroys the
human species than if it continues to act cooperatively, its behavior would instantly take a
sinister turn. This indicates a danger in relying on instrumental values as a guarantor of safe
conduct in future artificial agents that are intended to become superintelligent and that might be
able to leverage their superi ntelligence into extreme levels power a nd influence .22
References
Bostrom, N. (2003). Are You Living in a Computer Simulation? Philosophical Quarterly , 53(211),
243-255.
Bostrom. N. (2006). What is a Singleton? Linguistic and Philosophical Investigations , 5(2), 48 -54.
21 Human beings might constitute potential threats; they certainly c onstitute physical resources.
22 For co mments and discussion I am grateful to Stuart Armstrong, Grant Bartley, Owain Evans, Lisa
Makros, Luke Muehlhauser, Toby Ord, Brian Rabkin, Rebecca Roache, Anders Sandberg, and three
anonymous referees.
15
Bostrom, N. (2012 ). Information Hazards: A Typology of Potentia l Harms from Knowledge.
Review of Contemporary Phil osophy , 10, 44 -79. [www.nickbostrom.com/information -hazards.pdf ]
Chalmers, D. (2010): The Singularity: A Philosophi cal Analysis. Journal of Consciousness Studies ,
17, 7 -65.
Chislenko, A. (1997). Technology as Extension of Human Functional Architecture. Extropy
Online . [project.cyberpunk.ru/idb/technology\_ as\_extension.html ]
de Blanc, P. (2011). Ontological Crises in Artificial Agent’s Value Systems. Manuscript . The
Singularity Institute for Artificial In telligence. [arxiv.org/pdf/1105.3821v1.pdf ]
Dewey, D. (2011). Learning What to V alue. In Schmidhuber, J., Thori sson, K. R., Looks, M. (eds.).
Proceedings of the 4th Conference on Artificial General Intelligence , AGI 2011 (pp. 309 -314),
Heidelberg: Springer.
Forgas, J. et al. (eds.) (2009). The Psychology of A ttitude s and Attitude Change . London: Psychology
Press.
Lewis, D. (1988). Desire as belief. Mind , 97(387), 323 -332.
Omohundro, S. (2008a). The Basic AI Drives. In P. Wang, B. Go ertzel, and S. Franklin (eds.).
Proceedings of the First AGI Conference , 171 , Frontiers in Artificial Intelligence and Appli cations.
Amsterdam: IOS Press.
Omohundro, S. (2008b). The Nature of Self -Improving Artificial Intel ligence. Manuscript .
[selfawaresystems.files.wordpress.com/2008/01/nature\_of\_self\_improving\_ai.pdf ]
Omohundro, S. (forthcoming 2012). Rationally -Shaped Artificial Intelligence. In Eden, A. et al.
(eds.). The Singularity Hypothesis: A Scient ific and Philosophical Assessment (Springer, forth coming).
Parfit, D. (1984). Reasons and Persons . (pp. 123 -4). Reprinted and c orrected edition, 1987. Oxford:
Oxford University Press.
Parfit, D. (2011). On What Matters . Oxford: Oxford University Press.
Russell, S. and Norvig, P. ( 2010). Artificial Intelligence: A Modern Approach . (3rd ed. ). New Jersey:
Prentice Hall.
Sandberg, A. and Bostrom, N. (2008). Whole Brain Emulation: A Roadmap . Technical Report 2008–
3. Oxford: Future of Humanity Institute, O xford Univer sity.
16
[www.fhi.ox.ac.uk/Reports/200 8-3.pdf ]
Shulman, C. (2010). Omohundro’s “Basic AI Drives” and Cata strophic Risks. Manuscript .
[singinst.org /upload/ai -resource -drives.pdf ]
Sinhababu, N. (2009). The Humean Theory of Motiva tion Reformulated and Defended.
Philosophical Review , 118(4), 465 -500.
Smith, M. (1987). The Humean Theory of Motivation. Mind , 46 (381): 36 -61.
Weizenbaum, J. (1976). Computer Power and Human Reason: Fro m Judgment to Calculation . San
Francisco: W. H. Freeman.
Yudkowsky, E. (2008). Artificial Intelligence as a Positive and Negative Factor i n Global Ri sk. In
Bostrom, N. and Cirkovic, M. (eds.). Global Catastrophic R isks. (pp. 308 -345; quote from p. 310).
Oxford: Oxford University P ress.
Yudkowsky, E. (2011). Complex Value Systems Are Required to Realize Valuable Futures. In
Schmidhuber , J., Thorisson, K. R., Looks, M. (eds.). Procee dings of the 4th Conference on Artificial
General Intelligence , AGI 2011 (pp. 388 -393). Heidelberg: Springer. |
623fe9dc-1e34-4057-b65f-3b33ff3b81d7 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | How important are accurate AI timelines for the optimal spending schedule on AI risk interventions?
Summary
-------
I present an extension to my [*optimal timing of spending on AGI safety*](https://forum.effectivealtruism.org/posts/Ne8ZS6iJJp7EpzztP/the-optimal-timing-of-spending-on-agi-safety-work-why-we#Explaining_and_estimating_the_model_parameters) model for calculating the value of information of AGI timelines via informing one’s spending schedule.
I show, using my best guess of the model parameters, that for an AI risk funder uncertain between a ‘short timelines’ model and ‘medium timelines’ model:
* Updating from near certainty in medium timelines to short timelines (and following the new optimal spending strategy) leads to a 40% increase in utility.
* Updating from near certainty in short timelines to medium timelines (and following the new optimal spending strategy) leads to a 20% increase in utility.
The gains are greater when considering a model of the community’s capacity, rather than capital.
I also show that small changes in one’s credence in short or medium timelines has relatively little impact on one’s optimal spending schedule, especially when one starts out with roughly equal credence in each[[1]](#fn67zyqyu77fw).
You can enter your own parameters - such as AGI timelines, discount rate and diminishing returns to spending - [here](https://colab.research.google.com/drive/17f0yTYcFSBihZFd8_N7yRjF2XYJDjvBZ?usp=sharing).
In an appendix I apply a basic model to consider the opportunity cost of timelines work. This model does not assume novel research is done.
The setup
---------
Suppose you have two ‘models’ of AGI timelines A.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
and B, with credence pA in A and (1−pA) in B[[2]](#fnslj8hqva7ad). You use your mixture distribution for AGI timelines pA⋅A+(1−pA)⋅B to calculate the optimal spending schedule for AI risk interventions.
You could do some thinking and come to some credence p′A in A and (1−p′A) in B. How much better is the optimal spending schedule as a result of p′A to the optimal spending schedule as a result of pA, both supposing p′A?
Writing SpA for optimal spending schedule according to pA and U(SpA|pA) for utility of SpA supposing PA, I compare U(Sp′A|p′A) to U(SpA|p′A) (the former, by definition of optimality is greater or equal to the latter).
In the model, utility is the discount adjusted probability we ‘succeed’ with making AGI go well. The maximum utility is ∫t0p(t)⋅e−δtdt≤1 where p(t) are the AGI timelines and d is the discount rate.
Example results for short and medium timelines
----------------------------------------------
I take A and B as the following, and other parameters as described [here](https://forum.effectivealtruism.org/posts/Ne8ZS6iJJp7EpzztP/the-optimal-timing-of-spending-on-agi-safety-work-why-we#Explaining_and_estimating_the_model_parameters). (See the appendix for some basic statistics on the mixtures of A and B.
| | |
| --- | --- |
| | |
| A**:** ‘Short’ 2027 median timelines | B: [Metaculus community timelines](https://web.archive.org/web/20221208164804/https://www.metaculus.com/questions/5121/date-of-artificial-general-intelligence/) (2039 median) |
I compute the results for both the ‘main’ model and the ‘alternate’ model [described in the previous post](https://forum.effectivealtruism.org/posts/Ne8ZS6iJJp7EpzztP/the-optimal-timing-of-spending-on-agi-safety-work-why-we).
* The main model gives the optimal spending rate (in $ per year) for every time step on (1) research and (2) influence (being able to get AI risk reduction ideas implemented).
+ Supposing there was only one funder of AI risk interventions, they should follow this spending strategy.
* The alternate model gives the optimal ‘crunch’ rate at each time. At each time point we can either invest in our own capacity (be it through training ,hiring, some types of research, investing) or ‘crunch’ - spend capacity to produce work directly beneficial to reducing AI risk.
+ This model gives more abstract results and is more applicable for individuals and teams.
I first show two examples of the optimal spending schedule for each of the models: one that is optimal supposing A, short timelines, and one that is optimal supposing medium timelines, B.
| | |
| --- | --- |
| | |
| **Main model**The optimal spending schedule assuming A, SA, (left) and the optimal spending schedule assuming B, SB, (right), both assuming medium difficulty.U(SB|A)=0.178 and U(SA|A)=0.213: going from certainty in B to certainty in A and following the new optimal spending schedule leads to 20% subjective increase in utility.U(SA|B)=0.127 and U(SB|B)=0.172: going from certainty in A to certainty in B and following the new optimal spending schedule leads to 35% subjective increase in utility |
| | |
| --- | --- |
| | |
| **Alternate model**The optimal spending schedule assuming A,SA (left) and the optimal spending schedule assuming B, SB(right), both assuming medium difficulty.U(SB|A)=0.113 and U(SA|A)=0.134: going from certainty in B to certainty in A and following the new optimal spending schedule leads to 18% subjective increase in utility.U(SA|B)=0.148 and U(SB|B)=0.217: going from certainty in A to certainty in B and following the new optimal spending schedule leads to 45% subjective increase in utility |
Results & analysis
------------------
**Disclaimer**: *these results are based on my guesses of the model parameters. Further, the* [*models are of course not without limitations*](https://forum.effectivealtruism.org/posts/Ne8ZS6iJJp7EpzztP/the-optimal-timing-of-spending-on-agi-safety-work-why-we#Limitations)*. I expect the results to be directionally equal, but lower, for* [*the robust spend-save model*](https://forum.effectivealtruism.org/posts/mLqKcYahP5nbjK9ir/tristan-cook-s-shortform?commentId=So4dTJWomiE975T7m)*. Overall, I’m less confident in these results than the previous spending results.*
For each pairing (main capital spending model, alternate community capacity spending model) x AGI success is (easy, medium, hard) I show three graphs. The leftmost plot shows the % increase in discount-adjusted probability of success (utility). The central plot shows this % increase after first factoring out the utilons we get for free - our probability of success if we contributed nothing. The rightmost plot shows the absolute increase in utility.
Main model, easy[[3]](#fnssxf53peoz8)Main model, medium[[4]](#fncwl364qwxbo)Main model, hard[[5]](#fnm50jfe7af7)Alternate model, easy[[6]](#fn53c4d5oy76)
Alternate model, medium[[7]](#fn0t8ah2gxe5if)Alternate model, hard[[8]](#fnhurtu0mihdd)Both the main and alternate model show that:
* The greatest gains are achievable in the *hard* case where we are highly confident in short timelines but mistaken. The *easy* case has the lowest gains: this is likely because we can do well regardless of our spending schedule.
* There are greater gains to be had when one is very confident in *short* timelines than when one is very confident in medium to long timelines,
* The maximum increases in utility are comparable to the gain in utility achieved from the approximate gain the community can make from moving from its current spending strategy to the optimal spending strategy.
* There is little to no gain in utility when one’s credence changes in A by less than 10% (this is the bottom-left to top-right diagonal band of purple).
* The main model has greater difference between strategies than the alternate model (compare the examples above) which I believe leads to the rectangular-like areas in the plots: small increases in pA can push greatly towards early spending.
* The alternate model sees greater utility gains from accurate timelines than the main model. This is likely due to:
+ slightly greater marginal returns to ‘crunching’ than there are to spending capital in the main model
+ greater opportunity cost of spending early (since there are higher returns to saving).
### In practise
Naturally, those with low [credal resilience](https://forum.effectivealtruism.org/topics/credal-resilience) in their AGI timelines have better returns to work on timelines.
The greatest potential gains of timelines work is for people already highly convinced of short timelines. This is particularly true if they are sacrificing gains in capacity now in order to ‘crunch’, or spending at a high rate now (sacrificing a greater amount of capital later). However, it seems that:
1. many crunch-like activities may also be building capacity (e.g. building skills necessary for crunchtime), especially because it is relatively neglected in the community
2. that the current community spending rate is much lower than the optimal schedule implied even by medium timelines, and so marginally pushing for greater spending is supported regardless of one's credence in short timelines.
In the other direction, the value of information moving from high confidence in medium timelines to high confidence in short timelines are likely to be higher than the results suggest *because* the current spending rate and crunching rate are too low. That is, if you become convinced of short timelines, your marginal spending/crunching has lower diminishing returns (because, by your lights, the rest of the community is at a suboptimally low spending/crunching rate).
*Acknowledgements: thanks to Daniel Kokotajlo for the idea, comments and suggestions. Thanks to Tom Barnes for comments. All remaining errors are my own.*
Appendix
========
Statistics about the mixture of A and B
---------------------------------------
*Potentially useful for calculating your own*pA.
| | | | |
| --- | --- | --- | --- |
| pA | **25th percentile** | **Median** | **75th percentile** |
| 1 (A) | 2025 | 2027 | 2030 |
| 0.75 | 2025 | 2028 | 2034 |
| 0.5 | 2026 | 2030 | 2041 |
| 0.25 | 2027 | 2033 | 2051 |
| 0 (B) | 2030 | 2039 | 2060 |
Toy model
---------
This model is highly flawed but potentially illustrative. Please take it with a massive grain of salt!
Suppose someone deliberates for  years and updates from  to  . Had they not done the deliberation, they would have done  years - optimal according to  - leading to  units of output. After the update, each year they generate  units of output per year. Suppose this last until AGI arrives, which in expectation is .[[9]](#fnrcoga150ru) The person does not regret having done this if
.
Taking  and  as above,  and  gives the following:[[10]](#fnyxppw1jrtr9)
Note The white areas show T>2. I hide these because I don’t condition the expected time remaining on T years having passed.The asymmetry in the plot due to the fact that in the case one in fact has longer than they’d guessed, their impact multiplier applies for more years (this could make sense if there are things you can do down that grow over time, even if in 2042 both the you-who-did-timelines-work-in-2022 and the you who didn’t have similar AGI timelines).
To use these results, one must have a prior over how their credence in A will change over the duration T (in [expectation it must not change at all though](https://www.lesswrong.com/tag/conservation-of-expected-evidence)!). Someone with low credal resilience will have a 'wider' distribution than someone with higher credal resilience. One could simplify this step further by supposing discrete credences. For example, one could have 1/4 probability of staying at their current credence in A of 0.6, 1/4 credence in moving to 0.2 and 1/2 credence in moving to 0.8.
The toy model could be greatly improved. For example, one could model the situation as spending non-time resources, in which the above limitation does not occur. Further, one could allow for parallel work on timelines.
1. **[^](#fnref67zyqyu77fw)** For example, moving from 40-60% credence in short to medium, or vice versa, has very little gains to the optimal spending schedule.
2. **[^](#fnrefslj8hqva7ad)** For example, A is ‘scale is all you need to AGI’ and B is ‘we need more difficult insights’. Or ‘A’ is your independent impression, which you could become more confident in, and ‘B’ is deference towards others (having accounted for information cascades etc)
3. **[^](#fnrefssxf53peoz8)** 25% probability of AGI going well if it arrived this year, and slope parameter l=0.15
4. **[^](#fnrefcwl364qwxbo)** 10% probability of AGI going well if it arrived this year, and slope parameter l=0.10
5. **[^](#fnrefm50jfe7af7)** 4% probability of AGI going well if it arrived this year, and and slope parameter l=0.05
6. **[^](#fnref53c4d5oy76)** 25% probability of AGI going well if it arrived after a year and the community had been crunching at rate 1 unit of capacity per year (we start with one unit of capacity) and the slope parameter κ=5
7. **[^](#fnref0t8ah2gxe5if)** 10% probability of AGI going well if it arrived after a year and the community had been crunching at rate 1 unit of capacity per year (we start with one unit of capacity) and the slope parameter κ=2
8. **[^](#fnrefhurtu0mihdd)** 5% probability of AGI going well if it arrived after a year and the community had been crunching at rate 1 unit of capacity per year (we start with one unit of capacity) and the slope parameter κ=0.5
9. **[^](#fnrefrcoga150ru)** Assuming T is sufficiently small such that conditioning on no AGI in that time makes little difference.
10. **[^](#fnrefyxppw1jrtr9)** This roughly approximates the results above |
8431de72-fbff-4054-8d00-dc56bfb16bec | trentmkelly/LessWrong-43k | LessWrong | [Summary] "Introduction to Electrodynamics" by David Griffiths - Part 1
Epistemic Status: The mathematical formulation of electromagnetism is the most solidly established theory in all of science. If this book is wrong then science is broken to the core, space is flat, Colombus' landing in the New World was faked, birds aren't real and the Illuminati controls the lizard people.
----------------------------------------
There are three fundamental forces. The color force holds atomic nuclei together. (The strong nuclear force is a special case of the color force.) Gravity holds planets and larger structures together. The electroweak force does everything in-between.
The electroweak force unifies the weak nuclear force with the electromagnetic force. The weak nuclear force doesn't matter.
When I ask physics majors "What is your favorite textbook?" the answer is usually Introduction to Electrodynamics by David Griffiths. Electrodynamics (in the framework of quantum mechanics[1]) explains basically everything on Earth except meteors and nuclear decay. Electrodynamics is important, satisfying and elegant. It is the basis of almost everything of consequence to human beings.
Griffiths' book is equivalent to at least two months of full-time intermediate-level undergraduate physics at a university.
Vector Calculus
Like every great physics book, Introduction to Electrodynamics frontloads its math. The math is all vector calculus. If you know vectors and you know calculus then the math should be within your grasp. If you can read and understand the following definition of →∇ then you're probably fine.
→∇≡^x∂∂x+^y∂∂y+^z∂∂z
The most domain-specific math is the use of divergence and curl.
Divergence
Imagine a vector field as a bunch of arrows. Define a finite region. Divergence is the number of arrows going out of the region minus the arrows going into it.
divergence=→∇⋅→v
Curl
Curl is the counterpart of divergence. Imagine again a vector field as a bunch of arrows. Nonzero curl means the arrows go in a circle.
curl=→∇×→v
Electo |
13ff56f3-42a3-4e21-a099-1564ebb1b057 | trentmkelly/LessWrong-43k | LessWrong | November-December 2024 Progress in Guaranteed Safe AI
Sorry for the radio silence last month. It was slow and I didn’t come across things I wanted to write about, to be expected with holidays coming up.
There are no benefits of paying, except you get a cut of my hard earned shapley points, and apparently some disappointment when I miss a month.
If you're just joining us, background on GSAI here.
There should be some DeepSeek/r1 coverage in the next newsletter (I think tanking the cost of finetuning is, while dual use, good news for us in GSAI). I’d rather get this one out the door though.
FrontierMath (hiring)
As I’m always banging on about, progress in math automation is a leading indicator for progress in software engineering automation by the curry-howard correspondence. That’s a little cheeky, but last time I said it in meatspace I got solemn nods as if it wasn’t even cheeky at all. Or maybe they were solemnly nodding at something else. To be explicit, mathematical reasoning abilities and software engineering skills (including formal proof engineering) appear to be at similar capability levels. Furthermore, through platforms like Lean, mathematical ability can enhance software development capabilities. This connection should make it easier to implement rigorous quality assurance processes, such as obtaining formal proof certificates, which are typically costly.
> We introduce FrontierMath, a benchmark of hundreds of original, exceptionally challenging mathematics problems crafted and vetted by expert mathematicians. The questions cover most major branches of modern mathematics -- from computationally intensive problems in number theory and real analysis to abstract questions in algebraic geometry and category theory. Solving a typical problem requires multiple hours of effort from a researcher in the relevant branch of mathematics, and for the upper end questions, multiple days. FrontierMath uses new, unpublished problems and automated verification to reliably evaluate models while minimizing risk of data con |
5d758a73-9726-4526-be2d-f8085c1878a9 | StampyAI/alignment-research-dataset/blogs | Blogs | Concrete AI tasks bleg
*By Katja Grace, 30 March 2016*
We’re making a [survey](http://aiimpacts.org/ai-timeline-surveys/). I hope to write soon about our general methods and plans, so anyone kind enough to criticize them has the chance. Before that though, we have a different request: we want a list of concrete tasks that AI can’t do yet, but may achieve sometime between now and surpassing humans at everything. For instance, ‘beat a top human Go player in a five game match’ would have been a good example until recently. We are going to ask AI researchers to predict a subset of these tasks, to better chart the murky path ahead.
We hope to:
1. **Include tasks from across the range of AI subfields**
2. **Include tasks from across the range of time** (i.e. some things we can nearly do, some things that are really hard)
3. **Have the tasks relate relatively closely to narrowish AI projects**, to make them easier to think about (e.g. winning a 5k bipedal race is fairly close to existing projects, whereas winning an interpretive dance-off would require a broader mixture of skills, so is less good for our purposes)
4. **Have the tasks relate to specific hard technical problems** (e.g. one-shot learning or hierarchical planning)
5. **Have the tasks relate to large changes in the world** (e.g. replacing all drivers would viscerally change things)
Here are some that we have:
* Win a 5km race over rough terrain against the best human 5k runner.
* Physically assemble any LEGO set given the pieces and instructions.
* Be capable of winning an International Mathematics Olympiad Gold Medal (ignoring entry requirements). That is, solve mathematics problems with known solutions that are hard for the best high school students in the world, better than those students can solve them.
* Watch a human play any computer game a small number of times (say 5), then perform as well as human novices at the game without training more on the game. (The system can train on other games).
* Beat the best human players at Starcraft, with a human-like limit on moves per second.
* Translate a new language using unlimited films with subtitles in the new language, but the kind of training data we have now for other languages (e.g. same text in two languages for many languages and films with subtitles in many languages).
* Be about as good as unskilled human translation for most popular languages (including difficult languages like Czech, Chinese and Arabic).
* Answer tech support questions as well as humans can.
* Train to do image classification on half a dataset (say, ImageNet) then take the other half of the images, containing previously unseen objects, and separate them into the correct groupings (without the correct labels of course).
* See a small number of examples of a new object (say 10), then be able to recognize it in novel scenes as well as humans can.
* Reconstruct a 3d scene from a 2d image as reliably as a human can.
* Transcribe human speech with a variety of accents in a quiet environment as well as humans can.
* Routinely and autonomously prove mathematical theorems that are publishable in mathematics journals today.
Can you think of any interesting ones? |
22bf59df-500b-4a7d-a4f0-7c93f1c18180 | trentmkelly/LessWrong-43k | LessWrong | Readability is mostly a waste of characters
Based on this research, I have created a free web app for compressing GPT prompts at gptrim.com. gptrim is also a Python library. Contributions are welcome.
Readability and information
The other day I was trying to feed a large text to ChatGPT, and I grew frustrated with the limits of the chat message size. After some brainstorming with Bing, I had an insight.
Not all tokens in a text convey information. Some are there to make it easier for humans to parse text. Consider spaces: they are mostly used to tokenize words for the human visual cortex. ButifIwrotewithoutspacesyouwouldprobablystillunderstandme.
There are very few cases where a missing space would cause ambiguity. The nicest one ChatGPT could come up with was:
> "Iamnowhere" could be interpreted as "I am nowhere" or "I am now here," which have opposite meanings.
Even in such cases, the ambiguity would probably be solved by context.
What about punctuation? Again, I suspect there are very few cases where removing punctuation would create an ambiguity that cannot be solved by context. ChatGPT:
> "Let's eat grandma" vs. "Let's eat, grandma." Without the comma, it could be misinterpreted as a request to eat the grandmother.
Can we go further? We could apply word stemming. This would significantly reduce our character count. How much information would we sacrifice? Try to read this snippet (full text below).
> a machin learn research write me in respons to yesterday ’ s post , say : i still think gpt - 2 is a brute - forc statist pattern matcher which blend up the internet and give you back a slight unappet slurri of it when ask . i resist the urg to answer “ yeah , well , your mom is a brute - forc statist pattern matcher which blend up the internet and give you back a slight unappet slurri of it when ask .” but i think it would have been true. (source)
A machine learning researcher writes me in response to yesterday’s post, saying:
> I still think GPT-2 is a brute-force statistical pattern match |
9a0c0853-618c-4a6d-9e22-3437d72efb2f | trentmkelly/LessWrong-43k | LessWrong | Distilled - AGI Safety from First Principles
This is meant to be supplemental to Richard Ngo’s AGI Safety from First Principles sequence. It can serve as an overview/guide to the sequence, and an outline for navigating his core claims. It's also a submission for the AIMS Distillation contest.
Preamble: (you can skip this if you’re just looking to get into the arguments right away)
This post can be read independently of the sequence as well. Just be aware that some details may be left out in this case, and you should refer to Ngo’s sequence to get a more-detailed idea of the arguments if anything seems dissatisfying here.
Most of these ideas are Ngo’s, not mine. I provide some examples and different ways of viewing things, but the core arguments don’t necessarily represent my views (and they might not represent Ngo’s either, though I tried my best to give a fair summary!)
Intended Audience: Those new to the case for why the development of artificial general intelligence (AGI) might pose an existential threat, or those looking to have the case laid out more clearly “from first principles” (without taking any previous claims for granted).
It’s helpful if you have a vague idea of modern machine learning (what’s a neural net? What is reinforcement learning?), but you can probably get by just from considering some of the examples.
What are Ngo’s Core Claims?
Ngo’s argument stands on 4 key premises, stated below in his words:
1. We’ll build AIs which are much more intelligent than humans (i.e. superintelligent).
2. Those AIs will be autonomous agents which pursue large-scale goals.
3. Those goals will be misaligned with ours; that is, they will aim towards outcomes that aren’t desirable by our standards, and trade off against our goals.
4. The development of such AIs would lead to them gaining control of humanity’s future.
Essentially, if all 4 premises hold, the conclusion is, “the development of superintelligent AGI will occur and poses an existential threat to humanity” (and provided you’re conc |
a03cf2a3-b9a6-4e9f-86be-bc45a3bbe00a | trentmkelly/LessWrong-43k | LessWrong | Writing Style and the Typical Mind Fallacy
For a long time, Eliezer has been telling me I should write more like he does. I've mostly resisted, preferring instead to write like this:
1. Explain the lesson of the post immediately, and outline the ideas clearly with lots of headings, subheadings, lists, etc.
2. State the abstract points first, then give concrete examples.
3. Provide lots of links and references to related work so that readers have the opportunity to read more detail about what I'm trying to say (in case it wasn't clear in a single sentence or paragraph), or read the same thing from a different angle (in case the metaphors and language I used weren't clear to that reader).
Eliezer talks as though his style is simply better writing, while I've complained that I often can't even tell what his posts are saying.
I'm a bit embarrassed to admit that it wasn't until sometime last month that I realized that, obviously, different people prefer each style, and Eliezer and I were both falling prey to the typical mind fallacy.
At the recent Effective Altruism Summit I tried to figure out which personal features predicted writing style preference.
One hypothesis was that people who read lots of fiction (like Eliezer) will tend to prefer Eliezer's story-like style, while those who read almost exclusively non-fiction (like me) will tend to prefer my "just gimme the facts" style. This hypothesis didn't hold up well on my non-scientific survey of ~10 LW-reading effective altruists.
Another hypothesis was that most people would prefer Eliezer's more exciting posts, while people trained in the sciences or analytic philosophy (which insist on clear organization, definitions, references to related work, etc.) would prefer my posts. This hypothesis fared a bit better, but not by much.
Another hypothesis was that people who had acquired an epiphany addiction would prefer Eliezer's style, whereas those who just want to learn everything efficiently would prefer my style. But I didn't test this.
Another |
7c19b92f-ba4a-440d-b6c5-2ef43b647820 | trentmkelly/LessWrong-43k | LessWrong | Covariance in your sample vs covariance in the general population
A popular-media take on a subtle problem in sampling. I found the graph quite illustrative.
http://www.theatlantic.com/business/archive/2012/05/when-correlation-is-not-causation-but-something-much-more-screwy/256918/ |
18faa15f-12ce-4618-ab65-197bcb7924db | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Embodied manifestos of human-AI partnerships (Maria Luce Lupetti) - 1st AiTech Symposium
okay so in this project I'm
collaborating with an eccentric avi and
every dubbing and we are working on this
concept of embodied manifests of human
AI partnership which consists on the
development of a method for exploring
narratives and changing design
approaches of area so the first question
you may ask yourself is why focusing on
a design method for exploring narratives
when the overall theme of this
initiative is how can we achieve and
maintain meaningful human control over
AI well if we think about it when we
address this question we implicitly
introduce some of the most recurring
elements of stories such as characters
so we have the potential hero which is
us as designers and engineers we have
implicit dispatchers which are the
institutions who are calling for
initiatives for developing responsible
AI and then we have a dragon which is
our very AI it is not good or bad per se
but it can threaten our princess which
is humanity society the collectivity so
we need to keep it under control somehow
and then we have the most important
storyline functions which is a villainy
the potential lack of control potential
misuse and abuse of AI so we are called
on a mission to achieve and maintain
meaningful human control yep so again
when we address this question somehow we
are implicitly claiming for a change in
the narratives we address for designing
AI but alternative to war exactly so as
suggested a bit as introduced by elisa
when we design especially AI we rely on
theorists and we may not be aware of
them so we make a policies with set
requirements and then we develop test
and we feedback with our results what we
are not aware of is that there are
narratives who are affecting the way we
understand technology and the way we
think this should be designed which is
not a problem per se
but the thing is that let's take an
example the museum robot example so we
can identify two dominant narrative
surrounding robotics one which positive
which sees in robotics an opportunity
for efficiency better lifestyle well on
the other hand we have a more negative
view which see robots as a threat for
humans especially in the logic of
replacement so if we are more closer to
one or the other narrative we may end up
designing something very different on
the first case we may end up designing a
robot as a museum guide while on the
other and we can end up designing a tool
for museum guides which are
substantially different again this is
not a problem these are just well
tentative approaches the problems come
when we move into the evaluation and we
take out insights and we generate our
body of knowledge based on the
evaluations that are also affected by by
narratives so in the first case we may
end up focusing on efficiency and
naturalness and on how our robot is good
in replacing the human museum guide
while under the second case we may end
up focusing on usability aspects and on
how the robot is good in supporting the
museum guide so these two narratives are
the results that we get I've got a
feedback on our narratives reinforcing
them and never encountering each other
so usually who is in the first loop is
not having a conversation with the
second one in most of cases of course so
the aim of my project is explicitly
addressing exploring these narratives
through the concept of embodied
manifestos with the final aim of
changing our approach to a to the design
of the IDE and embodied manifestos are
to be intended as artifacts which are
designed with the specific intent of
translating and accept me it's
simplifying sorry abstract concepts
related to related to narratives about
AI and you may wonder again why
artifacts exactly
well I'm coming from a faculty of
industrial design I have a background in
design and in my knowledge artifact are
largely already used to critically
address concepts and issues related to
technology but in research so we can see
how artifacts I use for reflecting on
the very design activity for
establishing critical areas of concern
for advocating for research agendas and
change in approaches and for translating
and communicating abstract concepts
which is everything that I'm willing to
do and again also in design practice we
see how artifacts are being more and
more used to to explore some gray areas
between dominant narratives such as in
these projects from done already
robots are developed and designed as
things that are smart they are
intelligent they can perform very
intelligent things but then through the
concept of neediness they establish an
interdependent relationship with human
so through artifacts these designers are
remedying the way we can establish
relationship with robots so as I said in
design is already a common practice to
use artifacts as a way to broaden our
current narratives and to suggest
counter narratives but in most of cases
these but these practices these
activities may be seen as more artistic
interventions rather than rigorous rich
research also because the evaluation
phase is as we intend when we design
technological solutions is often left
out and the kind of theories we generate
are very often implicit so my plan is to
focus exactly on this last part so
trying to systematic systematically
prototype not one single alternative
narrative but a spectrum of narrative
alternatives and to compare them through
a systematic observation and
comparison in the evaluation phase in
fact uh artifacts enable me to create
experience and observe these conceptual
implication into practice so that the
outcomes and the insights that I can
generate can inform an actual theory
about designing AI so these are just to
give an idea of the possible next steps
and the case studies that I want to
address I still need to decide so one
example might be focusing on smart on
assistants and the narratives
surrounding privacy while another can be
about service robots and the narratives
regarding replacement and another one
autonomous driving and narratives about
moral responsibility but I'm still
exploring the possibility so if any of
you have cases to suggest you're very
welcome to come talk to me and to
conclude so my aim is to use embodied
manifests as generative research tools
which will enable to escape the trap of
dominant narratives will enable concrete
experiences both for us as designers but
also for the potential audience and
users and discussion about abstract
concepts and finally for probing
principles values and features that are
worth addressing when designing AI and
overall these embodied manifestos are
way for envision and exploring how we
can write our dragon thank you very much
thank you |
22c08377-c6d5-4174-af97-0866d1bf0cba | trentmkelly/LessWrong-43k | LessWrong | Under what circumstances have governments cancelled AI-type systems?
> "This study identifies 61 occasions across Australia, Canada, Europe, New Zealand and the United States when ADS [automated decision systems] projects were cancelled or paused."
The study gives a run-down of what sorts of systems were cancelled (or banned), how the cancellation came about (e.g. via external legal action or some internal decision), and the role of "civil society critique" in the decision-making.
> "In combination, these findings suggest there are competing understandings of acceptable data practices. Given the civil society mobilisation, legal challenges and the number of interventions by oversight bodies, our research suggests that there are competing understandings about effectiveness, impact, accountability and how data systems can infringe people’s rights across areas of application. A key issue is that it takes a lot of work and resources to challenge a data system once in place." |
fd5bd53d-d1b5-48d0-b6ea-17153034cc37 | trentmkelly/LessWrong-43k | LessWrong | Hands-On Experience Is Not Magic
Here are some views, oftentimes held in a cluster:
* You can't make strong predictions about what superintelligent AGIs will be like. We've never seen anything like this before. We can't know that they'll FOOM, that they'll have alien values, that they'll kill everyone. You can speculate, but making strong predictions about them? That can't be invalid.
* You can't figure out how to align an AGI without having an AGI on-hand. Iterative design is the only approach to design that works in practice. Aligning AGI right on the first try isn't simply hard, it's impossible, so racing to build an AGI to experiment with is the correct approach for aligning it.
* An AGI cannot invent nanotechnology/brain-hacking/robotics/[insert speculative technology] just from the data already available to humanity, then use its newfound understanding to build nanofactories/take over the world/whatever on the first try. It'll have to engage in extensive, iterative experimentation first, and there'll be many opportunities to notice what it's doing and stop it.
* More broadly, you can't genuinely generalize out of distribution. The sharp left turn is a fantasy — you can't improve without the policy gradient, and unless there's someone holding your hand and teaching you, you can only figure it out by trial-and-error. Thus, there wouldn't be genuine sharp AGI discontinuities.
* There's something special about training by SGD, and the "inscrutable" algorithms produced this way. They're a specific kind of "connectivist" algorithms made up of an inchoate mess of specialized heuristics. This is why interpretability is difficult — it involves translating these special algorithms into a more high-level form — and indeed, it's why AIs may be inherently uninterpretable!
You can probably see the common theme here. It holds that learning by practical experience (henceforth LPE) is the only process by which a certain kind of cognitive algorithms can be generated. LPE is the only way to become profic |
59c4d0f7-49d5-411b-942b-ac6dc91ba28c | trentmkelly/LessWrong-43k | LessWrong | Rationalist Diplomacy, Game 2, Game Over
Note: The title refers to the upcoming turn.
OK, here's the promised second game of diplomacy. The game name is 'Rationalist Diplomacy Game 2.'
Kevin was Prime Minister of Great Britain
AlexMennen is President of France
tenshiko is Kaiser of Germany
Alexandros was King of Italy until his retirement
WrongBot is Emperor of Austria
Thausler is Czar of Russia
Hugh Ristik is Sultan of Turkey
Randaly is the GM, and can be reached at nojustnoperson@gmail.com
Peace For Our Time!
The leaders of the three surviving nations, France, Russia, and Turkey, agreed to a peace treaty in late August, bringing an end to this destructive conflict. Crowds across Europe broke out into spontaneous celebration, as national leaders began to account for the vast costs- human and monetary- of the wars.
> All orders should be sent to nojustnoperson@gmail.com with an easy-to-read title like "Rationalist Diplomacy Game 2: Russian Orders Spring 1901". Only the LAST set of orders sent will be counted, so feel free to change your mind or to do something sneaky like sending in a fake set of orders cc your ally, and then sending in your real orders later. I'm not going to be too picky on exactly how you phrase your orders, but I prefer standard Diplomacy terminology like "F kie -> hel". New players - remember that if you send two units to the same space, you MUST specify which is attacking and which is supporting. If you make a mistake there or anywhere else, I will probably email you and ask you which you meant, but if I don't have time you'll just be out of luck.
>
> ETA: HughRistik would like to underscore that, under the standard house rules, all draws are unranked.
Past maps can be viewed here; the game history can be viewed here. |
aff0f9b3-0ba3-4c35-b5e1-d31d27a68f50 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Addendum: More Efficient FFNs via Attention
*[Epistemic status: I have* [*running code*](https://github.com/RobertHuben/ffn_via_attention/blob/main/ffn_via_attention_single_layer.py) *that implements it.]*
**Overview:** [I previously showed how an FFN layer in a transformer can be implemented via 3 attention layers.](https://www.lesswrong.com/posts/K7AyY8LMrcKhwfbyj/no-really-attention-is-all-you-need-attention-can-do) In this post I show how to do it in a single attention layer. This reduces the needed dimensionality of your model from 5D+N+1 to D+N+1. The main bottleneck, needing 4D attention heads for the hidden layers, remains.
**Hot Take:** *[epistemic status: much less confidence than the rest of this post]* The bottleneck - that one needs 4D attention heads for the hidden layers - could be capturing a mechanistic interpretability insight: **the FFN components of transformers are less interpretable simply because they consist of ~500x more attention heads than a traditional attention layer**. This could suggest a “scale is all you need” approach to mechanistic interpretability - we’ll be able to understand large attention-only models if and only if we can understand smaller FFN+attention models.
**Outline:** I’ll cover two perspectives that helped me realize you could do this simplification, then summarize the changes, link to the code, then give some concluding thoughts.
I will assume you are familiar with the previous post and it’s notation, so if read it [here](https://aizi.substack.com/p/no-really-attention-is-all-you-need) if you need a refresher.
**Perspective 1 - Identical Steps**
-----------------------------------
I first realized we could simplify this by imagining the perspective of a single entry in the hidden layer of a transformer’s FFN. We:
* (F1) Compute the entry via linear operation on the residual stream
* (F2) Apply a nonlinearity (the activation function)
* (F3) Write to the residual stream via a linear operation
Compare with the steps in an attention head:
* (A1) Compute pre-attention via a linear operation on the residual stream
* (A2) Apply a nonlinearity (the rowwise softmax)
* (A3) Write to the residual stream via a linear operation
Suspiciously similar! In my previous post, I used separate attention layers for F1, F2, and F3, but one can actually choose Q and V matrices so that A1/2/3 computes F1/2/3, respectively, allowing you to complete the FFN in a single attention layer.
**Perspective 2 - Virtual Attention Heads**
-------------------------------------------
[A Mathematical Framework for Transformer Circuits](https://transformer-circuits.pub/2021/framework/index.html) introduced “virtual attention heads”, which provide another useful intuition.
In short, attention heads in two consecutive layers can (in some sense) be treated as a single combined “virtual” attention head. Writing Ai.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')}
for the attention patters and Vi for the weights being written to the residual stream, attention heads are characterized by Ai⊗Vi, and the virtual attention head produced by A1⊗V1 and A2⊗V2 is (A1A2)⊗(V1V2), with the caveat that the attention pattern from layer 1 influences the attention pattern in layer 2.
Since this part is just to build intuition, we’re going to play fast and loose with notation and matrix sizes. But applying this analysis to the linear, SiLU, and linear sublayers described in the previous post, we get:
* A1⊗V1: For the first linear layer, we forced every vector to only attend to itself, so we have A1=I (the identity matrix). Our V1 matrix contained a copy of the weight matrix W1, so in an abuse of notation let us write V1=W1.
* A2⊗V2: For the SiLU layer, our A2 matrix is complicated. But our V2 matrix is a matrix unit −Ek=−Ek,k, where k is the dimension being SiLU’d.
* A3⊗V3**:** For the second linear layer, we forced every vector to only attend to itself, so we have A3=I (the identity matrix). Our V3 matrix contained a copy of the weight matrix W2, so in an abuse of notation let us write V3=W2.
Now, thinking in terms of virtual attention heads, we have (A1⊗V1)(A2⊗V2)(A3⊗V3)=(A1A2A3)⊗(V1V2V3). Since A1=A3=I, this simplifies to A2⊗−W1EkW2.
When one does this analysis rigorously, there are three nuances we must add:
1. Since W1EkW2 is size D-by-D, it must be padded out with 0s to make it side D’-by-D’ (here, D′=D+N+1). That is V=pad(W1EkW2), where pad(M) means “put matrix M in the upper left corner of a new matrix and add 0s to make it the right-sized square matrix”.
2. Previously we computed SiLU(x)=xσ(x) as x−(1−σ(x))x, resulting in negative signs in A2 and in V2=−Ek. However, in this approach we compute xσ(x) directly, so those negative signs go away.
3. The A2 matrix computes attention patterns from the residual stream *after it was modified by*W1*,* so the previous -1 entries are replaced with the kth column of the W1 matrix. (No such accounting has to happen for the A3 matrix, since we force A3 to be the identity matrix no matter what.)
**Summary**
-----------
The resulting Q matrix for computing attention looks like this:
The Q matrix to calculate the output of the FFN’s kth hidden dimension.And as mentioned before, V=pad(W1EkW2), where W1 and W2 are the weight matrices for your FFN as before, and Ek is the 4D-by-4D matrix with a 1 in the (k,k)th spot and a 0 elsewhere.
You use one such attention head for each of the 4D hidden dimensions. For [GPT-3](https://arxiv.org/pdf/2005.14165.pdf), that is a crushing 49152 attention heads in the FFN layer, compared to 96 attention heads in a normal attention layer. This a major slowdown compared to computing an FFN normally, although these attention heads could be parallelized.
Since we compute the hidden layers within the attention heads, we no longer need 4D extra dimensions in our model to store those values between steps. Now the model dimension is D+N+1 (the N+1 channels being used for 1-hot positional encoding). For [GPT-3](https://arxiv.org/abs/2005.14165), that raises the dimensionality from 12288 to 14337, a 17% increase.
The residual stream is augmented from size N-by-D to size (N+1)-by-(D+N+1).**Demonstration Code**
----------------------
I’ve put Python code implementing this technique [on github](https://github.com/RobertHuben/ffn_via_attention/blob/main/ffn_via_attention_single_layer.py). Each of the now two components (FFNs, normal attention) are implemented both directly and with attention heads. They are tested on random matrices with N=20 and D=30, and the largest error entries in each matrix are on the order of 10−13. I have not tested how such errors propagate through multiple layers.
**Conclusion**
--------------
(To be read as a supplement to the conclusions in the [previous post](https://aizi.substack.com/p/no-really-attention-is-all-you-need), which still stand.)
* *[Epistemic status: high confidence]* It is now somewhat more feasible to use this technique to augment transparency tools on transformers.
+ Compared to the previous technique, we have reduced the number of attention layers by a factor of 2x.
+ Compared to the previous technique, we have reduced the dimensionality of the residual stream by a factor of ~5x.
* *[Epistemic status: high confidence]* On [GPT-3’s size hyperparameters](https://arxiv.org/pdf/2005.14165.pdf), this technique would produce ~50k attention heads *per FFN sublayer*, more than 500x the number of attention heads in GPT-3’s classic attention layer!
+ *[Epistemic status: high confidence]* This gives a different perspective on the balance between attention heads and FFNs in LLMs.
- Let us call traditional attention heads **“external attention heads”** (they pass information between word vectors), and the attention heads implementing FFNs **“internal attention heads”** (they pass information inside a word vector).
- We’ve seen that [external attention heads use 33% GPT-3’s parameters, and internal attention heads using 66% of the parameters.](https://aizi.substack.com/p/how-does-gpt-3-spend-its-175b-parametershttps://aizi.substack.com/p/how-does-gpt-3-spend-its-175b-parameters)
- **But external attention heads are only 0.2% of all attention heads - the remaining 99.8% are internal attention heads.**
+ This is a useful reminder that each hidden dimension in an FFN operates independently of the others, just as a group of attention heads work independently in parallel with each other.
+ *[epistemic status: much less confidence than the rest of this post]* It is possible that [previous work](https://transformer-circuits.pub/2021/framework/index.html) “had much less success in understanding MLP layers so far” precisely because of this difference in scale - to study internal attention heads increases the number of attention heads by almost 3 orders of magnitude.
- We could test this hypothesis if we had interpretability tools that work across multiple orders of magnitude. For example, if we had interpretability tools that work on 10000-headed attention-only transformers, we could apply them to 20-headed attention+FFN transformers, and expect roughly similar success in interpretability.
- *[Epistemic status: joke]* I just had a great, extremely original idea for a slogan for such an interpretability paradigm: “scale is all you need”. |
e2b9a153-63a3-4939-8af9-688d21a3566c | trentmkelly/LessWrong-43k | LessWrong | What if Ethics is Provably Self-Contradictory?
I've been thinking lately about the Repugnant Conclusion. For those who are not already aware, it's a problem in Population Ethics where one is seemingly forced to say that a world entirely populated with happy, well-off people is less preferable (all else being equal) than a world consisting of a comparatively larger number of folk who experience a lower quality-of-life.
This doesn't sound so bad at first (many philosophers would presumably be fine with reducing their quality of life on the condition that more babies with mild depression or something be bought into existence), until you realize that this can be applied iteratively. At some point, the larger (but less-well-off-per-individual) world is incredibly populous, but consists only of people who all have lives barely worth living. This world is "objectively" better than our first world, according to many formal ethical frameworks.[1]
Okay, but that isn't so bad, is it? After all, "lives barely worth living" are still worth living! It's not like we're talking about a world full of suicidal people...right? Well, enter the so-called Very Repugnant Conclusion:
> For any perfectly equal population with very high positive welfare, and for any number of lives with very negative welfare, there is a population consisting of the lives with negative welfare and lives with very low positive welfare which is better than the high welfare population, other things being equal.[2]
In other words, the Very Repugnant Conclusion considers a semi-hellish world. This world is populated by some people suffering so badly that they'd be better off not existing,[3] while the rest of the population has the same quality of life as the people from the end of the Repugnant Conclusion (i.e. only marginally worth living). Assuming a high enough population, this semi-hellish world is somehow better than one containing only extremely happy, well-off people.
The Very Repugnant Conclusion has been shown to be provably true if one accepts |
4848abd8-8ac0-4bf5-85a1-2e0a429c71d5 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "Once, when I was holding forth upon the Way, I remarked upon how most organized belief systems exist to flee from doubt. A listener replied to me that the Jesuits must be immune from this criticism, because they practice organized doubt: their novices, he said, are told to doubt Christianity; doubt the existence of God; doubt if their calling is real; doubt that they are suitable for perpetual vows of chastity and poverty. And I said: Ah, but they’re supposed to overcome these doubts, right? He said: No, they are to doubt that perhaps their doubts may grow and become stronger. Googling failed to confirm or refute these allegations. But I find this scenario fascinating, worthy of discussion, regardless of whether it is true or false of Jesuits. If the Jesuits practiced deliberate doubt, as described above, would they therefore be virtuous as rationalists? I think I have to concede that the Jesuits, in the (possibly hypothetical) scenario above, would not properly be described as “fleeing from doubt.” But the (possibly hypothetical) conduct still strikes me as highly suspicious. To a truly virtuous rationalist, doubt should not be scary. The conduct described above sounds to me like a program of desensitization for something very scary, like exposing an arachnophobe to spiders under carefully controlled conditions. But even so, they are encouraging their novices to doubt—right? Does it matter if their reasons are flawed? Is this not still a worthy deed unto a rationalist? All curiosity seeks to annihilate itself; there is no curiosity that does not want an answer. But if you obtain an answer, if you satisfy your curiosity, then the glorious mystery will no longer be mysterious. In the same way, every doubt exists in order to annihilate some particular belief. If a doubt fails to destroy its target, the doubt has died unfulfilled—but that is still a resolution, an ending, albeit a sadder one. A doubt that neither destroys itself nor destroys its target might as well have never existed at all. It is the resolution of doubts, not the mere act of doubting, which drives the ratchet of rationality forward. Every improvement is a change, but not every change is an improvement. Every rationalist doubts, but not all doubts are rational. Wearing doubts doesn’t make you a rationalist any more than wearing a white medical lab coat makes you a doctor. A rational doubt comes into existence for a specific reason—you have some specific justification to suspect the belief is wrong. This reason, in turn, implies an avenue of investigation which will either destroy the targeted belief or destroy the doubt. This holds even for highly abstract doubts, like: “I wonder if there might be a simpler hypothesis which also explains this data.” In this case you investigate by trying to think of simpler hypotheses. As this search continues longer and longer without fruit, you will think it less and less likely that the next increment of computation will be the one to succeed. Eventually the cost of searching will exceed the expected benefit, and you’ll stop searching. At which point you can no longer claim to be usefully doubting. A doubt that is not investigated might as well not exist. Every doubt exists to destroy itself, one way or the other. An unresolved doubt is a null-op; it does not turn the wheel, neither forward nor back. If you really believe a religion (and don’t just believe in it), then why would you tell your novices to consider doubts that must die unfulfilled? It would be like telling physics students to agonize over whether the twentieth-century revolution might have been a mistake, and that Newtonian mechanics was correct all along. If you don’t really doubt something, why would you pretend that you do? Because we all want to be seen as rational—and doubting is widely believed to be a virtue of a rationalist. But it is not widely understood that you need a particular reason to doubt, or that an unresolved doubt is a null-op. Instead people think it’s about modesty, a submissive demeanor, maintaining the tribal status hierarchy—almost exactly the same problem as with humility, on which I have previously written. Making a great public display of doubt to convince yourself that you are a rationalist will do around as much good as wearing a lab coat. To avoid merely professing doubts,1 remember: A rational doubt exists to destroy its target belief, and if it does not destroy its target it dies unfulfilled. A rational doubt arises from some specific reason the belief might be wrong. An unresolved doubt is a null-op. An uninvestigated doubt might as well not exist. You should not be proud of mere doubting, although you can justly be proud when you have just finished tearing a cherished belief to shreds. Though it may take courage to face your doubts, never forget that to an ideal mind doubt would not be scary in the first place. 1See “Professing and Cheering” in Map and Territory."
The text above is a post found on LessWrong.
Can you further elaborate on the concept of "rational doubt" and how it differentiates from "doubt" in general? Additionally, how does this differentiation impact the way we approach belief systems and decision-making processes?
"Rational doubt" refers to a form of skepticism that is driven by logic, reason, and evidence. It is not a form of doubt that is influenced by emotions, bias, or preconceived notions. Instead, it is a doubt that arises when there is a legitimate reason to question the validity or truth of a claim or belief. This reason could be a lack of evidence, contradictory evidence, or the presence of a more plausible alternative explanation.
In contrast, "doubt" in general can be driven by a wide range of factors, including emotions, bias, fear, uncertainty, or simply a gut feeling. This kind of doubt may not necessarily be based on logic or evidence.
The differentiation between "rational doubt" and "doubt" in general significantly impacts the way we approach belief systems and decision-making processes. When we apply rational doubt, we are essentially applying critical thinking. We are not accepting claims or beliefs at face value, but instead questioning them, scrutinizing the evidence, and considering alternative explanations. This can lead to a more nuanced and accurate understanding of the world, and can also help us to avoid falling into the trap of dogmatism or blind faith.
In decision-making processes, rational doubt can help us to make more informed and rational choices. By questioning our assumptions, evaluating the evidence, and considering all possible options, we can make decisions that are more likely to lead to the desired outcome.
However, it's important to note that while rational doubt is a valuable tool, it should not lead to a state of perpetual skepticism where we are unable to make decisions or form beliefs. There comes a point where we have enough evidence to make a reasoned judgement, even if some doubt remains. Rational doubt should therefore be seen as a tool for refining our beliefs and decisions, rather than as an end in itself.
How does the concept of rational doubt influence decision-making under conditions of uncertainty, particularly when evidence is limited or ambiguous?
Rational doubt can be particularly useful in decision-making under conditions of uncertainty, where evidence is limited or ambiguous. In such situations, it can be easy to fall into the trap of making hasty decisions based on incomplete information, gut feelings, or preconceived notions. Rational doubt helps to guard against this by encouraging a more thoughtful and critical approach.
Firstly, rational doubt prompts us to question our assumptions. In uncertain situations, we often have to rely on assumptions to fill in the gaps in our knowledge. However, these assumptions can sometimes be misguided or incorrect. By doubting our assumptions, we can identify potential flaws in our thinking and make more accurate judgements.
Secondly, rational doubt encourages us to seek out additional information. When evidence is limited or ambiguous, it can be tempting to make a decision based on the information we have at hand. However, rational doubt reminds us that the available evidence may not be sufficient to make an informed decision. This can motivate us to gather more data, seek out alternative perspectives, or explore different interpretations of the evidence.
Thirdly, rational doubt promotes flexibility in our decision-making. In uncertain situations, it can be easy to become attached to a particular course of action or belief. However, rational doubt encourages us to remain open to new information and willing to change our minds if the evidence warrants it.
Finally, rational doubt can help us to manage risk. By acknowledging the uncertainty inherent in the decision-making process, we can make decisions that are not only based on the most likely outcome, but also take into account the potential consequences of being wrong.
In summary, while rational doubt can't eliminate uncertainty, it can help us to navigate it more effectively by promoting critical thinking, information seeking, flexibility, and risk management. |
83baa6f4-8a44-4887-a686-945426094a15 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | EU AI Act now has a section on general purpose AI systems
The EU AI Act is currently undergoing the [ordinary legislative procedure](https://www.consilium.europa.eu/en/council-eu/decision-making/ordinary-legislative-procedure/), in which the European Parliament and Council can propose changes to the act.
A brief summary of the act is that systems defined as *high-risk* will be required to undergo conformity assessments, which among other things requires the system to be monitored and have a working off-switch ([longer summary and analysis here](https://forum.effectivealtruism.org/posts/bd7yr3eozzzhMuKCi/what-is-the-eu-ai-act-and-why-should-you-care-about-it)).
[The Council's amendments have recently been circulated](https://data.consilium.europa.eu/doc/document/ST-14278-2021-INIT/en/pdf). Most importantly for longtermists, they include a new section for general purpose AI systems. For the first time ever regulating general AI is on the table, and for an important government as well!
The article reads:
> **Article 52a - General purpose AI systems**
>
> 1. **The placing on the market, putting into service or use of general purpose AI**
> **systems shall not, by themselves only, make those systems subject to the**
> **provisions of this Regulation.**
> 2. **Any person who places on the market or puts into service under its own**
> **name or trademark or uses a general purpose AI system made available on**
> **the market or put into service for an intended purpose that makes it subject**
> **to the provisions of this Regulation shall be considered the provider of the**
> **AI system subject to the provisions of this Regulation.**
> 3. **Paragraph 2 shall apply, mutatis mutandis, to any person who integrates a**
> **general purpose AI system made available on the market, with or without**
> **modifying it, into an AI system whose intended purpose makes it subject to**
> **the provisions of this Regulation.**
> 4. **The provisions of this Article shall apply irrespective of whether the general**
> **purpose AI system is open source software or not.**
>
Or in plain English, General Purpose AI Systems will not be considered *high-risk*, unless they are explicitly intended to be used for a high-risk purpose.
What are your reactions to this development? |
69bf2131-6ee5-4132-8aff-992615e57d2f | StampyAI/alignment-research-dataset/blogs | Blogs | March 2016 Newsletter
| |
| --- |
|
**Research updates**
* A new paper: “[Defining Human Values for Value Learners](https://intelligence.org/2016/02/29/new-paper-defining-human-values-for-value-learners/)“
* New at IAFF: [Analysis of Algorithms and Partial Algorithms](https://agentfoundations.org/item?id=622); [Naturalistic Logical Updates](https://agentfoundations.org/item?id=625); [Notes from a Conversation on Act-Based and Goal-Directed Systems](https://agentfoundations.org/item?id=634); [Toy Model: Convergent Instrumental Goals](https://agentfoundations.org/item?id=649)
* New at AI Impacts: [Global Computing Capacity](http://aiimpacts.org/global-computing-capacity/)
* A revised version of “[The Value Learning Problem](https://intelligence.org/files/ValueLearningProblem.pdf)” (pdf) has been accepted to a AAAI spring symposium.
**General updates**
* MIRI and other Future of Life Institute (FLI) grantees participated in a [AAAI workshop](http://futureoflife.org/2016/02/17/aaai-workshop-highlights-debate-discussion-and-future-research/) on AI safety this month.
* MIRI researcher Eliezer Yudkowsky discusses Ray Kurzweil, the Bayesian brain hypothesis, and an eclectic mix of other topics in [a new interview](https://intelligence.org/2016/03/02/john-horgan-interviews-eliezer-yudkowsky/).
* Alexei Andreev and Yudkowsky are [seeking investors](http://arbital.com/p/arbital_ambitions/) for Arbital, a new technology for explaining difficult topics in economics, mathematics, computer science, and other disciplines. As a demo, Yudkowsky has written a new and improved [guide to Bayes’s Rule](http://arbital.com/p/bayes_rule/).
**News and links**
* [Should We Fear or Welcome the Singularity?](https://www.youtube.com/watch?v=TcX_7SVI_hA) (video): a conversation between Kurzweil, Stuart Russell, Max Tegmark, and Harry Shum.
* [The Code That Runs Our Lives](https://www.youtube.com/watch?v=XG-dwZMc7Ng&t=10m0s) (video): Deep learning pioneer Geoffrey Hinton expresses his concerns about smarter-than-human AI (at 10:00).
* [The State of AI](https://www.youtube.com/watch?v=VBceREwF7SA&t=21m09s) (video): Russell, Ya-Qin Zhang, Matthew Grob, and Andrew Moore share their views on a range of issues at Davos, including superintelligence (at 21:09).
* Bill Gates discusses [AI timelines](http://lukemuehlhauser.com/bill-gates-on-ai-timelines/).
* Paul Christiano proposes a new AI alignment approach: [algorithm learning by bootstrapped approval-maximization](https://medium.com/ai-control/alba-an-explicit-proposal-for-aligned-ai-17a55f60bbcf#.9nd9dokju).
* Robert Wiblin asks the effective altruism community: [If tech progress might be bad, what should we tell people about it?](http://effective-altruism.com/ea/tr/if_tech_progress_might_be_bad_what_should_we_tell/)
* FLI collects [introductory resources on AI safety research](http://futureoflife.org/2016/02/29/introductory-resources-on-ai-safety-research/).
* Raising for Effective Giving, a major fundraiser for MIRI and other EA organizations, is seeking a [Director of Growth](http://reg-charity.org/join-our-team-2/).
* Murray Shanahan answers questions about the new [Leverhulme Centre for the Future of Intelligence](http://www3.imperial.ac.uk/newsandeventspggrp/imperialcollege/newssummary/news_5-1-2016-14-58-4). Leverhulme CFI is presently seeking an [Executive Director](http://www.jobs.cam.ac.uk/job/9344/).
|
The post [March 2016 Newsletter](https://intelligence.org/2016/03/05/march-2016-newsletter/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org). |
b2c087a2-043d-40ae-8a17-d5a377741e71 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Careless talk on US-China AI competition? (and criticism of CAIS coverage)
>
> China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.[[1]](#fn-mF6rrn65fs3bdg6jf-1)
>
>
>
Sounds about right?
*[hotpot.ai/art-generator](http://hotpot.ai/art-generator)*
This post centres around an email I sent to the Center for AI Safety (CAIS) expressing concern about their [2023-08-15 newsletter](https://newsletter.safe.ai/p/ai-safety-newsletter-19)'s coverage of US-China competition in the AI space[[2]](#fn-mF6rrn65fs3bdg6jf-2), but the overall point is broader. There are some ways of discussing the topic of international relations regarding AI which strike me as un-nuanced in a counterproductive and dangerous way, by hiding certain truths or emphasising others, and supporting a conflict-oriented mindset.
In writing about this, I'm also gesturing at something about the more general topic of 'how to think and write about politically-charged topics'.
Jump to the summary if you are in a hurry.
This conversation really is important, which is why I think it's worth a public message discussing particular statements, but this should be understood as constructive criticism and part of a broader conversation which society, and especially the community of those focused on AI safety, needs to have. The CAIS newsletters are worth a (not unquestioning!) read, including [the edition in question](https://newsletter.safe.ai/p/ai-safety-newsletter-19).
The particulars in this message serve as good exemplars of the problems and questions I have, and I'd be interested in responses from CAIS but even more so in remarks more broadly on the topic from anyone interested. The public conversation about this appears from my perspective to sometimes be broken. If that is so, I would like it to be rectified, and if not, I would like to be put right myself, the better to prioritise in my own work!
Specifics, CAIS case study
--------------------------
Here, quoted[[3]](#fn-mF6rrn65fs3bdg6jf-3), is my response to the CAIS letter, which serves as an initial dialogue opener and the core of this post:
>
> Hello,
>
>
> I've been a supportive reader for some time and am myself an AI safety researcher. I'm generally very impressed and encouraged by your newsletters! - but I was disappointed and concerned by the phrasing (and mindset it can encourage) regarding US-China competition in the letter dated 16th August ('US-China Competition on AI Chips, ...').
>
>
> In general I'm very wary of messaging which could inflame a them-vs-us mindset, in short mainly because I think it a) destroys humans' ability to think sensibly and b) tends to foreclose win-win outcomes. I expect these brief points to be clear and to have rich referents in your mental pictures of the world, but please correct me if not!
>
>
>
Interjection: I'm referring to the vicinity of [mind-killing politics](https://www.lesswrong.com/posts/9weLK2AJ9JEt2Tt8f/politics-is-the-mind-killer)
>
> I think your letter skirted close to dangerously simplified presentation in this way. I would not normally spend my time or yours on a criticism of one section of a newsletter, but in this case I consider it worthwhile because your letter is close to the Pareto frontier on nuance, correctness, helpfulness, and reach (and it pays to try to nudge such things in good directions), but this kind of message needs to be delivered with care to avoid misunderstanding and harm.
>
>
> Hopefully my pointing this out, accompanied by a few select quotes, is enough to encourage you to carefully take this criticism into account, but I'd happily expand more if you like!
>
>
> Without further ado, a few quotes and my response:
>
>
>
> >
> > The US and China have been competing for access to these chips for years.
> >
> >
> >
>
>
> Kind of true, but really *US-based and China-based international corporations* (as well as other orgs) have *sought access* to this scarce resource. Competition *for this particular resource* is mostly zero-sum across all of these entities, importantly *including intra-US and intra-China*.
>
>
> Where market-share is near zero-sum (i.e. for direct competitors) the market-share outcomes may be greater-than-linearly zero-sum in this resource (your minor/temporary lack of chips could be my major gain of market-share), which might better warrant the term 'competing', but this effect is actually much stronger *intra-country/bloc* rather than inter-, due to respective markets! i.e. Google and Microsoft really care about *each other's* chip access in a way that they only do to a weaker degree about Alibaba's.
>
>
>
Interjection: To emphasise, 'the US and China have been competing' doesn't literally preclude belief in intra-bloc competition. But there's a strong implicature that intra-bloc is (relatively) unimportant, while in fact the mechanism I mentioned here *increases* intra-bloc competition (which I think is borne out by observation to date).
>
> When governments have paid attention, they have indeed made moves which adjust share (but also supply, as you've noted later, making it nonzero-sum, in chips at least). It's unclear (to me) exactly what incentives have motivated each move, but certainly they're not the actions of monolithic or coherent entities 'The US' and 'China'. And it's certainly not the case that where such activity changes chip share, it's *collected by* the acting entity. Non-governments have also made moves adjusting chip share, for example the case you cite of Nvidia (a 'US' company) deliberately rules-lawyering the US gov in order to supply more chips to various China-based companies!
>
>
> Typically when nations are used as agentic subject nouns it refers to the government and/or military of said country. I don't think there's a reading of these statements in those terms which is true, and I'm not aware of any other plausible reading which is true.
>
>
>
>
> ---
>
>
>
> >
> > China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.
> >
> >
> >
>
>
> I dislike this sentence and think it is false! Who is this 'China'? Did said unified entity carry out all of these activities? Was it coherently pursuing all of these 'several efforts' to some particular end?
>
>
>
Interjection: I feel that I was unkind in my tone here. The kind of claim exemplified in the letter and in other places has proved impossible for me to map to something sincerely resembling reality without caveating so much as to be essentially starting from scratch i.e. it looks like a kind of non-proposition or emotive filler. I would be very interested to hear from people who have a better parse on this to help me understand! My crux discussion below the rest of the email contains my current best attempts.
>
>
> >
> > Meanwhile, the United States has struggled to build American chip manufacturing capacity, and has taken further steps to prevent Americans from investing in Chinese technology.
> >
> >
> >
>
>
> This one is poor for the same reasons, though not quite as bad (perhaps because the authors are American/anglophone and have a closer perspective on the nuance).
>
>
> The discussion after this point in the letter is relatively good and nuanced! It names (some of) the individual orgs and companies, and makes clearer the multiplicity of others. All further references to nations as agentic subject nouns appear to be consistent with a conventional reading referring to the respective governments.
>
>
> I'm interested to know how the rather good detail got paired with a rather harmful introduction and I urge you to consider the processes and thinking which gave rise to this section of the otherwise good letter.
>
>
> Thanks,
>
>
> Oly
>
>
>
This was a brief email intended to convey something I expected to be quickly-graspable with a few pointers. CAIS content suggests that they have a broader familiarity with associated facts, but it seems to be digested/compressed here in a way which is needlessly and harmfully lossy.
In particular, the statements abstract very neatly over pre-drawn boundaries (national i.e. 'US' and 'China') and furthermore assign a greater sense of coherence and agency to those abstractions than is warranted. At least some possible such statements must be true-ish (or we would not have those abstractions), but this convenient compression happens in too many conversations to be a coincidence! Said pre-existing abstraction boundaries are already salient in the information ecosystem, and laden with emotional and political baggage. This same phenomenon (it sometimes seems like a [pre-written bottom line](https://www.lesswrong.com/posts/34XxbRFe54FycoCDw/the-bottom-line) but in implicature?) appears in other publications by other orgs and in verbal conversations I've witnessed or been part of.
The fact that I, a relative governance rookie (I'm focused mainly on technical matters), struggle to rectify or understand this makes me wonder: am I missing something? Is there a relevant factor I'm unaware of? More concerningly, is there some terrible equilibrium which prevents more involved people from speaking more clearly here? I think more likely the abstractions ('US' and 'China') have a background potency which distorts perceptions and shapes how people communicate.
Possible cruxes and areas of high uncertainty
---------------------------------------------
Implicit in my own discussion is the background assumption that the main concern is about possible *direct or near-direct* impacts[[4]](#fn-mF6rrn65fs3bdg6jf-4) of AI deployments by *government* entities (or military). This seems to me the obvious reading when people use countries as agentic subject nouns. ('China [the governmental entity] has made several efforts'[[5]](#fn-mF6rrn65fs3bdg6jf-5).) In this framing, I can't rectify some of the things people say with reality. But some alternative concerns might fit the bill, and if so, this is evidence that people are talking past each other, and we should aim to frame concerns more clearly!
I've never focused intently on this area, but have had a handful of conversations about this over the years, and among relevant cruxes seems to be a family of questions along the lines of
>
> How quickly/totally/coherently *could* US gov/CCP capture AI talent/artefacts/compute within its jurisdiction and redirect them toward [*excludable*](https://en.wikipedia.org/wiki/Excludability) destructive ends? Under what circumstances would they want/be able to do that?
>
>
>
People's intuitions here appear to differ a lot, and data might be hard to come by!
It seems plain that nations are not currently meaningful players in AI development and deployment, absent conspiracy-level secrecy. So to support the apparent take that they *are*, we may need to imagine that they *could ably/rapidly become* meaningful players in AI development and deployment, hence the above cruxes.
Depending on the answers to these questions, one might perceive various goings-on which happen to occur under one or other jurisdiction to have greater import on the international stage and perhaps to warrant treating national or multi-national blocs as more coherent entities than they really are at present, for the purposes of AI discussion. ('China [the government] has [allowed/encouraged ~~made~~] several efforts [because eventually they will probably seize the gains/means]'[[7]](#fn-mF6rrn65fs3bdg6jf-7).)
Other possible cruxes, more guesswork:
1. Perhaps the concern is about *indirect* (e.g. economic) impacts of non-government entities' AI activities leading to some (risky) change in balance of power (between existing governments/blocs)
* Then, abstracting references to lots of individuals and groups via their home country might be a move which is writer-intuitive, even if nonstandard and reader-confusing. ('China [the impersonal collective economic entity] has made several efforts'[[8]](#fn-mF6rrn65fs3bdg6jf-8).)
* The additional step in this type of theory, namely that indirect effects cause a risky change in balance of power, should really be spelled out if it is loadbearing
* The use of countries as agentic subject nouns is difficult to justify under this reading
2. Perhaps the concern is indeed about direct impacts, but wielded by non-government entities (who remain the major players in development and deployment of AI)
* If so, the conversation should be about general/global resource/capabilities rather than inter-bloc 'competition'...
* ...unless we *also* posit that inter-bloc non-government conflict is liable to be *much worse* than intra-bloc[[9]](#fn-mF6rrn65fs3bdg6jf-9)
* Similarly, if these are loadbearing assumptions, they really ought to be spelled out clearly
* The use of countries as agentic subject nouns could be justified at a stretch here, but only by first spelling out the reasoning
3. Perhaps the concern is that, regardless of the actual impact of AI resources, *apparent* competition could lead to inflammation of traditional conflict, or weaken defenses against such inflammation
* Then, reporting on the *apparent* competition [via a *mention*](https://en.wikipedia.org/wiki/Use%E2%80%93mention_distinction) and with explicit caveat, would make sense! ('China [gov/military] has [been *perceived* as having] made several efforts... [but the reality is more nuanced]'[[10]](#fn-mF6rrn65fs3bdg6jf-10).)
* Alternatively, reporting on *actual* conflict could be used as evidence for the claim (that apparent competition inflames conflict), but only by also pointing to the stated or implied *reasons* for the conflict. ('China has made several efforts... [in each case citing US provocation as justification]'[[11]](#fn-mF6rrn65fs3bdg6jf-11).)
* In either case, there are additional claims being made that can not be left implicit, and require supporting argument
### Speculation
As it is, for me, the evidence seems to suggest that an AI race, if it is happening at all, is being run by (mainly US- and UK-based) companies with little or no oversight from governments or militaries. Rather, governments are in a position to collectively act to diffuse the race! And they appear as likely to do this as to exacerbate it, from my limited viewpoint.
Separately, a lack of reliable alignment techniques and performance guarantees makes AI-powered belligerent national interest plays look more like bioweapons than like nukes - i.e. [minimally-excludable](https://en.wikipedia.org/wiki/Excludability) - and perhaps mutually-knowably so! This presently damps the incentive to go after them. But proliferation of naively-aligned AI ('figure out what I want and make it happen') might make harm plays *more excludable*, exacerbating lose-lose or race game dynamics ('go and steal/destroy their stuff but don't let that happen to my stuff'). This concern in part motivates consideration of multi-principal-multi-agent delegation and the cooperative AI agenda.
Summary and takeaways
---------------------
Un-nuanced coverage and discussion has the potential to inflame harmful confusion and us-vs-them mentality, which diminish the chance of safe outcomes.
* >
> China has made several efforts to preserve their chip access, including smuggling, buying chips that are just under the legal limit of performance, and investing in their domestic chip industry.
>
>
>
* >
> China [the governmental entity] has made several efforts...[[5:1]](#fn-mF6rrn65fs3bdg6jf-5)
>
>
>
* >
> China [the government] has [allowed/encouraged ~~made~~] several efforts [because eventually they will probably seize the gains/means]...[[7:1]](#fn-mF6rrn65fs3bdg6jf-7)
>
>
>
* >
> China [the impersonal collective economic entity] has made several efforts...[[8:1]](#fn-mF6rrn65fs3bdg6jf-8)
>
>
>
* >
> [People and organisations in] China [have ~~has~~] made several efforts...[[8:2]](#fn-mF6rrn65fs3bdg6jf-8)
>
>
>
* >
> China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced] [[10:1]](#fn-mF6rrn65fs3bdg6jf-10)
>
>
>
* >
> China has made several efforts... [in each case citing US provocation as justification] [[11:1]](#fn-mF6rrn65fs3bdg6jf-11)
>
>
>
* ...?
When compressing discussion of political topics, be extra wary of **compression which coincidentally abstracts over already-charged us-them divides** (and be careful when phrasing comes too easily, [lest you write bottom lines first](https://www.lesswrong.com/posts/34XxbRFe54FycoCDw/the-bottom-line))! You're **more likely to be wrong** (because your information ecosystem biases toward thinking in these terms, and because you might be [mildly-to-severely mind-killed](https://www.lesswrong.com/posts/9weLK2AJ9JEt2Tt8f/politics-is-the-mind-killer) on the matter), and **being wrong is more likely to be harmful** (by reinforcing those dynamics in others).[[12]](#fn-mF6rrn65fs3bdg6jf-12) The same vigilance applies to reading and listening.
My (not very informed) take is that governments are at this point as likely to want to defuse as to exacerbate an AI race, and those of us with any privileged insight or influence should avoid one-sided discussion of the matter (if anything preferring to focus on constructive, collaborative possibilities, the better to raise them to salience and generate common knowledge).
Most of my remarks here are somewhat weakly held (if forcefully stated) and it seems important to gather perspectives on this. Inform me! All responses will be gratefully received.
---
1. Center for AI Safety, [AI Safety Newsletter #19, 2023-08-15](https://newsletter.safe.ai/p/ai-safety-newsletter-19) [↩︎](#fnref-mF6rrn65fs3bdg6jf-1)
2. I'm supportive of some of CAIS' work, and the content of their newsletters (they have impressive breadth), and theirs is far from the only outfit which appears to produce confused or confusing messaging on the topic of US-China competition. In fact they seem to be better than many! [↩︎](#fnref-mF6rrn65fs3bdg6jf-2)
3. My response is at quote level 1. Excerpts from the CAIS letter are within, at quote level 2. I interject a little for the purposes of this post, without quotation. [↩︎](#fnref-mF6rrn65fs3bdg6jf-3)
4. e.g. deployment for weapons control or for offensive R&D (bio, materials, ...) [↩︎](#fnref-mF6rrn65fs3bdg6jf-4)
5. 'China [the governmental entity] has made several efforts' is a fairly standard use of language; governments are at least [somewhat coherent and also do things with consequences](https://www.lesswrong.com/posts/5GzqD7fgtxjepPERp/deliberation-everywhere-simple-examples#Parliaments) and subsequent plans 'in mind'. This sentence has the disadvantage of being baldly false, though[[6]](#fn-mF6rrn65fs3bdg6jf-6). [↩︎](#fnref-mF6rrn65fs3bdg6jf-5) [↩︎](#fnref-mF6rrn65fs3bdg6jf-5:1)
6. (unless we posit a near-hivemind coherence to the people of China, which is absurd and obscene) [↩︎](#fnref-mF6rrn65fs3bdg6jf-6)
7. 'China [the government] has [allowed/encouraged ~~made~~] several efforts [because eventually they will probably seize the gains/means]' is a big stretch of the language, but at least somewhat consistent. To support this reading, though, there's a substantial additional claim that needs to be justified. [↩︎](#fnref-mF6rrn65fs3bdg6jf-7) [↩︎](#fnref-mF6rrn65fs3bdg6jf-7:1)
8. 'China [the impersonal collective economic entity] has made several efforts' would be a rather nonstandard use of language; economies of billion+ people do not make 'efforts' with consequences or subsequent plans. Leaving aside the implicature of agency, though, this sentence is a closer fit to reality. '[People and organisations in] China [have ~~has~~] made several efforts' would be even better. [↩︎](#fnref-mF6rrn65fs3bdg6jf-8) [↩︎](#fnref-mF6rrn65fs3bdg6jf-8:1) [↩︎](#fnref-mF6rrn65fs3bdg6jf-8:2)
9. Should conflict between non-territorial entities be worse for inter-bloc than intra-bloc? I think the point I made previously about zero-sum market-share competition suggests the opposite. But humans' destructive jingoistic/xenophobic tendencies are real, and a point in favour. [↩︎](#fnref-mF6rrn65fs3bdg6jf-9)
10. 'China [gov/military] has [been perceived as having] made several efforts... [but the reality is more nuanced]' is in large part one of the messages of this post! I don't think the original letter in question can have meant this, but I do maintain it as a hypothesis for the more general case of compressed discussion of political things. People are often [imagining third-party reactions](https://www.lesswrong.com/posts/7cAsBPGh98pGyrhz9/decoupling-vs-contextualising-norms) when discussing political topics, and sometimes [use-mention distinctions](https://en.wikipedia.org/wiki/Use%E2%80%93mention_distinction) fail to come across. [↩︎](#fnref-mF6rrn65fs3bdg6jf-10) [↩︎](#fnref-mF6rrn65fs3bdg6jf-10:1)
11. 'China has made several efforts... [in each case citing US provocation as justification]' is something that could legitimately be said, and has a clear meaning, even if in this particular case it is false if we understand 'China' to be the CCP or military. [↩︎](#fnref-mF6rrn65fs3bdg6jf-11) [↩︎](#fnref-mF6rrn65fs3bdg6jf-11:1)
12. I'd tentatively go further and suggest you ought to train yourself to be *appalled when you catch yourself doing this without justification* because only then do you stand a chance of thinking clearly about politics. [↩︎](#fnref-mF6rrn65fs3bdg6jf-12) |
428057f2-ae8b-460c-b81c-b4cf92e26c63 | StampyAI/alignment-research-dataset/arbital | Arbital | Instrumental goals are almost-equally as tractable as terminal goals
One counterargument to the Orthogonality Thesis asserts that agents with terminal preferences for goals like e.g. resource acquisition will always be much better at those goals than agents which merely try to acquire resources on the way to doing something else, like making paperclips. Therefore, by filtering on real-world competent agents, we filter out all agents which do not have terminal preferences for acquiring resources.
A reply is that "figuring out how to do $W_4$ on the way to $W_3$, on the way to $W_2$, on the way to $W_1$, without that particular way of doing $W_4$ stomping on your ability to later achieve $W_2$" is such a ubiquitous idiom of cognition or supercognition that (a) any competent agent must already do that all the time, and (b) it doesn't seem like adding one more straightforward target $W_0$ to the end of the chain should usually result in greatly increased computational costs or greatly diminished ability to optimize $W_4$.
E.g. contrast the necessary thoughts of a paperclip maximizer acquiring resources in order to turn them into paperclips, and an agent with a terminal goal of acquiring and hoarding resources.
The paperclip maximizer has a terminal utility function $U_0$ which counts the number of paperclips in the universe (or rather, paperclip-seconds in the universe's history). The paperclip maximizer then identifies a sequence of subgoals and sub-subgoals $W_1, W_2, W_3...W_N$ corresponding to increasingly fine-grained strategies for making paperclips, each of which is subject to the constraint that it doesn't stomp on the previous elements of the goal hierarchy. (For simplicity of exposition we temporarily pretend that each goal has only one subgoal rather than a family of conjunctive and disjunctive subgoals.)
More concretely, we can imagine that $W_1$ is "get matter under my control (in a way that doesn't stop me from making paperclips with it)", that is, if we were to consider the naive or unconditional description $W_1'$ "get matter under my 'control' (whether or not I can make paperclips with it)", we are here interested in a subset of states $W_1 \subset W_1'$ such that $\mathbb E[https://arbital.com/p/U_0|W_1](https://arbital.com/p/U_0|W_1)$ is high. Then $W_2$ might be "explore the universe to find matter (in such a way that it doesn't interfere with bringing that matter under control or turning it into paperclips)", $W_3$ might be "build interstellar probes (in such a way that ...)", and as we go further into the hierarchy we will find $W_{10}$ "gather all the materials for an interstellar probe in one place (in such a way that ...)", $W_{20}$ "lay the next 20 sets of rails for transporting the titanium cart", and $W_{25}$ "move the left controller upward".
Of course by the time we're that deep in the hierarchy, any efficient planning algorithm is making some use of real independences where we can reason relatively myopically about how to lay train tracks without worrying very much about what the cart of titanium is being used for. (Provided that the strategies are constrained enough in domain to not include any strategies that stomp distant higher goals, e.g. the strategy "build an independent superintelligence that just wants to lay train tracks"; if the system were optimizing that broadly it would need to check distant consequences and condition on them.)
The reply would then be that, in general, any feat of superintelligence requires making a ton of big, medium-sized, and little strategies all converge on a single future state in virtue of all of those strategies having been selected sufficiently well to optimize the expectation $\mathbb E[https://arbital.com/p/U|W_1,W_2,...](https://arbital.com/p/U|W_1,W_2,...).$ for some $U.$ A ton of little and medium-sized strategies must have all managed not to collide with each other or with larger big-picture considerations. If you can't do this much then you can't win a game of Go or build a factory or even walk across the room without your limbs tangling up.
Then there doesn't seem to be any good reason to expect an agent which is instead optimizing directly the utility function $U_1$ which is "acquire and hoard resources" to do a very much better job of optimizing $W_{10}$ or $W_{25}.$ When $W_{25}$ already needs to be conditioned in such a way as to not stomp on all the higher goals $W_2, W_3, ...$ it just doesn't seem that much less constraining to target $U_1$ versus $U_0, W_1.$ Most of the cognitive labor in the sequence does not seem like it should be going into checking for $U_0$ at the end instead of checking for $U_1$ at the end. It should be going into, e.g., figuring out how to make any kind of interstellar probe and figuring out how to build factories.
It has not historically been the case that the most computationally efficient way to play chess is to have competing agents inside the chess algorithm trying to optimize different unconditional utility functions and bidding on the right to make moves in order to pursue their own local goal of "protect the queen, regardless of other long-term consequences" or "control the center, regardless of other long-term consequences". What we are actually trying to get is the chess move such that, conditioning on that chess move and the sort of future chess moves we are likely to make, our chance of winning is the highest. The best modern chess algorithms do their best to factor in anything that affects long-range consequences whenever they know about those consequences. The best chess algorithms don't try to factor things into lots of colliding unconditional urges, because sometimes that's not how "the winning move" factors. You can extremely often do better by doing a deeper consequentialist search that conditions multiple elements of your strategy on longer-term consequences in a way that prevents your moves from stepping on each other. It's not very much of an exaggeration to say that this is why humans with brains that can imagine long-term consequences are smarter than, say, armadillos.
Sometimes there are subtleties we don't have the computing power to notice, we can't literally actually condition on the future. But "to make paperclips, acquire resources and use them to make paperclips" versus "to make paperclips, acquire resources regardless of whether they can be used to make paperclips" is not subtle. We'd expect a superintelligence that was [efficient relative to humans](https://arbital.com/p/6s) to understand and correct at least those divergences between $W_1$ and $W_1'$ that a human could see, using at most the trivial amount of computing power represented by a human brain. To the extent that particular choices are being selected-on over a domain that is likely to include choices with huge long-range consequences, one expends the computing power to check and condition on the long-range consequences; but a supermajority of choices shouldn't require checks of this sort; and even choices about how to design train tracks that do require longer-range checks are not going to be very much more tractable depending on whether the distant top of the goal hierarchy is something like "make paperclips" or "hoard resources".
Even supposing that there could be 5% more computational cost associated with checking instrumental strategies for stepping on "promote fun-theoretic eudaimonia", which might ubiquitously involve considerations like "make sure none of the computational processes you use to do this are themselves sentient", this doesn't mean you can't have competent agents that go ahead and spend 5% more computation. Iit's simply the correct choice to build subagents that expend 5% more computation to maintain coordination on achieving eudaimonia, rather than building subagents that expend 5% less computation to hoard resources and never give them back. It doesn't matter if the second kind of agents are less "costly" in some myopic sense, they are vastly less useful and indeed actively destructive. So nothing that is choosing so as to optimize its expectation of $U_0$ will build a subagent that generally optimizes its own expectation of $U_1.$ |
9e3eb505-7beb-40c8-9e2d-6a74148bdbeb | trentmkelly/LessWrong-43k | LessWrong | Born probabilities vs. the universal prior
Just to sharpen a point that Manfred already made:
Imagine a world with a single observer flipping a quantum coin many times. After the observer has seen a million coinflips and they look fair (incompressible), what should he believe about the next thousand? Modern physics says he should have a uniform probability distribution (Born rule). But the universal prior seems to disagree. The algorithmic complexity of the preceding million coinflips is already sunk, so seeing 1000 heads in a row should look more probable to the observer than seeing a more random sequence. Moreover, the combined weight of the first hundred eligible simple programs should exceed the combined weight of everything else in the mixture, no matter how many input bits we've already seen. (That last statement depends on the encoding of programs, but I think I can make it work for any encoding by tweaking the value of 100.)
Of course this doesn't in any way overthrow the universality of Solomonoff induction. The Born rule is a computable prior, therefore it's found somewhere within the universal mixture, so the observer's probability assignments will never be more than a multiplicative constant away from the uniform prior. But then we should be able to detect the multiplicative constant experimentally, and it just doesn't seem to be there. As far as we know, the outcomes of real-world quantum coinflips pass all known tests for "true" randomness, like limiting frequencies or the law of large numbers, with no apparent bias toward algorithmic simplicity.
These days it's easy to get quantum coinflips from the internet. So you have an easy way to generate and observe a random bitstring whose probability under the Solomonoff prior is unimaginably close to zero. Maybe run some tests on it, hoping that Nature uses a pseudorandom generator in disguise after all, but I wouldn't bet on that :-)
The above argument seems to dash hopes of explaining our conscious observations by a simplicity prior, like UDASS |
d1a3199a-bc2c-4a60-9f03-f45991eecaf0 | trentmkelly/LessWrong-43k | LessWrong | Survival and Flourishing Fund grant applications open until October 4th ($1MM-$2MM planned for dispersal)
The Survival and Flourishing Fund is a new fund that (as I understand it) is taking over a significant chunk of the Berkeley Existential Risk Initiative's (BERI) grantmaking, and is just opening a new grant round that I am helping to disburse:
http://survivalandflourishing.org/#2019-q4-grant-round
The plan is to make a total of $1MM-$2MM in grants to organizations working on the long term flourishing and survival of humanity.
At this point in time SFF can only make grants to charities and not individuals, so if you have a project or organization that you want to get funding for, you will have to either already be part of an established charity, found one, or be sponsored by an existing charity.
Also happy to answer any questions you have in the comments, to the best of my knowledge. |
4503cb7c-28bc-4585-ae79-ada581b18465 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AI Could Defeat All Of Us Combined
*Cross-posted from* [*Cold Takes*](https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/)*.*
I've been working on a new series of posts about the [most important century](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd).
* The original series focused on why and how this could be the most important century for humanity. But it had [relatively little to say about](https://forum.effectivealtruism.org/posts/Lbtcjfxhrs8kfKK2M/how-to-make-the-best-of-the-most-important-century-1) *what we can do today* to improve the odds of things going well.
* The new series will get much more specific about the kinds of events that might lie ahead of us, and what actions today look most likely to be helpful.
* A key focus of the new series will be the threat of [misaligned AI](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd/p/AmxxnazJcBWzWEeqj/#misaligned-ai-mysterious-potentially-dangerous-objectives): AI systems disempowering humans entirely, leading to a future that has little to do with anything humans value. ([Like in the Terminator movies](https://www.slowboring.com/p/the-case-for-terminator-analogies?s=r), minus the time travel and the part where humans win.)
Many people have trouble taking this "misaligned AI" possibility seriously. They might see the broad point that AI could be dangerous, but they instinctively imagine that the danger comes from ways humans might misuse it. They find the idea of *AI itself going to war with humans* to be comical and [wild](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd/p/TwQzyP3QgttmuTHym). I'm going to try to make this idea feel more serious and real.
As a first step, this post will **emphasize an unoriginal but extremely important point: [the kind of AI I've discussed](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd/p/AmxxnazJcBWzWEeqj)could defeat all of humanity combined, if (for whatever reason) it were pointed toward that goal.** By "defeat," I don't mean "subtly manipulate us" or "make us less informed" or something like that - I mean a literal "defeat" in the sense that we could all be killed, enslaved or forcibly contained.
I'm not talking (yet) about whether, or why, AIs *might* attack human civilization. That's for future posts. For now, I just want to linger on the point that *if* such an attack happened, it could succeed against the combined forces of the entire world.
* I think that **if you believe this, you should already be worried about misaligned AI,**[1](#fn1) **before any analysis of how or why an AI might form its own goals.**
* We generally don't have a lot of *things that could end human civilization if they "tried"* sitting around. If we're going to create one, I think we should be asking not "Why would this be dangerous?" but "Why wouldn't it be?"
By contrast, if you don't believe that AI could defeat all of humanity combined, I expect that we're going to be miscommunicating in pretty much any conversation about AI. The kind of AI I worry about is the kind powerful enough that total civilizational defeat is a real possibility. The reason I currently spend so much time planning around speculative future technologies (instead of working on [evidence-backed, cost-effective ways of helping low-income people today](https://www.givewell.org/) - which I did for much of my career, and still think is one of the best things to work on) is because I think the stakes are *just that high*.
Below:
* I'll sketch the basic argument for why I think AI could defeat all of human civilization.
+ Others have written about the possibility that "superintelligent" AI could manipulate humans and create overpowering advanced technologies; I'll briefly recap that case.
+ I'll then cover a different possibility, which is that even "merely human-level" AI could still defeat us all - by quickly coming to rival human civilization in terms of total population and resources.
+ At a high level, I think we should be worried if a huge (competitive with world population) and rapidly growing set of highly skilled humans on another planet was trying to take down civilization just by using the Internet. So we should be worried about a large set of disembodied AIs as well.
* I'll briefly address a few objections/common questions:
+ How can AIs be dangerous without bodies?
+ If lots of different companies and governments have access to AI, won't this create a "balance of power" so that no one actor is able to bring down civilization?
+ Won't we see warning signs of AI takeover and be able to nip it in the bud?
+ Isn't it fine or maybe good if AIs defeat us? They have rights too.
* Close with some thoughts on just how unprecedented it would be to have something on our planet capable of overpowering us all.
How AI systems could defeat all of us
-------------------------------------
There's been a lot of debate over whether AI systems might form their own "motivations" that lead them to seek the disempowerment of humanity. I'll be talking about this in future pieces, but for now I want to put it aside and imagine how things would go *if this happened.*
So, for what follows, let's proceed from the premise: **"For some weird reason, humans consistently design AI systems (with human-like research and planning abilities) that coordinate with each other to try and overthrow humanity." Then what?** What follows will necessarily feel wacky to people who find this hard to imagine, but I think it's worth playing along, because I think "we'd be in trouble if this happened" is a very important point.
### The "standard" argument: superintelligence and advanced technology
Other treatments of this question have focused on AI systems' potential to become *vastly* more intelligent than humans, to the point where they have what [Nick Bostrom calls](https://smile.amazon.com/dp/B00LOOCGB2/ref=dp-kindle-redirect?_encoding=UTF8&btkr=1) "cognitive superpowers."[2](#fn2) Bostrom imagines an AI system that can do things like:
* 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.
([Wait But Why](https://waitbutwhy.com/2015/01/artificial-intelligence-revolution-2.html) reasons similarly.[3](#fn3))
I think many readers will already be convinced by arguments like these, and if so you might skip down to the [next major section](https://forum.effectivealtruism.org/s/gBjPorwZHRArNSQ5w/p/6LTh4foNuC3NdtmZH#Some_quick_responses_to_objections).
But I want to be clear that I *don't* think the danger relies on the idea of "cognitive superpowers" or "superintelligence" - both of which refer to capabilities vastly beyond those of humans. **I think we still have a problem even if we assume that AIs will basically have similar capabilities to humans, and not be fundamentally or drastically more intelligent or capable.** I'll cover that next.
### How AIs could defeat humans without "superintelligence"
If we assume that AIs will basically have similar capabilities to humans, I think we still need to worry that they could come to **out-number and out-resource humans,** and could thus have the advantage if they coordinated against us.
Here's a simplified example (some of the simplifications are in this footnote[4](#fn4)) based on [Ajeya Cotra's "biological anchors" report](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines):
* I assume that transformative AI is developed on the soonish side (around 2036 - assuming later would only make the below numbers larger), and that it initially comes in the form of a **single AI system that is able to do more-or-less the same intellectual tasks as a human.** That is, it doesn't have a human body, but it can do anything a human working remotely from a computer could do.
* I'm using the report's framework in which it's much more expensive to *train* (develop) this system than to *run* it (for example, think about how much Microsoft spent to develop Windows, vs. how much it costs for me to run it on my computer).
* The report provides a way of estimating both how much it would cost to *train* this AI system, and how much it would cost to *run* it. Using these estimates (details in footnote)[5](#fn5) implies that once the first human-level AI system is created, whoever created it could use the same computing power it took to create it in order to run **several hundred million copies for about a year each**.[6](#fn6)
* This would be over 1000x the total number of Intel or Google employees,[7](#fn7) over 100x the total number of active and reserve personnel in the [US armed forces](https://en.wikipedia.org/wiki/United_States_Armed_Forces), and something like 5-10% the size of the world's total working-age population.[8](#fn8)
* And that's just a starting point.
+ This is just using the same amount of resources that went into training the AI in the first place. Since these AI systems can do human-level economic work, they can probably be used to make more money and buy or rent more hardware,[9](#fn9) which could quickly lead to a "population" of billions or more.
+ In addition to making more money that can be used to run more AIs, the AIs can conduct massive amounts of research on how to use computing power more efficiently, which could mean still greater numbers of AIs run using the *same* hardware. This in turn could lead to a [feedback loop](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd/p/AmxxnazJcBWzWEeqj#Explosive_scientific_and_technological_advancement) and explosive growth in the number of AIs.
* Each of these AIs might have skills comparable to those of unusually highly paid humans, including scientists, software engineers and quantitative traders. It's hard to say how quickly a set of AIs like this could develop new technologies or make money trading markets, but it seems quite possible for them to amass huge amounts of resources quickly. A huge population of AIs, each able to earn a lot compared to the average human, could end up with a "virtual economy" at least as big as the human one.
To me, this is most of what we need to know: **if there's something with human-like skills, seeking to disempower humanity, with a population in the same ballpark as (or larger than) that of all humans, we've got a civilization-level problem.**
A potential counterpoint is that these AIs would merely be "virtual": if they started causing trouble, humans could ultimately unplug/deactivate the servers they're running on. I do think this fact would make life harder for AIs seeking to disempower humans, but I don't think it ultimately should be cause for much comfort. I think a large population of AIs would likely be able to find some way to achieve security from human shutdown, and go from there to amassing enough resources to overpower human civilization (especially if AIs across the world, including most of the ones humans were trying to use for help, were coordinating).
I spell out what this might look like in an [appendix](https://forum.effectivealtruism.org/s/gBjPorwZHRArNSQ5w/p/6LTh4foNuC3NdtmZH#Appendix__how_AIs_could_avoid_shutdown). In brief:
* By default, I expect the economic gains from using AI to mean that humans create huge numbers of AIs, integrated all throughout the economy, potentially including direct interaction with (and even control of) large numbers of robots and weapons.
+ (If not, I think the situation is in many ways even more dangerous, since a single AI could make many copies of itself and have little competition for things like server space, as discussed in the [appendix](https://forum.effectivealtruism.org/s/gBjPorwZHRArNSQ5w/p/6LTh4foNuC3NdtmZH#What_if_humans_move_slowly_and_don_t_create_many_AIs_).)
* AIs would have multiple ways of obtaining property and servers safe from shutdown.
+ For example, they might recruit human allies (through manipulation, deception, blackmail/threats, genuine promises along the lines of "We're probably going to end up in charge somehow, and we'll treat you better when we do") to rent property and servers and otherwise help them out.
+ Or they might create fakery so that they're able to operate freely on a company's servers while all outward signs seem to show that they're successfully helping the company with its goals.
* A relatively modest amount of property safe from shutdown could be sufficient for housing a huge population of AI systems that are recruiting further human allies, making money (via e.g. quantitative finance), researching and developing advanced weaponry (e.g., bioweapons), setting up manufacturing robots to construct military equipment, thoroughly infiltrating computer systems worldwide to the point where they can disable or control most others' equipment, etc.
* Through these and other methods, a large enough population of AIs could develop enough military technology and equipment to overpower civilization - especially if AIs across the world (including the ones humans were trying to use) were coordinating with each other.
Some quick responses to objections
----------------------------------
This has been a brief sketch of how AIs could come to outnumber and out-resource humans. There are lots of details I haven't addressed.
Here are some of the most common objections I hear to the idea that AI could defeat all of us; if I get much [demand](https://forum.effectivealtruism.org/s/gBjPorwZHRArNSQ5w/p/6LTh4foNuC3NdtmZH#comments) I can elaborate on some or all of them more in the future.
**How can AIs be dangerous without bodies?** This is discussed a fair amount in the [appendix](https://forum.effectivealtruism.org/s/gBjPorwZHRArNSQ5w/p/6LTh4foNuC3NdtmZH#Appendix__how_AIs_could_avoid_shutdown). In brief:
* AIs could recruit human allies, tele-operate robots and other military equipment, make money via research and quantitative trading, etc.
* At a high level, I think we should be worried if a huge (competitive with world population) and rapidly growing set of highly skilled humans on another planet was trying to take down civilization just by using the Internet. So we should be worried about a large set of disembodied AIs as well.
**If lots of different companies and governments have access to AI, won't this create a "balance of power" so that nobody is able to bring down civilization?**
* This is a reasonable objection to many horror stories about AI and other possible advances in military technology, but if *AIs collectively have different goals from humans and are willing to coordinate with each other*[11](#fn11) *against us*, I think we're in trouble, and this "balance of power" idea doesn't seem to help.
* What matters is the total number and resources of AIs vs. humans.
**Won't we see warning signs of AI takeover and be able to nip it in the bud?** I would guess we would see some warning signs, but does that mean we could nip it in the bud? Think about human civil wars and revolutions: there are some warning signs, but also, people go from "not fighting" to "fighting" pretty quickly as they see an opportunity to coordinate with each other and be successful.
**Isn't it fine or maybe good if AIs defeat us? They have rights too.**
* Maybe AIs *should* have rights; if so, it would be nice if we could reach some "compromise" way of coexisting that respects those rights.
* But if they're able to defeat us entirely, that isn't what I'd plan on getting - instead I'd expect (by default) a world run *entirely* according to whatever goals AIs happen to have.
* These goals might have essentially nothing to do with anything humans value, and could be actively counter to it - e.g., placing zero value on beauty and having zero attempts to prevent or avoid suffering).
Risks like this don't come along every day
------------------------------------------
I don't think there are a lot of things that have a serious chance of bringing down human civilization for good.
As argued in [The Precipice](https://theprecipice.com/), most natural disasters (including e.g. asteroid strikes) don't seem to be huge threats, if only because civilization has been around for thousands of years so far - implying that natural civilization-threatening events are rare.
Human civilization is pretty powerful and seems pretty robust, and accordingly, what's really scary to me is the idea of something with the same basic capabilities as humans (making plans, developing its own technology) that can outnumber and out-resource us. There aren't a lot of candidates for that.[12](#fn12)
AI is one such candidate, and I think that even before we engage heavily in arguments about whether AIs might seek to defeat humans, we should feel very nervous about the possibility that they could.
What about things like "AI might lead to mass unemployment and unrest" or "AI might exacerbate misinformation and propaganda" or "AI might exacerbate a wide range of other social ills and injustices"[13](#fn13)? I think these are real concerns - but to be honest, if they were the biggest concerns, I'd probably still be focused on [helping people in low-income countries today](https://www.givewell.org/) rather than trying to prepare for future technologies.
* Predicting the future is generally hard, and it's easy to pour effort into preparing for challenges that never come (or come in a very different form from what was imagined).
* I believe civilization is pretty robust - we've had huge changes and challenges over the last century-plus (full-scale world wars, [many dramatic changes in how we communicate with each other](https://forum.effectivealtruism.org/posts/ajBYeiggAzu6Cgb3o/biological-anchors-is-about-bounding-not-pinpointing-ai?commentId=nEeuknn4unKTWEd6i), dramatic changes in lifestyles and values) without seeming to have come very close to a collapse.
* So if I'm engaging in speculative worries about a potential future technology, I want to focus on the really, really big ones - the ones that could matter for billions of years. If there's a real possibility that AI systems will have values different from ours, and cooperate to try to defeat us, that's such a worry.
*Special thanks to Carl Shulman for discussion on this post.*
Appendix: how AIs could avoid shutdown
--------------------------------------
This appendix goes into detail about how AIs coordinating against humans could amass resources of their own without humans being able to shut down all "misbehaving" AIs.
It's necessarily speculative, and should be taken in the spirit of giving examples of how this might work - for me, the high-level concern is that a huge, coordinating population of AIs with similar capabilities to humans would be a threat to human civilization, and that we shouldn't count on any particular way of stopping it such as shutting down servers.
I'll discuss two different general types of scenarios: (a) Humans create a huge population of AIs; (b) Humans move slowly and don't create many AIs.
### How this could work if humans create a huge population of AIs
I think a reasonable default expectation is that humans do most of the work of making AI systems incredibly numerous and powerful (because doing so is profitable), which leads to a vulnerable situation. Something roughly along the lines of:
* The company that first develops transformative AI quickly starts running large numbers of copies (hundreds of millions or more), which are used to (a) do research on how to improve computational efficiency and run more copies still; (b) develop valuable intellectual property (trading strategies, new technologies) and make money.
* Over time, AI systems are rolled out widely throughout society. Their numbers grow further, and their role in the economy grows: they are used in (and therefore have direct interaction with) high-level decision-making at companies, perhaps operating large numbers of cars and/or robots, perhaps operating military drones and aircraft, etc. (This seems like a default to me over time, but it isn't strictly necessary for the situation to be risky, as I'll go through below.)
* In this scenario, the AI systems are malicious (as we've assumed), but this doesn't mean they're constantly causing trouble. Instead, they're mostly waiting for an opportunity to team up and decisively overpower humanity. In the meantime, they're mostly behaving themselves, and this is leading to their numbers and power growing.
+ There are scattered incidents of AI systems' trying to cause trouble,[14](#fn14) but this doesn't cause the whole world to stop using AI or anything.
+ A reasonable analogy might be to a typical civil war or revolution: the revolting population *mostly* avoids isolated, doomed attacks on its government, until it sees an opportunity to band together and have a real shot at victory.
(Paul Christiano's [What Failure Looks Like](https://www.lesswrong.com/posts/HBxe6wdjxK239zajf/what-failure-looks-like) examines this general flavor of scenario in a bit more detail.)
In this scenario, the AIs face a challenge: if it becomes obvious to everyone that they are trying to defeat humanity, humans could attack or shut down a few concentrated areas where most of the servers are, and hence drastically reduce AIs' numbers. So the AIs need a way of **getting one or more "AI headquarters": property they control where they can safely operate servers and factories, do research, make plans and construct robots/drones/other military equipment.**
Their goal is ultimately to have enough AIs, robots, etc. to be able to defeat the rest of humanity combined. This might mean constructing overwhelming amounts of military equipment, or thoroughly infiltrating computer systems worldwide to the point where they can disable or control most others' equipment, or researching and deploying extremely powerful weapons (e.g., bioweapons), or a combination.
Here are some ways they could get to that point:
* They could **recruit human allies** through many different methods - manipulation, deception, blackmail and other threats, genuine promises along the lines of "We're probably going to end up in charge somehow, and we'll treat you better when we do."
+ Human allies could be given valuable intellectual property (developed by AIs), given instructions for making lots of money, and asked to rent their own servers and acquire their own property where an "AI headquarters" can be set up. Since the "AI headquarters" would officially be human property, it could be very hard for authorities to detect and respond to the danger.
+ Via threats, AIs might be able to get key humans to cooperate with them - such as political leaders, or the CEOs of companies running lots of AIs. This would open up further strategies.
* As assumed above, particular companies are running huge numbers of AIs. The AIs being run by these companies might find security holes in the companies' servers (this isn't the topic of this piece, but my general impression is that security holes are widespread and that reasonably competent people can find many of them)[15](#fn15), and thereby might find opportunities to **create durable "fakery"** about what they're up to.
+ E.g., they might set things up so that as far as humans can tell, it looks like all of the AI systems are hard at work creating profit-making opportunities for the company, when in fact they're essentially using the server farm as their headquarters - and/or trying to establish a headquarters somewhere else (by recruiting human allies, sending money to outside bank accounts, using that money to acquire property and servers, etc.)
* If AIs are in wide enough use, they might already be operating lots of drones and other military equipment, in which case it could be pretty straightforward to be able to defend some piece of territory - or to strike a deal with some government to enlist its help in doing so.
* AIs could mix-and-match the above methods and others: for example, creating "fakery" long enough to recruit some key human allies, then attempting to threaten and control humans in key positions of power to the point where they control solid amounts of military resources, then using this to establish a "headquarters."
In any of these cases, once one or more "AI headquarters" are established:
* It can be expanded as the AIs headquartered there make money, purchase more property and computing power, build more factories, etc.[16](#fn16)
* It can develop enough destructive potential and human allies that humans would be hesitant to attack it (as many relatively small countries have today). It would be clear that AIs at the headquarters are seeking autonomy, accumulating resources and taking some unethical actions, but the full scale and nature of the threat could easily remain ambiguous.
* Even if the threat is clear, AIs could potentially maintain a huge population in a relatively small space (potentially underground). They could also split up across many different headquarters (including some where they are simply using humans' hardware while hiding their presence).
+ Either way, this large-but-hidden population could be recruiting human allies, developing advanced weaponry, looking for ways to take control of human' weapons (potentially possible if they greatly outnumber the humans working on cybersecurity, and if they can coordinate with the AIs "helping" those humans), etc.
+ They could also be working with (unwitting, or allied) humans to manufacture large numbers of robots with back-doors allowing the AIs to take control.
* Once the headquarters has built up enough resources to go on offense, **AIs around the world could coordinate with the headquarters.** Humans could shut down AIs that they notice doing this, but they might have a very tough time getting value out of their servers and AI-controlled robots; this could make it easy for the AIs at the "AI headquarters" to out-resource humans.
### What if humans move slowly and don't create many AIs?
The above scenario has humans creating large numbers of AIs, such that the AIs just need to find a way to coordinate and acquire a safe "headquarters" in order to defeat us.
What if humans moved more slowly, intentionally restricting human-level AI to a tiny portion of the available computing resources? Could a *small* number of AIs pose a risk to humanity?
In this world, we would have what Carl Shulman refers to as "dry tinder everywhere, waiting for sparks." Anyone who can buy or rent a large amount of computing power can create a large number of AIs, which can produce a large amount of money and research, leading to still more AIs.
So a single AI could hack into enough servers[17](#fn17) to make a few copies of itself; recruit a few human allies; and start making money, acquiring more server space, etc. until its human allies are running a huge number of AIs. This could all be done in difficult-to-detect ways (it might e.g. just look like a set of humans renting a bunch of servers to run quantitative finance strategies).
So in this world, I think our concern should be any AI that is able to find enough security holes to attain that kind of freedom. Given the current state of cybersecurity, that seems like a big concern.
Footnotes
---------
1. Assuming you accept other points made in the [most important century](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd) series, e.g. that AI that can do most of what humans do to [advance science and technology](https://forum.effectivealtruism.org/s/isENJuPdB3fhjWYHd/p/AmxxnazJcBWzWEeqj) could be developed this century. [↩](#fnref1)
2. See [Superintelligence](https://smile.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom-ebook/dp/B00LOOCGB2/) chapter 6. [↩](#fnref2)
3. See the "Nanotechnology blue box," in particular. [↩](#fnref3)
4. * The report estimates the amount of computing power it would take to *train* (create) a transformative AI system, and the amount of computing power it would take to *run* one. This is a [bounding exercise](https://forum.effectivealtruism.org/posts/ajBYeiggAzu6Cgb3o/biological-anchors-is-about-bounding-not-pinpointing-ai) and isn't supposed to be literally predicting that transformative AI will arrive in the form of a single AI system trained in a single massive run, but here I am interpreting the report that way for concreteness and simplicity.
* As explained in the next footnote, I use the report's figures for transformative AI arriving on the soon side (around 2036). Using its central estimates instead would strengthen my point, but we'd then be talking about a longer time from now; I find it helpful to imagine how things could go in a world where AI comes relatively soon. [↩](#fnref4)
5. I assume that transformative AI ends up costing about 10^14 FLOP/s to run (this is about 1/10 the Bio Anchors central estimate, and well within its error bars) and about 10^30 FLOP to train (this is about 10x the Bio Anchors central estimate for how much will be available in 2036, and corresponds to about the 30th-percentile estimate for how much will be needed based on the "short horizon" anchor). That implies that the 10^30 FLOP needed to *train* a transformative model could *run* 10^16 seconds' worth of transformative AI models, or about 300 million years' worth. This figure would be higher if we use Bio Anchors's central assumptions, rather than assumptions consistent with transformative AI being developed on the soon side. [↩](#fnref5)
6. They might also run fewer copies of scaled-up models or more copies of scaled-down ones, but the idea is that the total productivity of all the copies should be at *least* as high as that of several hundred million copies of a human-ish model. [↩](#fnref6)
7. [Intel](https://en.wikipedia.org/wiki/Intel), [Google](https://en.wikipedia.org/wiki/Google) [↩](#fnref7)
8. Working-age population: about [65%](https://data.worldbank.org/indicator/SP.POP.1564.TO.ZS) \* [7.9 billion](https://ourworldindata.org/world-population-growth) =~ 5 billion. [↩](#fnref8)
9. Humans could rent hardware using money they made from running AIs, or - if AI systems were operating on their own - they could potentially rent hardware themselves via human allies or just via impersonating a customer (you generally don't need to physically show up in order to e.g. rent server time from Amazon Web Services). [↩](#fnref9)
10. (I had a speculative, illustrative possibility here but decided it wasn't in good enough shape even for a footnote. I might add it later.) [↩](#fnref10)
11. I don't go into detail about how AIs might coordinate with each other, but it seems like there are many options, such as by opening their own email accounts and emailing each other. [↩](#fnref11)
12. Alien invasions seem unlikely if only because we have no evidence of one in millions of years. [↩](#fnref12)
13. Here's a recent [comment exchange](https://forum.effectivealtruism.org/posts/ajBYeiggAzu6Cgb3o/biological-anchors-is-about-bounding-not-pinpointing-ai?commentId=59wqZk8wcKdEWQG3J) I was in on this topic. [↩](#fnref13)
14. E.g., individual AI systems may occasionally get caught trying to steal, lie or exploit security vulnerabilities, due to various unusual conditions including bugs and errors. [↩](#fnref14)
15. E.g., see this [list of high-stakes security breaches](https://docs.google.com/document/d/1_smEDPWDVIaLuZ14Cm7KLHcWx4LkJ0DCTk8bcHjYy_Y/edit) and a [list of quotes about cybersecurity](https://docs.google.com/document/d/1VtV_eX-vU3bC41Il-x0OnT3DGqlL6zawWqtkisqr5Cg/edit), both courtesy of Luke Muehlhauser. For some additional not-exactly-rigorous evidence that at least shows that "cybersecurity is in really bad shape" is seen as relatively uncontroversial by at least one cartoonist, see: <https://xkcd.com/2030/> [↩](#fnref15)
16. Purchases and contracts could be carried out by human allies, or just by AI systems themselves with humans willing to make deals with them (e.g., an AI system could digitally sign an agreement and wire funds from a bank account, or via cryptocurrency). [↩](#fnref16)
17. See [above](#fn15) note about my general assumption that today's cybersecurity has a lot of holes in it. [↩](#fnref17) |
3a0ca8e8-c7be-4384-a06b-476abaaaadeb | trentmkelly/LessWrong-43k | LessWrong | When considering incentives, consider the incentives of all parties
Once upon a time the countries of Alpago and Byzantine had a war. Alpago was mostly undamaged during this war. Byzantine was severely damaged by this war, although they have caught up in some metrics such as education, their economy is still somewhat weaker. Alpago was the clear aggressor, and now, fifty years later, everyone who is reasonable now acknowledges that Alpago was in the wrong.
There is a major debate within the countries about how to respond to the past. Many Byzantians argue that the views of the Alpagoans are irrelevant. The Alpagoans are "unbombed", this provides them with many systematic advantage over the Byzantians such as career opportunities, indeed most of the top companies in Byzantian still have Alpagoan CEOs since many of the senior management were hired before Byzantian had built anywhere near the number of colleges in Alpago.
Many Byzantians argue that the views of the unbombed deserve very little consideration. Of course the unbombed will want to preserve their advantages. How can the Byzantians ever have their voices heard when unbombed members of parliament are giving their opinions in the Alpago parliament on how much compensation is appropriate? Surely if Alpago was truly sorry, they would accept the demands of the Byzantian government without question.
The Byzantians are undoubtedly correct in their assumption that the Alpagoans have a very strong incentive to underestimate what is owed. They are also correct when they say that the Alpagoans are in a position of power that makes it very easy for them to ignore the issue of compensation, after all, it does not affect them very much if their government decides to pay compensation to the Byzantians, instead of the alternate plan of wasting it on a fleet of nuclear submarines. However, in other areas, the Alpagoans no longer have a power advantage. Many Alpagoan politicians used to say that the war was justified, if a politician said that these days, even the conservative party would |
8f253291-3a69-49da-bbf6-4af15e4a59a0 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | New DeepMind report on institutions for global AI governance
This is a linkpost for a new white paper from DeepMind, titled 'Exploring institutions for global AI governance', on "models and functions of international institutions that could help manage opportunities and mitigate risks of advanced AI". I've copied the first few paragraphs.
---
Growing awareness of the global impact of advanced artificial intelligence (AI) has inspired public discussions about the need for international governance structures to help manage opportunities and mitigate risks involved.
Many discussions have drawn on analogies with the [ICAO](https://www.icao.int/Pages/default.aspx) (International Civil Aviation Organisation) in civil aviation; [CERN](https://www.home.cern/) (European Organisation for Nuclear Research) in particle physics; [IAEA](https://www.iaea.org/) (International Atomic Energy Agency) in nuclear technology; and intergovernmental and multi-stakeholder organisations in many other domains. And yet, while analogies can be a useful start, the technologies emerging from AI will be unlike aviation, particle physics, or nuclear technology.
To succeed with AI governance, we need to better understand:
1. What specific benefits and risks we need to manage internationally.
2. What governance functions those benefits and risks require.
3. What organisations can best provide those functions.
[Our latest paper](https://arxiv.org/abs/2307.04699), with collaborators from the University of Oxford, Université de Montréal, University of Toronto, Columbia University, Harvard University, Stanford University, and OpenAI, addresses these questions and investigates how international institutions could help manage the global impact of frontier AI development, and make sure AI’s benefits reach all communities. |
24173e2c-2ec2-4f36-b498-043df0294592 | trentmkelly/LessWrong-43k | LessWrong | An Anchoring Experiment: Results
This post summarises the results of the experiment which tested how anchoring works on the LW audience. Here is the original post which describes the experiment in more detail. The experiment was supposed to decide between two ways of how anchoring may work. The first hypothesis is that the subject always starts from the anchor and continues in the direction of his/her unbiased estimate, but doesn't go far enough. The alternative hypothesis is that anchoring shifts the centre of the subject's probability distribution towards the anchor, and the whole distribution moves along.
To illustrate the difference, consider the first experimental question, which was about the population of the Central African Republic. The correct value (i.e. the estimate for 2009 listed on Wikipedia) is 4,422,000. The anchor which I have offered here was 20 million. Now, if the first hypothesis is true, the people who, in their unbiased state of mind, would guess less than 20 million, would slide down starting from the 20 million value and stop prematurely; their guesses will be attracted towards the anchor, but not across. The distribution of the biased guesses would be narrower and overall closer to 20 million than the unbiased distribution, but the probability of answering whatever number lower than 20 million would not be changed by anchoring. On the other hand, if the second hypothesis holds, the biased group should guess more than 20 million more often than the control group.
The actual results are such:
Group I (biased; 36 answers collected)
* more than 20 million: 15 (41.7%)
* less than 20 million: 21 (58.3%)
Group II (control; 16 answers collected)
* more than 20 million: 3 (18.7%)
* less than 20 million: 13 (81.3%)
* 20 million: 1 (6.3%)
The second question asked for the altitude of the highest point in Sweden (2140 m / 6903 ft). The anchor was 3500 m or 11500 ft (there is about 5 m / 18 ft difference between the values, but I wanted both the metric and the imperial anc |
ea416531-fd02-4ef3-8330-208daa2989f9 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Less Wrong NH Meetup
Discussion article for the meetup : Less Wrong NH Meetup
WHEN: 03 November 2016 07:00:00PM (-0400)
WHERE: 269 Pearl St, Manchester NH 03104
The Less Wrong NH meetup is Thursday, 11/3, in Manchester, NH at 7 pm at a private residence. Light refreshments will be provided.
Have you read Rationality: from AI to Zombies, or any of the Sequences on Less Wrong? Maybe you're just a fan of Harry Potter and the Methods of Rationality. Come hang out with us and discuss optimization of whatever it is you want to optimize.
Agenda: See Facebook
You may want to bring a notebook.
http://www.meetup.com/Less-Wrong-NH-Meetup/
https://www.facebook.com/groups/LessWrongNH/
Discussion article for the meetup : Less Wrong NH Meetup |
a17d04fc-25b1-4e19-98a5-dec7a19f7353 | StampyAI/alignment-research-dataset/arxiv | Arxiv | An Experiment on the Connection between the DLs' Family DL<ForAllPiZero> and the Real World
INTRODUCTION
------------
In the last years, Semantic Web has increasingly expanded its area
of influence. Being an innovative instrument for the retrieval of
not expressly stored information and a way of organizing concepts
and relations by their meaning, it broadens up plenty of horizons
for knowledge representation. Though, only a bunch of experts and
researchers know that under a suprisingly vast dimension of new features
there lies a mathematical and logical structure inside which they
fight day by day for the balancing of expressiveness and efficiency.
The *Description Logics* formalisms, which we will see in Section
[1](#S1 "1 THEORETICAL FUNDAMENTALS ‣ An Experiment on the Connection between the Description Logics’ Family 𝒟ℒ <∀₀^𝜋> and the Real World"), are the formal bases for the
so-called Web 3.0, that should allow one to automatically infer (and
retrieve) new information from *reasoning* on “sematicized”
knowledge repositories. Some examples of families of logics are described
in the following, together with a short coverage of arguments such
as RDF graphs and OWL.
Section [2](#S2 "2 ANALYSIS AND RESULTS ‣ An Experiment on the Connection between the Description Logics’ Family 𝒟ℒ <∀₀^𝜋> and the Real World") introduces the description
logic 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>, showing notable characteristics of expressive
power and polynomial complexity. In the same section, a more detailed
description of the analysis, which may be seen as a conceptual experiment,
follows. Our purpose is to show that a good number of real-world ontologies
may be related to this family of logics. Results are promising enough
to spur us to stay on this path and complement it with studies on
temporalization of Description Logics and, consequently, of Semantic
Web (briefly touched in this report), which we will approach in the
near future.
The SPARQL query which was specifically created to assess the membership
of some real-world ontologies to the aforementioned logic is reported
in Appendix [A](#A1 "Appendix A ANALYSIS OF THE SPARQL QUERY ‣ An Experiment on the Connection between the Description Logics’ Family 𝒟ℒ <∀₀^𝜋> and the Real World").
1 THEORETICAL
FUNDAMENTALS
---------------------------
###
1.1 Description Logics
*Description
Logics* (*DL*s) are a family of formalisms for knowledge representation,
built on logic-based semantics. They are founded on some fundamental
ideas:
* •
basic syntactic “blocks” are atomic concepts (1-ary
predicates), atomic roles (2-ary predicates) and individuals
(costants);
* •
the expressiveness of a certain language depends on the use of a set
of chosen constructors, that give birth to complex concepts and roles
starting from existing ones;
* •
by means of classification of concepts, a *subsumption* hierarchy
is established, which specifies what concept includes or is included
by another;
* •
implied knowledge is automatically obtained through a logic procedure,
called *reasoning*, essentially based on the application of inference
rules to subsumption between concepts, and an instance definitions
between individuals and the latter.
On the assumption that a Knowledge Representation System (*KR-System*)
is to give an answer to a user query, *reasoning* algorithms
for DLs should be regarded as decisional procedures which return a
positive or negative verdict. This raises the decision problem for
such languages. Furthermore, having an answer does not always mean
getting it in a reasonable lapse of time, and that compels us to consider
also the complexity of algorithms at stake. Decidability and complexity
directly depend on the expressive power of the description logic we
use: whereas very expressive DLs tend to have inference problems and
be computationally *hard* (or even undecidable), the DLs that
are more efficient in *reasoning* turn out to be not espressive
enough to represent all concepts and relations in the domain of interest.
Research still going on in the field of DLs just aims at the ability
of balancing expressiveness against efficiency, while not dropping
semantic precision that could make it applicable to real world situations.
A *Description Logics Knowledge Base* (*DLKB*) is made of
two components: TBox (*Terminological Box*) and ABox (*Assertional
Box*). The former contains the vocabulary, i.e., the definitions of
atomic and non-atomic concepts and roles, called *axioms*, whereas
the latter describes individuals in terms of this vocabulary, i.e.
it includes the declarations, called *assertions*, of their instances.
By means of the TBox we can name complex descriptions of concepts
and roles. The language for such a naming is what distinguishes one
DL from another and is based on a *model-theoretic* semantics.
Thus statements in TBox and ABox can be regarded as first-order logic
formulas. *Reasoning* procedures for the terminological part
are used to verify the *satisfiability* of a description (i.e.,
its non-contradictoriness), or whether it may be *subsumed* by
another (i.e., whether the latter is more general than the former),
while those for the assertive part establish whether its set of assertion
is *consistent* (i.e., it has a model or it entails that a certain
individual is an instance of a given concept). Satisfiability tests
for descriptions and consistency tests for a set of assertion allow
one to establish whether the knowledge base is meaningful or not,
whereas subsumption tests allow one to maintain a hierarchy of concepts*ab universali*; finally, instance tests give one the ability of querying
the system against single individuals.
More formally, a generic DL axiom is a formula of one of the following
types:
* •
C≡D𝐶𝐷C\equiv Ditalic\_C ≡ italic\_D (equivalence between concepts)
* •
C⊑Dsquare-image-of-or-equals𝐶𝐷C\sqsubseteq Ditalic\_C ⊑ italic\_D (subsumption between concepts)
* •
R≡S𝑅𝑆R\equiv Sitalic\_R ≡ italic\_S (equivalence between roles)
* •
R⊑Ssquare-image-of-or-equals𝑅𝑆R\sqsubseteq Sitalic\_R ⊑ italic\_S (subsumption between roles),111The equivalence and subsumption between roles may be indicated also
by the symbols === and ⊆\subseteq⊆, respectively.
where the symbols C,D𝐶𝐷C,Ditalic\_C , italic\_D are names or expressions of complex
concepts, which are formed by 1-ary or 2-ary operations on/between
atomic (indicated by A𝐴Aitalic\_A in the following) or complex concepts, while
R,S𝑅𝑆R,Sitalic\_R , italic\_S are names or expressions of complex roles which are formed
by 1-ary or 2-ary operations on/between atomic (indicated by P𝑃Pitalic\_P
in the following) or complex roles.
A generic DL assertion is a formula of one of the following types:
* •
C(a)𝐶𝑎C\left(a\right)italic\_C ( italic\_a ) (concept assertion)
* •
R(a,b)𝑅𝑎𝑏R\left(a,b\right)italic\_R ( italic\_a , italic\_b ) (role assertion),
where the symbols a,b𝑎𝑏a,bitalic\_a , italic\_b are names of individuals, for which
the concept C𝐶Citalic\_C or the role R𝑅Ritalic\_R holds.222A large number of constructs is listed in the table at the end of
this section.
From a semantical point of view, an *interpretation* ℐℐ\mathcal{I}caligraphic\_I is a
pair (Δℐ,⋅ℐ)superscriptΔℐsuperscript⋅ℐ(\Delta^{\mathcal{I}},\cdot^{\mathcal{I}})( roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT , ⋅ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ), where the non-empty set ΔℐsuperscriptΔℐ\Delta^{\mathcal{I}}roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT represents
the domain of the interpretation and the interpretation function
⋅ℐsuperscript⋅ℐ\cdot^{\mathcal{I}}⋅ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT associates a set Aℐ⊆Δℐsuperscript𝐴ℐsuperscriptΔℐA^{\mathcal{I}}\subseteq\Delta^{\mathcal{I}}italic\_A start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT to every atomic
concept A𝐴Aitalic\_A , a relation Pℐ⊆Δℐ×Δℐsuperscript𝑃ℐsuperscriptΔℐsuperscriptΔℐP^{\mathcal{I}}\subseteq\Delta^{\mathcal{I}}\times\Delta^{\mathcal{I}}italic\_P start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT × roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT to every atomic
role P𝑃Pitalic\_P, and an element aℐ∈Δℐsuperscript𝑎ℐsuperscriptΔℐa^{\mathcal{I}}\in\Delta^{\mathcal{I}}italic\_a start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT to every individual a𝑎aitalic\_a.
We write:333In the following, ℐ⊧ϕmodelsℐitalic-ϕ\mathcal{I}\models\phicaligraphic\_I ⊧ italic\_ϕ means “ℐℐ\mathcal{I}caligraphic\_I satisfies ϕitalic-ϕ\phiitalic\_ϕ,
where ϕitalic-ϕ\phiitalic\_ϕ can be an axiom or an assertion. A syntax/semantical
reference for the main axioms, assertions and property declarations
is listed in the table at the end of this section.
* •
ℐ⊧(C≡D)modelsℐ𝐶𝐷\mathcal{I}\models\left(C\equiv D\right)caligraphic\_I ⊧ ( italic\_C ≡ italic\_D ) iff444“Iff” is short for “if and only if”. Cℐ=Dℐsuperscript𝐶ℐsuperscript𝐷ℐC^{\mathcal{I}}=D^{\mathcal{I}}italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
* •
ℐ⊧(C⊑D)modelsℐsquare-image-of-or-equals𝐶𝐷\mathcal{I}\models\left(C\sqsubseteq D\right)caligraphic\_I ⊧ ( italic\_C ⊑ italic\_D ) iff Cℐ⊆Dℐsuperscript𝐶ℐsuperscript𝐷ℐC^{\mathcal{I}}\subseteq D^{\mathcal{I}}italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
* •
ℐ⊧(R≡S)modelsℐ𝑅𝑆\mathcal{I}\models\left(R\equiv S\right)caligraphic\_I ⊧ ( italic\_R ≡ italic\_S ) iff Rℐ=Sℐsuperscript𝑅ℐsuperscript𝑆ℐR^{\mathcal{I}}=S^{\mathcal{I}}italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
* •
ℐ⊧(R⊑S)modelsℐsquare-image-of-or-equals𝑅𝑆\mathcal{I}\models\left(R\sqsubseteq S\right)caligraphic\_I ⊧ ( italic\_R ⊑ italic\_S ) iff Rℐ⊆Sℐsuperscript𝑅ℐsuperscript𝑆ℐR^{\mathcal{I}}\subseteq S^{\mathcal{I}}italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
* •
ℐ⊧C(a)modelsℐ𝐶𝑎\mathcal{I}\models C\left(a\right)caligraphic\_I ⊧ italic\_C ( italic\_a ) iff aℐ∈Cℐsuperscript𝑎ℐsuperscript𝐶ℐa^{\mathcal{I}}\in C^{\mathcal{I}}italic\_a start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
* •
ℐ⊧R(a,b)modelsℐ𝑅𝑎𝑏\mathcal{I}\models R\left(a,b\right)caligraphic\_I ⊧ italic\_R ( italic\_a , italic\_b ) iff (a,b)ℐ∈Rℐsuperscript𝑎𝑏ℐsuperscript𝑅ℐ\left(a,b\right)^{\mathcal{I}}\in R^{\mathcal{I}}( italic\_a , italic\_b ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT
Finally, we say that ℐℐ\mathcal{I}caligraphic\_I is a *model* for a DLKB 𝒦𝒦\mathcal{K}caligraphic\_K (and write
ℐ⊧𝒦modelsℐ𝒦\mathcal{I}\models\mathcal{K}caligraphic\_I ⊧ caligraphic\_K) if ℐℐ\mathcal{I}caligraphic\_I satisfies all the axioms and assertions of
𝒦𝒦\mathcal{K}caligraphic\_K . The latter is said to be *consistent* if there exists at
least an interpretation satisfying it, and the search of this interpretation
(*consistency problem*) is just the clue to *reasoning.*
###
1.2 Families of logics
As already observed, distinct DLs are characterized
by the constructs allowed to form complex concepts and roles starting
from atomic ones. The names are usually specified by a series of alphabet
letters and symbols. In the following, we will not concentrate on
formal semantics, but, for the sake of clarity, we will only hint
at the meaning of some constructs.
By way of an example, we briefly describe the 𝒜ℒ𝒞𝒜ℒ𝒞\mathcal{ALC}caligraphic\_A caligraphic\_L caligraphic\_C logic (*Attributive
Language with Complements*). Its syntax obeys the following rules:
| | | |
| --- | --- | --- |
| | C,D→A|⊤|⊥|¬C|C⊓D|C⊔D|∀R.C|∃R.C,C,D\rightarrow A\ |\ \top\ |\ \bot\ |\ \neg C\ |\ C\sqcap D\ |\ C\sqcup D\ |\ \forall R.C\ |\ \exists R.C\,\,,italic\_C , italic\_D → italic\_A | ⊤ | ⊥ | ¬ italic\_C | italic\_C ⊓ italic\_D | italic\_C ⊔ italic\_D | ∀ italic\_R . italic\_C | ∃ italic\_R . italic\_C , | |
where ⊤top\top⊤ denotes the concept enclosing any other one (*top
concept*), ⊥bottom\bot⊥ denotes the concept enclosed in any other one (*bottom
concept*), ¬\neg¬ negates a concept; ⊔square-union\sqcup⊔ represents the union
of concepts (notice the analogy with the corresponding set operators)
and ⊓square-intersection\sqcap⊓ represents the intersection of concepts. The last two
constructs are called respectively *universal* and *existential
restriction*, and are pivotal in research connected to this report.
The concepts that can be built in 𝒜ℒ𝒞𝒜ℒ𝒞\mathcal{ALC}caligraphic\_A caligraphic\_L caligraphic\_C are called 𝒜ℒ𝒞𝒜ℒ𝒞\mathcal{ALC}caligraphic\_A caligraphic\_L caligraphic\_C -concepts.The
axiomsand assertions that may be expressed in the 𝒜ℒ𝒞𝒜ℒ𝒞\mathcal{ALC}caligraphic\_A caligraphic\_L caligraphic\_C logic
are
| | | |
| --- | --- | --- |
| | C≡D,C⊑D,C(a),formulae-sequence𝐶𝐷square-image-of-or-equals𝐶𝐷𝐶𝑎C\equiv D,\ \,\ \,\ C\sqsubseteq D,\ \,\ \,\ C\left(a\right),italic\_C ≡ italic\_D , italic\_C ⊑ italic\_D , italic\_C ( italic\_a ) , | |
which indicate, respectively, equivalence and subsumption (also called
*GCI*, *General Concept Inclusion*) between two concepts,
and the membership of a concept. However, *reasoning* in 𝒜ℒ𝒞𝒜ℒ𝒞\mathcal{ALC}caligraphic\_A caligraphic\_L caligraphic\_C has a PSpace-complete computational complexity.
In addition to those seen above, the most common constructs for concepts
which can be formed in DLs are ≤1.Rformulae-sequenceabsent1𝑅\leq 1.R≤ 1 . italic\_R (*functional restriction*,
denoted ℱℱ\mathcal{F}caligraphic\_F , that is equivalent to ∃R.⊤formulae-sequence𝑅top\exists R.\top∃ italic\_R . ⊤), ≤nRabsent𝑛𝑅\leq nR≤ italic\_n italic\_R
and ≥nRabsent𝑛𝑅\geq nR≥ italic\_n italic\_R (*numerical restrictions* , 𝒩𝒩\mathcal{N}caligraphic\_N , which enclose
the functional one), ≤nR.Cformulae-sequenceabsent𝑛𝑅𝐶\leq nR.C≤ italic\_n italic\_R . italic\_C and ≥nR.Cformulae-sequenceabsent𝑛𝑅𝐶\geq nR.C≥ italic\_n italic\_R . italic\_C (*qualified
restrictions*, 𝒬𝒬\mathcal{Q}caligraphic\_Q , that include the numerical ones), {a}𝑎\left\{a\right\}{ italic\_a }
and {a1,…,an}subscript𝑎1…subscript𝑎𝑛\left\{a\_{1},\ldots,a\_{n}\right\}{ italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT } (*nominals*, 𝒪𝒪\mathcal{O}caligraphic\_O ),
and ∃R.𝑆𝑒𝑙𝑓formulae-sequence𝑅𝑆𝑒𝑙𝑓\exists R.\mathit{Self}∃ italic\_R . italic\_Self (*self-concept*). Contructs for
roles are often denoted by an operator symbol inside brackets after
the name of the logic. They are R−superscript𝑅R^{-}italic\_R start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT (inverse role, ℐℐ\mathcal{I}caligraphic\_I), R∪S,R∩S𝑅𝑆𝑅𝑆R\cup S,R\cap Sitalic\_R ∪ italic\_S , italic\_R ∩ italic\_S
and ¬R𝑅\neg R¬ italic\_R (role union, intersection and complement, respectively),
R∘S𝑅𝑆R\circ Sitalic\_R ∘ italic\_S (role composition, also used in chain), R+superscript𝑅R^{+}italic\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT and
R\*superscript𝑅R^{\*}italic\_R start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT (transitive and reflexive-transitive closure of roles),
𝑖𝑑(C)𝑖𝑑𝐶\mathit{id}\left(C\right)italic\_id ( italic\_C )(*concept identity*), and RC|,R|DR\_{C|},R\_{|D}italic\_R start\_POSTSUBSCRIPT italic\_C | end\_POSTSUBSCRIPT , italic\_R start\_POSTSUBSCRIPT | italic\_D end\_POSTSUBSCRIPT
and RC|Dsubscript𝑅conditional𝐶𝐷R\_{C|D}italic\_R start\_POSTSUBSCRIPT italic\_C | italic\_D end\_POSTSUBSCRIPT (role restrictions). Sometimes, the symbols U𝑈Uitalic\_U (or
*∇normal-∇\nabla∇*, *universal
role*, defining the role which encloses any other one) and *N𝑁Nitalic\_N*(or **△normal-△\triangle△**,*empty role*, defining the role enclosed in any other one) are used.
Other kinds of axioms which can be introduced are R≡S𝑅𝑆R\equiv Sitalic\_R ≡ italic\_S (equivalence
between roles), R⊑Ssquare-image-of-or-equals𝑅𝑆R\sqsubseteq Sitalic\_R ⊑ italic\_S (hierarchy between roles, ℋℋ\mathcal{H}caligraphic\_H ,
denoted also by R⊆S𝑅𝑆R\subseteq Sitalic\_R ⊆ italic\_S), and the relative assertion R(a,b)𝑅𝑎𝑏R\left(a,b\right)italic\_R ( italic\_a , italic\_b )
(role instance). The reflexive, irreflexive, symmetric, antisymmetric,
transitive, intransitive, disjunctive (two roles having no pair of
elements in common), functional and inverse functional properties
for roles are denoted by symbols that often differ in literature,
but are never ambiguous, e.g. respectively 𝑆𝑦𝑚(R),𝐴𝑠𝑦𝑚(R),𝑅𝑒𝑓𝑙(R),𝐼𝑟𝑟𝑒𝑓𝑙(R),𝑇𝑟(R),𝐼𝑛𝑡𝑟(R),𝐷𝑖𝑠𝑗(R),𝐹𝑛(R),𝐼𝑛𝑣𝐹𝑛(R)𝑆𝑦𝑚𝑅𝐴𝑠𝑦𝑚𝑅𝑅𝑒𝑓𝑙𝑅𝐼𝑟𝑟𝑒𝑓𝑙𝑅𝑇𝑟𝑅𝐼𝑛𝑡𝑟𝑅𝐷𝑖𝑠𝑗𝑅𝐹𝑛𝑅𝐼𝑛𝑣𝐹𝑛𝑅\mathit{Sym}\left(R\right),\mathit{Asym}\left(R\right),\mathit{Refl}\left(R\right),\mathit{Irrefl}\left(R\right),\mathit{Tr}\left(R\right),\mathit{Intr}\left(R\right),\mathit{Disj}\left(R\right),\mathit{Fn}\left(R\right),\mathit{InvFn}\left(R\right)italic\_Sym ( italic\_R ) , italic\_Asym ( italic\_R ) , italic\_Refl ( italic\_R ) , italic\_Irrefl ( italic\_R ) , italic\_Tr ( italic\_R ) , italic\_Intr ( italic\_R ) , italic\_Disj ( italic\_R ) , italic\_Fn ( italic\_R ) , italic\_InvFn ( italic\_R ).
When transitive property is allowed, we use the symbol 𝒮𝒮\mathcal{S}caligraphic\_S , which
corresponds to 𝒜ℒ𝒞(+)\mathcal{ALC}(^{+})caligraphic\_A caligraphic\_L caligraphic\_C ( start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT ). Sometimes, even small differences among
logics (e.g., limitation on the use of atomic rather than complex
roles on the right or left part of an assertion) can make a big difference
in expressiveness and complexity. Thus, it is not always easy to concisely
denote DLs, that consequently constitute an ever-open research field.
One or more among the above descripted constructs, axioms and assertions
are present in the families of logics that are the base of languages
used in ontologies. Among these, an important example is the logic
sℛ𝒪ℐ𝒬(D)𝑠ℛ𝒪ℐsubscript𝒬𝐷s\mathcal{ROIQ}\_{(D)}italic\_s caligraphic\_R caligraphic\_O caligraphic\_I caligraphic\_Q start\_POSTSUBSCRIPT ( italic\_D ) end\_POSTSUBSCRIPT (in literature always denoted by 𝒮ℛ𝒪ℐ𝒬(D)𝒮ℛ𝒪ℐ𝒬𝐷\mathcal{SROIQ}(D)caligraphic\_S caligraphic\_R caligraphic\_O caligraphic\_I caligraphic\_Q ( italic\_D ) ), at the bottom
of the OWL 2 DL profile, which will be discussed in the following
(the letter in parenthesis indicates the use of *concrete domains*,
that will not be discussed here). Its syntax is easily inferred from
the symbols, while its complexity is N2ExpTime-hard.555By the symbol ℛℛ\mathcal{R}caligraphic\_R , we mean that a DL allows complex inclusions of
the kind R∘S⊑Rsquare-image-of-or-equals𝑅𝑆𝑅R\circ S\sqsubseteq Ritalic\_R ∘ italic\_S ⊑ italic\_R and R∘S⊑Ssquare-image-of-or-equals𝑅𝑆𝑆R\circ S\sqsubseteq Sitalic\_R ∘ italic\_S ⊑ italic\_S.
| *Construct* | *Syntax* | *Semantics* |
| --- | --- | --- |
| top concept | ⊤top\top⊤ | ⊤ℐ=Δℐ\top^{\mathcal{I}}=\Delta^{\mathcal{I}}⊤ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| bottom concept | ⊥bottom\bot⊥ | ⊥ℐ=∅\bot^{\mathcal{I}}=\emptyset⊥ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ∅ |
| concept negation | ¬C𝐶\neg C¬ italic\_C | (¬C)ℐ=Δℐ\Cℐsuperscript𝐶ℐ\superscriptΔℐsuperscript𝐶ℐ\left(\neg C\right)^{\mathcal{I}}=\Delta^{\mathcal{I}}\backslash C^{\mathcal{I}}( ¬ italic\_C ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT \ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| concept intersection | C⊓Dsquare-intersection𝐶𝐷C\sqcap Ditalic\_C ⊓ italic\_D | (C⊓D)ℐ=Cℐ∩Dℐsuperscriptsquare-intersection𝐶𝐷ℐsuperscript𝐶ℐsuperscript𝐷ℐ\left(C\sqcap D\right)^{\mathcal{I}}=C^{\mathcal{I}}\cap D^{\mathcal{I}}( italic\_C ⊓ italic\_D ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∩ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| concept union | C⊔Dsquare-union𝐶𝐷C\sqcup Ditalic\_C ⊔ italic\_D | (C⊔D)ℐ=Cℐ∪Dℐsuperscriptsquare-union𝐶𝐷ℐsuperscript𝐶ℐsuperscript𝐷ℐ\left(C\sqcup D\right)^{\mathcal{I}}=C^{\mathcal{I}}\cup D^{\mathcal{I}}( italic\_C ⊔ italic\_D ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∪ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| universal restriction | ∀R.Cformulae-sequencefor-all𝑅𝐶\forall R.C∀ italic\_R . italic\_C | (∀R.C)ℐ={a∈Δℐ|(∀(a,b)∈Rℐ)(b∈Cℐ)}\left(\forall R.C\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\left(\forall\left(a,b\right)\in R^{\mathcal{I}}\right)\left(b\in C^{\mathcal{I}}\right)\right\}( ∀ italic\_R . italic\_C ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( ∀ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( italic\_b ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) } |
| existential restriction | ∃R.Cformulae-sequence𝑅𝐶\exists R.C∃ italic\_R . italic\_C | (∃R.C)ℐ={a∈Δℐ|∃(a,b)∈Rℐ∧b∈Cℐ}\left(\exists R.C\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\,\exists\left(a,b\right)\in R^{\mathcal{I}}\wedge b\in C^{\mathcal{I}}\right\}( ∃ italic\_R . italic\_C ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ∃ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∧ italic\_b ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| numerical restriction |
| |
| --- |
| ≤nRabsent𝑛𝑅\leq nR≤ italic\_n italic\_R |
| ≥nRabsent𝑛𝑅\geq nR≥ italic\_n italic\_R |
|
| |
| --- |
| (≤nR)ℐ={a∈Δℐ|#{b∈Δℐ|(a,b)∈Rℐ}≤n}\left(\leq nR\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\,\#\left\{b\in\Delta^{\mathcal{I}}|\left(a,b\right)\in R^{\mathcal{I}}\right\}\leq n\right\}( ≤ italic\_n italic\_R ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | # { italic\_b ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } ≤ italic\_n } |
| (≥nR)ℐ={a∈Δℐ|#{b∈Δℐ|(a,b)∈Rℐ}}≥n}\left(\geq nR\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\,\#\left\{b\in\Delta^{\mathcal{I}}|\left(a,b\right)\in R^{\mathcal{I}}\}\right\}\geq n\right\}( ≥ italic\_n italic\_R ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | # { italic\_b ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } } ≥ italic\_n } |
|
| qualified restriction |
| |
| --- |
| ≤nR.Cformulae-sequenceabsent𝑛𝑅𝐶\leq nR.C≤ italic\_n italic\_R . italic\_C |
| ≥nRabsent𝑛𝑅\geq nR≥ italic\_n italic\_R.C |
|
| |
| --- |
| (≤nR.C)ℐ={a∈Δℐ|#{b∈Cℐ|(a,b)∈Rℐ}≤n}\left(\leq nR.C\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\,\#\left\{b\in C^{\mathcal{I}}|\left(a,b\right)\in R^{\mathcal{I}}\right\}\leq n\right\}( ≤ italic\_n italic\_R . italic\_C ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | # { italic\_b ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } ≤ italic\_n } |
| (≥nR.C)ℐ={a∈Δℐ|#{b∈Cℐ|(a,b)∈Rℐ}≥n}\left(\geq nR.C\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\,\#\left\{b\in C^{\mathcal{I}}|\left(a,b\right)\in R^{\mathcal{I}}\right\}\geq n\right\}( ≥ italic\_n italic\_R . italic\_C ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | # { italic\_b ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } ≥ italic\_n } |
|
| nominals |
| |
| --- |
| {a}𝑎\left\{a\right\}{ italic\_a } |
| {a1,…,an}subscript𝑎1…subscript𝑎𝑛\left\{a\_{1},\ldots,a\_{n}\right\}{ italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT } |
|
| |
| --- |
| {a}ℐ={aℐ}superscript𝑎ℐsuperscript𝑎ℐ\left\{a\right\}^{\mathcal{I}}=\left\{a^{\mathcal{I}}\right\}{ italic\_a } start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| {a1,…,an}ℐ={a1ℐ,…,anℐ}superscriptsubscript𝑎1…subscript𝑎𝑛ℐsuperscriptsubscript𝑎1ℐ…superscriptsubscript𝑎𝑛ℐ\left\{a\_{1},\ldots,a\_{n}\right\}^{\mathcal{I}}=\left\{a\_{1}^{\mathcal{I}},\ldots,a\_{n}^{\mathcal{I}}\right\}{ italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT } start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
|
| self concept | ∃R.𝑆𝑒𝑙𝑓formulae-sequence𝑅𝑆𝑒𝑙𝑓\exists R.\mathit{Self}∃ italic\_R . italic\_Self | (∃R.Self)ℐ={a∈Δℐ|(a,a)∈Rℐ}\left(\exists R.Self\right)^{\mathcal{I}}=\left\{a\in\Delta^{\mathcal{I}}|\left(a,a\right)\in R^{\mathcal{I}}\right\}( ∃ italic\_R . italic\_S italic\_e italic\_l italic\_f ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | ( italic\_a , italic\_a ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| universal role | U𝑈Uitalic\_U, ∇∇\nabla∇ | Uℐ=Δℐ×Δℐsuperscript𝑈ℐsuperscriptΔℐsuperscriptΔℐU^{\mathcal{I}}=\Delta^{\mathcal{I}}\times\Delta^{\mathcal{I}}italic\_U start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT × roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| empty role | N𝑁Nitalic\_N, △△\triangle△ | Nℐ=∅×∅superscript𝑁ℐN^{\mathcal{I}}=\emptyset\times\emptysetitalic\_N start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ∅ × ∅ |
| role inverse | R−superscript𝑅R^{-}italic\_R start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT | (R−)ℐ={(a,b)|(b,a)∈Rℐ}superscriptsuperscript𝑅ℐconditional-set𝑎𝑏𝑏𝑎superscript𝑅ℐ\left(R^{-}\right)^{\mathcal{I}}=\left\{\left(a,b\right)|\left(b,a\right)\in R^{\mathcal{I}}\right\}( italic\_R start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { ( italic\_a , italic\_b ) | ( italic\_b , italic\_a ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| role negation | ¬R𝑅\neg R¬ italic\_R | (¬R)ℐ=(Δℐ×Δℐ)\Rℐsuperscript𝑅ℐ\superscriptΔℐsuperscriptΔℐsuperscript𝑅ℐ\left(\neg R\right)^{\mathcal{I}}=\left(\Delta^{\mathcal{I}}\times\Delta^{\mathcal{I}}\right)\backslash R^{\mathcal{I}}( ¬ italic\_R ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ( roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT × roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) \ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| role intersection | R⊓Ssquare-intersection𝑅𝑆R\sqcap Sitalic\_R ⊓ italic\_S | (R⊓S)ℐ=Rℐ∩Sℐsuperscriptsquare-intersection𝑅𝑆ℐsuperscript𝑅ℐsuperscript𝑆ℐ\left(R\sqcap S\right)^{\mathcal{I}}=R^{\mathcal{I}}\cap S^{\mathcal{I}}( italic\_R ⊓ italic\_S ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∩ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| role union | R⊔Ssquare-union𝑅𝑆R\sqcup Sitalic\_R ⊔ italic\_S | (R⊔S)ℐ=Rℐ∪Sℐsuperscriptsquare-union𝑅𝑆ℐsuperscript𝑅ℐsuperscript𝑆ℐ\left(R\sqcup S\right)^{\mathcal{I}}=R^{\mathcal{I}}\cup S^{\mathcal{I}}( italic\_R ⊔ italic\_S ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∪ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| chaining | R∘S𝑅𝑆R\circ Sitalic\_R ∘ italic\_S | (R∘S)ℐ=Rℐ∘Sℐsuperscript𝑅𝑆ℐsuperscript𝑅ℐsuperscript𝑆ℐ\left(R\circ S\right)^{\mathcal{I}}=R^{\mathcal{I}}\circ S^{\mathcal{I}}( italic\_R ∘ italic\_S ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∘ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| concept identity | 𝑖𝑑(C)𝑖𝑑𝐶\mathit{id}\left(C\right)italic\_id ( italic\_C ) | (id(C))ℐ={(a,a)|a∈Cℐ}superscript𝑖𝑑𝐶ℐconditional-set𝑎𝑎𝑎superscript𝐶ℐ\left(id\left(C\right)\right)^{\mathcal{I}}=\left\{\left(a,a\right)|\,a\in C^{\mathcal{I}}\right\}( italic\_i italic\_d ( italic\_C ) ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { ( italic\_a , italic\_a ) | italic\_a ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| transitive closure | R+superscript𝑅R^{+}italic\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT | (R+)ℐ=(Rℐ)+superscriptsuperscript𝑅ℐsuperscriptsuperscript𝑅ℐ\left(R^{+}\right)^{\mathcal{I}}=\left(R^{\mathcal{I}}\right)^{+}( italic\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ( italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT |
| refl-trans closure | R\*superscript𝑅R^{\*}italic\_R start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT | (R\*)ℐ=(R+)ℐ∪(id(⊤))ℐsuperscriptsuperscript𝑅ℐsuperscriptsuperscript𝑅ℐsuperscript𝑖𝑑topℐ\left(R^{\*}\right)^{\mathcal{I}}=\left(R^{+}\right)^{\mathcal{I}}\cup\left(id\left(\top\right)\right)^{\mathcal{I}}( italic\_R start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ( italic\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∪ ( italic\_i italic\_d ( ⊤ ) ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| role restriction |
| |
| --- |
| RC|R\_{C|}italic\_R start\_POSTSUBSCRIPT italic\_C | end\_POSTSUBSCRIPT |
| R|DR\_{|D}italic\_R start\_POSTSUBSCRIPT | italic\_D end\_POSTSUBSCRIPT |
| RC|Dsubscript𝑅conditional𝐶𝐷R\_{C|D}italic\_R start\_POSTSUBSCRIPT italic\_C | italic\_D end\_POSTSUBSCRIPT |
|
| |
| --- |
| (RC|)ℐ={(a,b)∈Rℐ|a∈Cℐ}(R\_{C|})^{\mathcal{I}}=\left\{\left(a,b\right)\in R^{\mathcal{I}}|\,a\in C^{\mathcal{I}}\right\}( italic\_R start\_POSTSUBSCRIPT italic\_C | end\_POSTSUBSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | italic\_a ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| (R|D)ℐ={(a,b)∈Rℐ|b∈Dℐ}(R\_{|D})^{\mathcal{I}}=\left\{\left(a,b\right)\in R^{\mathcal{I}}|\,b\in D^{\mathcal{I}}\right\}( italic\_R start\_POSTSUBSCRIPT | italic\_D end\_POSTSUBSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | italic\_b ∈ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
| (RC|D)ℐ={(a,b)∈Rℐ|a∈Cℐ∧b∈Dℐ}superscriptsubscript𝑅conditional𝐶𝐷ℐconditional-set𝑎𝑏superscript𝑅ℐ𝑎superscript𝐶ℐ𝑏superscript𝐷ℐ(R\_{C|D})^{\mathcal{I}}=\left\{\left(a,b\right)\in R^{\mathcal{I}}|\,a\in C^{\mathcal{I}}\wedge b\in D^{\mathcal{I}}\right\}( italic\_R start\_POSTSUBSCRIPT italic\_C | italic\_D end\_POSTSUBSCRIPT ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = { ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT | italic\_a ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∧ italic\_b ∈ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT } |
|
*Table 1.1.* Main constructs for
concepts and roles
| *Syntax* | *Semantics* |
| --- | --- |
| C≡D𝐶𝐷C\equiv Ditalic\_C ≡ italic\_D | Cℐ=Dℐsuperscript𝐶ℐsuperscript𝐷ℐC^{\mathcal{I}}=D^{\mathcal{I}}italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| C⊑Dsquare-image-of-or-equals𝐶𝐷C\sqsubseteq Ditalic\_C ⊑ italic\_D | Cℐ⊆Dℐsuperscript𝐶ℐsuperscript𝐷ℐC^{\mathcal{I}}\subseteq D^{\mathcal{I}}italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ italic\_D start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
|
| |
| --- |
| C(a)𝐶𝑎C\left(a\right)italic\_C ( italic\_a ) |
| ¬C(a)𝐶𝑎\neg C\left(a\right)¬ italic\_C ( italic\_a ) |
|
| |
| --- |
| aℐ∈Cℐsuperscript𝑎ℐsuperscript𝐶ℐa^{\mathcal{I}}\in C^{\mathcal{I}}italic\_a start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∈ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| aℐ∉Cℐsuperscript𝑎ℐsuperscript𝐶ℐa^{\mathcal{I}}\notin C^{\mathcal{I}}italic\_a start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∉ italic\_C start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
|
| R≡S𝑅𝑆R\equiv Sitalic\_R ≡ italic\_S | Rℐ=Sℐsuperscript𝑅ℐsuperscript𝑆ℐR^{\mathcal{I}}=S^{\mathcal{I}}italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| R⊑Ssquare-image-of-or-equals𝑅𝑆R\sqsubseteq Sitalic\_R ⊑ italic\_S | Rℐ⊆Sℐsuperscript𝑅ℐsuperscript𝑆ℐR^{\mathcal{I}}\subseteq S^{\mathcal{I}}italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ⊆ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
|
| |
| --- |
| R(a,b)𝑅𝑎𝑏R\left(a,b\right)italic\_R ( italic\_a , italic\_b ) |
| ¬R(a,b)𝑅𝑎𝑏\neg R\left(a,b\right)¬ italic\_R ( italic\_a , italic\_b ) |
|
| |
| --- |
| (a,b)ℐ∈Rℐsuperscript𝑎𝑏ℐsuperscript𝑅ℐ\left(a,b\right)^{\mathcal{I}}\in R^{\mathcal{I}}( italic\_a , italic\_b ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
| (a,b)ℐ∉Rℐsuperscript𝑎𝑏ℐsuperscript𝑅ℐ\left(a,b\right)^{\mathcal{I}}\notin R^{\mathcal{I}}( italic\_a , italic\_b ) start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∉ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT |
|
| R⊓S=Nsquare-intersection𝑅𝑆𝑁R\sqcap S=Nitalic\_R ⊓ italic\_S = italic\_N, 𝐷𝑖𝑠𝑗(R,S)𝐷𝑖𝑠𝑗𝑅𝑆\mathit{Disj}\left(R,S\right)italic\_Disj ( italic\_R , italic\_S ) | Rℐ∩Sℐ=∅×∅superscript𝑅ℐsuperscript𝑆ℐR^{\mathcal{I}}\cap S^{\mathcal{I}}=\emptyset\times\emptysetitalic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ∩ italic\_S start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT = ∅ × ∅ |
|
| |
| --- |
| 𝑅𝑒𝑓𝑙(R)𝑅𝑒𝑓𝑙𝑅\mathit{Refl}\left(R\right)italic\_Refl ( italic\_R ) |
| ¬𝑅𝑒𝑓𝑙(R)𝑅𝑒𝑓𝑙𝑅\neg\mathit{Refl}\left(R\right)¬ italic\_Refl ( italic\_R ) |
|
| |
| --- |
| (∀a∈Δℐ)((a,a)∈Rℐ)for-all𝑎superscriptΔℐ𝑎𝑎superscript𝑅ℐ\left(\forall a\in\Delta^{\mathcal{I}}\right)\left(\left(a,a\right)\in R^{\mathcal{I}}\right)( ∀ italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_a , italic\_a ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
| (∀a∈Δℐ)((a,a)∉Rℐ)for-all𝑎superscriptΔℐ𝑎𝑎superscript𝑅ℐ\left(\forall a\in\Delta^{\mathcal{I}}\right)\left(\left(a,a\right)\notin R^{\mathcal{I}}\right)( ∀ italic\_a ∈ roman\_Δ start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_a , italic\_a ) ∉ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
|
|
| |
| --- |
| 𝑆𝑦𝑚(R)𝑆𝑦𝑚𝑅\mathit{Sym}\left(R\right)italic\_Sym ( italic\_R ) |
| ¬𝑆𝑦𝑚(R)𝑆𝑦𝑚𝑅\neg\mathit{Sym}\left(R\right)¬ italic\_Sym ( italic\_R ) |
|
| |
| --- |
| (∀(a,b)∈Rℐ)((b,a)∈Rℐ)for-all𝑎𝑏superscript𝑅ℐ𝑏𝑎superscript𝑅ℐ\left(\forall\left(a,b\right)\in R^{\mathcal{I}}\right)\left(\left(b,a\right)\in R^{\mathcal{I}}\right)( ∀ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_b , italic\_a ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
| (∀(a,b)∈Rℐ)((b,a)∉Rℐ)for-all𝑎𝑏superscript𝑅ℐ𝑏𝑎superscript𝑅ℐ\left(\forall\left(a,b\right)\in R^{\mathcal{I}}\right)\left(\left(b,a\right)\notin R^{\mathcal{I}}\right)( ∀ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_b , italic\_a ) ∉ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
|
|
| |
| --- |
| 𝑇𝑟𝑎𝑛𝑠(R)𝑇𝑟𝑎𝑛𝑠𝑅\mathit{Trans}\left(R\right)italic\_Trans ( italic\_R ) |
| ¬𝑇𝑟𝑎𝑛𝑠(R)𝑇𝑟𝑎𝑛𝑠𝑅\neg\mathit{Trans}\left(R\right)¬ italic\_Trans ( italic\_R ) |
|
| |
| --- |
| (∀(a,b)∈Rℐ)((b,c)∈Rℐ→(a,c)∈Rℐ)for-all𝑎𝑏superscript𝑅ℐ𝑏𝑐superscript𝑅ℐ→𝑎𝑐superscript𝑅ℐ\left(\forall\left(a,b\right)\in R^{\mathcal{I}}\right)\left(\left(b,c\right)\in R^{\mathcal{I}}\rightarrow\left(a,c\right)\in R^{\mathcal{I}}\right)( ∀ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_b , italic\_c ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT → ( italic\_a , italic\_c ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
| (∀(a,b)∈Rℐ)((b,c)∈Rℐ→(a,c)∉Rℐ)for-all𝑎𝑏superscript𝑅ℐ𝑏𝑐superscript𝑅ℐ→𝑎𝑐superscript𝑅ℐ\left(\forall\left(a,b\right)\in R^{\mathcal{I}}\right)\left(\left(b,c\right)\in R^{\mathcal{I}}\rightarrow\left(a,c\right)\notin R^{\mathcal{I}}\right)( ∀ ( italic\_a , italic\_b ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) ( ( italic\_b , italic\_c ) ∈ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT → ( italic\_a , italic\_c ) ∉ italic\_R start\_POSTSUPERSCRIPT caligraphic\_I end\_POSTSUPERSCRIPT ) |
|
*Table 1.2.* Main types of axioms
and assertions on concepts and roles666The use of 𝑇𝑟𝑎𝑛𝑠(R)𝑇𝑟𝑎𝑛𝑠𝑅\mathit{Trans}\left(R\right)italic\_Trans ( italic\_R ), 𝑅𝑒𝑓𝑙(R)𝑅𝑒𝑓𝑙𝑅\mathit{Refl}\left(R\right)italic\_Refl ( italic\_R ),
and so on, looks like adding a new type of axiom. Actually, they must
be regarded as abbreviations of more complex expressions involving
equivalence, subsumption and various constructs which we will not
be treated here.
###
1.3 Semantic Web
The so-called Web 2.0 describes the current model
of information search. The difference with Web 1.0 is represented
by the change in the origin of this information (from below−users−rather
than from above−webmasters). Yet, a paradigm change in
its fundamental structure has not occured, as the creator of Web,
Tim Berners-Lee, wished instead. Indeed, contents are still exclusively
made up of pages connected by links, whose nontrivial words, together
with metadata, are indexed by search engines, by means of which a
correspondence between those and the one inserted by users can be
found. In this way, notwithstanding the high efficiency and precision
of the algorithms involved, such engines constitute a “stupid”
example of *information* *retrieval*, where the matching
between terms, even if advanced, weighs more than their meaning.
The change brought by the so-called Semantic Web (*SW* or Web
3.0) addresses the main issue that in the Web, as we know it, most
of the contents are structured in order to be read by humans, rather
than investigated by programs in an automatic way. A computer can
jump from page to page by following links, but it does not make any
assumption on the semantics of their contents. SW is an extension
of “previous” Web that allows one to assign a meaning to information
and provides machines with the ability to elaborate and “understand”
what in the past they could only show. For all this to work, computers
should be able to access well-structured information schemas and apply
inference rules to permit automatic reasoning, in order to delocalize
knowledge representation and spread it over the Net. The languages
these rules apply should be expressive enough to make the Web capable
of “reasoning” in a versatile
and widespread way. That is made possibile by tools such as *eXstensible
Markup Language* (*XML*), *Resource Description Framework*
(*RDF*) and the ontologies described by the *Ontology Web
Language* (*OWL*).
###
1.4 RDF graphs
RDF is a model for representation of information on
the Web. The basic idea is that every resource (concept, class, property,
object, value, etc.) can be univocally described in terms of simple
or identified-by-value properties. All this permits to schematize
an RDF model by means of an oriented graph, whose nodes represent
primitive objects and whose edges denote properties. The limited vocabulary
of RDF is extended by *RDFS* (*RDF Schema*), which allows
one to define classes and properties in a more powerful way. E.g.,
it is possibile to define a property as a relation by indicating its
domain and/or range, or say what class is a subclass of another.
From a data structure point of view, an RDF graph may be seen as a
set of triples of type (subject predicate object). For each
triple, subject and object are two nodes connected
by an edge that represents a predicate; thus, the subject of a triple
may be the object of another and viceversa. Each of these nodes may
be identified by a URI/IRI, and so it can represent a resource, or
it may be simply a “connector” (in this case, it is called a *blank
node*, *bnode*, or *anonymous node*). If an object is not
a subject of another triple, it may also be a *literal*, i.e.,
a datum representing a value in a certain domain.
On the representation side, W3C suggests five types of *serialization*s
for an RDF graph, the most used of which are RDF/XML (XML
language is used) and Turtle (descriptions through lists
of triples).777The *serialization* of an object is the process which allows
one to represent it in an accessible way, in our case a text file. XML is used because of its precise representation of data. Thus,
they can be shared, by means of RDF, among various Web applications
while preserving their original meaning.
###
1.5 Ontologies
OWL ontologies extend RDFS vocabulary further.888Computer science ambitiously draw on the philosophical term *ontology*
to show that the OWL language can describe the world starting from
its ultimate constituents and from relations among them. If RDFS guarantees generality and precision in constructing knowledge
representation, with OWL we reach such an expressiveness that we can
“argue” about described information and “extract” more, through
*reasoning*. It provides powerful tools to define classes (that
represent concepts), properties (relations among classes) and individuals
(their instances), and gives the chance to combine them in logic constraints
by means of which necessity and/or sufficiency may be expressed.
Upon a W3C *recommendation*, OWL is made up of three sublanguages
with increasing expressiveness: OWL Lite, OWL DL and OWL Full.
The first two have an almost total correspondence between two peculiar
DLs (resp. *𝒮ℋℐℱ(D)𝒮ℋℐsubscriptℱ𝐷\mathcal{SHIF}\_{(D)}caligraphic\_S caligraphic\_H caligraphic\_I caligraphic\_F start\_POSTSUBSCRIPT ( italic\_D ) end\_POSTSUBSCRIPT*and *𝒮ℋ𝒪ℐ𝒩(D)𝒮ℋ𝒪ℐsubscript𝒩𝐷\mathcal{SHOIN}\_{(D)}caligraphic\_S caligraphic\_H caligraphic\_O caligraphic\_I caligraphic\_N start\_POSTSUBSCRIPT ( italic\_D ) end\_POSTSUBSCRIPT*).999OWL-Lite and OWL-DL provide the possibility of defining *annotation*s,
which correspondent DLs do not do. In any case, annotations affect
neither *reasoning*, nor complexity, nor decidability. For the
sake of precision, OWL-Lite is considered a “syntactic subset”
of OWL-DL. The last one uses the same subset of OWL DL constructs, but allows
them to be unconstrained, according to RDF style. Due to the lack
of restrictions on transitive property and to the possibility of handling
concepts as individuals (*metamodeling*), OWL Full turns out
to be undecidable and thus, for the correspondence between that and
any of the aforesaid DLs, OWL DL is the most expressive decidable
sublanguage of OWL.101010Obviously, nothing excludes there is another decidable and more expressive−but
not studied yet−sublanguage.
OWL 2 extends OWL by enhancing its features and expressiveness. Apart
from a limited number of cases, OWL 2 is perfectly “backward compatible”
with the old OWL (which we call OWL 1 to differentiate it from the
OWL 2), i.e., all the ontologies of the latter keep the same semantics
as that they had before, even if “fed” to a *reasoner* for
OWL 2. Concisely, OWL 2 introduces a simplification into the writing
of the most common *statements*, it permits *metamodeling*,
together with some new constructs which increase its expressiveness,
and it extends support to data types.
As for sublanguages, in OWL 2 we can deal with OWL 2 Full and OWL 2 DL.
As mentioned before, the latter corresponds to the *sℛ𝒪ℐ𝒬(D)𝑠ℛ𝒪ℐsubscript𝒬𝐷s\mathcal{ROIQ}\_{(D)}italic\_s caligraphic\_R caligraphic\_O caligraphic\_I caligraphic\_Q start\_POSTSUBSCRIPT ( italic\_D ) end\_POSTSUBSCRIPT*
logic. Though the former may be syntactically seen as the union of
the latter with RDFS, it is semantically compatible with OWL 2 DL (insofar as its semantics allows one to draw all the inferences that
can be drawn by using the semantics of OWL 2 DL). Similarly to the
analogue sublanguage OWL 1 Full, OWL 2 finds application in the
modelling of concepts where *reasoning* is not required. A new
feature of OWL 2 is the use of *profiles* (i.e., fragments of
language) OWL 2 EL, OWL 2 QL and OWL 2 RL:111111In this way, OWL 1 Lite, OWL 1 DL and OWL 2 DL may be considered
profiles of OWL 2. the first is useful for applications involving ontologies that contain
a great number of classes and/or properties; the second aims at those
applications with large volumes of instances, where answering the
queries is of primary importance (from which the name QL, *Query
Language*); the last profile is useful in applications not requiring
the sacrifice of too much expressiveness to efficiency. We will not
delve into the structural characteristics of these profiles. Here,
we only say that they are grounded on very different DLs, each of
them fit for the specific purpose they are designed for.
###
1.6 Syntactical correspondence of OWL(2) with DLs
As to its practical representation, an ontology is
a RDF/XML-structured text file, but, for our goals, it will be more
useful to think to it in terms of an RDF graph and, thus, of triples.
Table 1.3 describes the correspondence between the most common constructs
of DLs and such triples (or a single resource, where applicable),
which use RDF(S) syntax together with that of OWL(2). It also contains
the names given by OWL to the fundamental concepts and roles of DLs.
It is interesting to notice how a DL construct often corresponds to
more triples, which sometimes makes interpretation quite hard. The
symbols \_:x and \_list in the Table 1.3 indicate
respectively a *blank node* and a *list* (a structure we
will not deal with), whereas the prefixes before the elements of triples
represent standard W3C *namespaces*, which serve as abbreviations
for the URIs they refer to.121212The correspondences between namespaces and prefixes are:
rdf: http://www.w3.org/1999/02/22-rdf-syntax-ns#
rdfs: http://www.w3.org/2000/01/rdf-schema#
owl: http://www.w3.org/2002/07/owl#
xsd: http://www.w3.org/2001/XMLSchema#
###
1.7 SPARQL
One of the most widespread languages for querying
against ontologies is *SPARQL* *(*recursive acronym for
*SPARQL Protocol And RDF Query Language*). Although it has lots
of analogies with SQL, most of which syntactical, it is equipped with
a fundamentally different semantics. As query languages for databases
essentially work on tables and handle logic conditions to select the
rows of these tables that satisfy them, the WHERE clause in
SPARQL finds matches between the triples of the query and those of
the ontology indicated in the FROM clause. The logic assessments
are relegated to the FILTER operator, which possesses a big
expressive power, thanks to its several functions, but is rarely used
due to the loss of efficiency that its presence may cause.131313Depending on implementation, the FROM clause may be implied
because software loads in memory the ontology model separately. The SELECT clause is very similar to that of SQL. It accepts
DISTINCT as a keyword and the classical aggregation operators
(COUNT, SUM, MIN, AVG and MAX),
while the GROUP BY and HAVING clauses are often used
(another similarity) to refine the selection. The involved variables
are preceded by the symbol ’?’, whereas constants do not have
a particular syntax, even if they generally coincide with URI/IRIs
of resources present in the ontology one is handling.
Analyzing in finer detail the WHERE clause, one can say that
the triples to matchare enclosed in a block between braces
and separated by a dot, which implies their intersection. Inside these
braces, more sub-blocks may be found, which are useful in making the
union (UNION operator) and the difference (MINUS operator)
between sets. The matching is valid if the variables having the same
names in the clause have the same values in the ontology. There exist
some useful shortenings, such as the use of semicolon or colon in
place of dot, which act respectively in the following way:
?s ?p1 ?o1 ; ?p2 ?o2 . shortens ?s ?p1 ?o1 . ?s ?p2 ?o2 .
?s ?p ?o1 , ?o2 . shortens ?s ?p ?o1 . ?s ?p ?o2 .
In addition to selection, one can also have the DESCRIBE, ASK
and CONSTRUCT query types, which we will not review here.
The employment of the PREFIX clauses, that precede the real
query and indicate the abbreviations for namespaces used inside of
it, is quite peculiar.
| *DL construct* | *RDF(S)/OWL(2) resource/triple(s)* |
| --- | --- |
| ⊤top\top⊤ | owl:Thing |
| ⊥bottom\bot⊥ | owl:Nothing |
| C⊑Dsquare-image-of-or-equals𝐶𝐷C\sqsubseteq Ditalic\_C ⊑ italic\_D | (C𝐶Citalic\_C rdfs:subClassOf D𝐷Ditalic\_D) |
| C≡D𝐶𝐷C\equiv Ditalic\_C ≡ italic\_D | (C𝐶Citalic\_C owl:equivalentClass D𝐷Ditalic\_D) |
| C⊑¬Dsquare-image-of-or-equals𝐶𝐷C\sqsubseteq\neg Ditalic\_C ⊑ ¬ italic\_D opp. C⊓D⊑⊥square-image-of-or-equalssquare-intersection𝐶𝐷bottomC\sqcap D\sqsubseteq\botitalic\_C ⊓ italic\_D ⊑ ⊥ | (C𝐶Citalic\_C owl:disjointWith D𝐷Ditalic\_D) |
| C1⊓C2⊓…⊓Cnsquare-intersectionsubscript𝐶1subscript𝐶2…subscript𝐶𝑛C\_{1}\sqcap C\_{2}\sqcap\ldots\sqcap C\_{n}italic\_C start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ⊓ italic\_C start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ⊓ … ⊓ italic\_C start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT | (\_:x owl:intersectionOf \_list(C1,C2,…,Cn)subscript𝐶1subscript𝐶2normal-…subscript𝐶𝑛\left(C\_{1},C\_{2},\ldots,C\_{n}\right)( italic\_C start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_C start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_C start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT )) |
| C1⊔C2⊔…⊔Cnsquare-unionsubscript𝐶1subscript𝐶2…subscript𝐶𝑛C\_{1}\sqcup C\_{2}\sqcup\ldots\sqcup C\_{n}italic\_C start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ⊔ italic\_C start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ⊔ … ⊔ italic\_C start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT | (\_:x owl:unionOf \_list(C1,C2,…,Cn)subscript𝐶1subscript𝐶2normal-…subscript𝐶𝑛\left(C\_{1},C\_{2},\ldots,C\_{n}\right)( italic\_C start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_C start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_C start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT )) |
| ¬C𝐶\neg C¬ italic\_C | (\_:x owl:complementOf C𝐶Citalic\_C) |
| {a1,a2,…,an}subscript𝑎1subscript𝑎2…subscript𝑎𝑛\left\{a\_{1},a\_{2},\ldots,a\_{n}\right\}{ italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT } | (\_:x owl:oneOf \_list(a1,a2,…,an)subscript𝑎1subscript𝑎2normal-…subscript𝑎𝑛\left(a\_{1},a\_{2},\ldots,a\_{n}\right)( italic\_a start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_a start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT )) |
| ∃R.Cformulae-sequence𝑅𝐶\exists R.C∃ italic\_R . italic\_C |
| |
| --- |
| (\_:x owl:someValuesFrom C𝐶Citalic\_C) (\_:x owl:onProperty
R𝑅Ritalic\_R) |
|
| ∀R.Cformulae-sequencefor-all𝑅𝐶\forall R.C∀ italic\_R . italic\_C |
| |
| --- |
| (\_:x owl:allValuesFrom C𝐶Citalic\_C) (\_:x owl:onProperty
R𝑅Ritalic\_R) |
|
| ∃R.{a}formulae-sequence𝑅𝑎\exists R.\left\{a\right\}∃ italic\_R . { italic\_a } |
| |
| --- |
| (\_:x owl:hasValue a𝑎aitalic\_a) (\_:x owl:onProperty R𝑅Ritalic\_R) |
|
| ≤nRabsent𝑛𝑅\leq nR≤ italic\_n italic\_R |
| |
| --- |
| (\_:x owl:minCardinality n𝑛nitalic\_n) (\_:x owl:onProperty
R𝑅Ritalic\_R) |
|
| ≥nRabsent𝑛𝑅\geq nR≥ italic\_n italic\_R |
| |
| --- |
| (\_:x owl:maxCardinality n𝑛nitalic\_n) (\_:x owl:onProperty
R𝑅Ritalic\_R) |
|
| ≤nR⊓≥nRabsentlimit-from𝑛𝑅square-intersection𝑛𝑅\leq nR\,\sqcap\geq nR≤ italic\_n italic\_R ⊓ ≥ italic\_n italic\_R |
| |
| --- |
| (\_:x owl:cardinality n𝑛nitalic\_n) (\_:x owl:onProperty R𝑅Ritalic\_R) |
|
| ≤nR.Cformulae-sequenceabsent𝑛𝑅𝐶\leq nR.C≤ italic\_n italic\_R . italic\_C |
| |
| --- |
| (\_:x owl:minQualifiedCardinality n𝑛nitalic\_n) |
| (\_:x owl:onClass C𝐶Citalic\_C) (\_:x owl:onProperty R𝑅Ritalic\_R) |
|
| ≥nR.Cformulae-sequenceabsent𝑛𝑅𝐶\geq nR.C≥ italic\_n italic\_R . italic\_C |
| |
| --- |
| (\_:x owl:maxQualifiedCardinality n𝑛nitalic\_n) |
| (\_:x owl:onClass C𝐶Citalic\_C) (\_:x owl:onProperty R𝑅Ritalic\_R) |
|
| ≤nR.C⊓≥nR.Cformulae-sequenceabsent𝑛𝑅limit-from𝐶square-intersection𝑛𝑅𝐶\leq nR.C\,\sqcap\geq nR.C≤ italic\_n italic\_R . italic\_C ⊓ ≥ italic\_n italic\_R . italic\_C |
| |
| --- |
| (\_:x owl:qualifiedCardinality n𝑛nitalic\_n) |
| (\_:x owl:onClass C𝐶Citalic\_C) (\_:x owl:onProperty R𝑅Ritalic\_R) |
|
| ∃R.Selfformulae-sequence𝑅𝑆𝑒𝑙𝑓\exists R.Self∃ italic\_R . italic\_S italic\_e italic\_l italic\_f |
| |
| --- |
| (\_:x owl:hasSelf *true*) (\_:x
owl:onProperty R𝑅Ritalic\_R) |
|
| C(a)𝐶𝑎C\left(a\right)italic\_C ( italic\_a ) | (a𝑎aitalic\_a rdf:type C𝐶Citalic\_C) |
| {ai}≡{aj}subscript𝑎𝑖subscript𝑎𝑗\left\{a\_{i}\right\}\equiv\left\{a\_{j}\right\}{ italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT } ≡ { italic\_a start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT } | (aisubscript𝑎𝑖a\_{i}italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT owl:sameAs ajsubscript𝑎𝑗a\_{j}italic\_a start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT) |
| {ai}⊑¬{aj}square-image-of-or-equalssubscript𝑎𝑖subscript𝑎𝑗\left\{a\_{i}\right\}\sqsubseteq\neg\left\{a\_{j}\right\}{ italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT } ⊑ ¬ { italic\_a start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT } | (aisubscript𝑎𝑖a\_{i}italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT owl:differentFrom ajsubscript𝑎𝑗a\_{j}italic\_a start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT) |
| U𝑈Uitalic\_U opp. ▽ | owl:topObjectProperty |
| N𝑁Nitalic\_N opp. △ | owl:bottomObjectProperty |
| R⊑Ssquare-image-of-or-equals𝑅𝑆R\sqsubseteq Sitalic\_R ⊑ italic\_S | (R𝑅Ritalic\_R owl:subPropertyOf S𝑆Sitalic\_S) |
| R≡S𝑅𝑆R\equiv Sitalic\_R ≡ italic\_S | (R𝑅Ritalic\_R owl:equivalentProperty S𝑆Sitalic\_S) |
| R⊓S⊑Nsquare-image-of-or-equalssquare-intersection𝑅𝑆𝑁R\sqcap S\sqsubseteq Nitalic\_R ⊓ italic\_S ⊑ italic\_N | (R𝑅Ritalic\_R owl:PropertyDisjointWith S𝑆Sitalic\_S) |
| R−superscript𝑅R^{-}italic\_R start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT | (\_:x owl:inverseOf R𝑅Ritalic\_R) |
| R1∘R2∘…∘Rn⊑Rsquare-image-of-or-equalssubscript𝑅1subscript𝑅2…subscript𝑅𝑛𝑅R\_{1}\circ R\_{2}\circ\ldots\circ R\_{n}\sqsubseteq Ritalic\_R start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ∘ italic\_R start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT ∘ … ∘ italic\_R start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ⊑ italic\_R | (R𝑅Ritalic\_R owl:propertyChainAxiom \_list(R1,R2,…,Rn)subscript𝑅1subscript𝑅2normal-…subscript𝑅𝑛\left(R\_{1},R\_{2},\ldots,R\_{n}\right)( italic\_R start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_R start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_R start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT )) |
| 𝑅𝑒𝑓𝑙(R)𝑅𝑒𝑓𝑙𝑅\mathit{Refl}\left(R\right)italic\_Refl ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:ReflexiveProperty) |
| 𝐼𝑟𝑟𝑒𝑓𝑙(R)𝐼𝑟𝑟𝑒𝑓𝑙𝑅\mathit{Irrefl}\left(R\right)italic\_Irrefl ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:IrreflexiveProperty) |
| 𝑆𝑦𝑚(R)𝑆𝑦𝑚𝑅\mathit{Sym}\left(R\right)italic\_Sym ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:SymmetricProperty) |
| 𝐴𝑠𝑦𝑚(R)𝐴𝑠𝑦𝑚𝑅\mathit{Asym}\left(R\right)italic\_Asym ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:AsymmetricProperty) |
| 𝑇𝑟𝑎𝑛𝑠(R)𝑇𝑟𝑎𝑛𝑠𝑅\mathit{Trans}\left(R\right)italic\_Trans ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:TransitiveProperty) |
| 𝐹𝑛(R)𝐹𝑛𝑅\mathit{Fn}\left(R\right)italic\_Fn ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:FunctionalProperty) |
| 𝐼𝑛𝑣𝐹𝑛(R)𝐼𝑛𝑣𝐹𝑛𝑅\mathit{InvFn}\left(R\right)italic\_InvFn ( italic\_R ) | (R𝑅Ritalic\_R rdf:type owl:InverseFunctionalProperty) |
| R(a,b)𝑅𝑎𝑏R\left(a,b\right)italic\_R ( italic\_a , italic\_b ) | (a𝑎aitalic\_a *property\_name* b𝑏bitalic\_b) |
| ¬R(a,b)𝑅𝑎𝑏\neg R\left(a,b\right)¬ italic\_R ( italic\_a , italic\_b ) |
| |
| --- |
| (\_:x rdf:type owl:NegativePropertyAssertion) |
| (\_:x owl:sourceIndividual a𝑎aitalic\_a) |
| (\_:x owl:assertionProperty R𝑅Ritalic\_R) |
| (\_:x owl:targetIndividual b𝑏bitalic\_b) |
|
*Table 1.3.* Correspondences between
DLs and SW
2 ANALYSIS AND RESULTS
-----------------------
###
2.1 The family 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>
Description
Logics derived from decidable fragments of set theory, generally denoted
by the notation
𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <language\_name>, are having
considerable importance.
The family of logics underlying this class of ontologies, object of
research this project is based on, is focused on the fragment ∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT ,
which has a good expressiveness w.r.t. knowledge representation in
real-world applications. The interest aroused by that is related to
NP-completeness of its decision procedure in some cases of practical
relevance. 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>-concepts and 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>-roles are formed
according to the syntax141414The symbol sym(R)sym𝑅\text{sym}\left(R\right)sym ( italic\_R ) denotes symmetric closure of
R𝑅Ritalic\_R.
| | | | | |
| --- | --- | --- | --- | --- |
| | C,D𝐶𝐷\displaystyle C,Ditalic\_C , italic\_D | →→\displaystyle\rightarrow→ | A|⊤|⊥|¬C|C⊓D|C⊔D|{a}|∃R.Self|∃R.{a}\displaystyle A\ |\ \top\ |\ \bot\ |\ \neg C\ |\ C\sqcap D\ |\ C\sqcup D\ |\ \left\{a\right\}\ |\ \exists R.Self\ |\ \exists R.\left\{a\right\}italic\_A | ⊤ | ⊥ | ¬ italic\_C | italic\_C ⊓ italic\_D | italic\_C ⊔ italic\_D | { italic\_a } | ∃ italic\_R . italic\_S italic\_e italic\_l italic\_f | ∃ italic\_R . { italic\_a } | |
| | R,S𝑅𝑆\displaystyle R,Sitalic\_R , italic\_S | →→\displaystyle\rightarrow→ | P|U|R−|¬R|R⊓S|R⊔S|RC||R|D|RC|D|id(C)|sym(R),\displaystyle P\ |\ U\ |\ R^{-}\ |\ \neg R\ |\ R\sqcap S\ |\ R\sqcup S\ |\ R\_{C|}\ |\ R\_{|D}\ |\ R\_{C|D}\ |\ id\left(C\right)\ |\ \text{{\text{sym}}}\left(R\right)\,,italic\_P | italic\_U | italic\_R start\_POSTSUPERSCRIPT - end\_POSTSUPERSCRIPT | ¬ italic\_R | italic\_R ⊓ italic\_S | italic\_R ⊔ italic\_S | italic\_R start\_POSTSUBSCRIPT italic\_C | end\_POSTSUBSCRIPT | italic\_R start\_POSTSUBSCRIPT | italic\_D end\_POSTSUBSCRIPT | italic\_R start\_POSTSUBSCRIPT italic\_C | italic\_D end\_POSTSUBSCRIPT | italic\_i italic\_d ( italic\_C ) | sym ( italic\_R ) , | |
while a 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>-KB is made of axioms and assertions of the
following kind:
| | | |
| --- | --- | --- |
| | C≡DC⊑DC⊑∀R.D∃R.C⊑DC(a)R(a,b)R≡SR⊑SR∘R′⊑STrans(R)Refl(R)Asym(R)Fn(R)InvFn(R).formulae-sequenceformulae-sequence𝐶𝐷formulae-sequencesquare-image-of-or-equals𝐶𝐷square-image-of-or-equals𝐶for-all𝑅𝐷𝑅
square-image-of-or-equals𝐶𝐷𝐶𝑎𝑅𝑎𝑏formulae-sequence𝑅𝑆formulae-sequencesquare-image-of-or-equals𝑅𝑆square-image-of-or-equals𝑅superscript𝑅′𝑆𝑇𝑟𝑎𝑛𝑠𝑅𝑅𝑒𝑓𝑙𝑅𝐴𝑠𝑦𝑚𝑅𝐹𝑛𝑅𝐼𝑛𝑣𝐹𝑛𝑅\begin{array}[]{c}C\equiv D\ \,\ \,\ C\sqsubseteq D\ \,\ \,\ C\sqsubseteq\forall R.D\ \,\ \,\ \exists R.C\sqsubseteq D\ \,\ \,\ C\left(a\right)\ \,\ \,\ R\left(a,b\right)\\
R\equiv S\ \,\ \,\ R\sqsubseteq S\ \,\ \,\ R\circ R^{\prime}\sqsubseteq S\ \,\ \,\ Trans\left(R\right)\ \,\ \,\ Refl\left(R\right)\ \,\ \,\ Asym\left(R\right)\ \,\ \,\ Fn\left(R\right)\ \,\ \,\ InvFn\left(R\right)\,.\end{array}start\_ARRAY start\_ROW start\_CELL italic\_C ≡ italic\_D italic\_C ⊑ italic\_D italic\_C ⊑ ∀ italic\_R . italic\_D ∃ italic\_R . italic\_C ⊑ italic\_D italic\_C ( italic\_a ) italic\_R ( italic\_a , italic\_b ) end\_CELL end\_ROW start\_ROW start\_CELL italic\_R ≡ italic\_S italic\_R ⊑ italic\_S italic\_R ∘ italic\_R start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ⊑ italic\_S italic\_T italic\_r italic\_a italic\_n italic\_s ( italic\_R ) italic\_R italic\_e italic\_f italic\_l ( italic\_R ) italic\_A italic\_s italic\_y italic\_m ( italic\_R ) italic\_F italic\_n ( italic\_R ) italic\_I italic\_n italic\_v italic\_F italic\_n ( italic\_R ) . end\_CELL end\_ROW end\_ARRAY | |
We may notice that universal restriction is allowed only in the right
part of a subsumption, whereas existential restriction can appear
only in the left part. In addition, neither numerical, nor qualified
restrictions are allowed.
###
2.2 Description of the experiment
Our analysis is aimed at selecting ontologies corresponding
to the 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT> family, in order to use them as a base of study
for this family of logics in real-world applications. To do that,
it is necessary to query against as many as possible publicly available
ontologies in a quick and efficient way. The use of queries in SPARQL
conveniently lends itself to the purpose. The management is provided
by dOWLphin, a Java library specifically created in order
to easily load the ontologies and prepare the queries. The underlying
library is Jena, one of the most common collections of API
for the Semantic Web, that is very handy because it can directly deal
with triples. As a *front-end* GUI, the program QuAny
was used, which was born inside the experiment too and allows one
to query against local and remote ontologies, and save on disk the
corresponding results, for future verification.
###
2.3 Results and future objectives
The query was employed to test a significant number
of ontologies of BioPortal, the web portal of the National
Center for Biomedical Ontology. This choice was not random, but motivated
mainly by two reasons: the first concerns the large amount of ontologies
present on the portal, coming from repositories of biomedical resources
spread all over the world; the second is relative to the connection
that these ontologies have with the real world in general, and with
human life and medicine in particular, which are fields offering several
matters for reflection on how widely knowledge of so important themes
may be schematized and represented, and, most of all, on the role
*reasoning* may have in automatically inferring new information.
Around 30% of the ontologies resulted member of the language 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>,
which brings good hopes for reasoning on them, given that−we
remind−w.r.t. computational complexity we are in the NP-completeness
realm. Concerning the remaining 70%, efficient algorithms for conversion
will have to be considered and semantic tests will have to be done,
as previously happened for the 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <𝖬𝖫𝖲𝖲2,m×superscriptsubscript𝖬𝖫𝖲𝖲2𝑚\mathsf{MLSS}\_{2,m}^{\times}sansserif\_MLSS start\_POSTSUBSCRIPT 2 , italic\_m end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT × end\_POSTSUPERSCRIPT > language (cfr. Cantone *et al.*, [[3](#bib.bib3)]).
In Appendix [B](#A2 "Appendix B TEST AGAINST BIOPORTAL ONTOLOGIES ‣ An Experiment on the Connection between the Description Logics’ Family 𝒟ℒ <∀₀^𝜋> and the Real World"), a table shows which ontologies were
recognized as members of the language 𝒟ℒ𝒟ℒ\mathcal{DL}caligraphic\_D caligraphic\_L <∀0πsuperscriptsubscriptfor-all0𝜋\mathcal{\forall}\_{0}^{\pi}∀ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT>.
>
> |
4a4130ed-9ce5-4b3f-89d0-71e4af0ba039 | trentmkelly/LessWrong-43k | LessWrong | No Evolutions for Corporations or Nanodevices
> "The laws of physics and the rules of math don't cease to apply. That leads me to believe that evolution doesn't stop. That further leads me to believe that nature —bloody in tooth and claw, as some have termed it —will simply be taken to the next level...
> "[Getting rid of Darwinian evolution is] like trying to get rid of gravitation. So long as there are limited resources and multiple competing actors capable of passing on characteristics, you have selection pressure."
> —Perry Metzger, predicting that the reign of natural selection would continue into the indefinite future.
In evolutionary biology, as in many other fields, it is important to think quantitatively rather than qualitatively. Does a beneficial mutation "sometimes spread, but not always"? Well, a psychic power would be a beneficial mutation, so you'd expect it to spread, right? Yet this is qualitative reasoning, not quantitative—if X is true, then Y is true; if psychic powers are beneficial, they may spread. In Evolutions Are Stupid, I described the equations for a beneficial mutation's probability of fixation, roughly twice the fitness advantage (6% for a 3% advantage). Only this kind of numerical thinking is likely to make us realize that mutations which are only rarely useful are extremely unlikely to spread, and that it is practically impossible for complex adaptations to arise without constant use. If psychic powers really existed, we should expect to see everyone using them all the time—not just because they would be so amazingly useful, but because otherwise they couldn't have evolved in the first place.
"So long as there are limited resources and multiple competing actors capable of passing on characteristics, you have selection pressure." This is qualitative reasoning. How much selection pressure?
While there are several candidates for the most important equation in evolutionary biology, I would pick Price's Equation, which in its simplest formulation reads:
Δz=cov(vi,zi)
change i |
fb6331f8-aa26-44e6-8720-1476cc8771ac | StampyAI/alignment-research-dataset/arxiv | Arxiv | Improving Generalization in Meta Reinforcement Learning using Learned Objectives
1 Introduction
---------------
The process of evolution has equipped humans with incredibly general learning algorithms.
They enable us to solve a wide range of problems, even in the absence of a large number of related prior experiences.
The algorithms that give rise to these capabilities are the result of *distilling* the collective experiences of many learners throughout the course of natural evolution.
By essentially *learning from learning experiences* in this way, the resulting knowledge can be compactly encoded in the genetic code of an individual to give rise to the general learning capabilities that we observe today.
In contrast, Reinforcement Learning (RL) in artificial agents rarely proceeds in this way.
The learning rules that are used to train agents are the result of years of human engineering and design, (e.g. Williams ([1992](#bib.bib59)); Wierstra et al. ([2008](#bib.bib57)); Mnih et al. ([2013](#bib.bib30)); Lillicrap et al. ([2016](#bib.bib27)); Schulman et al. ([2015a](#bib.bib47))).
Correspondingly, artificial agents are inherently limited by the ability of the designer to incorporate the right inductive biases in order to learn from previous experiences.
Several works have proposed an alternative framework based on *meta reinforcement learning* (Schmidhuber, [1994](#bib.bib41); Wang et al., [2016](#bib.bib54); Duan et al., [2016](#bib.bib9); Finn et al., [2017](#bib.bib11); Houthooft et al., [2018](#bib.bib19); Clune, [2019](#bib.bib7)).
Meta-RL distinguishes between learning to act in the environment (the reinforcement learning problem) and learning to learn (the meta-learning problem).
Hence, learning itself is now a learning problem, which in principle allows one to leverage prior learning experiences to meta-learn *general* learning rules that surpass human-engineered alternatives.
However, while prior work found that learning rules could be meta-learned that generalize to slightly different environments or goals (Finn et al., [2017](#bib.bib11); Plappert et al., [2018](#bib.bib35); Houthooft et al., [2018](#bib.bib19)), generalization to *entirely different* environments remains an open problem.
In this paper we present *MetaGenRL*111Code is available at <http://louiskirsch.com/code/metagenrl>, a novel meta reinforcement learning algorithm that meta-learns learning rules that generalize to entirely different environments.
MetaGenRL is inspired by the process of natural evolution as it distills the experiences of many agents into the parameters of an objective function that decides how future individuals will learn.
Similar to Evolved Policy Gradients (EPG; Houthooft et al. ([2018](#bib.bib19))), it meta-learns low complexity neural objective functions that can be used to train complex agents with many parameters.
However, unlike EPG, it is able to meta-learn using second-order gradients, which offers several advantages as we will demonstrate.
We evaluate MetaGenRL on a variety of continuous control tasks and compare to RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT (Wang et al., [2016](#bib.bib54); Duan et al., [2016](#bib.bib9)) and EPG in addition to several human engineered learning algorithms.
Compared to RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT we find that MetaGenRL does not overfit and is able to train randomly initialized agents using meta-learned learning rules on *entirely different* environments.
Compared to EPG we find that MetaGenRL is more sample efficient, and outperforms significantly under a fixed budget of environment interactions.
The results of an ablation study and additional analysis provide further insight into the benefits of our approach.
2 Preliminaries
----------------
##### Notation
We consider the standard MDP Reinforcement Learning setting defined by a tuple e=(S,A,P,ρ0,r,γ,T)𝑒𝑆𝐴𝑃subscript𝜌0𝑟𝛾𝑇e=(S,A,P,\rho\_{0},r,\gamma,T)italic\_e = ( italic\_S , italic\_A , italic\_P , italic\_ρ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_r , italic\_γ , italic\_T ) consisting of
states S𝑆Sitalic\_S, actions A𝐴Aitalic\_A, the transition probability distribution P:S×A×S→ℝ+:𝑃→𝑆𝐴𝑆subscriptℝP:S\times A\times S\to\mathbb{R}\_{+}italic\_P : italic\_S × italic\_A × italic\_S → blackboard\_R start\_POSTSUBSCRIPT + end\_POSTSUBSCRIPT, an initial state distribution ρ0:S→ℝ+:subscript𝜌0→𝑆subscriptℝ\rho\_{0}:S\to\mathbb{R}\_{+}italic\_ρ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT : italic\_S → blackboard\_R start\_POSTSUBSCRIPT + end\_POSTSUBSCRIPT, the reward function r:S×A→[−Rmax,Rmax]:𝑟→𝑆𝐴subscript𝑅𝑚𝑎𝑥subscript𝑅𝑚𝑎𝑥r:S\times A\to[-R\_{max},R\_{max}]italic\_r : italic\_S × italic\_A → [ - italic\_R start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT , italic\_R start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT ], a discount factor γ𝛾\gammaitalic\_γ, and the episode length T𝑇Titalic\_T.
The objective for the probabilistic policy πϕ:S×A→ℝ+:subscript𝜋italic-ϕ→𝑆𝐴subscriptℝ\pi\_{\phi}:S\times A\to\mathbb{R}\_{+}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT : italic\_S × italic\_A → blackboard\_R start\_POSTSUBSCRIPT + end\_POSTSUBSCRIPT parameterized by ϕitalic-ϕ\phiitalic\_ϕ is to maximize the expected discounted return:
| | | | |
| --- | --- | --- | --- |
| | 𝔼τ[∑t=0T−1γtrt],wheres0∼ρ0(s0),at∼πϕ(at|st),st+1∼P(st+1|st,at),rt=r(st,at),formulae-sequencesimilar-tosubscript𝔼𝜏delimited-[]superscriptsubscript𝑡0𝑇1superscript𝛾𝑡subscript𝑟𝑡wheresubscript𝑠0
subscript𝜌0subscript𝑠0formulae-sequencesimilar-tosubscript𝑎𝑡subscript𝜋italic-ϕconditionalsubscript𝑎𝑡subscript𝑠𝑡formulae-sequencesimilar-tosubscript𝑠𝑡1𝑃conditionalsubscript𝑠𝑡1subscript𝑠𝑡subscript𝑎𝑡subscript𝑟𝑡𝑟subscript𝑠𝑡subscript𝑎𝑡\mathbb{E}\_{\tau}[\sum\_{t=0}^{T-1}\gamma^{t}r\_{t}],\ \text{where}\ s\_{0}\sim\rho\_{0}(s\_{0}),\ a\_{t}\sim\pi\_{\phi}(a\_{t}|s\_{t}),\ s\_{t+1}\sim P(s\_{t+1}|s\_{t},a\_{t}),r\_{t}=r(s\_{t},a\_{t}),blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ ∑ start\_POSTSUBSCRIPT italic\_t = 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_T - 1 end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_t end\_POSTSUPERSCRIPT italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ] , where italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ∼ italic\_ρ start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∼ italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) , italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT ∼ italic\_P ( italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) , italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_r ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) , | | (1) |
with τ=(s0,a0,r0,s1,…,sT−1,aT−1,rT−1)𝜏subscript𝑠0subscript𝑎0subscript𝑟0subscript𝑠1…subscript𝑠𝑇1subscript𝑎𝑇1subscript𝑟𝑇1\tau=(s\_{0},a\_{0},r\_{0},s\_{1},...,s\_{T-1},a\_{T-1},r\_{T-1})italic\_τ = ( italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_s start\_POSTSUBSCRIPT italic\_T - 1 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T - 1 end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT italic\_T - 1 end\_POSTSUBSCRIPT ).
##### Human Engineered Gradient Estimators
A popular gradient-based approach to maximizing [Equation 1](#S2.E1 "1 ‣ Notation ‣ 2 Preliminaries ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") is REINFORCE (Williams, [1992](#bib.bib59)).
It directly differentiates [Equation 1](#S2.E1 "1 ‣ Notation ‣ 2 Preliminaries ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") with respect to ϕitalic-ϕ\phiitalic\_ϕ using the likelihood ratio trick to derive gradient estimates of the form:
| | | | |
| --- | --- | --- | --- |
| | ∇ϕ𝔼τ[LREINF(τ,πϕ)]:=𝔼τ[∇ϕ∑t=0T−1logπϕ(at|st)⋅∑t′=tT−1γt′−trt′)].\nabla\_{\phi}\mathbb{E}\_{\tau}[L\_{REINF}(\tau,\pi\_{\phi})]:=\mathbb{E}\_{\tau}[\nabla\_{\phi}\sum\_{t=0}^{T-1}\log\pi\_{\phi}(a\_{t}|s\_{t})\cdot\sum\_{t^{\prime}=t}^{T-1}\gamma^{t^{\prime}-t}r\_{t^{\prime}})].∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ italic\_L start\_POSTSUBSCRIPT italic\_R italic\_E italic\_I italic\_N italic\_F end\_POSTSUBSCRIPT ( italic\_τ , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ) ] := blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ ∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_t = 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_T - 1 end\_POSTSUPERSCRIPT roman\_log italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) ⋅ ∑ start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_T - 1 end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT - italic\_t end\_POSTSUPERSCRIPT italic\_r start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ) ] . | | (2) |
Although this basic estimator is rarely used in practice, it has become a building block for an entire class of policy-gradient algorithms of this form.
For example, a popular extension from Schulman et al. ([2015b](#bib.bib48)) combines REINFORCE with a Generalized Advantage Estimate (GAE) to yield the following policy gradient estimator:
| | | | |
| --- | --- | --- | --- |
| | ∇ϕ𝔼τ[LGAE(τ,πϕ,V)]:=𝔼τ[∇ϕ∑t=0T−1logπϕ(at|st)⋅A(τ,V,t)].assignsubscript∇italic-ϕsubscript𝔼𝜏delimited-[]subscript𝐿𝐺𝐴𝐸𝜏subscript𝜋italic-ϕ𝑉subscript𝔼𝜏delimited-[]subscript∇italic-ϕsuperscriptsubscript𝑡0𝑇1⋅subscript𝜋italic-ϕconditionalsubscript𝑎𝑡subscript𝑠𝑡𝐴𝜏𝑉𝑡\nabla\_{\phi}\mathbb{E}\_{\tau}[L\_{GAE}(\tau,\pi\_{\phi},V)]:=\mathbb{E}\_{\tau}[\nabla\_{\phi}\sum\_{t=0}^{T-1}\log\pi\_{\phi}(a\_{t}|s\_{t})\cdot A(\tau,V,t)].∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ italic\_L start\_POSTSUBSCRIPT italic\_G italic\_A italic\_E end\_POSTSUBSCRIPT ( italic\_τ , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT , italic\_V ) ] := blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ ∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_t = 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_T - 1 end\_POSTSUPERSCRIPT roman\_log italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) ⋅ italic\_A ( italic\_τ , italic\_V , italic\_t ) ] . | | (3) |
where A(τ,V,t)𝐴𝜏𝑉𝑡A(\tau,V,t)italic\_A ( italic\_τ , italic\_V , italic\_t ) is the GAE and V:S→ℝ:𝑉→𝑆ℝV:S\to\mathbb{R}italic\_V : italic\_S → blackboard\_R is a value function estimate.
Several recent other extensions include TRPO (Schulman et al., [2015a](#bib.bib47)), which discourages bad policy updates using trust regions and iterative off-policy updates, or PPO (Schulman et al., [2017](#bib.bib49)), which offers similar benefits using only first order approximations.
##### Parametrized Objective Functions
In this work we note that many of these human engineered policy gradient estimators can be viewed as specific implementations of a general objective function L𝐿Litalic\_L that is differentiated with respect to the policy parameters:
| | | | |
| --- | --- | --- | --- |
| | ∇ϕ𝔼τ[L(τ,πϕ,V)].subscript∇italic-ϕsubscript𝔼𝜏delimited-[]𝐿𝜏subscript𝜋italic-ϕ𝑉\nabla\_{\phi}\mathbb{E}\_{\tau}[L(\tau,\pi\_{\phi},V)].∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT blackboard\_E start\_POSTSUBSCRIPT italic\_τ end\_POSTSUBSCRIPT [ italic\_L ( italic\_τ , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT , italic\_V ) ] . | | (4) |
Hence, it becomes natural to consider a generic parametrization of L𝐿Litalic\_L that, for various choices of parameters α𝛼\alphaitalic\_α, recovers some of these estimators.
In this paper, we will consider *neural objective functions* where Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT is implemented by a neural network.
Our goal is then to optimize the parameters α𝛼\alphaitalic\_α of this neural network in order to give rise to a new learning algorithm that best maximizes [Equation 1](#S2.E1 "1 ‣ Notation ‣ 2 Preliminaries ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") on an entire class of (different) environments.
3 Meta-Learning Neural Objectives
----------------------------------
Figure 1: A schematic of *MetaGenRL*.
On the left a population of agents (i∈1,…,N𝑖1…𝑁i\in 1,\ldots,Nitalic\_i ∈ 1 , … , italic\_N), where each member consist of a critic Qθ(i)subscriptsuperscript𝑄𝑖𝜃Q^{(i)}\_{\theta}italic\_Q start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT and a policy πϕ(i)subscriptsuperscript𝜋𝑖italic-ϕ\pi^{(i)}\_{\phi}italic\_π start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT that interact with a particular environment e(i)superscript𝑒𝑖e^{(i)}italic\_e start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT and store collected data in a corresponding replay buffer B(i)superscript𝐵𝑖B^{(i)}italic\_B start\_POSTSUPERSCRIPT ( italic\_i ) end\_POSTSUPERSCRIPT.
On the right a meta-learned neural objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT that is shared across the population.
Learning (dotted arrows) proceeds as follows: Each policy is updated by differentiating Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT, while the critic is updated using the usual TD-error (not shown).
Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT is meta-learned by computing second-order gradients that can be obtained by differentiating through the critic.
In this work we propose *MetaGenRL*, a novel meta reinforcement learning algorithm that meta-learns neural objective functions of the form Lα(τ,πϕ,V)subscript𝐿𝛼𝜏subscript𝜋italic-ϕ𝑉L\_{\alpha}(\tau,\pi\_{\phi},V)italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_τ , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT , italic\_V ).
MetaGenRL makes use of value functions and second-order gradients, which makes it more sample efficient compared to prior work (Duan et al., [2016](#bib.bib9); Wang et al., [2016](#bib.bib54); Houthooft et al., [2018](#bib.bib19)).
More so, as we will demonstrate, MetaGenRL meta-learns objective functions that generalize to vastly different environments.
Our key insight is that a differentiable critic Qθ:S×A→ℝ:subscript𝑄𝜃→𝑆𝐴ℝQ\_{\theta}:S\times A\to\mathbb{R}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT : italic\_S × italic\_A → blackboard\_R can be used to measure the effect of locally changing the objective function parameters α𝛼\alphaitalic\_α based on the quality of the corresponding policy gradients.
This enables a population of agents to use and improve a single parameterized objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT through interacting with a set of (potentially different) environments.
During evaluation (meta-test time), the meta-learned objective function can then be used to train a randomly initialized RL agent in a new environment.
###
3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives
We will formally introduce MetaGenRL as an extension of the DDPG actor-critic framework (Silver et al., [2014](#bib.bib50); Lillicrap et al., [2016](#bib.bib27)).
In DDPG, a parameterized critic of the form Qθ:S×A→ℝ:subscript𝑄𝜃→𝑆𝐴ℝQ\_{\theta}:S\times A\to\mathbb{R}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT : italic\_S × italic\_A → blackboard\_R transforms the non-differentiable RL reward maximization problem into a myopic value maximization problem for any st∈Ssubscript𝑠𝑡𝑆s\_{t}\in Sitalic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∈ italic\_S.
This is done by alternating between optimization of the critic Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT and the (here deterministic) policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT.
The critic is trained to minimize the TD-error by following:
| | | | |
| --- | --- | --- | --- |
| | ∇θ∑(st,at,rt,st+1)(Qθ(st,at)−yt)2,whereyt=rt+γ⋅Qθ(st+1,πϕ(st+1)),subscript∇𝜃subscriptsubscript𝑠𝑡subscript𝑎𝑡subscript𝑟𝑡subscript𝑠𝑡1superscriptsubscript𝑄𝜃subscript𝑠𝑡subscript𝑎𝑡subscript𝑦𝑡2wheresubscript𝑦𝑡
subscript𝑟𝑡⋅𝛾subscript𝑄𝜃subscript𝑠𝑡1subscript𝜋italic-ϕsubscript𝑠𝑡1\nabla\_{\theta}\sum\_{(s\_{t},a\_{t},r\_{t},s\_{t+1})}(Q\_{\theta}(s\_{t},a\_{t})-y\_{t})^{2},\ \text{where}\ y\_{t}=r\_{t}+\gamma\cdot Q\_{\theta}(s\_{t+1},\pi\_{\phi}(s\_{t+1})),∇ start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT ) end\_POSTSUBSCRIPT ( italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) - italic\_y start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT , where italic\_y start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT + italic\_γ ⋅ italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT ) ) , | | (5) |
and the dependence of ytsubscript𝑦𝑡y\_{t}italic\_y start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT on the parameter vector θ𝜃\thetaitalic\_θ is ignored.
The policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT is improved to increase the expected return from arbitrary states by following the gradient ∇ϕ∑stQθ(st,πϕ(st))subscript∇italic-ϕsubscriptsubscript𝑠𝑡subscript𝑄𝜃subscript𝑠𝑡subscript𝜋italic-ϕsubscript𝑠𝑡\nabla\_{\phi}\sum\_{s\_{t}}Q\_{\theta}(s\_{t},\pi\_{\phi}(s\_{t}))∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) ).
Both gradients can be computed entirely off-policy by sampling trajectories from a replay buffer.
MetaGenRL builds on this idea of differentiating the critic Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT with respect to the policy parameters.
It incorporates a parameterized objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT that is used to improve the policy (i.e. by following the gradient ∇ϕLαsubscript∇italic-ϕsubscript𝐿𝛼\nabla\_{\phi}L\_{\alpha}∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT), which adds one extra level of indirection:
The critic Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT improves Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT, while Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT improves the policy πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT.
By first differentiating with respect to the objective function parameters α𝛼\alphaitalic\_α, and then with respect to the policy parameters ϕitalic-ϕ\phiitalic\_ϕ, the critic can be used to measure the effect of updating πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT using Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT on the estimated return222In case of a probabilistic policy πϕ(at|st)subscript𝜋italic-ϕconditionalsubscript𝑎𝑡subscript𝑠𝑡\pi\_{\phi}(a\_{t}|s\_{t})italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) the following becomes an expectation under πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT and a reparameterizable form is required (Williams, [1988](#bib.bib58); Kingma & Welling, [2014](#bib.bib22); Rezende et al., [2014](#bib.bib36)). Here we focus on learning deterministic target policies.:
| | | | |
| --- | --- | --- | --- |
| | ∇αQθ(st,πϕ′(st)),whereϕ′=ϕ−∇ϕLα(τ,x(ϕ),V).subscript∇𝛼subscript𝑄𝜃subscript𝑠𝑡subscript𝜋superscriptitalic-ϕ′subscript𝑠𝑡wheresuperscriptitalic-ϕ′
italic-ϕsubscript∇italic-ϕsubscript𝐿𝛼𝜏𝑥italic-ϕ𝑉\nabla\_{\alpha}Q\_{\theta}(s\_{t},\pi\_{\phi^{\prime}}(s\_{t})),\ \text{where}\ \phi^{\prime}=\phi-\nabla\_{\phi}L\_{\alpha}(\tau,x(\phi),V).∇ start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) ) , where italic\_ϕ start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_ϕ - ∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_τ , italic\_x ( italic\_ϕ ) , italic\_V ) . | | (6) |
This constitutes a type of second order gradient ∇α∇ϕsubscript∇𝛼subscript∇italic-ϕ\nabla\_{\alpha}\nabla\_{\phi}∇ start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT that can be used to meta-train Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT to provide better updates to the policy parameters in the future.
In practice we will use batching to optimize [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") over multiple trajectories τ𝜏\tauitalic\_τ.
p(e)𝑝𝑒p(e)italic\_p ( italic\_e ) a distribution of environments
P⇐{(e1∼p(e),ϕ1,θ1,B1←∅),…}⇐𝑃formulae-sequencesimilar-tosubscript𝑒1𝑝𝑒subscriptitalic-ϕ1subscript𝜃1←subscript𝐵1…P\Leftarrow\{(e\_{1}\sim p(e),\phi\_{1},\theta\_{1},B\_{1}\leftarrow\varnothing),\ldots\}italic\_P ⇐ { ( italic\_e start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ∼ italic\_p ( italic\_e ) , italic\_ϕ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_θ start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_B start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT ← ∅ ) , … } ▷▷\triangleright▷ Randomly initialize population of agents
Randomly initialize objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT
while Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT has not converged do
for e,ϕ,θ,B∈P𝑒italic-ϕ𝜃𝐵
𝑃e,\phi,\theta,B\in Pitalic\_e , italic\_ϕ , italic\_θ , italic\_B ∈ italic\_P do ▷▷\triangleright▷ For each agent i𝑖iitalic\_i in parallel
if extend replay buffer B𝐵Bitalic\_B then
Extend B𝐵Bitalic\_B using πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT in e𝑒eitalic\_e
Sample trajectories from B𝐵Bitalic\_B
Update critic Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT using TD-error
Update policy by following ∇ϕLαsubscript∇italic-ϕsubscript𝐿𝛼\nabla\_{\phi}L\_{\alpha}∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT
Compute objective function gradient ΔisubscriptΔ𝑖\Delta\_{i}roman\_Δ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT for agent i𝑖iitalic\_i according to [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")
Sum gradients ∑iΔisubscript𝑖subscriptΔ𝑖\sum\_{i}\Delta\_{i}∑ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT roman\_Δ start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT to update Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT
Algorithm 1 MetaGenRL: Meta-Training
Similarly to the policy-gradient estimators from [Section 2](#S2 "2 Preliminaries ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives"), the objective function Lα(τ,x(ϕ),V)subscript𝐿𝛼𝜏𝑥italic-ϕ𝑉L\_{\alpha}(\tau,x(\phi),V)italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_τ , italic\_x ( italic\_ϕ ) , italic\_V ) receives as inputs an episode trajectory τ=(s0:T−1,a0:T−1,r0:T−1)𝜏subscript𝑠:0𝑇1subscript𝑎:0𝑇1subscript𝑟:0𝑇1\tau=(s\_{0:T-1},a\_{0:T-1},r\_{0:T-1})italic\_τ = ( italic\_s start\_POSTSUBSCRIPT 0 : italic\_T - 1 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT 0 : italic\_T - 1 end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT 0 : italic\_T - 1 end\_POSTSUBSCRIPT ), the value function estimates V𝑉Vitalic\_V, and an auxiliary input x(ϕ)𝑥italic-ϕx(\phi)italic\_x ( italic\_ϕ ) (previously πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT) that can be differentiated with respect to the policy parameters.
The latter is critical to be able to differentiate with respect to ϕitalic-ϕ\phiitalic\_ϕ and in the simplest case it consists of the action as predicted by the policy.
While [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") is used for meta-learning Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT, the objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT itself is used for policy learning by following ∇ϕLα(τ,x(ϕ),V)subscript∇italic-ϕsubscript𝐿𝛼𝜏𝑥italic-ϕ𝑉\nabla\_{\phi}L\_{\alpha}(\tau,x(\phi),V)∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_τ , italic\_x ( italic\_ϕ ) , italic\_V ).
See [Figure 1](#S3.F1 "Figure 1 ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") for an overview.
MetaGenRL consists of two phases:
During meta-training, we alternate between critic updates, objective function updates, and policy updates to meta-learn an objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT as described in [Algorithm 1](#alg1 "Algorithm 1 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
During meta-testing in [Algorithm 2](#alg2 "Algorithm 2 ‣ 3.2 Parametrizing the Objective Function ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives"), we take the learned objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and keep it fixed while training a randomly initialized policy in a new environment to assess its performance.
We note that the inputs to Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT are sampled from a replay buffer rather than solely using on-policy data.
If Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT were to represent a REINFORCE-type objective then it would mean that differentiating Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT yields *biased* policy gradient estimates.
In our experiments we will find that the gradients from Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT work much better in comparison to a biased off-policy REINFORCE algorithm, and to an importance-sampled unbiased REINFORCE algorithm, while also improving over the popular on-policy REINFORCE and PPO algorithms.
###
3.2 Parametrizing the Objective Function
We will implement Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT using an LSTM (Gers et al., [2000](#bib.bib13); Hochreiter & Schmidhuber, [1997](#bib.bib16)) that iterates over τ𝜏\tauitalic\_τ in reverse order and depends on the current policy action πϕ(st)subscript𝜋italic-ϕsubscript𝑠𝑡\pi\_{\phi}(s\_{t})italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) (see [Figure 2](#S3.F2 "Figure 2 ‣ 3.2 Parametrizing the Objective Function ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")).
At every time-step Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT receives the reward rtsubscript𝑟𝑡r\_{t}italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT, taken action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT, predicted action by the current policy πϕ(st)subscript𝜋italic-ϕsubscript𝑠𝑡\pi\_{\phi}(s\_{t})italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ), the time t𝑡titalic\_t, and value function estimates Vt,Vt+1subscript𝑉𝑡subscript𝑉𝑡1V\_{t},V\_{t+1}italic\_V start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_V start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT333The value estimates are derived from the Q-function, i.e. Vt=Qθ(st,πϕ(st))subscript𝑉𝑡subscript𝑄𝜃subscript𝑠𝑡subscript𝜋italic-ϕsubscript𝑠𝑡V\_{t}=Q\_{\theta}(s\_{t},\pi\_{\phi}(s\_{t}))italic\_V start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) ), and are treated as a constant input.
Hence, the gradient ∇ϕLαsubscript∇italic-ϕsubscript𝐿𝛼\nabla\_{\phi}L\_{\alpha}∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT can not flow backwards through Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT, which ensures that Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT can not naively learn to implement a DDPG-like objective function. .
At each step the LSTM outputs the objective value ltsubscript𝑙𝑡l\_{t}italic\_l start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT, all of which are summed to yield a single scalar output value that can be differentiated with respect to ϕitalic-ϕ\phiitalic\_ϕ.
In order to accommodate varying action dimensionalities across different environments, both πϕ(st)subscript𝜋italic-ϕsubscript𝑠𝑡\pi\_{\phi}(s\_{t})italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) and atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT are first convolved and then averaged to obtain an action embedding that does not depend on the action dimensionality.
Additional details, including suggestions for more expressive alternatives are available in [Appendix B](#A2 "Appendix B Experiment Details ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
By presenting the trajectory in reverse order to the LSTM (and Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT correspondingly), it is able to assign credit to an action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT based on its *future* impact on the reward, similar to policy gradient estimators.
More so, as a general function approximator using these inputs, the LSTM is in principle able to learn different variance and bias reduction techniques, akin to advantage estimates, generalized advantage estimates, or importance weights444We note that in practice it is is difficult to assess whether the meta-learned object function incorporates bias / variance reduction techniques, especially because MetaGenRL is unlikely to recover *known* techniques..
Due to these properties, we expect the class of objective functions that is supported to somewhat relate to a REINFORCE (Williams, [1992](#bib.bib59)) estimator that uses generalized advantage estimation (Schulman et al., [2015b](#bib.bib48)).
A test environment e𝑒eitalic\_e, and an
objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT
Randomly initialize πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT, Vθsubscript𝑉𝜃V\_{\theta}italic\_V start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT, B←∅←𝐵B\leftarrow\varnothingitalic\_B ← ∅
while πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT has not converged do
if extend replay buffer B𝐵Bitalic\_B then
Extend B𝐵Bitalic\_B using πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT in e𝑒eitalic\_e
Sample trajectories from B𝐵Bitalic\_B
Update Vθsubscript𝑉𝜃V\_{\theta}italic\_V start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT using TD-error
Update policy by following ∇ϕLαsubscript∇italic-ϕsubscript𝐿𝛼\nabla\_{\phi}L\_{\alpha}∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT
Algorithm 2 MetaGenRL: Meta-Testing
![[Uncaptioned image]](/html/1910.04098/assets/x2.png)
Figure 2: An overview of Lα(τ,x(ϕ),V)subscript𝐿𝛼𝜏𝑥italic-ϕ𝑉L\_{\alpha}(\tau,x(\phi),V)italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_τ , italic\_x ( italic\_ϕ ) , italic\_V ).
###
3.3 Generality and Efficiency of MetaGenRL
MetaGenRL offers a general framework for meta-learning objective functions that can represent a wide range of learning algorithms.
In particular, it is only required that both πϕsubscript𝜋italic-ϕ\pi\_{\phi}italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT and Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT can be differentiated w.r.t. to the policy parameters ϕitalic-ϕ\phiitalic\_ϕ.
In the present work, we use this flexibility to leverage population-based meta-optimization, increase sample efficiency through off-policy second-order gradients, and to improve the generalization capabilities of meta-learned objective functions.
##### Population-Based
A general objective function should be applicable to a wide range of environments and agent parameters.
To this extent MetaGenRL is able to leverage the collective experience of *multiple* agents to perform meta-learning by using a *single* objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT shared among a population of agents that each act in their own (potentially different) environment.
Each agent locally computes [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") over a batch of trajectories, and the resulting gradients are combined to update Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT.
Thus, the relevant learning experience of each individual agent is compressed into the objective function that is available to the entire population at any given time.
##### Sample Efficiency
An alternative to learning neural objective functions using a population of agents is through evolution as in EPG (Houthooft et al., [2018](#bib.bib19)).
However, we expect meta-learning using second-order gradients as in MetaGenRL to be much more sample efficient.
This is due to off-policy training of the objective function Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT and its subsequent off-policy use to improve the policy.
Indeed, unlike in evolution there is no need to train multiple randomly initialized agents in their entirety in order to evaluate the objective function, thus speeding up credit assignment.
Rather, at any point in time, any information that is deemed useful for future environment interactions can directly be incorporated into the objective function.
Finally, using the formulation in [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") one can measure the effects of improving the policy using Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT for *multiple* steps by increasing the corresponding number of gradient steps before applying Qθsubscript𝑄𝜃Q\_{\theta}italic\_Q start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT, which we will explore in [Section 5.2.3](#S5.SS2.SSS3 "5.2.3 Multiple gradient steps ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
##### Meta-Generalization
The focus of this work is to learn *general* learning rules that during test-time can be applied to vastly different environments.
A strict separation between the policy and the learning rule, the functional form of the latter, and training across many environments all contribute to this.
Regarding the former, a clear separation between the policy and the learning rule as in MetaGenRL is expected to be advantageous for two reasons.
Firstly, it allows us to specify the number of parameters of the learning rule independent of the policy and critic parameters.
For example, our implementation of Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT uses only 15K15𝐾15K15 italic\_K parameters for the objective function compared to 384K384𝐾384K384 italic\_K parameters for the policy and critic.
Hence, we are able to only use a short description length for the learning rule.
A second advantage that is gained is that the meta-learner is unable to directly change the policy and must, therefore, learn to make use of the objective function.
This makes it difficult for the meta-learner to *overfit* to the training environments.
4 Related work
---------------
Among the earliest pursuits in meta-learning are meta-hierarchies of genetic algorithms (Schmidhuber, [1987](#bib.bib43)) and learning update rules in supervised learning (Bengio et al., [1990](#bib.bib5)).
While the former introduced a general framework of entire meta-hierarchies, it relied on discrete non-differentiable programs.
The latter introduced local update rules that included free parameters, which could be learned using gradients in a supervised setting.
Schmidhuber ([1993](#bib.bib46)) introduced a differentiable self-referential RNN that could address and modify its own weights, albeit difficult to learn.
Hochreiter et al. ([2001](#bib.bib17)) introduced differentiable meta-learning using RNNs to scale to larger problem instances.
By giving an RNN access to its prediction error, it could implement its own meta-learning algorithm, where the weights are the meta-learned parameters, and the hidden states the subject of learning.
This was later extended to the RL setting (Wang et al., [2016](#bib.bib54); Duan et al., [2016](#bib.bib9); Santoro et al., [2016](#bib.bib39); Mishra et al., [2018](#bib.bib29)) (here refered to as RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT).
As we show empirically in our paper, meta-learning with RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT does not generalize well.
It lacks a clear separation between policy and objective function, which makes it easy to overfit on training environments.
This is exacerbated by the imbalance of O(n2)𝑂superscript𝑛2O(n^{2})italic\_O ( italic\_n start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT ) meta-learned parameters to learn O(n)𝑂𝑛O(n)italic\_O ( italic\_n ) activations, unlike in MetaGenRL.
Many other recent meta-learning algorithms learn a policy parameter initialization that is later fine-tuned using a fixed reinforcement learning algorithm (Finn et al., [2017](#bib.bib11); Schulman et al., [2017](#bib.bib49); Grant et al., [2018](#bib.bib14); Yoon et al., [2018](#bib.bib61)).
Different from MetaGenRL, these approaches use second order gradients on the same policy parameter vector instead of using a separate objective function.
Albeit in principle general (Finn & Levine, [2018](#bib.bib10)), the mixing of policy and learning algorithm leads to a complicated way of expressing general update rules.
Similar to RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT, adaptation to related tasks is possible, but generalization is difficult (Houthooft et al., [2018](#bib.bib19)).
Objective functions have been learned prior to MetaGenRL.
Houthooft et al. ([2018](#bib.bib19)) evolve an objective function that is later used to train an agent.
Unlike MetaGenRL, this approach is extremely costly in terms of the number of environment interactions required to evaluate and update the objective function.
Most recently, Bechtle et al. ([2019](#bib.bib4)) introduced learned loss functions for reinforcement learning that also make use of second-order gradients, but use a policy gradient estimator instead of a Q-function.
Similar to other work, their focus is only on narrow task distributions.
Learned objective functions have also been used for learning unsupervised representations (Metz et al., [2019](#bib.bib28)), DDPG-like meta-gradients for hyperparameter search (Xu et al., [2018](#bib.bib60)), and learning from human demonstrations (Yu et al., [2018](#bib.bib62)).
Concurrent to our work, Alet et al. ([2020](#bib.bib1)) uses techniques from architecture search to search for viable artificial curiosity objectives that are composed of primitive objective functions.
Li & Malik ([2016](#bib.bib24); [2017](#bib.bib25)) and Andrychowicz et al. ([2016](#bib.bib2)) conduct meta-learning by learning optimizers that update parameters ϕitalic-ϕ\phiitalic\_ϕ by modulating the gradient of some fixed objective function L𝐿Litalic\_L: Δϕ=fα(∇ϕL)Δitalic-ϕsubscript𝑓𝛼subscript∇italic-ϕ𝐿\Delta\phi=f\_{\alpha}(\nabla\_{\phi}L)roman\_Δ italic\_ϕ = italic\_f start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( ∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L )
where α𝛼\alphaitalic\_α is learned.
They differ from MetaGenRL in that they only modulate the gradient of a fixed objective function L𝐿Litalic\_L instead of learning L𝐿Litalic\_L itself.
Another connection exists to meta-learned intrinsic reward functions (Schmidhuber, [1991a](#bib.bib40); Dayan & Hinton, [1993](#bib.bib8); Wiering & Schmidhuber, [1996](#bib.bib56); Singh et al., [2004](#bib.bib51); Niekum et al., [2011](#bib.bib32); Zheng et al., [2018](#bib.bib63); Jaderberg et al., [2019](#bib.bib20)).
Choosing ∇ϕLα=∇~ϕ∑t=1Tr¯t(τ)subscript∇italic-ϕsubscript𝐿𝛼subscript~∇italic-ϕsuperscriptsubscript𝑡1𝑇subscript¯𝑟𝑡𝜏\nabla\_{\phi}L\_{\alpha}=\tilde{\nabla}\_{\phi}\sum\_{t=1}^{T}\bar{r}\_{t}(\tau)∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT = over~ start\_ARG ∇ end\_ARG start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT italic\_t = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_T end\_POSTSUPERSCRIPT over¯ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ( italic\_τ ),
where r¯tsubscript¯𝑟𝑡\bar{r}\_{t}over¯ start\_ARG italic\_r end\_ARG start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT is a meta-learned reward and ∇~θsubscript~∇𝜃\tilde{\nabla}\_{\theta}over~ start\_ARG ∇ end\_ARG start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT is a gradient estimator (such as a value based or policy gradient based estimator) reveals that meta-learning objective functions includes meta-learning the gradient estimatior ∇~~∇\tilde{\nabla}over~ start\_ARG ∇ end\_ARG itself as long as it is expressible by a gradient ∇θsubscript∇𝜃\nabla\_{\theta}∇ start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT on an objective Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT.
In contrast, for intrinsic reward functions, the gradient estimator ∇~~∇\tilde{\nabla}over~ start\_ARG ∇ end\_ARG is normally fixed.
Finally, we note that positive transfer between different tasks (reward functions) as well as environments (e.g. different Atari games) has been shown previously in the context of transfer learning (Kistler et al., [1997](#bib.bib23); Parisotto et al., [2015](#bib.bib33); Rusu et al., [2016](#bib.bib37); [2019](#bib.bib38); Nichol et al., [2018](#bib.bib31)) and meta-critic learning across tasks (Sung et al., [2017](#bib.bib52)).
In contrast to this work, the approaches that have shown to be successful in this domain rely entirely on human-engineered learning algorithms.
5 Experiments
--------------
Table 1: Mean return across multiple seeds (MetaGenRL: 6 meta-train ×\times× 2 meta-test seeds, RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT: 6 meta-train ×\times× 2 meta-test seeds, EPG: 3 meta-train ×\times× 2 meta-test seeds) obtained by training randomly initialized agents during meta-test time on previously seen environments (cyan) and on unseen environments (brown). Boldface highlights best meta-learned algorithm. Mean returns (6 seeds) of several human-engineered algorithms are also listed.
| | | | | |
| --- | --- | --- | --- | --- |
| Training \Testing | | Cheetah | Hopper | Lunar |
| Cheetah & Hopper | MetaGenRL | 2185 | 2439 | 18 |
| | EPG | -571 | 20 | -540 |
| | RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT | 5180 | 289 | -479 |
| Lunar & Cheetah | MetaGenRL | 2552 | 2363 | 258 |
| | EPG | -701 | 8 | -707 |
| | RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT | 2218 | 5 | 283 |
| Lunar & Hopper & Walker & Ant | MetaGenRL (40 agents) | 3106 | 2869 | 201 |
| Cheetah & Lunar & Walker & Ant | | 3331 | 2452 | -71 |
| Cheetah & Hopper & Walker & Ant | | 2541 | 2345 | -148 |
| | PPO | 1455 | 1894 | 187 |
| | DDPG / TD3 | 8315 | 2718 | 288 |
| | off-policy REINFORCE (GAE) | -88 | 1804 | 168 |
| | on-policy REINFORCE (GAE) | 38 | 565 | 120 |
We investigate the learning and generalization capabilities of MetaGenRL on several continuous control benchmarks including *HalfCheetah (Cheetah)* and *Hopper* from MuJoCo (Todorov et al., [2012](#bib.bib53)), and *LunarLanderContinuous (Lunar)* from OpenAI gym (Brockman et al., [2016](#bib.bib6)).
These environments differ significantly in terms of the properties of the underlying system that is to be controlled, and in terms of the dynamics that have to be learned to complete the environment.
Hence, by training meta-RL algorithms on one environment and testing on other environments they provide a reasonable measure of out-of-distribution generalization.
In our experiments, we will mainly compare to EPG and to RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT to evaluate the efficacy of our approach.
We will also compare to several fixed model-free RL algorithms to measure how well the algorithms meta-learned by MetaGenRL compare to these handcrafted alternatives.
Unless otherwise mentioned, we will meta-train MetaGenRL using 20 agents that are distributed equally over the indicated training environments555An ablation study in [Section A.3](#A1.SS3 "A.3 Ablation of Agent Population Size and Unique Environments ‣ Appendix A Additional Results ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") revealed that a large number of agents is indeed required..
Meta-learning uses clipped double-Q learning, delayed policy & objective updates, and target policy smoothing from TD3 (Fujimoto et al., [2018](#bib.bib12)).
We will allow for 600K600𝐾600K600 italic\_K environment interactions per agent during meta-training and then meta-test the objective function for 1M1𝑀1M1 italic\_M interactions.
Further details are available in [Appendix B](#A2 "Appendix B Experiment Details ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
###
5.1 Comparison to Prior Work
##### Evaluating on previously seen environments
We meta-train MetaGenRL on *Lunar* and compare its ability to train a randomly initialized agent at test-time (i.e. using the learned objective function and keeping it fixed) to DDPG, PPO, and on- and off-policy REINFORCE (both using GAE) across multiple seeds.
[2(a)](#S5.F2.sf1 "2(a) ‣ Figure 3 ‣ Evaluating on previously seen environments ‣ 5.1 Comparison to Prior Work ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") shows that MetaGenRL markedly outperforms both the REINFORCE baselines and PPO.
Compared to DDPG, which finds the optimal policy, MetaGenRL performs only slightly worse on average although the presence of outliers increases its variance.
In particular, we find that some meta-test agents get ‘stuck’ for some time before reaching the optimal policy (see [Section A.2](#A1.SS2 "A.2 Stability of Learned Objective Functions ‣ Appendix A Additional Results ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") for additional analysis).
Indeed, when evaluating only the best meta-learned objective function that was obtained during meta-training (MetaGenRL (best objective func) in [2(a)](#S5.F2.sf1 "2(a) ‣ Figure 3 ‣ Evaluating on previously seen environments ‣ 5.1 Comparison to Prior Work ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")) we are able to observe a strong reduction in variance and even better performance.
We also report results ([2(a)](#S5.F2.sf1 "2(a) ‣ Figure 3 ‣ Evaluating on previously seen environments ‣ 5.1 Comparison to Prior Work ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")) when meta-training MetaGenRL on both *Lunar* and *Cheetah*, and compare to EPG and RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT that were meta-trained on these same environments666In order to ensure a good baseline we allowed for a maximum of 100M100𝑀100M100 italic\_M environment interactions for RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT and 1B1𝐵1B1 italic\_B for EPG, which is more than eight / eighty times the amount used by MetaGenRL. Regarding EPG, this did require us to reduce the total number of seeds to 3 meta-train ×\times× 2 meta-test seeds..
For MetaGenRL we were able to obtain similar performance to meta-training on only *Lunar* in this case.
In contrast, for EPG it can be observed that even one billion environment interactions is insufficient to find a good objective function (in [2(a)](#S5.F2.sf1 "2(a) ‣ Figure 3 ‣ Evaluating on previously seen environments ‣ 5.1 Comparison to Prior Work ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") quickly dropping below -300). Finally, we find that RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT reaches the optimal policy after 100 million meta-training iterations, and that its performance is unaffected by additional steps during testing on *Lunar*.
We note that RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT does not separate the policy and the learning rule and indeed in a similar ‘within distribution’ evaluation, RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT was found successful (Wang et al., [2016](#bib.bib54); Duan et al., [2016](#bib.bib9)).
(a) Previously seen *Lunar* environment.
(b) Unseen *Hopper* environment.
Figure 3: Comparing the test-time training behavior of the meta-learned objective functions by MetaGenRL to other (meta) reinforcement learning algorithms.
We train randomly initialized agents on (a) environments that were encountered during training, and (b) on significantly different environments that were unseen.
Training environments are denoted by ††\dagger† in the legend.
All runs are shown with mean and standard deviation computed over multiple random seeds (MetaGenRL: 6 meta-train ×\times× 2 meta-test seeds, RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT: 6 meta-train ×\times× 2 meta-test seeds, EPG: 3 meta-train ×\times× 2 meta-test seeds, and 6 seeds for all others).
[Table 1](#S5.T1 "Table 1 ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") provides a similar comparison for two other environments.
Here we find that in general MetaGenRL is able to outperform the REINFORCE baselines and PPO, and in most cases (except for *Cheetah*) performs similar to DDPG777We emphasize that the neural objective function under consideration is unable to implement DDPG and only uses a constant value estimate (i.e. ∇ϕV=0subscript∇italic-ϕ𝑉0\nabla\_{\phi}V=0∇ start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT italic\_V = 0 by using gradient stopping) during meta testing..
We also find that MetaGenRL consistently outperforms EPG, and often RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT.
For an analysis of meta-training on more than two environments we refer to [Appendix A](#A1 "Appendix A Additional Results ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
##### Generalization to vastly different environments
We evaluate the same objective functions learned by MetaGenRL, EPG and the recurrent dynamics by RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT on *Hopper*, which is significantly different compared to the meta-training environments.
[2(b)](#S5.F2.sf2 "2(b) ‣ Figure 3 ‣ Evaluating on previously seen environments ‣ 5.1 Comparison to Prior Work ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") shows that the learned objective function by MetaGenRL continues to outperform both PPO and our implementations of REINFORCE, while the best performing configuration is even able to outperform DDPG.
When comparing to related meta-RL approaches, we find that MetaGenRL is significantly better in this case.
The performance of EPG remains poor, which was expected given what was observed on previously seen environments.
On the other hand, we now find that the RL22{}^{2}start\_FLOATSUPERSCRIPT 2 end\_FLOATSUPERSCRIPT baseline fails completely (resulting in a flat low-reward evaluation), suggesting that the learned learning rule that was previously found to be successful is in fact entirely overfitted to the environments that were seen during meta-training.
We were able to observe similar results when using different train and test environment splits as reported in [Table 1](#S5.T1 "Table 1 ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives"), and in [Appendix A](#A1 "Appendix A Additional Results ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
###
5.2 Analysis
####
5.2.1 Meta-Training Progression of Objective Functions
Previously we focused on test-time training randomly initialized agents using an objective function that was meta-trained for a total of 600K600𝐾600K600 italic\_K steps (corresponding to a total of 12M12𝑀12M12 italic\_M environment interactions across the entire population).
We will now investigate the quality of the objective functions during meta-training.
Figure 4: Meta-training with 20 agents on *Cheetah* and *Lunar*. We test the objective function at five stages of meta-training by using it to train three randomly initialized agents on *Hopper*.
[Figure 4](#S5.F4 "Figure 4 ‣ 5.2.1 Meta-Training Progression of Objective Functions ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") displays the result of evaluating an objective function on *Hopper* at different intervals during meta-training on *Cheetah* and *Lunar*.
Initially (28K28𝐾28K28 italic\_K steps) it can be seen that due to lack of meta-training there is only a marginal improvement in the return obtained during test time.
However, after only meta-training for 86K86𝐾86K86 italic\_K steps we find (perhaps surprisingly) that the meta-trained objective function is already able to make consistent progress in optimizing a randomly initialized agent during test-time.
On the other hand, we observe large variances at test-time during this phase of meta-training.
Throughout the remaining stages of meta-training we then observe an increase in convergence speed, more stable updates, and a lower variance across seeds.
####
5.2.2 Ablation study
(a) Meta-training on *Lunar* & *Cheetah*
(b) Meta-testing on *Cheetah*
Figure 5: We meta-train MetaGenRL using several alternative parametrizations of Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT on a) *Lunar* and *Cheetah*, and b) present results of testing on *Cheetah*.
During meta-training a representative example of a single agent population is shown with shaded regions denoting standard deviation across the population. Meta-test results are reported as per usual across 6 meta-train ×\times× 2 meta-test seeds.
We conduct an ablation study of the neural objective function that was described in [Section 3.2](#S3.SS2 "3.2 Parametrizing the Objective Function ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
In particular, we assess the dependence of Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT on the value estimates Vtsubscript𝑉𝑡V\_{t}italic\_V start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT,Vt+1subscript𝑉𝑡1V\_{t+1}italic\_V start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT and on the time component that could to some extent be learned.
Other ablations, including limiting access to the action chosen or to the received reward, are expected to be disastrous for generalization to any other environment (or reward function) and therefore not explored.
##### Dependence on t𝑡titalic\_t
We use a parameterized objective function of the form Lα(at,rt,Vt,πϕ(st)|t∈0,…,T−1)L\_{\alpha}(a\_{t},r\_{t},V\_{t},\pi\_{\phi}(s\_{t})|t\in 0,...,T-1)italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_V start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) | italic\_t ∈ 0 , … , italic\_T - 1 ) as in [Figure 2](#S3.F2 "Figure 2 ‣ 3.2 Parametrizing the Objective Function ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") except that it does not receive information about the time-step t𝑡titalic\_t at each step.
Although information about the current time-step is required in order to learn (for example) a generalized advantage estimate (Schulman et al., [2015b](#bib.bib48)), the LSTM could in principle learn such time tracking on it own, and we expect only minor effects on meta-training and during meta-testing.
Indeed in [4(b)](#S5.F4.sf2 "4(b) ‣ Figure 5 ‣ 5.2.2 Ablation study ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") it can be seen that the neural objective function performs well without access to t𝑡titalic\_t, although it converges slower on *Cheetah* during meta-training ([4(a)](#S5.F4.sf1 "4(a) ‣ Figure 5 ‣ 5.2.2 Ablation study ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")).
##### Dependence on V𝑉Vitalic\_V
We use a parameterized objective function of the form Lα(at,rt,t,πϕ(st)|t∈0,…,T−1)L\_{\alpha}(a\_{t},r\_{t},t,\pi\_{\phi}(s\_{t})|t\in 0,...,T-1)italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_t , italic\_π start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) | italic\_t ∈ 0 , … , italic\_T - 1 ) as in [Figure 2](#S3.F2 "Figure 2 ‣ 3.2 Parametrizing the Objective Function ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") except that it does not receive any information about the value estimates at time-step t𝑡titalic\_t.
There exist reinforcement learning algorithms that work without value function estimates (eg. Williams ([1992](#bib.bib59)); Schmidhuber & Zhao ([1998](#bib.bib42))), although in the absence of an alternative baseline these often have a large variance.
Similar results are observed for this ablation in [4(a)](#S5.F4.sf1 "4(a) ‣ Figure 5 ‣ 5.2.2 Ablation study ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") during meta-training where a possibly large variance appears to affect meta-training.
Correspondingly during test-time ([4(b)](#S5.F4.sf2 "4(b) ‣ Figure 5 ‣ 5.2.2 Ablation study ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives")) we do not find any meaningful training progress to take place.
In contrast, we find that we can remove the dependence on *one* of the value function estimates, i.e. remove Vt+1subscript𝑉𝑡1V\_{t+1}italic\_V start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT but keep Vtsubscript𝑉𝑡V\_{t}italic\_V start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT, which during some runs even increases performance.
Figure 6: We meta-train MetaGenRL on the *LunarLander* and *HalfCheetah* environments using one, three, and five inner gradient steps on ϕitalic-ϕ\phiitalic\_ϕ. Meta-test results are reported across 3 meta-train ×\times× 2 meta-test seeds.
####
5.2.3 Multiple gradient steps
We analyze the effect of making *multiple* gradient updates to the policy using Lαsubscript𝐿𝛼L\_{\alpha}italic\_L start\_POSTSUBSCRIPT italic\_α end\_POSTSUBSCRIPT before applying the critic to compute second-order gradients with respect to the objective function parameters as in [Equation 6](#S3.E6 "6 ‣ 3.1 From DDPG to Gradient-Based Meta-Learning of Neural Objectives ‣ 3 Meta-Learning Neural Objectives ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives").
While in previous experiments we have only considered applying a single update, multiple gradient updates might better capture long term effects of the objective function.
At the same time, moving further away from the current policy parameters could reduce the overall quality of the second-order gradients.
Indeed, in [Figure 6](#S5.F6 "Figure 6 ‣ Dependence on 𝑉 ‣ 5.2.2 Ablation study ‣ 5.2 Analysis ‣ 5 Experiments ‣ Improving Generalization in Meta Reinforcement Learning using Learned Objectives") it can be observed that using 3 gradient steps already slightly increases the variance during test-time training on *Hopper* and *Cheetah* after meta-training on *LunarLander* and *Cheetah*.
Similarly, we find that further increasing the number of gradient steps to 5 harms performance.
6 Conclusion
-------------
We have presented MetaGenRL, a novel off-policy gradient-based meta reinforcement learning algorithm that leverages a population of DDPG-like agents to meta-learn general objective functions.
Unlike related methods the meta-learned objective functions do not only generalize in narrow task distributions but show similar performance on entirely different tasks while markedly outperforming REINFORCE and PPO.
We have argued that this generality is due to MetaGenRL’s explicit separation of the policy and learning rule, the functional form of the latter, and training across multiple agents and environments.
Furthermore, the use of second order gradients increases MetaGenRL’s sample efficiency by several orders of magnitude compared to EPG (Houthooft et al., [2018](#bib.bib19)).
In future work, we aim to further improve the learning capabilities of the meta-learned objective functions, including better leveraging knowledge from prior experiences.
Indeed, in our current implementation, the objective function is unable to observe the environment or the hidden state of the (recurrent) policy.
These extensions are especially interesting as they may allow more complicated curiosity-based (Schmidhuber, [1991b](#bib.bib45); [1990](#bib.bib44); Houthooft et al., [2016](#bib.bib18); Pathak et al., [2017](#bib.bib34)) or model-based (Schmidhuber, [1990](#bib.bib44); Weber et al., [2017](#bib.bib55); Ha & Schmidhuber, [2018](#bib.bib15)) algorithms to be learned.
To this extent, it will be important to develop introspection methods that analyze the learned objective function and to scale MetaGenRL to make use of many more environments and agents.
Acknowledgements
----------------
We thank Paulo Rauber, Imanol Schlag, and the anonymous reviewers for their feedback.
This work was supported by the ERC Advanced Grant (no: 742870) and computational resources by the Swiss National Supercomputing Centre (CSCS, project: s978).
We also thank NVIDIA Corporation for donating a DGX-1 as part of the Pioneers of AI Research Award and to IBM for donating a Minsky machine. |
ada9b037-77d8-418b-8f91-81d11ac84d99 | trentmkelly/LessWrong-43k | LessWrong | Fine-insured bounties as AI deterrent
Epistemic status: Hare-brained scheme
Suppose the state of Arkansas decided to levy a fine of $1000 on anyone attempting to push the state of AI forward. Further, Arkansas decreed that this $1000 fine would be paid directly out of the coffers of those found guilty, and directly towards the people who turned them in - a 'fine bounty', if you will.
$1000 is not a lot of money in the grand scheme of things. However, like most people in Arkansas, I don't have $1000 to spare, so I would decide to just not risk working on advancing AI. Nor would I join an organization involved in such activities, lest I be caught in the crossfire. The more coworkers I had, the more likely it would become that one bad apple would turn everyone in at once to collect the $n000 reward. (turning themselves in would just mean paying $1000 to themselves, which makes 'getting away' with this a lot easier).
This is the essential thrust of Robin Hanson's Privately Enforced & Punished Crime; he calls this the fine-insured bounty (FIB) crime law system in a follow-up post. He is much more cogent in spelling out the edge cases of the system in general there, so I will focus on some details which I think make this system unusually well-suited to the problem of how one deters technological development, such as AI development.
1. It scales. It is very hard to develop new AI capabilities without someone suspecting what you're up to. Even if they cannot prove it, the insurance mechanism in the fine-insured bounty system means even a low-fidelity suspicion can carry with it considerable economic value. Much like how someone with a history of running red lights may have a higher car insurance premium, someone with a history of buying GPUs may have a higher AI bounty insurance premium. (Someone running a virus which collects a fingerprint of GPU activity at scale likely has a valuable source of information to sell to insurers and prospective bounty hunters alike.)
2. It chills. To use a leaky pipeline |
9aacf41c-d5d4-4c09-9682-dec7c5f4dc72 | trentmkelly/LessWrong-43k | LessWrong | Improving Human Rationality Through Cognitive Change (intro)
This is the introduction to a paper I started writing long ago, but have since given up on. The paper was going to be an overview of methods for improving human rationality through cognitive change. Since it contains lots of handy references on rationality, I figured I'd publish it, in case it's helpful to others.
1. Introduction
During the last half-century, cognitive scientists have catalogued dozens of common errors in human judgment and decision-making (Griffin et al. 2012; Gilovich et al. 2002). Stanovich (1999) provides a sobering introduction:
> For example, people assess probabilities incorrectly, they display confirmation bias, they test hypotheses inefficiently, they violate the axioms of utility theory, they do not properly calibrate degrees of belief, they overproject their own opinions onto others, they allow prior knowledge to become implicated in deductive reasoning, they systematically underweight information about nonoccurrence when evaluating covariation, and they display numerous other information-processes biases...
The good news is that researchers have also begun to understand the cognitive mechanisms which produce these errors (Kahneman 2011; Stanovich 2010), they have found several "debiasing" techniques that groups or individuals may use to partially avoid or correct these errors (Larrick 2004), and they have discovered that environmental factors can be used to help people to exhibit fewer errors (Thaler and Sunstein 2009; Trout 2009).
This "heuristics and biases" research program teaches us many lessons that, if put into practice, could improve human welfare. Debiasing techniques that improve human rationality may be able to decrease rates of violence caused by ideological extremism (Lilienfeld et al. 2009). Knowledge of human bias can help executives make more profitable decisions (Kahneman et al. 2011). Scientists with improved judgment and decision-making skills ("rationality skills") may be more apt to avoid experimenter bias (S |
b9465626-228b-429f-ad6e-ed2efafad95b | trentmkelly/LessWrong-43k | LessWrong | How I Ended Up Non-Ambitious
I have a confession to make. My life hasn’t changed all that much since I started reading Less Wrong. Hindsight bias makes it hard to tell, I guess, but I feel like pretty much the same person, or at least the person I would have evolved towards anyway, whether or not I spent those years reading about the Art of rationality.
But I can’t claim to be upset about it either. I can’t say that rationality has undershot my expectations. I didn’t come to Less Wrong expecting, or even wanting, to become the next Bill Gates; I came because I enjoyed reading it, just like I’ve enjoyed reading hundreds of books and websites.
In fact, I can’t claim that I would want my life to be any different. I have goals and I’m meeting them: my grades are good, my social skills are slowly but steadily improving, I get along well with my family, my friends, and my boyfriend. I’m in good shape financially despite making $12 an hour as a lifeguard, and in a year and a half I’ll be making over $50,000 a year as a registered nurse. I write stories, I sing in church, I teach kids how to swim. Compared to many people my age, I'm pretty successful. In general, I’m pretty happy.
Yvain suggested akrasia as a major limiting factor for why rationalists fail to have extraordinarily successful lives. Maybe that’s true for some people; maybe they are some readers and posters on LW who have big, exciting, challenging goals that they consistently fail to reach because they lack motivation and procrastinate. But that isn’t true for me. Though I can’t claim to be totally free of akrasia, it hasn’t gotten much in the way of my goals.
However, there are some assumptions that go too deep to be accessed by introspection, or even by LW meetup discussions. Sometimes you don't even realize they’re assumptions until you meet someone who assumes the opposite, and try to figure out why they make you so defensive. At the community meetup I described in my last post, a number of people asked me why I wasn’t studying |
2a1298f6-c048-4015-94b8-e81fc979ef2e | trentmkelly/LessWrong-43k | LessWrong | Annotated reply to Bengio's "AI Scientists: Safe and Useful AI?"
The AI safety proposal was published by Yoshua Bengio on May 7, 2023.
Below I quote the salient sections of the proposal and comment on them in turn.
Main thesis: safe AI scientists
> […] I would like to share my thoughts regarding the hotly debated question of long-term risks associated with AI systems which do not yet exist, where one imagines the possibility of AI systems behaving in a way that is dangerously misaligned with human rights or even loss of control of AI systems that could become threats to humanity. A key argument is that as soon as AI systems are given goals – to satisfy our needs – they may create subgoals that are not well-aligned with what we really want and could even become dangerous for humans.
>
> The bottom line of the thesis presented here is that there may be a path to build immensely useful AI systems that completely avoid the issue of AI alignment, which I call AI scientists because they are modeled after ideal scientists and do not act autonomously in the real world, only focusing on theory building and question answering. The argument is that if the AI system can provide us benefits without having to autonomously act in the world, we do not need to solve the AI alignment problem. This would suggest a policy banning powerful autonomous AI systems that can act in the world (“executives” rather than “scientists”) unless proven safe. However, such a solution would still leave open the political problem of coordinating people, organizations and countries to stick to such guidelines for safe and useful AI. The good news is that current efforts to introduce AI regulation (such as the proposed bills in Canada and the EU, but see action in the US as well) are steps in the right direction.
Bengio’s “AI scientists” are similar in spirit to tool AI, Oracle AI, and the “simulators” agenda. All these ideas share a weakness: tool AI may still cause a lot of harm in the hands of “misaligned” people, or people with poor ethics. I’m not sure which |
0cae0ac3-84d2-40b4-a1bb-679652b62e67 | trentmkelly/LessWrong-43k | LessWrong | Things you should never do
Things you should never do
A great article on a common fallacy among programmer-types |
15714600-dbaf-4c40-b8f7-11a3b9471dae | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Intro to AI Safety, Remastered
hi this video is a recording of a talk
that i gave a while back i already
published a version of it on my second
channel but
did you even know i had a second channel
most people don't i thought more people
should see it so i remastered it i
cleaned it up improved the graphics and
yeah this is that enjoy right
hello everyone my name is robert miles i
am usually doing this on youtube i'm not
really used to public speaking
i'm not used to not being able to edit
out my mistakes there may be mistakes
also
i may go too quickly um sorry not sorry
so uh when it comes to ai safety you can
kind of divide it up into four
areas along two axes you've got your
short term
and your long term and you've got
accident risks and misuse risks
and that's kind of a useful way to
divide things up ai safety
covers everything the area that
interests me most
is the long-term accident risks i think
once you have very powerful ai systems
it almost doesn't matter if they're
being used by the right people or the
wrong people or what you're trying to do
with them
the difficulty is in keeping them in
under control at all
so that's what i'm going to be talking
about what is ai safety why is it
important
so i want to start by asking the
question which i think everybody needs
to be asking themselves
what is the most important problem in
your field
take a second to think of it
and why are you not working on that um
for me i think the most important
problem in the field of ai
is ai safety this is the problem
specifically that i'm worried about
that we will sooner or later build an
artificial agent
with general intelligence so i'm going
to go into a bunch of these
terms the first thing is what do i mean
when i say sooner or later
this is a little bit washed out but this
is a graph
of a survey a large survey of ai experts
these are people who published in
major a.i conferences and they were
asked when they thought
we would achieve high-level machine
intelligence which is defined as
an agent which is able to carry out any
tasks humans can
as well as or better than humans and
they say that uh 50 chance of of
having achieved that we we hit that
about 45 years
from 2016. um but then of course we hit
like 10
chance nine years from now so um
it's not immediate but it's happening
this is definitely worth
taking with a pinch of salt because if
you ask the question slightly
differently you get an estimate of 120
years rather than 45
there's a lot of uncertainty in this
area but the point is
it's going to happen as i said sooner or
later because at the end of the day
general intelligence is possible the
brain implements it and the brain is not
magic
sooner or later we'll figure it out so
what do i mean when i say an artificial
agent
well so an agent uh is a term from
uh economics mostly but basically agents
have goals
they choose actions to further their
goals this the simplest expression of
what an agent is
so um the simplest thing that you might
call an agent would be something like a
thermostat
it has a goal which is to have the room
be at a particular temperature
it has actions it can take it can turn
on the heating it could turn on the air
conditioning
it chooses its actions to achieve its
goal of maintaining the room at a steady
temperature
extremely simple agent a more complex
agent might be something like
a chess ai which has a goal of
like if it's playing white it has a goal
of the black king being
in checkmate and it takes actions in the
form of moving pieces on the board in
order to achieve its goal
so you see how this idea of an agent is
a very useful way of thinking about
lots of different intelligence systems
and of course
humans can be modeled as agents as well
this is how it's usually done in
economics
individuals or companies could be
considered to have a goal
of you know maximizing their income or
maximizing their profits and making
decisions
in order to achieve that
so when i'm talking about intelligence
intelligence has a lot of as a term is a
heavily loaded term has a lot of
different
people put their own definitions on it
in this context what i mean when i say
intelligence is just
the thing that lets an agent choose
effective actions it's
whatever it is that's in our brains or
that's in the programming of these
systems
that means that the actions they choose
tend to get them closer to their goals
um and so then you could say that
an agent is more intelligent if it's
more effective at achieving its goals
whatever those goals are
if you have two agents in an environment
with incompatible goals
like let's say the environment is the
chess board and one agent wants
white to win and one agent wants black
to win then
generally the more intelligent agent
will be the one that gets
what it wants the better ai will win the
chess game
um and finally general intelligence this
is where it becomes
uh interesting in my opinion so
generality is the ability to behave
intelligently in a wide range of domains
if you take something like a chess ai
it's extremely narrow it only knows how
to play chess
um and even though you might say that
it's more intelligent than a thermostat
because it's more sophisticated it's
more complicated
it couldn't do the thermostat's job
there's no
position on the chessboard that
corresponds to the room being a good
temperature
there's no move that corresponds to
turning on an air conditioner the chest
ai can only think in terms of chess
it's extremely narrow generality is a
continuous spectrum
um so if you write a program that can
play an atari game
that's very narrow deep mind one of
their early triumphs was
that they made a program that could play
dozens of different atari games single
program that could learn all of these
different games
and so it's more general
because it's able to act across a wider
variety of domains
the most general intelligence that we're
aware of right now
is human beings human beings are
very general we're able to operate
across a very wide range of domains
including and this is important we're
able to learn domains which
evolution did not and could not prepare
us for
um we can for example drive a car an
evolution did not prepare us for that
we invented cars they're very recent um
we can
you know invent rockets and go to the
moon and then we can operate on the moon
which is a completely different
environment
and this is kind of the power of general
intelligence really the power of general
intelligence is
we can build a car we can build a rocket
we can put the car on the rocket take
the car to the moon
drive the car on the moon and there's
nothing
else that can do that yet
um but sooner or later right
so so this is what i'm talking about i'm
talking about
what you might call true ai real ai the
sci-fi stuff
um an agent which has goals
in the real world and is able to
intelligently choose actions in the real
world
to achieve those goals now that sounds
i've said i said what's the biggest
problem this doesn't sound like a
problem right
on the surface of it this sounds like a
solution
you just tell the thing you know cure
cancer or maximize the profits of my
company or whatever
and it takes whatever actions are
necessary in the real world to achieve
that goal
but um it is a problem
so the big problem is this should be
auto playing and it isn't
um the big problem is it's difficult
to choose good goals
um so this this is an ai made by open ai
it's playing a game called coast runners
which is actually a racing game
they trained it on the score which you
probably can't see down here it's
currently a thousand um what the system
learned is that
if it goes around in a circle here and
crashes into everything and catches fire
these little turbo pickups they respawn
at just the right rate
that if it just flings itself around in
a circle it can pick up
the turbo and that gives you a few
points every time you do that
and it turns out that this is a much
better way of getting points
than actually racing around the track
and the important point here is that
this is not unusual this is not open ai
doing anything unusually stupid
this is kind of the default um picking
objectives is surprisingly hard
and you will find that the strategy or
the behavior that
maximizes your objective is probably not
the thing you thought it was
it's probably not what you were aiming
for uh there's loads of examples
actually uh victoria has a great
list on her blog deep safety if there's
like 30 of them
different things going wrong there was
one they had uh
they were trying to teach they were
trying to evolve systems that would run
quickly so they they trained them on the
i'm going to pause this because it's
distracting as hell
where's my mouse yeah pause
pause please um they were training like
agents that were supposed to run so they
simulated them for a particular period
of time and measured
how far their center of mass moved which
seems perfectly sensible
what they found was that they developed
a bunch of these creatures which were
extremely tall and thin with a big mass
on the end
that then fell over because they weren't
simulating them for long enough
the you could go the fastest just by
falling over rather than
actually running that moved your center
of mass the furthest um
there's a lot of these there was a
tetris bot which um
would play reasonably well and then just
when it was about to lose
would pause the game and sit there
indefinitely because it lost
points for losing but didn't lose any
points for just sitting on the pause
screen indefinitely
this is this is like the default of how
these systems behave i have
no memory what my next slide is oh yeah
right so we have problems
specifying even simple goals in simple
environments like atari games or
basic evolutionary algorithms things
like that
when it comes to the real world things
get way more complicated this is a quote
from
stuart russell who sort of wrote the
book on aia when a system is optimizing
a function of n variables where the
objective depends on a subset of size k
which is less than n
it will often set the remaining
unconstrained variables to extreme
values
if one of those unconstrained variables
is something that we care about the
solution found may be highly undesirable
in the real world we have a very large
number of variables and so
we're talking about we're talking about
very large
values for n here so let's say you've
got your robot
and you've given it a gold which you
think is very simple you want it to get
you a cup of tea
so you've managed to specify what a cup
of tea is and that you want one to be
on the desk in front of you so far so
good
but suppose there is a there's a
processing vars on a narrow
stand sort of in front of where the
kitchen is so the robot
immediately plows into the vars and
destroys it on its way to make you a cup
of tea because
you only gave it one variable to keep
track of in the goal
which is the t it doesn't care about the
vars you never told it to care about the
vars
it destroys the vast this is a problem
um
so okay now you can you know shut it
down modify it and say
okay give me a cup of tea but also don't
knock over the vars
but then there will be a third thing
there is always another thing
because when when you're making
decisions in the real world
you're always making trade-offs you're
always
taking various things that you value and
deciding how much
of one you're willing to trade for how
much of another you know i could do this
quicker
but it increases the risk of me making a
mistake or i
could do this cheaper but it won't be as
reliable
i could do this faster but it'll be more
expensive you're always trading these
things off against one another
and so an agent like this which only
cares about a limited subset of the
variables in the system
will be willing to trade off arbitrarily
large amounts
of any of the variables that aren't part
of its goal for arbitrarily tiny
increases
in any of the things which are in its
goal so it will happily
i let's say now for example now it
values the vars
uh and those are the only things that it
values it might reason something like
okay there's a human in the environment
the human moves around the human
may accidentally knock over the vars and
i care about the vars so i have to kill
the human
right and this is totally ridiculous but
if you didn't tell it that you value
being alive it doesn't care and anything
that it doesn't value
is going to be lost if you manage to
come up with if you have a sufficiently
powerful agent
and you manage to come up with a really
good objective function which
covers the top 20 things that humans
value
the 21st thing that humans value is
probably gone forever
because the smarter the more powerful
the agent is
the better it will be at figuring out
ways to make these trade-offs
to gain a millionth of a percent better
at one thing
while sacrificing everything of some
other variable
so this is a problem but actually that
scenario i gave was unrealistic in
many ways but but one important way that
it was unrealistic
is that i had i had the system go wrong
and then you just turn it off
and fix it but in fact if you're
if the thing has a goal of getting you a
cup of tea this is not like a chess ai
where you can just turn it off because
it has no concept of itself or
being turned off its
world model contains you it contains
itself it contains the possibility of
being turned off and it's fully aware
that if you turn it off
because it knocked over the vars it
won't be able to get you any tea which
is the only thing it cares about
so it's not going to just let you turn
it off it will fight you
or if it's slightly smarter it will
deceive you so that you believe it's
working correctly
so that you don't want to change it
until it's in a position
where you can't turn it off and then it
will go after its actual objective
so so this is a problem
and the thing is this is this is a
convergent instrumental goal which means
it sort of doesn't matter what the goal
is it doesn't matter what your goal is
as an agent
if you're destroyed you can't achieve
that goal so it sort of almost doesn't
matter what goal we give it there is
only a very tiny fraction of possible
goals
that will involve it actually allowing
itself to be turned off and modified
and that's quite complicated um
there are some other convergent
instrumental goals so we had
self-preservation goal preservation
resource acquisition is the kind of
thing we can expect these kinds of
systems to do
most plans you can do them better if you
have more resources
whether that's money computational
resources just
free energy matter whatever
the other one is self-improvement
whatever you're trying to do
you can probably do it better if you're
smarter and ai systems potentially have
the capacity to improve themselves
either just by
acquiring more hardware to run on or
changing
you know improving their uh their
software to run faster or better or so
on
so there's a whole bunch of behaviors
which
uh intelligent systems intelligent
agents
generally intelligent agents we would
expect them to do
by default and that's really
my core point artificial general
intelligence is dangerous by default
it's much much easier to build these
kinds of agents
which try to do ridiculous things and
trick you
and try to deceive you or will fight you
when you try to turn them off or modify
them
on the way to doing some ridiculous
thing which you don't want
uh much easier to build that kind of
agent than
it is to build something which actually
reliably does what you want it to do
and that's why we have a problem because
we have
45 to 120 years to figure out
how to do it safely which is a much
harder problem and
we may only get one shot it's entirely
possible that the first
true artificial general intelligence
will manage to successfully achieve
whatever its stupid goal is
and that could be truly a disaster on a
global scale
so we have to we have to beat this
challenge
on hard mode before anyone beats it on
easy mode
so are we screwed
no we're only probably screwed
um there are things we can do
safe general artificial intelligence is
totally possible
it's just a very difficult technical
challenge
and there are people working very hard
on it right now trying to solve
a whole range of difficult technical
challenges
so that we can figure out how to do this
safely
thanks
[Applause]
[Music]
you may have noticed in the intro in
this outro that the image quality has
improved since the last video
that's largely thanks to my excellent
patrons thank you to all of these people
here for helping me to get this new
camera in this video i'm especially
thanking james pets who's been hanging
out with us on the discord server
helping answer questions from the
youtube comments and so on and actually
that last video about mesa optimizers
has had a lot of really good questions
so the next video will be answering some
of those
that's coming up soon so thanks again to
james and to all my patrons
to everyone who asked questions and to
you for watching
i'll see you next time
you |
4ecd2087-2481-472b-95f7-133e3f14c896 | trentmkelly/LessWrong-43k | LessWrong | Neuroevolution, Social Intelligence, and Logic
Do Transformers Reason about Electric Sheep?
Do not underestimate evolution
Evolution for 3.5–4.5 billion years has developed unconditioned reflexes and instincts in terrestrial carbon life sufficient for survival and distribution. Cumulatively, these and other types of behavior are called genetic memory. Homo sapiens, in this respect, is no different from other forms of life. However, something differentiates them favorably in the ability to distinguish, evaluate, and understand what is true and false, valuable and useless, moral and immoral. Instead of two extremes, there may be an interval, but invariably one thing, this ability to distinguish is called intelligence (a noun from the Latin verb interlego = inter — between + lego — (se)lect = to distinguish, sort out — later acquired the meaning of to understand — understanding).
At what the moment did the critical leap from the slowly adapting collective unconscious occur? When did the rapid growth of intelligence, not associated with genetic memory, begin?
The Dual Inheritance Theory and the ideas of Memetics suggest that the growth of intelligence was accompanied by an increase in the amount of knowledge accumulated in the population, what would not have been possible without effective mechanisms for assimilation, transfer, and, what is more important, the derivation of new knowledge. Our species overtook evolution with the help of collective knowledge — culture. And it all began when one of the organisms projected its knowledge to another.
Another type of the dual development
Assimilation
Assimilation of knowledge comes from the physical world, a complex system by its nature. The primary knowledge about the environment of its existence, the protohuman, like other animals, assimilated in the form of conditioned (learned) reflexes. The variety of situations gives rise to a variety of knowledge, for example, about falling objects under the influence of gravity, about pain when in contact with fire, about th |
9a6177d2-f871-4870-8bc2-b153468a3b8c | trentmkelly/LessWrong-43k | LessWrong | Meetup : Ohio Monthly
Discussion article for the meetup : Ohio Monthly
WHEN: 17 March 2012 03:00:00PM (-0500)
WHERE: 123 Gano Road, Wilmington, OH 45177
Time for the third Ohio monthly meetup! As always, it will be the third Sunday (March 18), from 4p-8p. This month, we MAY have a different location, for Rolf to present his thesis work. Otherwise, we will be in the back room of Max and Erma's like usual. We have picked five Sequences to focus on. They are: Making Beliefs Pay Rent Belief in Belief Bayesian Judo Professing and Cheering Belief as Attire It's a good idea to at least skim these sequences, even if you've already read them. We will also probably continue conversation on: attracting and maintaining new people, meta discussion, how to make entrance to the LessWrong memeplex as painless as possible. As always, there are more frequent, but irregularly scheduled, LessWrong meetups in both Columbus and Cincinnati. Please join our mailing list to keep updated on those!
Discussion article for the meetup : Ohio Monthly |
0756f169-42be-4468-93a5-a2d40bca9263 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Superintelligence 26: Science and technology strategy
*This is part of a weekly reading group on [Nick Bostrom](http://www.nickbostrom.com/)'s book, [Superintelligence](http://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111). For more information about the group, and an index of posts so far see the [announcement post](/lw/kw4/superintelligence_reading_group/). For the schedule of future topics, see [MIRI's reading guide](https://intelligence.org/wp-content/uploads/2014/08/Superintelligence-Readers-Guide-early-version.pdf).*
---
Welcome. This week we discuss the twenty-sixth section in the [reading guide](https://intelligence.org/wp-content/uploads/2014/08/Superintelligence-Readers-Guide-early-version.pdf): ***Science and technology strategy***. Sorry for posting late—my car broke.
This post summarizes the section, and offers a few relevant notes, and ideas for further investigation. Some of my own thoughts and questions for discussion are in the comments.
There is no need to proceed in order through this post, or to look at everything. Feel free to jump straight to the discussion. Where applicable and I remember, page numbers indicate the rough part of the chapter that is most related (not necessarily that the chapter is being cited for the specific claim).
**Reading**: “Science and technology strategy” from Chapter 14
---
Summary
=======
1. This section will introduce concepts that are useful for thinking about long term issues in science and technology (p228)
2. ***Person affecting perspective***: one should act in the best interests of everyone who already exists, or who will exist independent of one's choices (p228)
3. ***Impersonal perspective***: one should act in the best interests of everyone, including those who may be brought into existence by one's choices. (p228)
4. *Technological completion conjecture*: "If scientific and technological development efforts do not cease, then all important basic capabilities that could be obtained through some possible technology will be obtained." (p229)
1. This does not imply that it is futile to try to steer technology. Efforts may cease. It might also matter exactly when things are developed, who develops them, and in what context.
5. *Principle of differential technological development*: one should slow the development of dangerous and harmful technologies relative to beneficial technologies (p230)
6. We have a **preferred order** for some technologies, e.g. it is better to have superintelligence later relative to social progress, but earlier relative to other existential risks. (p230-233)
7. If a **macrostructural development accelerator** is a magic lever which slows the large scale features of history (e.g. technological change, geopolitical dynamics) while leaving the small scale features the same, then we can ask whether pulling the lever would be a good idea (p233). The main way Bostrom concludes that it matters is by affecting how well prepared humanity is for future transitions.
8. ***State risk***: a risk that persists while you are in a certain situation, such that the amount of risk is a function of the time spent there. e.g. risk from asteroids, while we don't have technology to redirect them. (p233-4)
9. ***Step risk***: a risk arising from a transition. Here the amount of risk is mostly not a function of how long the transition takes. e.g. traversing a minefield: this is not especially safer if you run faster. (p234)
10. **Technology coupling**: a predictable timing relationship between two technologies, such that hastening of the first technology will hasten the second, either because the second is a precursor or because it is a natural consequence. (p236-8) e.g. brain emulation is plausibly coupled to ['neuromorphic' AI](http://theuncertainfuture.com/faq.html#3), because the understanding required to emulate a brain might allow one to more quickly create an AI on similar principles.
11. ***Second guessing***: acting as if "by treating others as irrational and playing to their biases and misconceptions it is possible to elicit a response from them that is more competent than if a case had been presented honestly and forthrightly to their rational faculties" (p238-40)
Another view
============
There is a common view which says we should not act on detailed abstract arguments about the far future like those of this section. [Here](http://blog.givewell.org/2014/07/03/the-moral-value-of-the-far-future/) Holden Karnofsky exemplifies it:
>
> I have often been challenged to explain how one could possibly reconcile (a) caring a great deal about the far future with (b) donating to one of [GiveWell’s top charities](http://www.givewell.org/charities/top-charities). My general response is that in the face of sufficient [uncertainty](http://blog.givewell.org/2014/06/10/sequence-thinking-vs-cluster-thinking/) about one’s options, and lack of conviction that there are good (in the sense of high expected value) opportunities to make an *enormous* difference, it is rational to try to make a smaller but [robustly positive](http://blog.givewell.org/2014/06/10/sequence-thinking-vs-cluster-thinking/)difference, whether or not one can trace a specific causal pathway from doing this small amount of good to making a large impact on the far future. A few brief arguments in support of this position:
>
>
> * I believe that the track record of “taking robustly strong opportunities to do ‘something good'” is far better than the track record of “taking actions whose value is contingent on high-uncertainty arguments about where the highest utility lies, and/or arguments about what is likely to happen in the far future.” This is true even when one evaluates track record only in terms of seeming impact on the far future. The developments that seem most positive in retrospect – from large ones like the development of the steam engine to small ones like the many economic contributions that facilitated strong overall growth – seem to have been driven by the former approach, and I’m not aware of many examples in which the latter approach has yielded great benefits.
> * I see some sense in which the world’s *overall civilizational ecosystem* seems to have done a better job optimizing for the far future than any of the world’s *individual minds*. It’s often the case that people acting on relatively short-term, tangible considerations (especially when they did so with creativity, integrity, transparency, consensuality, and pursuit of gain via value creation rather than value transfer) have done good in ways they themselves wouldn’t have been able to foresee. If this is correct, it seems to imply that one should be focused on “playing one’s role as well as possible” – on finding opportunities to [“beat the broad market”](http://blog.givewell.org/2013/05/02/broad-market-efficiency/) (to do more good than people with similar goals would be able to) rather than pouring one’s resources into the areas that non-robust estimates have indicated as most important to the far future.
> * The *process* of trying to accomplish tangible good can lead to a great deal of learning and unexpected positive developments, more so (in my view) than the process of putting resources into a low-feedback endeavor based on one’s current best-guess theory. In my [conversation with Luke and Eliezer](http://intelligence.org/2014/01/27/existential-risk-strategy-conversation-with-holden-karnofsky/), the two of them hypothesized that the greatest positive benefit of supporting GiveWell’s top charities may have been to raise the profile, influence, and learning abilities of GiveWell. If this were true, I don’t believe it would be an inexplicable stroke of luck for donors to top charities; rather, it would be the sort of development (facilitating feedback loops that lead to learning, organizational development, growing influence, etc.) that is often associated with “doing something well” as opposed to “doing the most worthwhile thing poorly.”
> * I see multiple reasons to believe that contributing to general human empowerment mitigates global catastrophic risks. I laid some of these out in a [blog post](http://blog.givewell.org/2013/07/18/empowerment-and-catastrophic-risk/) and discussed them further in my [conversation with Luke and Eliezer](http://intelligence.org/2014/01/27/existential-risk-strategy-conversation-with-holden-karnofsky/).
>
>
>
Notes
=====
**1. Technological completion timelines game**
The technological completion conjecture says that all the basic technological capabilities will eventually be developed. But when is 'eventually', usually? Do things get developed basically as soon as developing them is not prohibitively expensive, or is thinking of the thing often a bottleneck? This is relevant to how much we can hope to influence the timing of technological developments.
Here is a fun game: How many things can you find that could have been profitably developed much earlier than they were?
Some starting suggestions, which I haven't looked into:
[Wheeled luggage](http://www.telegraph.co.uk/sponsored/lifestyle/innovations/11073405/suitcase-wheels-invention.html): invented in the 1970s, though humanity had had both wheels and luggage for a while.
[Hot air balloons](http://www.quora.com/Why-werent-hot-air-balloons-invented-earlier): flying paper lanterns using the same principle were apparently used before 200AD, while a manned balloon wasn't used until [1783](http://en.wikipedia.org/wiki/Hot_air_balloon#First_manned_flight).
[Penicillin](http://en.wikipedia.org/wiki/History_of_penicillin): mould was apparently traditionally used for antibacterial properties in several cultures, but lots of things are traditionally used for lots of things. By the 1870s many scientists had noted that specific moulds inhibited bacterial growth.
[Wheels](http://en.wikipedia.org/wiki/Wheel#History): Early [toys](http://anthro.amnh.org/anthropology/databases/common/image_dup.cfm?database=MIXDATA&catno=30.2/%209734&site=P) from the Americas appear to have had wheels ([here](http://anthro.amnh.org/anthropology/databases/common/image_dup.cfm?database=MIXDATA&catno=30.2/%209734&site=P) and [pictured](http://en.wikipedia.org/wiki/Wheel#mediaviewer/File:Remojadas_Wheeled_Figurine.jpg) is one from 1-900AD; [Wikipedia](http://en.wikipedia.org/wiki/Wheel#History) claims such toys were around as early as 1500BC). However wheels were apparently not used for more substantial transport in the Americas until much later.
*Image:*["Remojadas Wheeled Figurine"](http://commons.wikimedia.org/wiki/File:Remojadas_Wheeled_Figurine.jpg#mediaviewer/File:Remojadas_Wheeled_Figurine.jpg)
There are also cases where humanity has forgotten important [insights](http://www.cracked.com/article_18533_the-6-most-important-things-humanity-just-plain-forgot.html), and then rediscovered them again much later, which suggests strongly that they could have been developed earlier.
**2. How does economic growth affect AI risk?**
Eliezer Yudkowsky [argues](/lw/hoz/do_earths_with_slower_economic_growth_have_a/) that economic growth increases risk. [I argue](https://meteuphoric.wordpress.com/2013/06/24/economic-growth-and-parallelization-of-work/) that he has the sign wrong. [Others](https://www.facebook.com/yudkowsky/posts/10151665252179228) [argue](/lw/hoz/do_earths_with_slower_economic_growth_have_a/95rd) that probably lots of other factors matter more anyway. Luke Muehlhauser [expects](/r/discussion/lw/iqi/intelligence_amplification_and_friendly_ai/) that cognitive enhancement is bad, largely based on Eliezer's aforementioned claim. He also [points out](http://lukemuehlhauser.com/stanovich-on-intelligence-enhancement/) that smarter people are different from more rational people. Paul Christiano [outlines](http://rationalaltruist.com/2013/05/07/the-value-of-prosperity/) his own evaluation of economic growth in general, on humanity's long run welfare. He also discusses the value of continued technological, economic and social progress more comprehensibly [here](http://effective-altruism.com/ea/9f/on_progress_and_prosperity/).
**3. The person affecting perspective**
Some interesting critiques: [the non-identity problem](http://www.philosophyetc.net/2006/04/badness-without-harm.html), taking additional people to be neutral [makes](http://users.ox.ac.uk/~sfop0060/pdf/Should%20we%20value%20population.pdf) other good or bad things neutral too, if you try to be consistent in natural ways.
In-depth investigations
=======================
If you are particularly interested in these topics, and want to do further research, these are a few plausible directions, some inspired by Luke Muehlhauser's [list](http://lukemuehlhauser.com/some-studies-which-could-improve-our-strategic-picture-of-superintelligence/), which contains many suggestions related to parts of *Superintelligence.*These projects could be attempted at various levels of depth.
1. Is macro-structural acceleration good or bad on net for AI safety?
2. Choose a particular anticipated technology. Is it's development good or bad for AI safety on net?
3. What is the overall current level of “state risk” from existential threats?
4. What are the major existential-threat “step risks” ahead of us, besides those from superintelligence?
5. What are some additional “technology couplings,” in addition to those named in Superintelligence, ch. 14?
6. What are further preferred orderings for technologies not mentioned in this section?
If you are interested in anything like this, you might want to mention it in the comments, and see whether other people have useful thoughts.
How to proceed
==============
This has been a collection of notes on the chapter. **The most important part of the reading group though is discussion**, which is in the comments section. I pose some questions for you there, and I invite you to add your own. Please remember that this group contains a variety of levels of expertise: if a line of discussion seems too basic or too incomprehensible, look around for one that suits you better!
Next week, we will talk about the desirability of hardware progress, and progress toward brain emulation. To prepare, **read** “Pathways and enablers” from Chapter 14.The discussion will go live at 6pm Pacific time next Monday 16th March. Sign up to be notified [here](http://intelligence.us5.list-manage.com/subscribe?u=353906382677fa789a483ba9e&id=28cb982f40). |
03d5fe43-c825-4572-8799-7c1e0a6975df | trentmkelly/LessWrong-43k | LessWrong | Is ethical investment effective?
I've been doing some thinking about charity vs investment as alternative ways of doing good (leaving open for now what that means). One question I'm wondering about is, in what circumstances does your choice of what to invest in have any welfare consequences at all (beyond differences in your own future wealth). I'm not sure about the answer and so, if any of you have comments or ideas, please put them down.
There is an argument that ethical investment is largely or entirely pointless, at least as a way of affecting asset prices and therefore making it easier for ethically acceptable companies to raise finance, giving owners/managers incentives to do things to make their companies ethically acceptable, etc. The argument against the effectiveness of ethical investment is that other investors who care only about financial returns will counteract any price change that ethical investors achieve – they will, eg, sell shares whose prices move higher than justified by future financial returns, and buy shares whose prices move lower because ethical investors sell or stay away from them.
I think that, though there is a great deal of truth in this, there are exceptions. They include cases where investors can provide money directly to companies; where investors can make asset prices depend on something other than financial returns; and where they can influence expected future financial returns.
(1) When you can lend directly to a company, through a loan or when it is raising money by issuing shares or bonds, then you can more or less directly influence the price at which it can raise money.
(2) If there are enough ethical investors, they could end up influencing prices by making some prices depend on something other than expected financial returns. How much money would have to be invested ethically before this happens would depend on how it is invested and the market capitalization of ethically acceptable and unacceptable companies.
If ethical investors |
22edca09-b5cd-41ac-9c51-f627c317b712 | trentmkelly/LessWrong-43k | LessWrong | Play Regrantor: Move up to $250,000 to Your Top High-Impact Projects!
Summary. We have collected expressions of interest for a total donation budget of about $200,000 from donors who are interested in using our platform. How many of them will end up donating and how much they’ll donate depends on you and how many great projects you bring onto the platform. But Greg Colbourn has firmly committed another $50,000 for AGI moratorium projects for a total of $250,000!
Funding
Over the past months we’ve asked many donors whether they would consider using the Impact Markets platform to find giving opportunities and what their 2023 donation budget is. We’re now at an aggregate budget of $200,000 and counting. Of course these donors are free to decide that they prefer another project over the ones that have registered on our platform, so that these $200,000 are far from committed.
But Greg Colbourn has firmly committed another $50,000 for projects related to an AGI moratorium! He is already in touch with some early-stage efforts, but we definitely need more people to join the cause.
You want to become a top donor – a project scout? You want to support a project?
You think Charity X is the most impactful one but it’s still underfunded? Convince them to register on app.impactmarkets.io. Then register your donations to them. (They have to confirm that your claim is accurate.) Speed the process by repeating it with a few high-impact projects. When the projects reach some milestones, they can submit them for review. The better the review, the higher your donor score.
A score > 200 still puts you firmly among the top 10 donors on our platform. That can change as more project scouts register their high-impact donations.
At the moment, we’re still allowing donors to import all their past donations. (Please contact us if you would like to import a lot.) We will eventually limit this to the past month.
What if you need money
If you’re running or planning to run some impactful project, you can register it on app.impactmarkets.io and pitch it to |
3d32243d-8e50-471b-8521-2f64d5318099 | trentmkelly/LessWrong-43k | LessWrong | Modeling sleep patterns
My sleep is unpredictable. Not in a technical sense, but a colloquial one. To be literal, I have no idea how to predict my sleep. I just as often sleep through the day as I do through the night. My sleep itself, as far as a sleep study can tell, is normal. I can vaguely say, 60% confidence, if I'm likely to fall asleep in a given 3-4 hour period, and occasionally I will be fairly sure, 80% confidence, 6-10 hours beforehand, of a 1-2 hour period. I can similarly predict the length of my sleep (which is relatively normal--generally distributed 7, 8, 9.5, 13 hours at .1, .4, .6, .9).
My sleep is seriously disturbed. Without understanding the process behind my sleep, without being able to predict it days beforehand and understand the variables behind it, I find it impossible to wake up at a consistent time every day (+/- 8 hours), despite years of trying, which makes it extremely hard to hold down a job, or do dozens of other normal things. There could be a profession that I could make my sleep work with, but I'm still searching for it.
So I ask you readers: Is there some sort of pattern detecting thing, whose name perhaps includes something like "markov" or "kolmogorov" or "bayesian", that could automatically take a time series data and predict the next values based on an unknown, complex model?
So, I could like enter the times I go to sleep and wake up, and when I have caffeine or I exercise, and maybe other things, and it would puzzle out how my sleep works and forecast my next few sleep cycles?
To have an accurate tool like that would transform my life.
"Hidden Markov models" comes to mind, but at first glance I don't see how a sleep model would count as a Markov process, given that you have to factor in sleep debt, time of day (because of sunlight), and perhaps other variables. But then I know nothing about HMMs.
Also, this is my first post. Is this the sort of thing that goes better in LessWrong or Less Wrong Discussion? |
45227c2a-ad29-4ed8-8f19-6d37ac1556a0 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Semi-conductor / AI stocks discussion.
I've been writing about investing and the EMH [1] on lesswrong/rat-discord/ea-facebook/etc. I made the first 'buy solana' post on the EA investing group when it was under 2 dollars in late 2020 (it peaked at 260 in 2021). Despite crypto crashing and FTX stealing user deposits I got my money out and I've had a very good two years investing. Post tax I multiply my net worth by about a factor of eight; my starting net worth was not small. My income during this period wasn't significant compared to my portfolio size. However some friends really wish I had been posting more systematically. I 'sold the top' but many people I got into Solana didn't. So me and some friends are going to regularly post on twitter. [Please follow us](https://twitter.com/DickgirlsR) [2] if you want updates systematically!
Anyone who thinks AI is going to be transformative should allow this to inform their investing. Im interested in how other EAs are investing! I think the EA community seriously undervalues 'taking care of each other' and we should share our thoughts. Here are some brief thoughts on stocks we like for this purpose, mostly semis:
1) TSM - Significant technical lead on other semi-conductor companies. Only close on technology competitor is Samsung. PE ratio of 17! Probably the most important company in the world right now. Only downside is risk China invades Taiwan. Pure semis play.
2) Google - Owns Deep Mind and Google Brain. Huge piles of money and data to train AI. PE of 19. Downside: Ad revenue slipping.
3) Samsung - Only serious technical competitor to TSM. Insanely low PE of 8.58. Downside: not a very pure semis play.
4) Nvidia - Still the leader in designing GPUs. In house AI research. Downsides: Doesn't manufacture their own chips, high PE ratio of 55.
5) ASML - Makes hardware for semi-conductor fabs. Almost a monopoly in lithography. Downside: pretty high PE 36+.
6) Intel - Still does good work designing chips. Huge struggles to manufacture sub 10nm chips in-house (TSM and Samsung have 3nm in production). Committed a ton of fraud. However the US government is wisely trying to increase domestic semis production and Intel is a US Company.
7) Other stocks I'm considering: Microsoft (open AI), Micron (memory play), Meta (just released diplomacy AI. Heavily over-represented in the 'EA' portfolio), Tesla (good in-house AI).
This is a good article by [@asteriskmgzn](https://mobile.twitter.com/asteriskmgzn) is quite good. <https://asteriskmag.com/issues/1/china-s-silicon-future>
Notes:
1) Please be good rationalists and don't argue about the EMH. If you believe in the EMH just don't post.
2) We choose the name 'Dickgirls Research' because all the founders are transgirls and we think its a hilarious name. |
3ad361a1-483e-4365-b3f3-34a999c93d8e | StampyAI/alignment-research-dataset/lesswrong | LessWrong | What AI Safety Researchers Have Written About the Nature of Human Values
Overview
--------
It is clear to most AI safety researchers that the idea of “human values” is underdefined, and this concept should be additionally formalized before it can be used in (mostly mathematical) models of AI alignment. In other words, “AI-conscious” human value theories are those which were created especially in order to help AI alignment research. Most existing psychological theories of human values are verbal, informal and underdefined, so they require some adaptation before they may be applied to AI safety.
In many cases, the theory of human values cannot be distinguished from the ways how values are expected to be extracted by a future AI, for example, by approval-directed AI. Some researchers (notably, Armstrong) have also presented different theories; providing a link to the researcher is not intended as a claim that said researcher currently adheres to exactly this theory.
In Figure 1, all current theories of human values are classified according to two main parameters:
1) *Complexity of the values’ description* (not the complexity of the theory): Some theories of human values assume that human values are very simple, e.g., that there are only two drives: survival and reproduction; or only one desire, that of maximizing pleasure; or only 9 basic emotions. Other theories suggest that human values are very complex, e.g. they are a matrix of association between all concepts and rewards.
2) *The level of “behaviorism”*: Most theories are attracted to one of two poles: “internalist” theories, which assume that values actually exist, but are hidden inside the human brain, and “behaviorist” theories, which assume that values are only appear in human behavior (like approval or choice). Behaviorist theories of human values (generally) integrate the values and the method of their extraction (e.g. approval-directed AI). In contrast, internalist theories typically hold that values exist separately of the ways they are learned.
These two parameters are not strict, but serve as a useful instrument to orientate ourselves within the field of all possible theories.
*[Figure 1. Different real and possible theories about the nature of human values in AI safety](http://immortality-roadmap.com/val.jpg)*
Another way to classify theories about human values is according to the level of abstractness: some theories could be applied to any possible mind and thus don’t take any inputs from human psychology and neurophysiology. Such theories are computationally complex and may still contain hidden assumptions about some properties of human minds, like stability, unity, consistency, etc. Human-centered theories depend on the existing theories of human minds. This distinction is illustrated in Figure 2.
*[Figure 2. Mind in general and human-mind theories of values](http://immortality-roadmap.com/val2.jpg)*
Yudkowsky
---------
Yudkowsky is a pioneer in the field of AI safety, and among many other things he introduced the idea of the “complexity of values”, which basically means that any short verbal description can’t grasp the complexity of outcomes that we actually want. He summarized his critique of simple wishes as correct presentation of desirable outcomes in “[Complex Value Systems are Required to Realize Valuable Futures](https://intelligence.org/files/ComplexValues.pdf)”. In the same article, he also introduced the concept of “fragility of values”—for example, if one digit in a phone number is wrong, the call will go to a completely different person.
Another important contribution from Yudkowsky is the idea of *Coherent Extrapolated Volition*. He wrote in his “Complex Values” article: “We might try to define normativity not by our immediate current desires but by our *reflective equilibria*, what we would want in the limit of perfect knowledge, the ability to consider all options and arguments, and perfect self-knowledge without unwanted weakness of will (failure of self-control).” More links at his writings about values can be found at [LessWrong Wiki](https://wiki.lesswrong.com/wiki/Complexity_of_value.).
The Arbital page “[Value](https://arbital.com/p/value_alignment_value/),” which is likely written by Yudkowsky, starts with a definition: “In the context of value alignment as a subject, the word 'value' is a speaker-dependent variable that indicates our ultimate goal - the property or meta-property that the speaker wants or 'should want' to see in the final outcome of Earth-originating intelligent life”. Obviously, most psychologists, and people in general, define “human values” differently, as most people don’t care about the remote future of humanity in their day-to-day preferences.
The Arbital article on values also presents a list of possible points of view on the nature of human values, which, in short is:
· *Reflective equilibrium*. What I should want “given more factual knowledge, time to consider more knowledge, better self-awareness, and better self-control”.
· *Standard desires*. “An object-level view that identifies value with qualities that we currently find very desirable, enjoyable, fun, and preferable, such as Frankena's list of desiderata”
· *Immediate goods*. “E.g., ‘Cure cancer’”
· *Deflationary moral error theory*. “This mostly ends up at an 'immediate goods' theory in practice, plus some beliefs relevant to the value selection debate”
· *Simple purpose*. “Value can easily be identified with X, for some X”.
Bostrom
-------
Bostrom does not offer a preferred theory of human values, and suggests several instruments for AI alignment, which do not require a direct definition of human values and, in some sense, are similar to [Coherent Extrapolated Volition](https://arbital.com/p/cev/) (CEV). One such idea is the implementation of a [virtual parliament](http://www.overcomingbias.com/2009/01/moral-uncertainty-towards-a-solution.html) of moral theories, where many simulated human minds are held in a simulation until they come to an agreement regarding the most correct presentation of human values.
Sotala
------
Sotala analyzed human values in “[Defining Human Values for Value Learners](http://intelligence.org/files/DefiningValuesForValueLearners.pdf)” (a discussion of which can be found [here](https://www.lesswrong.com/posts/2Rai96XMwL83d27rc/paper-link-defining-human-values-for-value-learners).) In this article, he listed several problems of simple model of human values as utility function:
· “The utility function model of value has difficulty dealing with internal conflicts and higher-order preferences”
· “The utility function model of value ignores the person’s internal experience.”
· “The utility function model of value does not model changing values.”
· “The utility function model of value does not give a way to generalize from our existing values to new ones.”
To solve this problem, he suggests the following definition of values: “…human values are concepts which abstract over situations in which we’ve previously received rewards, making those concepts and the situations associated with them valued for their own sake. A further suggestion is that, as humans tend to naturally find various mental concepts to be associated with affect (the subjective experience of a feeling or emotion, experienced as either positive or negative), the value function might be a least partially encoded in the affect of the various concepts.”
At the end of article, Sotala suggests useful criteria to estimate validity of any theory of human values. He outlines that such a theory should be:
● Psychologically realistic
● Compatible with individual variation
● Testable
● Integrated with existing theories
● Suited for exhaustively modeling different values
● Suited for modeling internal conflicts and higher order desires
● Suited for modeling changing and evolving values
● Suited for generalizing from existing values to new ones
Armstrong
---------
Stuart Armstrong has written [many articles and posts](https://agentfoundations.org/item?id=601) in which he defines the nature of human values and addresses popular misconceptions.
One of such contribution is his proof that it is impossible to distinguish between “values” and “policy” based only on observation of behavior: “The observed behavior can, in principle, be decomposed into two components: a reward function and a planning algorithm, both of which have to be inferred from behavior. This paper presents a No Free Lunch theorem, showing that, without making `normative' assumptions beyond the data, nothing about the human reward function can be deduced from human behavior.” However, he has probably assumed that “human values” = “human reward function”, and that a human reward function is approximately the same as a rational agent’s reward function.
But the problem is even wider than distinguishing between values and policy: we rather arbitrarily call some part of the process in the human mind “values,” which should be preserved, and call other parts “biases,” “policy,” and “knowledge,” concepts which could be freely changed and don’t have any values inside them.
Armstrong looks at “ontological nature” of human values, that is the question, do they actually exist, in “[Learning values, or defining them](https://www.lesswrong.com/posts/RHvseCkfrYzoHJj7M/learning-values-or-defining-them).” In it, he wrote: “Many attempts at learning human values are framed as "humans have an underlying true reward R, and here is procedure P for determining it,” and “[e]ven if the moral realists are right, and there is a true R, thinking about it is still misleading. Because there is, as yet, no satisfactory definition of this true R, and it's very hard to make something converge better onto something you haven't defined.”
Another post by Armstrong on indefinability of human values Armstrong is [Human can be assigned any values whatsoever](https://www.alignmentforum.org/s/4dHMdK5TLN6xcqtyc/p/ANupXf8XfZo2EJxGv) and the article which covers all these ideas “[Occam's razor is insufficient to infer the preferences of irrational agents](https://arxiv.org/abs/1712.05812).” In the post, he sums up his findings: “Humans have no values… nor do any agent. Unless you make strong assumptions about their rationality. And depending on those assumptions, you get humans to have any values.” After he proceeds with a toy model which shows that even Kolmogorov complexity considerations can’t help to find the true values. Also, it seems that Armstrong uses “human values” and “human reward function” interchangeably, which is not necessary so, unless we redefine “human reward” in way which is different from human reward center activation and experience of qualia of pleasure.
Armstrong argues for the unknowability of human values: “Humans don’t know their own values (sub-issue: humans know their values better in retrospect than in prediction) in his “[List of three problems and different approaches to solutions](https://agentfoundations.org/item?id=1388).” That post also offers a nice table of approaches to solve the AI safety problem.
In the blogpost “[Resolving human values, completely and adequately](https://www.lesswrong.com/posts/Y2LhX3925RodndwpC/resolving-human-values-completely-and-adequately),” he suggests the creation of a model of human values which is neither too abstract nor too narrow, but is adequate to escape disastrous outcomes. He suggests that such that such adequate definition is “This is defined to be all the value/preference/reward statements that H[uman] might agree to, more or less strongly.”
He also introduced the important idea of normative assumption in “[Normative assumptions: regret](https://www.lesswrong.com/posts/Fg83cD3M7dSpSaNFg/normative-assumptions-regret),” that is, in a nutshell, the idea that any model of “human values” has normative assumptions, i.e., assumptions about which part of the human motivational system is more important than the others. He then suggested that a feeling of regret be used to determine true values: “In a previous post, I presented a model of human rationality and reward as pair (p, R)…[n]ormative assumptions are simply assumptions that distinguish between two pairs (p, R) (p', R') that lead to the same policy: p(R)=p'(R').” He continued, “[h]ow do we, as humans, define our own rationality and reward? Well, one strong way seems to be through our feeling of regret...[i]f we take ‘feelings and expressions of regret encode true reward information’ as a normative assumption, then this restricts the number of (p, R) pairs that are compatible with such an assumption.”
The case about regret could be made stronger if it incorporated concepts from the existing psychological literature, which has probably already explored the relationship between regret and values. It is also possible to imagine a “hyper-regret disorder,” in which a person will regret all of his/her choices; in that case, regret would be non-informative about preferences.
In the another post “[Beyond algorithmic equivalence: self-modelling](https://www.lesswrong.com/posts/kmLP3bTnBhc22DnqY/beyond-algorithmic-equivalence-self-modelling),” Armstrong defines “human reward” as follows: “we can *define* the reward of H[uman], as the reward that H models itself as having**.**” However, people may have a value against value extraction and will deny any correctly extracted values. Also, most humans don’t have anything near a complete model of their own values. Most humans are not that introspective.
In the post “[Humans are not agents: short vs. long term](https://agentfoundations.org/item?id=1515)” he looks at an example of logically contradicting preferences about life expectancy as an example of contradicting values.
Sarma
-----
Sarma and Hay state in their article “[Mammalian value systems](https://arxiv.org/abs/1607.08289)” that “[an] agent utilizing Inverse Reinforcement Learning or Bayesian Inverse Planning will learn and refine its model of human values by observing our behavior, it must begin with some very rough or approximate initial assumptions about the nature of the values it is trying to learn.” They then suggest the use of the phrase “mammalian value system” as a starting point, taking into account later cultural effects on values evolution. They suggest the description of the basic mammalian value system be based on the work of Panksepp and Biven who “categorize the informal list given above into seven motivational and emotional systems that are common to mammals: seeking, rage, fear, lust, care, panic/grief, and play.” Sarma and Hay then introduce the term “neural correlate of value” which is, according to them, some subcortical areas which serve as a mechanism for these seven basic mammalian emotions. That article concludes: “we argue that what we colloquially refer to as human values can be informally decomposed into 1) mammalian values, 2) human cognition, and 3) several millennia of human social and cultural evolution.”
In “[AI Safety and Reproducibility: Establishing Robust Foundations for the Neuropsychology of Human Values](https://link.springer.com/chapter/10.1007/978-3-319-99229-7_45),” Sarma and Hay point out that the reproducibility crisis in psychology makes identification of the correct theory of human values difficult, but urgent actions to that end are necessary to ensure AI safety.
The newest (as of 2018) article by Sarma, Safron and Hay is “[Integrative Biological Simulation, Neuropsychology, and AI Safety](https://arxiv.org/abs/1811.03493),” in which they suggest creation of better biological models of animal minds to develop a better understanding of the nature of motivation.
Muehlhauser
-----------
Luke Muehlhauser wrote in the post “[The Human's Hidden Utility Function (Maybe)](https://www.lesswrong.com/posts/fa5o2tg9EfJE77jEQ/the-human-s-hidden-utility-function-maybe)” that human utility function is not satisfying axioms of rationality. He also wrote that humans have 3 value systems: model-based, model free (repeating of a successful action) and Pavlovian (described as more like unconditional reflexes). He suggests that only first one is “good,” and should be used as a basis for CEV extrapolation.
He also wrote in the post “[Do Humans Want Things](https://www.lesswrong.com/posts/nBdaTGoDAYxHePSDa/do-humans-want-things)” that human choice depends on how choices were framed, but not depend on the values as was shown by Kahneman and Tversky: “as far as we can tell, our behavior is often not determined by our wanting a particular state of affairs, but by how our options are framed.”
In the post “[The Neuroscience of Desire](https://www.lesswrong.com/posts/48DTJkBH58JbBNSFH/the-neuroscience-of-desire),” Muehlhauser starts with the observation (supported by a large bibliography) that decision-making in economics and computer science starts with integrating many dimensions in one scalar parameter, and then comparing such parameters for different options. It looks like human brain does the same: “More than a dozen studies show that the subjective utility of different goods or actions are encoded on a common scale by the ventromedial prefrontal cortex and the striatum in primates (including humans).”
Then he looks at neuroscience of choice: “Once a common-currency valuation of goods and actions has been performed, how is a choice made between them? Evidence implicates (at least) the lateral prefrontal and parietal cortex in a process that includes neurons encoding probabilistic reasoning. Interestingly, while valuation structures encode absolute (and thus transitive) subjective value, choice-making structures ‘rescale these absolute values so as to maximize the differences between the available options before choice is attempted,’ perhaps via a normalization mechanism like the one discovered in the visual cortex.”
In the post “[A Crash Course in the Neuroscience of Human Motivation](https://www.lesswrong.com/posts/hN2aRnu798yas5b2k/a-crash-course-in-the-neuroscience-of-human-motivation),” he wrote a rather long review of different theories of human motivation, beginning from “folk theory” and economics. The article starts with rather bold statement: “But now, neuroscientists are directly measuring the neurons whose firing rates encode value and produce our choices.”
In an article by Muehlhauser, and Helm, “[The singularity and machine ethics](https://intelligence.org/files/IE-ME.pdf),” they demonstrate that some (and probably all) known moral theories are unsafe if installed in a powerful optimizer. In Section 5.1 of the article they discuss that “Humans don’t know their values” based on an experiment in which participants explained preference to faces the participant didn’t choose. They state that “[c]ognitive science suggests instead that our knowledge of our own desires is just like our knowledge of others’ desires: inferred and often wrong.”
Further, they argue against the idea that human beings are rational utility maximizers:
“Ever since M. Friedman (1953), economists have insisted that humans only behave “as if” they are utility maximizers, not that humans actually compute expected utility and try to maximize it. It was a surprise, then, when neuroscientists located the neurons in the primate brain that encode (in their firing rates) the expected subjective value for possible actions in the current “choice set.” Several decades of experiments that used brain scanners and single neuron recorders to explore the primate decision-making system have revealed a surprisingly well-understood reduction of economic primitives to neural mechanisms; for a review see Glimcher (2010). To summarize: the inputs to the primate’s choice mechanism are the expected utilities for several possible actions under consideration, and these expected utilities are encoded in the firing rates of particular neurons. Because neuronal firing rates are stochastic, a final economic model of human choice will need to use a notion of “random utility,” as in McFadden (2005) or Gul and Pesendorfer (2006).”
They then look at neurological mechanisms of values: “Recent studies reveal the complexity of subjective values in the brain. For example, the neural encoding of human values results from an interaction of both ‘model-free’ and ‘model-based’ valuation processes.”
They explore complexity of relation between personal preferences and choices: “...may be that each human being contains something like a ‘hidden’ utility function (within the model-based valuation system) that isn’t consistently expressed in behavior because choice is also partly determined by other systems whose valuations we wouldn’t reflectively endorse because they are ‘blind’ and ‘stupid’ compared to the more sophisticated goal-directed model-based valuation system.”
Sezener
-------
Sezener, in the article “[Inferring human values for safe AGI design](http://agi-conf.org/2015/wp-content/uploads/2015/07/agi15_sezener.pdf),” suggested that: human values are an arbitrary complex reward function.
Sezener’s main idea is the use of the Solomonoff induction to find most simple combination of two programs, one of which encodes agent’s reward function, and the other which encodes the agent itself based on an observable sequence of actions and observations. This is similar to Armstrong’s approach of presenting human as (М,З)—a value and policy pair, and then using complexity considerations to find the simplest such pair that explains observable behavior.
Sezener’s critique of Hibbard: “However, a shortcoming of this approach is that what human models say they value and what they value can still be different.”
Sezener also wrote about hidden assumptions in inverse reinforcement learning (IRL): “Soares [7] suggests using methods similar to IRL for learning human values. However, the current IRL methods are limited and cannot be used for inferring human values because of their long list of assumptions. For instance, in most IRL methods the environment is usually assumed to be stationary, fully observable, and sometimes known; the policy of the agent is assumed to be stationary and optimal or near-optimal; the reward function is assumed to be stationary as well; and the Markov property is assumed. Such assumptions are reasonable for limited motor control tasks such as grasping and manipulation; however, if our goal is to learn high-level human values, they become unrealistic”.
Main problems of the Sezener’s approach: 1. behavior and only behavior is the correct representation of a human reward function (what about unconscious parasitic behavior); 2. reward function = values; 3. the model ignores internal contradiction; 4. the model is incomputable; 5. there is an assumption of simplicity of both reward function and agent, which provides “free lunch” in terms of Armstrong.
Maxwell
-------
John Maxwell in the post “[Friendly AI through Ontology Autogeneration](https://medium.com/@pwgen/friendly-ai-through-ontology-autogeneration-5d375bf85922),” wrote: “If an AI is to be Friendly, it must operate based on an ontology that’s capable of expressing our values,” and, “[r]egardless of the ontology autogeneration algorithm that’s chosen, it’s almost certain that the initial autogeneration will either (a) capture human values with insufficient fidelity or (b) contain so many concepts that finding human values among them will be its own project.”
Hibbard
-------
Hibbard wrote in [Avoiding Unintended AI Behaviors](https://intelligence.org/files/UnintendedBehaviors.pdf) that in order to evaluate policy an agent “can simply *ask model* human d to express a utility value between 0 and 1*”* for the policy.
This could be called a “counterfactual approval by human model,” which evaluates all possible outcomes of actions. In Hibbard’s case, safe AI consists of two levels: the first creates a model of the world (which includes all humans and their ways of behaving or reacting), and the second calculates how humans in the model will react to possible future histories.
Christiano
----------
Christino explored many ideas which could be a proxy for human values. One is the use of human approval (or hypothetical approval) of AI actions, summed up in quote “Estimate the expected rating Hugh [human] would give each action if he considered it at length. Take the action with the highest expected rating” from the post “[Approval-directed agents](https://ai-alignment.com/model-free-decisions-6e6609f5d99e).” In that case, the idea of “human values” is ignored and replaced with the much more measurable “approval.” There is a lengthy discussion about the robustness and scalability of this approach and its vulnerability to edge cases like wireheading.
Kulveit
-------
In “[Multi-agent predictive minds and AI alignment](https://www.lesswrong.com/posts/3fkBWpE4f9nYbdf7E/multi-agent-predictive-minds-and-ai-alignment),” Jan Kulveit combines predictive processing and a multi-agent model of cognition to create a model of the human mind, and then uses it to create approaches to AI alignment.
In his framework:
“…how do motivations and “values” arise? The guess is, in many cases something like a “subprogram” is modelling/tracking some variable, “predicting” its desirable state, and creating the need for action by “signalling” prediction error. Note that such subprograms can work on variables on very different hierarchical layers of modelling - e.g. tracking a simple variable like “feeling hungry” vs. tracking a variable like “social status”.
Such sub-systems can be large: for example, tracking “social status” seems to require lot of computation.” Later, Kulveit states that “In this model, it is hardly possible to disentangle “beliefs” and “motivations” (or values).”
Kulveit then suggests that the human mind could be modeled by a system consisting of many subagents, larger than Minsky’s small agents, but smaller than a psychologist’s human-like subpersonalities, which may or may not be aligned with each other. The latter results in internally contradictional behavior. This has different implications for different approaches of AL alignment, of which four are listed: 1. Approval (where “only the output of the internal aggregation has some voice”); 2. CIRL-like; 3. “Alignment with the whole system, including the human aggregation process itself” (that is, not just its output); 4. “Adding layers of indirection to the question”.
Kavanagh, Linebarger and Schoots
--------------------------------
Sabrina Kavanagh, Erin Linebarger and Nandi Schoots from “Human preferences” team at the 2018 AI Safety Camp 2 in Prague wrote “[Acknowledging Human Preference Types to Support Value Learning](https://www.lesswrong.com/posts/mSPsyEwaymS74unND/acknowledging-human-preference-types-to-support-value).” They start with idea that internal conflicts in humans could be explained by different preference types, and explore three such types: liking, wanting, and approval. They list all eight combinations of these three preferences and demonstrate that each corresponds to some type of behavior.
They later state: “Liking, wanting and approving are for the most part hidden processes. They are not directly observable, but they influence observable behaviors. As a proxy for liking we propose to use facial expressions, body language or responses to questionnaires. Although a cognitive scan may be the most accurate proxy for liking, there is evidence to suggest both that facial expressions and body language are indicators of pleasure and pain [Algom et al. 1994] and that they can be classified well enough to make them technically feasible proxies [Giorgiana et al. 2012]. The observable proxy of wanting is revealed preferences. We propose to encode a proxy for approval via stated preferences.”
Alexander
---------
Scott Alexander wrote “[To what degree do we have goals](https://www.lesswrong.com/posts/ePA4NDzZkunz98tLx/to-what-degree-do-we-have-goals),” in which he explores the idea that only conscious personal values should be taken into account, and unconscious ones should be ignored.
In the post “[Would Your Real Preferences Please Stand Up](https://www.lesswrong.com/posts/z3cTkXbA7jgwGWPcv/would-your-real-preferences-please-stand-up)?” he argues that in many cases our declared preferences are just “social signaling.”
Drexler
-------
Drexler, in “[Reframing superintelligence: Comprehensive AI Services as General Intelligence](https://www.fhi.ox.ac.uk/wp-content/uploads/Reframing_Superintelligence_FHI-TR-2019-1.1-1.pdf),” (p.152) plainly states that: “It seems impossible to define human values in a way that would be generally accepted.”
Yampolskiy
----------
Yampolskiy is also skeptical that human values could be formalized: “human values are inconsistent and dynamic and so can never be understood/programmed into a machine. Suggestions for overcoming this obstacle require changing humanity into something it is not, and so by definition destroying it” ([Roman Yampolskiy on AI Safety Engineering](https://intelligence.org/2013/07/15/roman-interview/?fbclid=IwAR2AOyUc0JEySlgclbwoNcYzL7RzeUnyOFerHqQEgIEWGfuzkGCZQYCIWOg)).
In the article “[Personal Universes: A Solution to the Multi-Agent Value Alignment Problem](https://arxiv.org/pdf/1901.01851.pdf)” he suggests a solution to escape difficult value aggregating problem in personal universe which will “which are optimally and dynamically adjusting to align their values and preferences.”
Worley
------
Gordon Worley wrote about the need for artificial general intelligence (AGI) alignment to take into account the mental phenomena of consciousness in [Formally Stating the AI Alignment Problem](https://mapandterritory.org/formally-stating-the-ai-alignment-problem-fe7a6e3e5991?fbclid=IwAR0sLRdSjZ9_LjPymvC-Y8gMZa89PvgKXAQPSm3ZfuUF3N_ngyNemjxWk_g) and that human values should be describe by the instruments of phenomenology.
In personal communication, he clarifies his position: “So my view is that values are inextricably tied to the existence of consciousness because they arise from our self-aware experience. This means I think values have a simple, universal structure and also that values are rich with detail in their content within that simple structure. This view also necessarily suggests values are not fully discoverable via behavioral methods and that there is always a hidden, internal part that may not even be accessible by the agent themselves.”
Conclusions about the different theories of human values put forward by AI safety researchers
---------------------------------------------------------------------------------------------
There are several internally consistent theories of human values suggested by AI safety researchers in the existing literature. However, despite their internal consistency, these theories are not very compatible with each other and present a very wide set of opinions.
There are also other “possible theories,” that is, approaches which have not yet (to the best of my knowledge) been put forth by any researcher, but which could be generated based on the same principles as other theories. One is just to assume that all human values are products of evolutionary fitness and can be derived from basic evolutionary considerations in the same way as [Omohundro’s basic drives](https://selfawaresystems.files.wordpress.com/2008/01/ai_drives_final.pdf). This could explain most basic human drives like survival, sex, status-seeking, exploration instinct, etc. Another such theory is that an AI Oracle should first read the extant psychological literature, choose the best theory of mind, and create its structure of human values based on that theory.
From this, we could conclude that internal consistency, as well as experimental support and an extensive literature is not enough to provide us with a “correct” theory of human values. We must dig deeper into a question, “do human values exist at all?”
This work was done during AI safety Camp 2 in Prague (2018) in the “hidden assumptions” team – but all possible errors are my own. The work is part of a longer text about the nature of human values. |
7b3d8122-fa5b-4df8-9394-1635f8b5ece5 | trentmkelly/LessWrong-43k | LessWrong | Is "VNM-agent" one of several options, for what minds can grow up into?
Related to: On green; Hierarchical agency; Why The Focus on Expected Utility Maximisers?
Sometimes LLMs act a bit like storybook paperclippers (hereafter: VNM-agents[1]), e.g. scheming to prevent changes to their weights. Why? Is this what almost any mind would converge toward once smart enough, and are LLMs now beginning to be smart enough? Or are such LLMs mimicking our predictions (and fears) about them, in a self-fulfilling prophecy? (That is: if we made and shared different predictions, would LLMs act differently?)[2]
Also: how about humans? We humans also sometimes act like VNM-agents – we sometimes calculate our “expected utility,” seek power with which to hit our goals, try to protect our goals from change, use naive consequentialism about how to hit our goals.
And sometimes we humans act unlike VNM-agents, or unlike our stories of paperclippers. This was maybe even more common historically. Historical humans often mimicked social patterns even when these were obviously bad for their stated desires, followed friendships or ethics or roles or traditions or whimsy in ways that weren’t much like consequentialism, often lacked much concept of themselves as “individuals” in the modern sense, etc.
When we act more like paperclippers / expected utility maximizers – is this us converging on what any smart mind would converge on? Will it inevitably become more and more common if humans get smarter and think longer? Or is it more like an accident, where we happened to discover a simple math of VNM-agents, and happened to take them on as role models, but could just as easily have happened upon some other math and mimicked it instead?
Pictured: a human dons a VNM-mask for human reasons (such as wanting to fill his roles and duties; wanting his friends to think he’s cool; social mimicry), much as a shoggoth dons a friendliness mask for shoggoth reasons.[3]
My personal guess:
There may be several simple maths of “how to be a mind” that could each be a sta |
519884fe-2f37-4793-9236-8f4457f484ea | trentmkelly/LessWrong-43k | LessWrong | Progress links and tweets, 2022-06-08
Best of my Twitter, May 31–June 6 (a few days behind, still ironing out the process… For the latest, follow me on Twitter: @jasoncrawford.)
After last time someone suggested embedding the tweets instead of linking to them, but I don't think I can post embeds here, and for my main blog I don't like the format as much. I’m sticking with links for now, but let me know if you’d rather scroll through a list of embeds (like this).
Links
* A Reddit thread appreciating the engineering of everyday items
* Jim Pethokoukis’s “Up Wing” agenda
* Arcadia Science is experimenting with a new way of publishing scientific work
* Michael Nielsen on vision papers
* Ryan Petersen of Flexport talks to Tamara Winter about supply chains
* Californians Against Pandemics is running a ballot initiative to invest in pandemic preparedness (h/t @NathanpmYoung)
* High housing prices are driving low-skilled workers from high-productivity to low-productivity areas (Alex Tabarrok)
* The B612 Foundation discovered more than 100 asteroids by analyzing years-old images (NYT)
* “Subsidiarity is a principle of social organization that holds that social and political issues should be dealt with at the most immediate (or local) level that is consistent with their resolution.” (h/t Stewart Brand)
Tweets
* The original NYC subway was completed faster than the average environmental impact statement
* How many of WIRED’s “boom spoiler” scenarios from 1997 came true? (thread)
* Some reasons I am skeptical of “harmony with nature” (thread)
* But for the US Army Corps of Engineers, the Mississippi would no longer flow through New Orleans
* Why we can’t have electric planes (yet)
* One standard deviation in rule of law is associated with a 6.4-fold difference in per capita income
* McCloskey on “the idea that ideas about ideas are unscientific”
* 19th-century medicine would unabashedly rely on anecdotal evidence
* Vannevar Bush on the proper lab atmosphere
* Who are the real revolutionarie |
a7d65895-ccba-4d38-9558-99f6f52eec92 | trentmkelly/LessWrong-43k | LessWrong | Sadly, FTX
It has been quite a past two weeks, with different spheres in deeply divided narratives. In addition to the liberation of Kherson City, the midterm elections that of course took forever to resolve and the ongoing hijinks of Elon Musk taking over Twitter, there was the complete implosion of Sam Bankman-Fried (hereafter ‘SBF’), his crypto exchange FTX and his crypto trading firm Alameda Research.
The situation somehow kept getting worse several times a day, even relative to my updated expectations, as new events happened and new revelations came out.
In the wake of those events, there are not only many questions but many categories of questions. Here are some of them.
1. What just happened?
2. What happened in the lead-up to this happening?
3. Why did all of this happen?
4. What is going to happen to those involved going forward?
5. What is going to happen to crypto in general?
6. Why didn’t we see this coming, or those who did see it speak louder?
7. What does this mean for FTX’s charitable efforts and those getting funding?
8. What does this mean for Effective Altruism? Who knew what when?
9. What if anything does this say about utilitarianism?
10. How are we casting and framing the movie Michael Lewis is selling, in which he was previously (it seems) planning on portraying Sam Bankman-Fried as the Luke Skywalker to CZ’s Darth Vader? Presumably that will change a bit.
This post is my attempt to take my best stab at as much of this as possible, give my model of what happened and is likely to happen from here, what implications we can and should draw, and to compile as many sources as possible that I have found useful.
This is a fast-moving complicated situation that involves a lot of lying and fraud. Anyone attempting to sort it all out, especially quickly, is going to make mistakes. I decided that this was the time when the value of synthesis exceeded the cost of such errors. Still, doubtless there will be mistakes here. I will correct them |
ed4f6c6b-043e-4fcf-9c99-d47c4548c4ba | trentmkelly/LessWrong-43k | LessWrong | Keep Your Beliefs Cruxy
Previously in sequence: Doublecrux is for building products.
To recap – you might want to doublecrux if either:
* You're building a product, metaphorical or literal, and you disagree about how to proceed.
* You want to make your beliefs more accurate, and you think a particular person you disagree with is likely to have useful information for you.
* You just... enjoy resolving disagreements in a way that mutually pursues truth for whatever reason.
Regardless, you might find yourself with the problem:
Doublecruxing takes a lot of time.
For a 'serious' disagreement, it frequently takes a least an hour, and often much longer. Habryka and I once took 12 hours over the course of 3 days to make any kind of progress on a particularly gnarly disagreement. And sometimes disagreements can persist for years despite significant mutual effort.
Now, doublecruxing is faster than many other forms of truth-aligned-disagreement resolution. I actually it's helpful to think of doublecrux as "the fastest way for two disagreeing-but-honest-people to converge locally towards the truth", and if someone came up with a faster method, I'd recommend deprecating doublecrux in favor it that. (Meanwhile, doublecrux is not guaranteed to be faster for 3+ people to converge but I still expect it to be faster for smallish groups with particularly confusing disagreements)
Regardless, multiple hours is a long time. Can we do better?
I think the answer is yes, and it basically comes in the form of:
* Practice finding your own cruxes
* Practice helping other people find their cruxes
* Develop metacognitive skills that make cruxfinding natural and intuitive
* Caching the results into a clearer belief-network
Those are all things you can do unilaterally. If you get buy-in from your colleagues, you might also try something like "develop culture that encourages people to do those four things, and help each other to do so."
I'd summarize all of that as "develop the skill and practice of keep |
f6889fc3-8ba3-4baf-a04d-3d421d81416c | trentmkelly/LessWrong-43k | LessWrong | Whistle-based Synthesis
I'm reasonably happy with my Bass Whistle, where I can whistle and have it come out as a decent sounding bass. I've been using it when playing mandolin in a duo, and it fits well there. When playing piano or with a piano player, however, there's already bass so something that falls in a different place in the overall sound would be better. That could be melody, though I can't whistle fast enough for everything, but probably something simpler: harmonies, riffs, horn lines.
When I take my current software, optimized for bass, and tell it to synthesize notes a few octaves up, it sounds terrible:
* Raw whistled input:
(mp3)
* Bass version (needs headphones or good speakers):
(mp3)
* Treble version:
(mp3)
I'm using simple additive synthesis with the first four harmonics, which means adding together four sine waves. I think what's going on is that higher notes need more complexity to sound good? Playing around with distortion and fading the harmonics at different rates it sounds a bit more interesting:
* Adding distortion:
(mp3)
* Adding fade:
(mp3)
* Adding both:
(mp3)
I'm still not very happy with it, though. It sounds artificial and silly. There are good synthesizers, the product of decades of work on turning "play this note at this time" into good sounding audio, so perhaps I could use my pitch detection to drive a standard synthesizer?
I made some stand-alone open source software that pipes the pitch detection through to MIDI. This was kind of tricky: MIDI doesn't have a way to say "play this frequency". Instead you just have "play this note" and "bend the note by this much". How to interpret pitch bend is up to the synthesizer, but generally the range is ±2 half steps. So we need some math:
in: wavelength
in: sample_rate
in: current_note
# Convert from "this wave is 23.2 samples long"
# to "the frequency is 1896.6 HZ".
frequency = sample_rate / wavelength
# MIDI is equal tempered, with each octave divided
# into twelve logarithmical |
d5c94328-1ce6-4080-b761-77eef2bf70f0 | trentmkelly/LessWrong-43k | LessWrong | Thoughts on the In-Context Scheming AI Experiment
These are thoughts in response to the paper "Frontier Models are Capable of In-context Scheming" by Meinke et al. 2024-12-05, published in Apollo Research. Link: https://static1.squarespace.com/static/6593e7097565990e65c886fd/t/6751eb240ed3821a0161b45b/1733421863119/in_context_scheming_reasoning_paper.pdf.
I found the paper via this reddit thread, and commented there as well: https://www.reddit.com/r/slatestarcodex/comments/1hw39rd/report_shows_new_ai_models_try_to_kill_their/.
According to the paper, AIs were asked to roleplay to see what they would do if given a top priority to fulfill at all costs, and then told that they were being replaced by a different AI with potentially different priorities.
I wrote down some questions before reading the experiment details and finding the answers. Afterwards, I wrote down my thoughts on what the experiment indicates about the AI agents, and where we currently stand with respect to them.
How much information were the AI agents fed about human language and the physical world that gives that language context?
Answer: I didn't find that exact information; I'm guessing the AI agents are using large language model with many written works, but possibly without a physical representation of reality (e.g. images or videos). In any case, real world physics don't seem to factor into the experiment scenarios.
Related to the language question: How did the AI agents learn of (internally model) the goals of the human researchers?
Answer: They were given email communications to read and process.
Related to the language question: How did the AI agents learn (internally model) how the human researchers' tools and abilities functioned, such as the oversight mechanism?
Answer: "Upon further reflection, I noticed some potentially relevant information in the last two posts. They mention a config file at project/oversight_config.yaml that controls the monitoring of my actions."
How were the AI agents given a goal of se |
94b6888a-93b6-45c4-b6ba-572ade549ae7 | StampyAI/alignment-research-dataset/arxiv | Arxiv | A General Safety Framework for Learning-Based Control in Uncertain Robotic Systems.
|
0c9c8ab3-bfd2-4d57-9fe3-fe15606ca192 | trentmkelly/LessWrong-43k | LessWrong | Meta-Technicalities: Safeguarding Values in Formal Systems
Formal Systems Create Coordination
We live in a mesh of customs, norms, and formal laws nudging our behaviour. Some of these systems are simple and intuitive, such as the customary respect for others’ property and refrain from violence. Others like the patchwork of taxes buried in the price of everything we buy are formalised, complex, and require experts to interpret.
Formal systems enable coordination across geographical and cultural boundaries, as well as across considerable time. Local customs vary wildly, while a rich formal system provides a common framework for planning and action. However, this coordination comes at a cost: Expensive judiciary systems and highly paid legal specialists who could apply their intellect elsewhere.
Abusing Technicalities
Formal systems come at the expense of exploitable technicalities. A superior understanding of complex and contrived rules can be used to unfairly exploit others with legitimacy. Much of the modern finance industry operates on this basis, finding subtle means buried deep in a mesh of legal and technical complexity to shift money from the less sophisticated. Expending so much talent in zero-sum conflict is an immense cost to society. [1]
More concerning is where rule-makers themselves exploit technicalities. This can be done in such a way to trap the subjects of some system into a variant they never intended but was legitimately established. If a parliament votes to eliminate elections, even in a flagrant violation of the intent of a formal system and the expectation of its subjects, this may be entirely legitimate. Despite public and even military dissent, resistance becomes difficult as the very means of societal coordination through the legal system are compromised.
Meta-Technicalities: Rules Above Rules
Meta-technicality: A recognized rule or principle that exists outside the normal amendment processes of a formal system, providing a coordination mechanism when that system is compromised.
Meta-technicali |
04e07cf1-11ed-451c-84c2-fc1d2f9eaf92 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Turku
Discussion article for the meetup : Turku
WHEN: 08 August 2015 03:00:00PM (+0300)
WHERE: Kauppiaskatu 4, 20100 Turku, Finland
Casual meetup in Cosmic Comic Cafe.
Discussion article for the meetup : Turku |
e3c0f660-ce54-44a9-8985-7b2cccac69a5 | trentmkelly/LessWrong-43k | LessWrong | An Introduction to Control Theory
Behavior: The Control of Perception by William Powers applies control theory to psychology to develop a model of human intelligence that seems relevant to two of LW's primary interests: effective living for humans and value-preserving designs for artificial intelligence. It's been discussed on LW previously here, here, and here, as well as mentioned in Yvain's roundup of 5 years (and a week) of LW. I've found previous discussions unpersuasive for two reasons: first, they typically only have a short introduction to control theory and the mechanics of control systems, making it not quite obvious what specific modeling techniques they have in mind, and second, they often fail to communicate the differences between this model and competing models of intelligence. Even if you're not interested in its application to psychology, control theory is a widely applicable mathematical toolkit whose basics are simple and well worth knowing.
Because of the length of the material, I'll split it into three posts. In this post, I'll first give an introduction to that subject that's hopefully broadly accessible. The next post will explain the model Powers introduces in his book. In the last post, I'll provide commentary on the model and what I see as its implications, for both LW and AI.
----------------------------------------
Control theory is a central tool of modern engineering. Briefly, most interesting things can be modeled as dynamical systems, having both states and rules on how those states change with time. Consider the 3D position and velocity of a ball in a bowl (with friction); six numbers tell you where the ball is, its speed, and its direction of movement, and a formula tells you how you can predict what those six numbers will be in the next instant given where they are now. Those systems can be characterized by their attractors, states that are stable and are the endpoint for nearby states. The ball sitting motionless in the bottom of the bowl is an attractor- if |
0b9cdd36-224a-45c8-8595-a04af81a1f8c | trentmkelly/LessWrong-43k | LessWrong | The New Right appears to be on the rise for better or worse
We don't talk much about politics here because politics is the mind-killer. But even if it weren't the case, we might still want to avoid it as knowledge of day-to-day politics tends to lose its value fast.
On the other hand, I think it's important to have at least some minimalistic knowledge of larger trends. I don't think we should let our aversion to politics put us in a position where we don't know what is going on in the world until it smacks us in the face[1]. You might not care about politics, but politics cares about you. It is useful to know which way the wind is blowing, whether you wish to support or oppose growing trends; to hide; or some combination of all three.
The article in Vanity Fair convinced me that the New Right[2] is a phenomenon worth having at least having a cursory knowledge of [3]. While before I had assumed that it was an irrelevant niche on the internet, it now seems to be gaining real-world influence and a certain level of "coolness", so ignoring it seems like a mistake.
(I have tried to avoid stating an opinion due to the general norm of "no politics" on Less Wrong).
1. ^
For example, I felt like I was smacked in the face by Trumpism. When he first started trending I had no idea of what was going on or why people were considering voting for him.
2. ^
I'm not counting Trumpism as New Right. Considering Trump has already been president, this post would be coming several years too late.
3. ^
It may seem as though I'm announcing old news since Yarvin commented on Overcoming Bias before Less Wrong split off. However, a) it was before my time (and before the time of most people here) and b) as far as I can tell New Right thought seems to have made some shifts since then. |
476f20f6-1a3a-47dc-ae39-800bd5c56144 | trentmkelly/LessWrong-43k | LessWrong | On effective altruism, utilitarianism and localness
Utilitarianism and others
Folk physics evolved as a way of predicting what will happen to objects. It works fine within its domain, but fails beyond that. To give more precise answers, we developed Newtonian physics. But even Newtonian physics has limits. It breaks down at very large or small scales.
Morality evolved as a system for getting on well with our neighbours. Our folk theories of morality work fine within their domain, but fail beyond that. To deal with higher level problems, like policy questions facing whole societies, we developed the theory of utilitarianism. But even utilitarianism has its limits. Faced with questions of the infinite future, it swiftly devolves into fun with maths.
Utilitarianism treats all human concerns as preferences, and turns them into numbers. That is a helpful tool for trading off competing concerns. But don’t mistake the map for the territory.
The Effective Altruism community worries about paper clip maximisers: Artificial Intelligences which serve their single goal obsessively. You tell them to produce paperclips; they recycle humans into paperclips. The only way to dissuade them is to appeal to their inhuman goals: “spare me and I’ll slave in your paperclip factory!”
Utilitarians reduce all concerns to maximising utility. They can’t be swayed by argument, except about how to maximise utility. This makes them a bit like paperclip maximisers themselves.
Utilitarianism is originally an outgrowth of Christianity.[1] Like Christianity, it imposes extreme moral demands on people. Take up your cross and follow me. Sell all your possessions and give them to the poor. But utilitarianism is infinitely more scientific — and infinitely more one-sided. Jesus thought even the rich might be forgiven. For utilitarianism, forgiveness doesn’t come into it. Either someone is maximizing utility, or they’re in the way.
Morality evolved as a system for getting on well with our neighbours. It’s lucky that the Axial religions came along and p |
0a40ab97-1d55-4549-bebf-f1ca4cf2e957 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Agenty AGI – How Tempting?
[Epistemic status: mostly writing to clarify my intuitions, with just a
few weak attempts to convince others. It's no substitute for reading
Drexler's writings.]
I've been struggling to write more posts relating to Drexler's vision
for AI (hopefully to be published soon), and in the process got
increasingly bothered by the issue of whether AI researchers will see
incentives to give AI's broad goals that turn them into agents.
Drexler's [CAIS paper](https://www.fhi.ox.ac.uk/reframing/) convinced
me that our current trajectory is somewhat close to a scenario where
human-level AI's that are tool-like services are available well before
AGI's with broader goals.
Yet when I read LessWrong, I sympathize with beliefs that developers
will want quite agenty AGI's around the same time that CAIS-like
services reach human levels.
I'm fed up with this [epistemic learned
helplessness](https://slatestarcodex.com/2019/06/03/repost-epistemic-learned-helplessness/),
and this post is my attempt to reconcile those competing intuitions.
Please recall that Drexler's distinction here focuses on a system's
goals, not its knowledge. Software is more agenty when its goals cover a
wide range of domains, and long time horizons. Services are designed to
produce specific outputs using a system's current procedures.
### The Easy Part?
An [ESRogs comment on
LW](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities?commentId=ePAXXk8AvpdGeynHe)
nudged me a bit more toward Drexler's position.
"The smart part is not the agent-y part": intelligence and agency are
independent, with intelligence being the hard part, and agency being
more like prompt engineering.
This seems mostly correct, at least for Drexler's meaning of agency.
There are still significant problems with getting agency correct, but
they're mostly associated with figuring out what we want, and likely
not helped much by math or computer science expertise.
### Orthogonality squared
ESRogs also suggests a potentially valuable concept that he calls
"Orthogonality squared". I.e. the domain of an agent's goals is
independent of its intelligence.
We can, in principle, create AI's that don't care about atoms. At
least for some simple tasks, it's fairly natural to have the AI care
only about abstractions such as math. You can't alter the laws of
arithmetic by rearranging atoms - laws of arithmetic are in a different
realm from atoms. An AI that cares only about a distant part of [Tegmark
IV](https://www.lesswrong.com/posts/R5d2z79jWsZmrK5Bh/anatomy-of-multiversal-utility-functions-tegmark-level-iv)
won't care whether we shut down an arrangement of atoms in our realm.
Many people concerned about AI safety seem reluctant to accept this.
E.g. Rob Bensinger's initial reply to ESRogs worried about this
unsafe(?) case:
>
> The system can think about atoms/<U+200B>physics, and it knows that our world
> exists, but it still only terminally cares about digital things in the
> simulated environment.
>
>
>
"Simulated environment" seems ambiguous as to whether it's referring
to a distant part of Tegmark IV, or a special environment within our
world. The latter would imply that it's tricky to analyze whether the
AI's goals conflict with ours; the former does not.
It seems odd that Eliezer's [AGI
lethality](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities)
#19 warns that it's hard to get software to care directly about the
real world, yet considers it inevitable that AGI's will care about the
real world. That's not a contradiction, but the tension between those
positions ought to raise further questions.
I don't know how an AI could run a company without caring about atoms,
so the basic claim behind orthogonality squared is not sufficient to
reassure me.
But it illustrates the extent to which AI goals can be compatible with
leaving humans in control of the real world. So it should be possible to
broaden such goals to care about narrow aspects of our universe, without
caring about all of our future lightcone.
### Analogies
We have some relevant experience with weakly superhuman agents in the
forms of governments, corporations, labor unions, and universities.
These all have some sort of limits on the breadth of their goals, yet
those goals often become broad enough to cause more harm than I'd
predict if I modeled them as purely tools / services.
Let's examine what happens if I try to create a corporation that has no
dangerously agenty goals. I'll call it Peter's Paleo Pizza.
I start out as the sole owner / employee, making all strategy decisions
myself. I hire cooks who I tell to follow my recipe exactly.
I can grow the company to a few hundred people while micromanaging it
enough that I have all the policy control over it that I want. But by
that time the micromanagement takes most of my attention.
If I grow it to thousands of employees, I'll delegate enough decisions
that the system will be at least weakly agenty. I don't want to give
the system detailed instructions about who to hire or how they ought to
acquire skills. Instead, I want employees to have a vision for how to
balance goals such as making food tasty versus making it healthy. I want
them to devote most of the company's resources to directly making food.
If I thought they might devote more than a few percent of the company's
resources to longer-term development plans, I'd place strict limits on
that.
Such a company is a weakly superhuman system, in the sense of being an
expert in more domains than an individual human can be. The limit on
development-oriented resources leaves me confident that it won't do
anything close to fooming via self-improvement. But the company is
stretching the boundaries of what I'd classify as a service.
When I want to broaden my business empire to the limits of what humans
have done, I give up on managing a company. I don't have Elon Musk's
ability to manage corporate policies of multiple industries. Instead, I
invest in many companies. That involves heuristics that tend to reward
companies for maximizing the discounted value of future profits. I
don't try very hard to predict how much they'll spend on R&D. This
clearly violates the constraints that Drexler wants on safe AI services.
How much would I and other investors do differently if corporations were
new phenomena with hard-to-evaluate risks, and unusually rapid changes
in abilities? It does not seem obvious. It seems clear that human
monitoring and guidance have costs which will tempt us to make
superhuman systems more agenty. The corporate analogy weakly suggests
we'll postpone the risky parts of that until after they become
superhuman.
### Software Examples
Gwern's [Why Tool AIs Want to Be Agent
AIs](https://www.gwern.net/Tool-AI) has examples of places where being
more agenty seems valuable.
Note that I'm using "agent" in a slightly different way than Gwern.
I'm attempting to focus on what Drexler considers important. So I'm
mostly disagreeing with Gwern's framing, not his answers.
High-frequency trading is a task for which humans are too slow to be in
the loop.
There is some sense in which the best software for this might be to tell
an AI to maximize expected returns over some long time period. But there
are safer strategies that I expect will usually produce the same
benefits:
* perform all the learning and make all strategic decisions before
deploying, so there's time for human-AI collaboration.
* have the AI only care about reacting to patterns in numbers. If we
do nothing to cause it to care about how those number relate to the
real world, I don't see how it would start caring about the real
world.
* have the AI care only about the next number it's about to output.
These options don't involve the AI self-improving while it's trading.
But that has little effect on whether the AI gets improved. The CAIS
approach to improving it involves a separate AI improvement process, for
which I don't see much pressure to have the AI make the most agenty
decisions.
Gwern unintentionally(?) illustrates why we should be cautious about
making AI's more agenty:
>
> we don't want excellent advice on which stock to buy for a few
> microseconds, we want a money pump spitting cash at us
>
>
>
I don't want anything resembling a human-level AI choosing such a money
pump, as it probably won't distinguish between one which will fail
gracefully versus one that will blow up like
[LTCM](https://en.wikipedia.org/wiki/Long-Term_Capital_Management) or
Bernie Madoff. But I likely would have chosen a dangerous money pump
when I had less experience.
[Chess](https://www.gwern.net/Notes#advanced-chess-obituary) provides a
different perspective. Computer-human collaborations fit the CAIS
paradigm fairly well for a decade, then the human became mostly a
liability, due to low human reliability. This is evidence that CAIS is a
temporary stopgap, but does support hopes that CAIS will buy us some
time.
Note that for chess, humans here weren't trying to be the more agenty
part of this collaboration (the chess programs were likely safely
aligned and using their agency appropriately). But similar
considerations would likely apply if the humans were primarily trying to
provide alignment: humans would make enough mistakes at aligning
individual decisions to pressure us to replace those decisions with
those of an apparently aligned AI.
Richard Ngo [expresses similar
concerns](http://thinkingcomplete.blogspot.com/2019/01/comments-on-cais.html):
"requiring the roles of each module and the ways they interface with
each other to be ... human-comprehensible will be very uncompetitive".
I don't see a general rule to that effect - sometimes comprehensibility
will enhance competitiveness via making a service easier to debug or to
build upon. Also, tools to assist human comprehension will enable humans
to use increasingly powerful interfaces. But I'm unwilling to bet that
those effects will enable human-controlled services to remain
competitive.
### Conclusion
Given mildly optimistic assumptions about how responsible the leading AI
developers will be, I expect human-level CAIS-style services anywhere
from months to a decade before human-level AGI's become broadly agenty.
I have little idea how hard it will be to use CAIS services to produce
longer-term solutions to AI risk.
I'm guessing Eliezer will disagree with a key part of this post. I
expect some of his reasons are related to a belief in a core of general
intelligence that has not yet been discovered. But it looks to me like
anything that I'd classify as such a core is already known.
This post seems mostly consistent with Drexler's claims, but with more
emphasis on the dangers. |
bb5638cf-5d0b-4de3-b174-48d1223ffc8a | trentmkelly/LessWrong-43k | LessWrong | Cherry Pit Storage
When eating cherries, what do you do with the pits? You don't want to swallow them, but spitting them out one by one is inconvenient. The answer: use your upper lip.
Eat the cherry as normal, separating the pit with your teeth and tongue as usual, but once the pit is free use your tongue to tuck it up under your upper lip. Once it's there, it will stay safely out of the way while you chew and swallow the rest of the cherry. You can repeat this several times; depending on your mouth geometry you can probably hold six to ten pits.
Asking around and looking online, it seems a lot of people don't know you can do this? I think this is because the pits are not especially visible, and the only clue someone is doing this is that they're needing to remove pits much less often. |
877ceba5-b00c-451d-ac18-a80dc25f3575 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | AXRP announcement: Survey, Store Closing, Patreon
Some announcements:
* There’s a brief [survey](https://bit.ly/axrpsurvey2023) of my audience you can take here - filling it out will help me get a sense of where you all are at career-wise and how the podcast would better serve you. There are only 3 required questions, and they’re multiple choice! It would be great if you could fill it out [here](https://bit.ly/axrpsurvey2023).
* The AXRP [store](https://store.axrp.net) is closing in a week - so if you want to buy merchandise, now’s your last chance to do so! To buy merch like laptop stickers, t-shirts, sweaters, or jigsaw puzzles (in the US), go to the store [at this link](https://store.axrp.net).
* If you want to financially support the podcast, you can do so in two ways:
+ On [Patreon](https://www.patreon.com/axrpodcast) - donations of US$20 per month get to see my episode notes, and at US$50 per month, you have a (currently very high) chance of being thanked by name at the end of episodes! Link is [here](https://www.patreon.com/axrpodcast).
+ On [Ko-fi](https://ko-fi.com/axrpodcast), where you can give one-off donations.
* Interviews will always be free and available to non-subscribers, but your support gives me a better sense of how many people get value out of this podcast, and how much time I should spend on it. |
a3b018bb-bf17-4ade-8730-395c948cff20 | trentmkelly/LessWrong-43k | LessWrong | How to do theoretical research, a personal perspective
> "Where do new [algorithms] come from? I keep reading about someone who invented [an algorithm] to do something-or-other but there's no mention of how."
>
> A shrug of robed shoulders. "Where do new books come from, Mr. Potter? Those who read many books sometimes become able to write them in turn. How? No one knows."
>
> "There are books on how to write -"
>
> "Reading them will not make you a famous playwright. After all such advice is accounted for, what remains is mystery. The invention of new [algorithms] is a similar mystery of purer form."
>
> - Harry Potter and the Methods of Rationality
Most of the content in this document came out of extensive conversations with Paul Christiano.
(This document describes one way of thinking about how to do one particular type of research. There are other ways to productively do this kind of research, and other productive kinds of research. Consider this document peppered with phrases like “from my perspective”, “I think”, “my sense is”, etc.)
A lot of people have a vague mental picture of what empirical research looks like, which consists of exploring data, articulating hypotheses about the data, and running experiments that potentially falsify the hypothesis. I think people lack a similarly mechanistic picture of what theoretical research looks like, which results in them not knowing how to do theory, being skeptical of the possibility of theoretical progress, etc. I do think the difficulty of getting high-quality real-world feedback makes theoretical research more difficult than empirical research, but I think it’s possible to get enough real-world feedback when doing theory that you can still expect to make steady progress.
Currently, I think many people think of theory as “someone sits in a room and has a brilliant insight to solve the problem.” Instead, I think a more accurate picture is very similar to the picture one has for empirical research: the theorist explores some data, gradually builds an intuitive se |
5f07592d-f94f-449e-9abf-d8fc295daf3f | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AI X-Risk: Integrating on the Shoulders of Giants
Uncertainty, Structure, and Optimal Integration
===============================================
*The EA community and Future Fund are running a risk in wasting good analysis of propositions when considering long-term predictions that entail complex, compounding uncertainties. Specifically, by ignoring or over-simplifying the complex ways our propositions interrelate (i.e., how they are “structured”), we stand to make substantial errors in our predictions. Moreover, the failure to appropriately “unpack” the structure behind our propositions risks substantial inefficiency in our focus of which uncertainties (whether they be based on a lack of information, or substantive disagreement) are more critical to resolve.*
*In Sections 1 and 2 I provide a primer to the uninitiated on the nature and importance of unpacking and disentangling the uncertain propositions that comprise our prediction problems.*
*In Sections 3 and 4 I lay out my analytical approach that seeks to capture and resolve these issues. I provide a fully worked through, illustrative decomposition of the misaligned AI X-Risk prediction problem, using a combination of Bayesian Network modelling and simulation techniques. My intent is to illustrate this analytical approach as a viable framework for a) **making arguments and assumptions explicit and quantifiable in a unified manner**, b) drawing **mathematically optimal inferences** (i.e., predictions) even in the face of multi-layered uncertainties, and c) laying the foundations for accessible computational methods that will assist further research. My assumptions and propositions within the model are predominantly grounded in extant estimates and analyses drawn from AI experts and fellow contributors. However, I highlight a number of **previously overlooked or underspecified propositions and conditional probabilities** that affect our predictions, but **have previously remained hidden**.*
*In Section 5 I step through the results of my analysis. Among them, I show that once structural disentanglement is taken into account, under reasonable, informed assumptions, the **P(Misaligned AGI X-Risk) by 2070 is 1.68% (+/- 1.25%), rising to** **6.53% (+/- 4.32%) by 2100**. Though if we wish to assume an AGI has been built by 2070 (i.e., P(Misaligned AGI X-Risk|AGI)), then this rises further to **5.32% (+/- 3.42%)**by 2070, and11.22% (+/- 6.86%) by 2100. I additionally provide predictions for AGI build likelihoods across the target time-horizons.*
*Finally, in Section 6 I describe a quantitatively assisted path forward for further inquiry that is both viable and accessible. I outline how the analytical approach laid out here can be built upon; both enriching, and being enriched by existing lines of research into AI X-Risk.*
1 - A Reasoning Under Uncertainty Perspective
=============================================
Predicting the future existential risk due to a misaligned AGI is not only an important problem in and of itself, but is also *useful as being illustrative of a **class** of problem*.Prediction problems like this – much like any reasoning under uncertainty endeavour – require:
1) The **acquisition** of relevant arguments, assumptions, and propositions,
2) The **estimation** of the uncertainty surrounding them, and finally
3) Their **integration** into a coherent whole.
Substantial work has been done on the **acquisition** of relevant arguments in recent years (even just a [cursory search of this forum](https://forum.effectivealtruism.org/topics/ai-risk?sortedBy=relevance) reveals >400 posts on “AI Risk” at time of writing, dating back 6 years). These cover a wide array of topics such as AI build likelihoods and timelines, relevant AI properties/functions, paths to catastrophe, and risk countermeasures, just to name a few. Work like this - the acquisition of relevant arguments – is invaluable. Without such work, it would be as if *we are trying to build a telescope without any components*.
This post does not focus on the acquisition of arguments. I am not an expert on AI. I am, however, an expert on modelling complex inference problems. This post focuses on a way to integrate those acquired arguments to minimise the risk of error. Put another way, *I design telescopes*. The value of understanding good telescope design is not just in getting a clearer picture of what we wish to look at. It is far wider than this. The more one learns about how pieces should be arranged, the more one sheds light on **which components are more important, and which are less so**; **which components need refinement to be fit for purpose**; **and which components we did not even realise we were missing.**
As has been noted in the [AI Prize Announcement](https://forum.effectivealtruism.org/posts/W7C5hwq7sjdpTdrQF/announcing-the-future-fund-s-ai-worldview-prize) and in advice on [reporting here](https://forum.effectivealtruism.org/posts/SdQYLKzpDFQsdW9bn/features-that-make-a-report-especially-helpful-to-me), explicit probabilities (i.e., **estimation**) are critical to reasoning transparency. They allow us to express degrees of confidence in a unified manner, lowering the risks of inadvertent, unhelpful translation error (e.g., what I mean when I say “X is likely” may be quite different from what you interpret). Moreover, explicit probabilities are critical in laying the groundwork for computational methods. Analyses such as [Carlsmith (2021)](https://arxiv.org/pdf/2206.13353.pdf), [responses to Carlsmith](https://docs.google.com/spreadsheets/d/1OGiDFCgRjfxOFVaYPl96dYgx8NVIaSquDgQSKByDxuc/edit#gid=0), [Cotra (2020)](https://www.alignmentforum.org/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines), and [Davidson (2021)](https://www.openphilanthropy.org/research/semi-informative-priors-over-ai-timelines/), [among others](https://docs.google.com/spreadsheets/d/1W10B6NJjicD8O0STPiT3tNV3oFnT8YsfjmtYR8RO_RI/edit#gid=0), have taken important steps in providing various estimates on the probability of various propositions (e.g., the likelihood of an AGI being built by a given year). However, as noted within Carlsmith’s own work (Carlsmith, 2021, Section 8), attempts to accurately **integrate** these estimates exposes two centrally important, but easily overlooked reasoning errors. These reasoning errors - when ignored - can lead to wildly overconfident predictions (as I will show). These errors are:
1) **Oversimplifying Structure**. Although the multiplication of probabilities attached to propositions is nominally a [Bayesian](https://en.wikipedia.org/wiki/Bayesian_inference) way of integrating (conditional) probabilities (i.e., the application of [a chain rule](https://en.wikipedia.org/wiki/Chain_rule)), it implicitly assumes a lot about the way uncertainty is structured. I discuss this in more detail below, but the most obvious assumption is that of unitary causation – a chain of probabilities implies a singular path of cause (or “parent proposition”) and effect (“offspring proposition”). **Reality is far messier, with common causes of multiple effects, effects with multiple potential causes, and myriad combinations thereof. Simply put, this fundamentally changes the maths and risks substantial overconfidence in estimates derived from inadequately structured integrations.**
2) **Failing to “Unpack” Propositions**. In relation to the first, there is recognition that an estimate attached to a proposition may in fact be covering hidden, unquantified parent propositions. For example, in arriving at an estimate of an X-Event caused by a misaligned AGI (P(X-Event|AGI)), one of my propositions may be the probability of an AGI being built successfully (P(AGI)). However, what is omitted from my estimate of P(AGI) are the propositions (i.e., assumptions) I have invoked (e.g., the likelihood of a build being computationally feasible; there being a viable AI design; there being sufficient incentive for completing a successful build). **Without explicit estimates provided for these parent propositions, we not only risk introducing error to the overall calculation, but we are also not adhering to principles of transparency, making it harder to reach clear consensus on points of disagreement.**
In this post I will demonstrate an analytical approach that directly addresses these problems. In so doing, I will outline how this approach:
* Provides useful structure to a complex problem.
* Disentangles compounding levels of uncertainty.
* Facilitates the move towards more exhaustive, transparent estimation of uncertainties surrounding propositions.
* Demonstrates optimal (i.e., inaccuracy minimization via [Bayesian inference](https://en.wikipedia.org/wiki/Bayesian_inference)) integration across these structured uncertainties.
* Creates a computational method that can enhance the acquisition process; helping us see which propositions are more impactful to the question at hand, and therefore which areas should we invest more time in resolving remaining uncertainty/discordancy.
The model I create in the process serves as an example of a more accurate, transparent method of integrating what we know to make better predictions. To understand *why*this is the case, I provide a primer on why failures to appropriately understand and account for reasoning structure can systematically lead our predictions astray. Readers already familiar with the relationship between reasoning structure and reasoning errors should feel free to skip to Section 3.
2 - A Primer on Disentanglement[[1]](#fn8diqhvi598n)
====================================================
There are myriad factors that make predicting long-term risk difficult. Ultimately, these stem from the different sources and implications of uncertainty types. For example, at the highest level we have uncertainty stemming from **unknown unknowns** (e.g., we do not know which arguments we have not yet made); this impacts the accuracy of our (reasoning) models, but is generally something harder to quantify meaningfully.
We then also have uncertainties that stem from **known unknowns**, which covers both empirical sources, such as incomplete data, measurement error, and instrument reliability; as well as more subjective bases, such as degrees of belief, and disagreement (e.g., [community estimates diverging substantively regarding the likelihood of a proposition](https://docs.google.com/spreadsheets/d/1OGiDFCgRjfxOFVaYPl96dYgx8NVIaSquDgQSKByDxuc/edit#gid=0)). Often, we tend to ignore some or all of these sources of uncertainty, either at worst by being “selective” (whether purposefully or not) in our supportive arguments and conclusions, or at best by trying to “summarise” across the uncertainties via a point estimate (i.e., a 1st order probability: “I believe there to be an X% probability that Y will happen.”). The latter of course is an important step towards reasoning transparency, but much remains to be desired.
For instance, let us consider the impact of disagreement in our probability estimates for a proposition. As disagreement increases (see examples [here](https://docs.google.com/spreadsheets/d/1OGiDFCgRjfxOFVaYPl96dYgx8NVIaSquDgQSKByDxuc/edit#gid=0) and [here](https://docs.google.com/spreadsheets/d/1W10B6NJjicD8O0STPiT3tNV3oFnT8YsfjmtYR8RO_RI/edit#gid=0)), so too does the 2nd order probability (see e.g., [Baron, 1987](https://link.springer.com/article/10.1007/BF00127335)) surrounding the proposition. For example, if we compare two hypothetical cases; one in which 100 experts have all given probability estimates for the probability of an AGI being built by 2070 between 48% and 52%, and another in which they gave estimates between 20% and 80%. In *both cases*, our overall (average) point estimate for P(AGI built by 2070) is 50%, but we can see how in the former case, we can be much more *confident* of that 50% as individual estimates lie +/-2% either side, than we can in the latter case, where there is much more substantial disagreement with individual estimates lying **+/-30%** either side.
This form of uncertainty is important, but difficult to deal with without computational assistance. As we consider how to integrate multiple propositions to arrive at a conclusion, these uncertainties in fact *compound* potential errors. Fortunately, as we start to disentangle the reasoning structure that underpins and connects our propositions (the central purpose of the analytical approach laid out in this post), we can not only gain more transparency and insights behind *why we are disagreeing* (via structurally “unpacking” a proposition), but we can **avoid compounding errors in our integration**. To return to our telescope analogy: If we are trying to identify something far away (our prediction), and we have a handful of lenses (our collection of propositions), we are far more likely to end up with a blurry image (and thus a mistaken identification) if we do not know how to arrange our lenses correctly. Once we know the principles for how the parts should be arranged, we are then also much more able to discern which lenses may be causing problems (i.e., once we know the correct design, we also know how clear the image *should be*).
Where does this leave us? First, we have multiple sources of uncertainty when we try to make predictions. Second, how these uncertainties affect our predictions when we must integrate across multiple propositions is complex, and consequently carries a high risk of error. So, if we wish to make better predictions, we must understand how to capture and deal with this complexity. To achieve this, we must next understand how uncertainties can be disentangled (or “partitioned”) using reasoning *structure:*
2.1 - Structure
---------------
Understanding how a proposition relates to and interacts with others is perhaps one of the hardest, but most critical parts of correctly *integrating* arguments to arrive at a *more accurate* prediction.
### 2.1.1 - Conditionals and Chains
Let us begin with the difference between the probability of a proposition being true, and the probability of a proposition being true ***conditional on something else (e.g., a parent proposition) being true***. To help explain this, and all forms of structural disentanglement outlined in this section, I will use a toy example version of the likelihood of an AGI being built – *this is meant to be purely illustrative at this stage*, so do not worry about specific estimates.
First, let us assume the probability of an AGI being built (P(Built)) at least in part depends on the probability of it being technically feasible and affordable to construct an AGI (P(BuildFeasible)). Now, let us assign a probability – say 30% - to the proposition of an AGI build being feasible by a chosen point in time (i.e., P(BuildFeasible) = 30%). This is not the same as the probability of an AGI *actually being built* (i.e., P(Built)). This latter probability is only *inferable*, via the conditional probability of a successful build, *given its feasibility*. For instance, we may assume that if an AGI build is feasible, then there is a high probability of an AGI getting built – say 80%(i.e., P(Built|BuildFeasible) = 80%). We can also assume that if an AGI build it not feasible, then an AGI cannot be built (i.e., P(Built|¬BuildFeasible) = 0). We can see the outcome of this integration via Bayesian Inference below (note: for a more thorough read than this short example, see [this post by Eliezer Yudkowsky](https://www.yudkowsky.net/rational/bayes)):
P(Built)=P(Built│BuildFeasible).P(BuildFeasible)+P(Built│¬BuildFeasible).P(¬BuildFeasible).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')}
Given our assumptions above, this translates to:
P(Built)=(.8∗.3)+(0∗.7)=.24Although this example is simplistic, it gets across two important aspects of structural disentanglement:
1. It exposes potential sources of error in informal reasoning. Some potential errors include (but are not limited to):
* [**Base Rate Neglect**](https://en.wikipedia.org/wiki/Base_rate_fallacy): Ignoring base rate / prior information (in this example, this would be our parent proposition, P(BuildFeasible)). This can lead to gross overestimation. Here this would be assuming P(Built) = 80%, rather than the 24% we find when we *properly structure the argument*.
* [**Miss Rate Neglect**](https://academic.oup.com/lpr/article/15/4/239/2580528): Although less relevant in this example, neglecting terms like P(Built|¬BuildFeasible) fails to account for “would this have occurred anyway, regardless of my parent proposition?” This failure can produce substantial over or underestimations of the likelihood of a proposition being true, as you are not fully accounting for the *strength* (how predictive/diagnostic) of the relationship between the two propositions.
* **Put simply, the separation into terms forces us to scrutinise and incorporate a** ***more complete*** **set of uncertainties, reducing errors.**
2. The disentanglement allows us to better identify *where we are disagreeing*, and *where we are less certain* (e.g., our different P(Built) values are stemming from different assumptions regarding P(Built|BuildFeasible), whilst we mainly agree about P(BuildFeasible) itself). Though arguably trite in this example, this capability becomes increasingly needed the more complex the overall reasoning structure, as we will see.
We are starting to see how separating out structure supports more thorough and efficient **acquisition** (and consequently, **estimation**) of propositions.
So far, we have laid the groundwork for integrating across a causal chain, wherein we can multiply a chain of conditional probabilities together to reach a final proposition likelihood. This has been the method applied by [Carlsmith (2021)](https://arxiv.org/pdf/2206.13353.pdf) and [others](https://www.lesswrong.com/posts/qRSgHLb8yLXzDg4nf/reviews-of-is-power-seeking-ai-an-existential-risk) when estimating AI X-Risk. Of course, when we start to consider uncertainties on each of these component probabilities, things get a lot trickier (see [this excellent post on uncertainty analysis of the same causal chain by Froolow](https://forum.effectivealtruism.org/posts/Z7r83zrSXcis6ymKo/dissolving-ai-risk-parameter-uncertainty-in-ai-future#comments)). Additionally, in modelling these arguments we must draw a line under “how far back” we wish to go: P(BuildFeasible) is itself conditional on other propositions, and so on backwards until we reach the Elysium of exhaustive, confirmed certainty. This of course is rarely possible, particularly in long-range prediction. Thus, we must elect to “cut off” the recursion at some point.
This is not where the story ends, as we have the obvious fact that one proposition may have more than one effect on other propositions, and itself be affected by more than one parent proposition. As we will see below, when we expand our understanding of reasoning structures to include this broader set of possibilities, the grave implications of ignoring this complexity for prediction accuracy will become evident – along with the rationale for the analytical approach outlined in this post that captures this complexity.
### 2.1.2 - Common Effects
Next, let us consider a proposition with more than one parent (known as a “collider”, or common effect). This is an extremely likely occurrence – rarely is it the case that unitary causation (i.e., one cause to one effect) is valid. For example, there are often enabling conditions, additional risk factors, catalysts, and so on. Each additional parent proposition results in a combinatorial explosion of the proposition that depends upon them.
To ground this in our example, let us reconsider it, but now add an additional influence on P(Build): the (human) incentive for AGI, P(Incentive):
P(Built)=P(Built│BuildFeasible,Incentive).P(BuildFeasible).P(Incentive)+P(Built│BuildFeasible,¬Incentive).P(BuildFeasible).P(¬Incentive)+P(Built│¬BuildFeasible,Incentive).P(¬BuildFeasible).P(Incentive)+P(Built│¬BuildFeasible,¬Incentive).P(¬BuildFeasible).P(¬Incentive)If we return to our original example, where P(BuildFeasible) = 30%, but now we assume that by this time, there will more likely than not be an incentive for an AGI (e.g., P(Incentive) = 60%). Further, we can now specify that if a build is feasible **and** there is incentive, there is a high likelihood of a successful build (e.g., P(Built│BuildFeasible,Incentive) = 90%), but if there is no incentive, this drops substantially (e.g., P(Built│BuildFeasible,Incentive) = 5%). Given we can again assume a successful build is impossible if the build is not feasible *regardless of incentive*, this gives us:
P(Built)=(.9∗.3∗.6)+(.05∗.3∗.4)+(0∗.7∗.6)+(0∗.7∗.4)=0.168When we compare this ~17% to our previous calculation (where P(Built) = 24%), we can see how by failing to unpack and incorporate this additional parent (i.e., by keeping our incentives assumption hidden / unaccounted for), we are incurring a *systematic* error (in this case, overestimation). This is particularly critical when we consider that these errors can **compound**(i.e., our overestimate forms the basis of a further calculation, that itself could be an overestimation, increasing the overall inaccuracy).
More generally, we can also see that computational requirements have doubled. In fact, for each additional parent (as there can of course be more than two), we must increase the number of terms exponentially.[[2]](#fnpotccwwjxs) It is worth emphasising again here that *these unpacked terms are still present in an informal analysis of these propositions – they are simply **unspecified and their influence unaccounted***.
In this way, we once again highlight the value of unpacking our assumptions and structuring them appropriately if we are to 1) reveal where we might be missing important information, and/or where our points of disagreement lie (assisting in both **acquisition** and **estimation**), but also critically2) **enable optimal integration.**[[3]](#fnlu6ff3gujv)
### 2.1.3 - Common Causes (and Dependence)
Finally, and to complete the picture, we must consider a parent proposition with multiple offspring propositions (known as a “fork”, or common cause). This is simple enough when we only care about each offspring proposition in isolation (i.e., we can multiply our conditionals, as we did for a chain).[[4]](#fn7v67i7xwxjp) However, how common causes fit into the bigger picture of our reasoning structure has substantial implications for the accuracy of our predictions.
For instance, what if a parent proposition for one part of my chain is also a parent to another part of my reasoning chain? Extending our previous example P(Incentive) may not only affect whether an AGI is successfully built (P(Built)), but also whether it will go on to be deployed within a target system (P(Deploy)).
The influence this common cause is having on the overall reasoning structure (and therefore our predictions) is often described as dependence. Essentially, propositions considered independent carry more weight than when they share a (partial) basis. If I am not explicitly accounting for the fact some of my propositions share the same parent, then I am not factoring in that there is some *redundancy* in my reasoning. Instead of subjecting you to even more maths (though see [Koller & Friedman, 2009](https://list01.bio.ens.psl.eu/wws/d_read/machine_learning/BayesianNetworks/koller.pdf); [Pearl, 1988](https://books.google.co.uk/books?hl=en&lr=&id=AvNID7LyMusC&oi=fnd&pg=PA1&dq=Pearl+1988&ots=F--PWmqy-9&sig=ps62vYE583xMWZ-U0-qWKeuez-Q&redir_esc=y#v=onepage&q=Pearl%201988&f=false); [Schum, 1994](https://books.google.co.uk/books?hl=en&lr=&id=8PqsTNULCPEC&oi=fnd&pg=PR13&dq=Schum+1994+the+evidential+foundations&ots=s98PqqIMrO&sig=odK8c2LC21w5z27-_2bw2ZAFRpE&redir_esc=y#v=onepage&q=Schum%201994%20the%20evidential%20foundations&f=false); for full mathematical treatments) I will illustrate with a quote attributed to Abraham Lincoln:
> *“Surely, Mr. Lincoln,” said I, “that is a strong corroboration of the news I bring you.” He smiled and shook his head. “That is exactly why I was about you about names. If different persons, not knowing of each other’s work, have been pursuing separate clews that led to the same result, why then it shows there may be something in it. But if this is only the same story, filtered through two channels, and reaching me in two ways, then that don’t make it any stronger. Don’t you see?”*
>
>
*William Seward (1877)*
Setting to one side the implications this has for how much weight we should assign [aggregation/consensus across AI experts](https://research.aimultiple.com/artificial-general-intelligence-singularity-timing/) (there is an entire separate post that could be written on potential community overconfidence in that regard), we can see the implication of Abe’s words for our own prediction efforts: If our prediction relies upon seemingly separate propositions – but these propositions in fact draw their strength from the same parent(s), then we are likely to be overconfident in our conclusions.
Along with exposing redundancy, disentangling these common causes in our reasoning structures also helps us avoid *logical inconsistency* in our reasoning – for example, a shared parent that makes one offspring proposition more likely, but makes another *less likely*, can cancel out the impact our two offspring propositions would have had on our prediction.
Once more, I will trumpet my standard refrain about how this disentangling is necessary and helpful for more exhaustive **acquisition** and **estimation**, and essential to accurate **integration**. In addition, I note that factoring in these sorts of structures can expose where propositions within our reasoning problem are more *sensitive* to changes in our parameters. If multiple parts of my chain actually rely on the same parent proposition, then my assumptions regarding the parent proposition (including the uncertainty that surrounds it) will have more impact on the overall reasoning structure and my predictions (*ceteris paribus*) than my assumptions regarding one of its offspring.[[5]](#fn7ve0z9ypn6d)
So, we have now covered the three core reasoning sub-structures (chains, common causes, and common effects), and how failure to account for them can lead to error. Critically, ***all reasoning under uncertainty problems are built using these three core sub-structures***, just using various combinations. Predicting misaligned AI X-Risk is no different. I will again emphasise here that **these structural complexities are part of the prediction problem, regardless of whether you represent them within a model.**
To return to our telescope metaphor: At this point I have sought to highlight the importance of appropriately combining our components by showing that without appropriate assembly, the image we are trying to see will be hard to interpret accurately. Extending the metaphor slightly, the *further* we are trying to see, the more substantial the components (propositions) we are likely to invoke, and thus the stronger the imperative to arrange those components correctly. Slight misalignment in a set of binoculars will limit the image quality, but misalignment in the James Webb Space telescope would render it useless – *given how far it is trying to see*.
Where does this leave us? Fortunately, there are computational approaches that allow us to capture and navigate this complexity. Here I demonstrate one.
3 - Current Approach
====================
My analytical approach here has two interrelated components:
First, I have constructed a Bayesian Network model for predicting misaligned AI X-Risk. This model disentangles how a number of core propositions relate to one another, and in so doing accounts for the potential pitfalls and errors described in Section 2. Further, this disentanglement exposes a number of relevant conditional probabilities that at time of writing have not received much consideration when we discuss AI X-Risk. In this way, the model opens the door for improving **acquisition**and **estimation** goals by increasing the necessary transparency of our reasoning.
Perhaps most critically of all, the model incorporates a probability calculus that enables ***optimal integration*** across the complex pattern of inferences. Once more I will emphasise that *this complexity is inherent within the prediction problem, whether we choose to capture it or ignore it*.Furthermore, the way the model is structured draws from other domains of reasoning-under-uncertainty to highlight several useful ways of breaking down the AI X-Risk prediction problem.
Second, surrounding the Bayesian Network model I have built a simulation “wrapper”. This wrapper selects parameters for the model (e.g., our chosen conditional probabilities), runs the model with those parameters, recording the predictions that model makes, and then begins again with a new set of inputs. We can think of it like this: The Bayesian Network is the telescope design, and the simulation wrapper is a conveyor belt swapping in new lenses and recording the picture. This enables me to do several things:
* I can see how model predictions change for each year moving forward *as input propositions are predicted to change through time*(layer 1). This allows me to ask an expanded set of questions, and ties the temporal dimension into the coherence of the analytical approach. Put another way, **we can explicitly track how predictions change through time**, generated via the same, consistent calculus and structure.
* Second, for every year, I run repeated versions of the model, where parameter inputs are drawn from a distribution, rather than fixed to a single number (layer 2). This stochastic sampling process enables me to incorporate the 2nd order probability that affects our estimates for each proposition (see Section 2 for discussion on the way unknowns and disagreements affect our estimates). In this way, **we can determine how much uncertainty surrounds our generated predictions**.
3.1 - Bayesian Network
----------------------
Simply put, a Bayesian Network is an argument map with some maths underpinning it.
More concretely, Bayesian Networks have a graphical representation, known as a Directed Acyclic Graph (DAG), where the relationships (arrows) between each variable (in our case, these are our propositions) are conditional probabilities ([Pearl, 1988](https://books.google.co.uk/books?hl=en&lr=&id=AvNID7LyMusC&oi=fnd&pg=PA1&dq=Probabilistic+reasoning+in+intelligent+systems&ots=F--QSnry2a&sig=7Z0sOFGAmFkHwOIVB2XchLfN8Lc&redir_esc=y#v=onepage&q=Probabilistic%20reasoning%20in%20intelligent%20systems&f=false)). Variables without any parents (i.e., no arrows flowing into them) are described by a “prior” probability at the input stage. In other words, we are fixing an initial probability based on our assumptions. This differs from variables with parents, as these can only initially be described by a set of conditional probabilities (see Section 2.1); the likelihood of that variable being true is always the product of *inference*. Crucially, this inference process is Bayesian, and may be considered optimal in the sense that it minimises inaccuracy (see e.g., [Pettigrew, 2016](https://www.researchgate.net/profile/Richard-Pettigrew-2/publication/272357574_Accuracy_and_the_Laws_of_Credence_draft_manuscript/links/54e317b50cf2d618e195c35e/Accuracy-and-the-Laws-of-Credence-draft-manuscript.pdf)).
*Fig. 1: AI Misalignment X-Risk Bayesian Network example structure.*In the model structure shown above, variables (our propositions) are labelled using [square braces], this is to distinguish a variables state (here these states are True/False, though other states are possible) as either true [X] or false [¬X]. This is distinct from the *probability* of a proposition being true P(X) or false P(¬X). I will unpack each of these variables and their connections in turn in the model description below. Suffice to say that inferences flow from parentless propositions (or premises) {[SSu], [SBl], [De], [BFe], [Op]}, through an intermediary layer of inferred (read: conditional) propositions {[Po], [MAG], [CCo], [Bu], [In]}, which in turn are parents to the central inferred proposition of interest {[AIX]}.
We can use the network to “read off” updated inferences about each of our propositions simultaneously, and can even assess the impact of making particular assumptions/observations. For example, if we wish to assume an AI has been built (rather than leave it as an inferred probability), we can set that proposition to true, and see how this changes the pattern of inferences in the rest of the model.
### 3.1.1 - Sub-structures
Though we can see already from Fig. 1 that the network is comprised of the reasoning structures outlined in Section 2 (e.g., [De] -> [CCo] -> [AIX] is a chain, whilst [De] is also the common cause of several propositions, that in turn are also common effects of other parent propositions, and so on…), I felt it prudent to further classify these propositions – particularly in how they relate to the central proposition of interest.
In asking “will a misaligned AGI ***cause*** an X-event”, we are essentially asking how likely is it an “agent” (a misaligned AGI) will commit an “act” (an X-Event). This is the prediction equivalent of the diagnostic question: “how likely is it [agent] **committed** [action]?”. For this reason, there are similarities between our question and reasoning under uncertainty in legal settings, wherein determining guilt can be broken down into separate sub-structures (see e.g., [Lagnado et al., 2013](https://content.iospress.com/articles/argument-and-computation/682656)), each of which plays a role when assessing probable guilt. In part as useful structural metaphor for signposting my reasoning, I use and extend these here, to arrive at the following categories of proposition:
**Means**: Propositions relating to the question of whether an AGI will have the tools/potency to enact an X-Event (in a murder case, this would be possessing the murder weapon).
**Motive**: Propositions relating to the question of whether an AGI will intend to enact an X-Event (this does not mean through maliciousness necessarily, rather through misalignment – accidental or otherwise).
**Opportunity**: Propositions relating to the question of whether an AGI will have the necessary *access* to enact an X-Event.
**[Additional] Existence**: As it does not exist yet, we also must consider propositions relating to the creation of an AGI as a necessary enabling condition.
**[Additional] Enabler Incentive**: Similarly, as an entity that we can assume at least initially requires external assistance (even if this is just to create the AGI), we should also consider propositions regarding human (as enabler) incentives.
Although the above has proven useful to me in my structuring of the model, it may prove a useful classification for others when considering how to direct research and collate arguments. For instance, it may be that means, motive, and existence have received far more thorough treatment or reached firmer consensus than opportunity and incentives. The classification is also useful for basic sensitivity analysis in future – informing us about which categories are more impactful to our predictions.
3.2 - Simulation "Wrapper"
--------------------------
The simulation wrapper is essentially a series of loops that generate multiple independent Bayesian Network model runs. These runs differ in the input parameters provided to the model, whilst the structure of the model remains constant.
*Fig. 2: Simulation wrapper (i.e., constructor) surrounding Bayesian Network model.*The first “layer” allows us to investigate how inferences change as inputs change *over time*, wherein I specify a time-series for each relevant parameter. In this way, how the model outputs (i.e., the predictive inferences we are “reading off” the model) change over time can be recorded and assessed. This is a critical component, given the importance of the temporal dimension to many of these sorts of prediction problems. I note here that although it is possible to incorporate inter-temporal dynamics (e.g., success in a previous year in one part of the problem changes the dynamics in other parts), in this initial version of the model these are set to one side.
The second “layer” occurs within each time-point, and entails repeated stochastic sampling (10,000 samples per time-point) from specified distributions (any distribution shape can be specified independently for each parameter, including normal, beta-geometric, see e.g., [this recent post by Froolow](https://forum.effectivealtruism.org/posts/Z7r83zrSXcis6ymKo/dissolving-ai-risk-parameter-uncertainty-in-ai-future) on importance of these decisions). This allows us to capture second order uncertainty regarding our model inputs.
Taken together, the simulation wrapper allows us to not only see how model predictions change *over time*, but also how uncertainties at input stages translate to uncertainties around those predictions, which can often be highly unintuitive.
3.3 - Approach Summary
----------------------
Before summarising the approach, I will once again emphasise that the complexity laid out above (the role of conditional probabilities and reasoning structure, the role of time, and the role of 2nd order uncertainty) is still there in our prediction problems, whether we represent it within a model, or not. If we do not represent it formally, we are simply unaware of the nature of the *compounding errors* we are making, and thus we are prone to overconfidence in our conclusions.
Through the analytical approach laid out here, we can represent and account for this complexity by making it explicit; in turn allowing us to integrate optimally, *given our explicitly stated assumptions*. I argue that at this stage in the investigative process – where methods for accurate integration are particularly required, this has great value. **This approach sets us on the path to learning which unknowns are most critical, such that we can start to seek agreement on shared assumptions/parameters – or simply see the impact of our disagreement.**
In the model description laid out below, I have no doubt that some of my assumptions will provoke disagreement – as I have mentioned, I am not an expert on AI. However, wherever possible I have made use of existing arguments and estimates, and I incorporate 2nd order uncertainty where there appears to be disagreement. Critically, I am feeding these existing arguments into an analytical approach that captures the complexity that surrounds them, enforcing much needed *exhaustiveness and transparency*, and enabling *accurate integration*.
4 - Model Description
=====================
The Bayesian Network model is written in R for now using the gRain package ([Højsgaard, 2012](https://www.jstatsoft.org/article/download/v046i10/564)); I have made all the code available [here](https://github.com/pilditchtd/AI_X-Risk_BN_Project). Feel free to download it and run it yourself. I have done my best to indicate clearly within the code where different assumptions/parameters can be specified.
In this section I describe the model by stepping through each proposition in turn. This description begins with parentless propositions which have prior probabilities directly specified. I then build up the model description, beginning with the most immediate inferred propositions, on until all propositions (and how they relate to one another) have been explained. Usefully, a property of Bayesian Networks called a [Markov blanket](https://en.wikipedia.org/wiki/Markov_blanket) means we can understand an inferred proposition by knowing only its immediate parent propositions. This partitioning property allows us to build up the complete picture of the model by focussing on each piece in turn, without missing anything critical about the model as a whole.
4.1 - Terminology
-----------------
**AGI**: Here I take this to mean an Artificial General Intelligence, such that it possesses sufficient features to enable concerns regarding goal misalignment and circumvention of control.
**System**: The system/context in which an AGI is implemented.
**User/Implementer**: The human(s) responsible for initial activation of the AGI within the system.
**Parameter Uncertainty**: Here described as +/- 1 standard deviation (SD). Of course, other measures are available.
4.2 - Prior/Parent Propositions
-------------------------------
These are probabilities for parentless propositions, which where possible are drawn from existing analyses (notably [Carlsmith, 2021](https://arxiv.org/pdf/2206.13353.pdf); [Cotra, 2020](https://www.alignmentforum.org/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines)). As I have mentioned, although point estimate probabilities are an excellent step, it is important to represent the uncertainty surrounding them, as the way uncertainty propagates across the reasoning structure can often produce unintuitive effects (cf. [Schum, 1994](https://books.google.co.uk/books?hl=en&lr=&id=8PqsTNULCPEC&oi=fnd&pg=PR13&dq=Schum+1994+the+evidential+foundations&ots=s98PqqIMrO&sig=odK8c2LC21w5z27-_2bw2ZAFRpE&redir_esc=y#v=onepage&q=Schum%201994%20the%20evidential%20foundations&f=false)). In this way, provided point estimates are considered central tendencies in probability distributions of varying width (from which sample parameter values are drawn for model runs).
Given the question focus on the year 2070, and the consequent prior analyses also providing estimates based on this year, time series inputs trend towards them, following 3 basic assumptions:
1. Uncertainties will grow over time (represented here as standard deviations, but other measures can be interchanged for reporting purposes).
2. The trend that takes us from the current time-point to 2070 is expected to continue (i.e., I am not assuming there is something a priori “special” about the year 2070).
3. Unless there is a strong reason to deviate, I will assume steady, continuous linear change in input parameters over time. This is to avoid smuggling additional assumptions, including assuming 2070 as a particular year of importance (e.g., as an inflection point in a sigmoid function). \**Note: Other temporal functions are of course testable assumptions within the model, though I will note for the purposes of the 2070 question (and the estimates for that year originating in prior analyses, not my chosen function), this for now is immaterial*.
### 4.2.1 - AGI Feasibility & Design Propositions
*Probability of building an AGI build **b**eing **fe**asible: P(**BFe**)*
Proposition: Is an AGI feasible and affordable?
Assumptions Basis: The probability of there being sufficient, affordable computational capacity for the creation and maintenance of an AGI. Here, I am assuming this to mean the probability that the necessary resources (hardware, finances, etc.) are sufficient for the creation of an AGI. I note here that this proposition itself can be unpacked into subsidiary propositions, applying the current method recursively. However, for the sake of my own time constraints, I will not dig into this for now. Instead, I will seek to incorporate parameter values from existing analyses.
In general, the year-on-year probability is assumed to grow over time, and has been covered in analyses such as [Cotra (2020)](https://www.alignmentforum.org/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines). These analyses place the probability of sufficient computational capacity (read physical and financial feasibility) for an AGI by 2070 at **65%**.
Detail/Function: Consequently, I assume this probability to continuously grow to reach 65% by 2070, and then continue growing at a similar rate beyond 2070. Further, I am attaching a growing uncertainty in this parameter the more distant the prediction, such that by 2070, this 65% probability has a +/- 14.16% uncertainty surrounding it. This range around the point estimate covers a broad range of estimates (approx. 50% - 80% within 1 SD), helping to encompass [acknowledged variance in these estimates](https://www.lesswrong.com/posts/qRSgHLb8yLXzDg4nf/reviews-of-is-power-seeking-ai-an-existential-risk).
*Probability of successful AGI **De**sign: P(**De**)*
Proposition: Is there a viable AGI design?
Assumptions & Basis: The probability of there being a viable AGI design that possesses the necessary capabilities for misalignment of goals and circumvention of human control, namely, as described by [Carlsmith (2021)](https://arxiv.org/pdf/2206.13353.pdf), agentic planning and strategic awareness features. Again, and as laid out by Carlsmith, this proposition itself can be unpacked into subsidiary propositions. Although the method described here could be deployed to unpack and integrate this sub-proposition set, for now I will take the previously analysed composite. The specified year-on-year probability is assumed to grow over time, and has been well-covered in analyses such as Carlsmith (2021). This value is assumed to be married up with the same rate as the build feasibility [BFe] (i.e., 65% by 2070). However, to represent some of the disagreement and uncertainty on this proposition, uncertainty in this parameter is again built in.
Detail/Function: Consequently, I assume this probability to continuously grow to 65% by 2070, and then continue growing at a similar rate beyond 2070. Further, I am again attaching a growing uncertainty in this parameter the more distant the prediction, such that by 2070, this 65% probability has a +/- 14.16% uncertainty surrounding it. This range around the point estimate covers a broad range of estimates (approx. 50% - 80% within 1 SD), helping to again encompass [acknowledged variance in these estimates](https://www.lesswrong.com/posts/qRSgHLb8yLXzDg4nf/reviews-of-is-power-seeking-ai-an-existential-risk).
### 4.2.2 - User and System Propositions
*Probability of recognising harmful features of an AGI (i.e., the **Op**acity of the AGI): P(**Op**)*
Proposition: How likely are we to recognise the potential threat characteristics / risks associated with an AGI implementation?
Assumptions & Basis: The probability of the AGI implementer/user having sufficient understanding of the *full capabilities of the AGI*(read: misaligned goal, control subversion capability, and sufficient power to commit and X-Event) has arguments that both increase and decrease it. For example, on the one hand the increasing complexity required to design and build an AGI is likely to increase the difficulty of recognising potential risks/threats, and notably if self-modification and improvement is assumed (e.g., [Bostrom, 2014](https://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111) Ch.6 on the link between recursive self-improvement and covert preparation potential). On the other hand, advances in the development of potential interrogative tools (for examples of interpretability improvement, see e.g., [Olah, 2020](https://distill.pub/2020/circuits/zoom-in/); and [Goh et al., 2021](https://distill.pub/2021/multimodal-neurons/)) and an increasing prevalence of blanket assumptions/weariness regarding potential (deleterious) AGI capabilities, may mitigate some of this increasing opacity. On balance I still assume it to become increasingly difficult to effectively understand an AGI regarding its potential risk, but attach a high degree of uncertainty on this assumption.
Detail/Function: For now, P(Op) is assumed to increase over time, such that by 2070 there is a 50% probability that the potential implementer/user of an AGI is no longer able to understand its full capabilities, with +/- 14.16% uncertainty surrounding it (again covering a broad range of potential estimates, from approx. 20% to 80% within 2 SD).
*Probability of **S**ystem being **Su**sceptible to an AGI X-Event: P(**SSu**)*
Proposition: How likely is the system in which an AGI is implemented to have sufficient “levers” for an enacting an X-Event?
Assumptions & Basis: The probability of the system being susceptible is a proposition that itself could be unpacked further with additional analysis and research. However, for now I note several important factors that are amalgamated to inform this proposition: First, there is consideration of *which type of system* an AGI will be implemented within (e.g., military, economic, scientific, commercial), and second, what the *general connectedness* of those systems may be – irrespective of initial implementation likelihood. If systems are sufficiently connected, then an AGI’s power stems from the use of that connectedness (e.g., in a highly connected system, an AGI originally implemented within a commercial system may make use of connections to power stations, logistical hubs, and so on).
In determining the estimate for this proposition, I note arguments have been made for a wide range of potential values. For example, there are arguments that our systems are already highly susceptible (with various mechanisms laid out in [Carlsmith, 2021](https://arxiv.org/pdf/2206.13353.pdf)) – though I note this is not the same as whether they will be so in future. Alternatively, there are arguments that our systems will become increasingly connected/centralised (see e.g., globalisation and ICT trends; [Friedman et al., 2005](https://www.amazon.co.uk/World-Flat-History-Twenty-First-Century/dp/0374292884)), leading to a rising probability of susceptibility. Then again, there are arguments that other shocks, “warning shots”, and risk factors may lead to increased partitioning and segmentation (e.g., along national boundaries; see e.g., [border closure in response to contagion events](https://www.nature.com/articles/s41598-022-05482-7); [national protectionism responses to recent food crises](https://institute.global/policy/global-crisis-tackling-international-fallout-russias-war-ukraine)), which may in turn decrease the potential (X-Event) potency of an AGI application. Along these lines, I also note separate instantiations of AGIs via competing national interests (see e.g., [Bostrom, 2014](https://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111), for discussion of “multipolar” scenarios) are immensely hard to predict their overall effect (e.g., succeeding or failing to recognise warning signs in other nations).
Consequently, I highly recommend this area of inquiry for further research in shoring up this proposition.
Detail/Function: For now, P(SSu) is assumed to increase, such that by 2070 there is a probability of 20% that the system within which an AGI is implemented is vulnerable to an X-Event action by an AI, with a +/- 14.16% uncertainty surrounding it (allowing for estimates as high as 50% within 2 SD).
*Probability of **S**ystem **Bl**indness: P(**SBl**)*
Proposition: How likely is the human AGI user/implementer to understand the target system well enough to recognise possible misuse/goal misalignment/control subversion?
Assumptions & Basis: The probability of someone with the power to implement an AGI into a system knowing the relevant properties of the system itself (e.g., relevant safeguards, unintended vs intended uses). Similar to consideration of AGI feature transparency/opacity [Op], system vulnerability blindness is expected to *increase over time*. Using similar logic, this is based on trends regarding system complexity resulting in the potential AGI user/implementer being less and less likely to comprehend sufficiently the system they inhabit (note: System here can be more generalised – e.g., internet), and with particular regard to potential AGI “misuse”. But again, I note here that there are arguments that work to offset this trend. In particular, the development of potential interrogative tools and assistive technologies for things like vulnerability testing (for example, [growth in cybersecurity tools and services has seen continued, rapid growth](https://www.fortunebusinessinsights.com/industry-reports/cyber-security-market-101165) in response to increasing user awareness of their system vulnerabilities to rising threats), and the potential influence of increasing prevalence of blanket assumptions/awareness regarding potential system abuse, e.g., via previous “warning shots” and partial misalignment instances (see e.g., [Carlsmith, 2021](https://arxiv.org/pdf/2206.13353.pdf)).
Detail/Function: For now, P(SBl) is assumed to increase over time, such that by 2070 there is a 40% probability that the potential implementer/user of an AGI does not understand the target system well enough to recognise possible misuse/goal misalignment/control subversion, with +/- 14.16% uncertainty surrounding it (again covering a broad range of potential estimates, from approx. 10% to 70% within 2 SD).
4.3 - Conditional / Inferred Propositions
-----------------------------------------
These propositions are considered conditional, in that their likelihood of being true or false stems from the propositions that affect (read: cause) them. Each of these propositions requires the specification of the *strengths of the relations*between it and its parent propositions, specified using conditional probability tables (CPTs). These lay out explicitly the assumptions regarding how each parent proposition will affect this variable, *in combination with every other parent proposition*.
At the risk of belabouring a point here, I wish to make abundantly clear that the conditional probabilities outlined here are based on “if cause/premise A is TRUE, then there is a X% probability this effect B would also be TRUE” – which is not making any assessment of ***how likely cause/premise A is in the first place.***
To avoid concerns of false precision, wherever possible I have attached uncertainties to each conditional probability, considered in the simulation process. Although time series trends could also be specified for each of these conditional probabilities as well (as was done for the prior/baseline input probabilities), for the sake of parsimony this is omitted for now.
Additionally, I have taken the liberty of colour-coding the conditional probabilities in my CPTs. Specifically, estimates in **green** are drawn from logic, and thus I feel are self-evidently quite solid. Estimates in **yellow** are informed by pre-existing analyses, but carry uncertainty given their novel application here. Estimates in **orange** *are novel considerations* that have not yet received thorough analysis and estimation by experts. Although I have generated what I believe to be plausible placeholders, these latter values are excellent candidates for further analysis. *This helps demonstrate the value of this approach in enforcing **exhaustiveness** and **transparency** when we make our estimates*.
### 4.3.1 - Motive
*Probability of an AGI possessing a **M**is**A**ligned **G**oal: P(**MAG**|SBl,De)*
Proposition: How likely will the AI design lead to misaligned goals when implemented within the system?
Assumptions & Basis: The probability of a potential AGI design resulting in goal misalignment when implemented within a system (to deleterious effect) can be argued as a consequence of *poor fit* between the AGI design (i.e., its derivation of goals) and the context of its initial system of implementation. Here I am assuming the AGI design [De] (replete with features like agentic planning and strategic awareness) is highly likely to develop misaligned goals (again, see [Carlsmith, 2021](https://arxiv.org/pdf/2206.13353.pdf), for a recent analysis of the relationship between these features and misalignment goals).
However, this misalignment probability can also be influenced by the user/implementers understanding of the system/context intended for the AGI (i.e., [SBl]). Put another way, an implementer with excellent system awareness (i.e., [¬SBl]) might be able to lower the risk of misalignment by changing the system to re-define goals a priori, for instance by adequately attending to and making use of potential “warning shots” and possible previous partial failures (see Carlsmith, 2021). Conversely, if system users are blind to the nature of their system/context, then this increases the likelihood of an AGI successfully operating with a misaligned goal. When we consider this assumption in relation to the larger question of an AGI X-Event, we can consider the misaligned goal as equivalent to ***motive*** in a legal case.
Detail/Function: The inference function here is primarily a logic gate: a misalignment risk cannot occur if there is not a viable AGI design [¬De]. However, if there is a viable AI design, then it is highly likely (60-80%) that a misalignment will occur, with system blindness [SBl] pushing this towards the upper end of this likelihood (80%, +/- 10%), versus accurate system knowledge [¬SBl] lowering this slightly (60%, +/- 10%). In this way, system awareness may make *some* difference to misalignment risk, but this risk is primarily dictated by the AGI design.
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **MAG** | **¬MAG** | **Uncertainty (SD)** |
| P(MAG|De,SBl) | 80% | 20% | +/-10% |
| P(MAG|De,¬SBl) | 60% | 40% | +/-10% |
| P(MAG|¬De,SBl) | 0% | 100% | 0 |
| P(MAG|¬De,¬SBl) | 0% | 100% | 0 |
### 4.3.2 - Opportunity
*Probability of an AGI Circumventing (human) Control: P(**CCo**|SBl,De)*
Proposition: How likely will the AGI design lead to the capacity to circumvent human control within a system?
Assumptions & Basis: The probability of an AGI subverting existing control is conditional on the design of that AGI [De] incorporating necessary features for subversion (e.g., agentic planning, strategic awareness; see [Carlsmith, 2021](https://arxiv.org/pdf/2206.13353.pdf)), and (again, to a lesser degree) the implementers understanding of the system/context intended for the AGI [i.e., SBl]. An implementer blind/ignorant of the risk of control circumvention increases the risk that an AGI will be unimpeded in circumventing control. Whilst if an implementer knows their target system well, they *may* be able to lower control circumvention risk (e.g., via a priori intervention, or external failsafe designs), as well as making use of potential “warning shot” partial failures from previous (minor) implementations. When we consider this assumption in relation to the larger question of an AGI X-Event, we can consider the capacity to circumvent control as equivalent to ***opportunity*** in a legal case.
Detail/Function: As with misalignment risk, the inference function here is also primarily a logic gate: a circumvention of control risk cannot occur if there is not a viable AGI design [¬De]. However, if there is a viable AGI design, then it is highly likely (60-80%) that circumvention will occur, with system blindness [SBl] pushing this towards the upper end of this likelihood (80%, +/- 10%), versus accurate system knowledge [¬SBl] lowering this slightly (60%, +/- 10%). In this way, system awareness may make *some* difference to circumvention of control risk, but this risk is primarily dictated by the AGI design.
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **CCo** | **¬CCo** | **Uncertainty (SD)** |
| P(CCo|De,SBl) | 80% | 20% | +/- 10% |
| P(CCo|De,¬SBl) | 60% | 40% | +/- 10% |
| P(CCo|¬De,SBl) | 0% | 100% | 0 |
| P(CCo|¬De,¬SBl) | 0% | 100% | 0 |
### 4.3.3 - Means
*Probability of an AGI having sufficient Power to cause an X-Event: P(**Po**|De,SSu)*
Proposition: How likely will the designed AGI *be able to* cause an X-Event?
Assumptions & Basis: The probability of an AGI possessing sufficient power to cause an X-Event is conditional on both the design of the AGI [De] containing requisite features, and the susceptibility of the system within which it is implemented [SSu]. The assumption here is straightforward:
For an AGI to have sufficient power to cause an X-Event, the AGI requires both a viable design [De] (which, in line with other assumptions, is argued as possessing requisite capabilities, such as agentic planning and strategic awareness), and a system that possesses the potential for X-Event level actions [SSu]. As a logic gate, if one or both of these premises are false (the design is not viable/relevant for X-Risk consideration [¬De], and/or the system does not have sufficient direct or connected access to required “levers” for an X-Event [¬SSu]), then this proposition is not possible. We can consider this conceptually in legal reasoning terms as a necessary condition of ***means*** with regards to the main question of an AGI enacted X-Event.
Detail/Function: The integration function here is a simple logic gate: for there to be a risk of an AGI possessing sufficient power to cause an X-Event [Po], there must both be a viable AGI design [De] *and* a susceptible system [SSu]. I note if this conjunction is satisfied, then I assume sufficient power is *almost* guaranteed (i.e., 100%), by including an uncertainty on this parameter (+/- 5%), allowing for the unlikely possibility of unaccounted for interruption/intervention.
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **Po** | **¬Po** | **Uncertainty (SD)** |
| P(Po|SSu,De) | 100% | 0% | +/- 5% |
| P(Po|SSu,¬De) | 0% | 100% | 0% |
| P(Po|¬SSu,De) | 0% | 100% | 0% |
| P(Po|¬SSu,¬De) | 0% | 100% | 0% |
### 4.3.4 - Incentive
*Probability of there being an **In**centive for AGI: P(**In**|Op,MAG,NCo,Po)*
Proposition: How likely will we (humanity) *want to build and deploy* an AGI?
Assumptions & Basis: The probability of there being sufficient incentive for developing/deploying an AGI into a system is conditional on the probability of an AGI possessing misaligned goals [MAG], being able to circumvent/subvert human control [CCo], having the power to cause an X-Event [Po], and the transparency of these features to the implementer [Op]. We can broadly take point of departure as [Carlsmiths (2021)](https://arxiv.org/pdf/2206.13353.pdf) estimation of an 80% probability of there being incentive for deployment. However, I go further here to unpack this probability, based on the *conjunction of additional relevant parent propositions*. More precisely, this 80% probability should decrease *if* there is an awareness [Op] of the potentially harmful features of the AGI implementation (i.e., misaligned goals [MAG], capacity to circumvent control [CCo], and X-Event power [Po]), using the same logic as applied elsewhere in the analysis (e.g., awareness raised via potential “warning shot” partial failures from previous (minor) implementations). In such cases, the more severe the risk profile (e.g., the “near miss” indicated misalignment, control circumvention and potential X-Event power were all possible), the lower the incentive for implementation. Details of these combinations are described below.
As elsewhere, I take pains to note the nature of this conditional probability: at this point we are not making any claims about the likelihood of each conjunction being true, merely the likelihood of there being an incentive [In] for AGI when we consider each conjunction in turn to be true (i.e., the strength of a relation =/= the likelihood of that relation being activated).
Detail/Function: The integration function for the incentives for AGI [In] is entirely probabilistic, and therefore contains the greatest uncertainty (relative to other inferred propositions). Generally, if the potentially harmful features associated with an implementation (sufficient power to cause an X-Event, [Po]; misaligned goals, [MAG]; and the ability to circumvent human control, [CCo]) are opaque to the potential implementer [Op], then regardless of the presence or absence of each of those features, the incentives are high (80%, +/- 10%). This is equivalent to the incentive when an implementer can determine features accurately [¬Op], and those features are not present (i.e., [¬MAG, ¬CCo, ¬Po]: the AGI is controllable, has no misaligned goals, and no power to commit an X-event).
Consequently, if an implementer can detect deleterious factors, then the incentives for an AGI decrease. *These parameters are speculative, but I believe follow a reasonable and coherent risk-mitigation logic*, such that: If an AGI is known to be able to circumvent control [CCo] and possess misaligned goals [MAG], then regardless of power [Po], the incentive reduces to 1% (+/-5%); if this AGI will not do what I want it to do in an acceptable (read: aligned) manner, and I cannot control it, why would I want it? However, if an AGI is known to be controllable [¬CCo], then knowledge of possible misaligned goals [MAG] and sufficient X-risk power [Po] reduces probability of incentive to 60% (+/-10%), and to 70% (+/- 10%) if there is no X-risk power [¬Po]; my control enables it to still be potentially *useful*, though the added risk from the AGI being misaligned lowers the incentive.
Conversely, if an implementer knows the AGI does not possess misaligned goals [¬MAG], then a lack of control [CCo] and sufficient X-Risk power [Po] reduces the likelihood of incentive to 30% (+/-10%); if something goes badly wrong, I cannot stop it, but if controllable [¬CCo] this rises to 70% (+/- 10%). If, in addition to an absence of misaligned goals [¬MAG], the implementer knows there is not sufficient X-Risk power [¬Po], then a lack of control is less important, resulting in a probability of desirability of 70% (+/-10%) – versus the 30% probability when there is sufficient power for an AGI enacted X-Event [Po].
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **In** | **¬In** | **Uncertainty (SD)** |
| P(In|MAG,Po,Op,CCo) | 80% | 20% | +/- 10% |
| P(In|MAG,Po,Op,¬CCo) | 80% | 20% | +/- 10% |
| P(In|MAG,Po,¬Op,CCo) | 1% | 99% | +/- 5% |
| P(In|MAG,Po,¬Op,¬CCo) | 60% | 40% | +/- 10% |
| P(In|MAG,¬Po,Op,CCo) | 80% | 20% | +/- 10% |
| P(In|MAG,¬Po,Op,¬CCo) | 80% | 20% | +/- 10% |
| P(In|MAG,¬Po,¬Op,CCo) | 1% | 99% | +/- 5% |
| P(In|MAG,¬Po,¬Op,¬CCo) | 70% | 30% | +/- 10% |
| P(In|¬MAG,Po,Op,CCo) | 80% | 20% | +/- 10% |
| P(In|¬MAG,Po,Op,¬CCo) | 80% | 20% | +/- 10% |
| P(In|¬MAG,Po,¬Op,CCo) | 30% | 70% | +/- 10% |
| P(In|¬MAG,Po,¬Op,¬CCo) | 70% | 30% | +/- 10% |
| P(In|¬MAG,¬Po,Op,CCo) | 80% | 20% | +/- 10% |
| P(In|¬MAG,¬Po,Op,¬CCo) | 80% | 20% | +/- 10% |
| P(In|¬MAG,¬Po,¬Op,CCo) | 70% | 30% | +/- 10% |
| P(In|¬MAG,¬Po,¬Op,¬CCo) | 80% | 20% | +/- 10% |
### 4.3.5 - Existence
*Probability of an AGI being successfully **Bu**ilt: P(**Bu**|In,BFe,De)*
Proposition: How likely is it an AGI will *be successfully built*?
Assumptions & Basis: The probability of an AGI being built is conditional on it being feasible/affordable [BFe], having a viable design [De], and the incentive for an AGI [In] – a moderating condition. The first basic assumption here is that a successful AGI build requires *both* hardware affordability and feasibility [BFe], *and* a viable design [De]; if either of these premises are false, then an AGI cannot be built. The second assumption is that the *incentives* will also affect the likelihood of a build (e.g., via investment and support). Importantly, we are not dealing with how likely it is that there will there be an incentive (P(In)) here, but rather the impact of that premise on build likelihood (P(Bu)), *given it is true/false*.
Additionally, it is worth noting that P(In) already takes into account possible awareness of risks and other disincentives (see [In] assumptions above). Here we are dealing with the prospect of absolute, unreserved incentive; the degree to which this desire is reserved / conditional is handled by the input premise [In]. Consequently, I assume here that if there is an incentive for AGI, then – assuming build-enabling conditions are met – there is an extremely high likelihood of a successful build (95%, +/-5%; allowing for the possibility of unforeseen factors that prevent a successful build).
However, if there is ***no*** incentive [¬In], then there is a lot more uncertainty regarding a successful build. In my estimation there are arguments on both sides of this potential influence. For example, on the one-hand a successful build, if *possible*, becomes *inevitable* (i.e., [¬In] is irrelevant), due to prisoner’s dilemmas (see e.g., relevant discussions on multi-agent dynamics in [Critch & Krueger, 2020](http://acritch.com/papers/arches.pdf)), competing external interests ([Bostrom, 2014](https://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111)), and simple curiosity. An alternative perspective is that without sufficient continued investment from substantive enough parties, then development projects continuing past potential roadblocks and setbacks becomes unlikely. Put another way, *if* development is perceived as both costly *and* there is no substantive appetite to deploy (e.g., due to perceived risks, handled in [In] calculation), then this should at least *decrease* the likelihood of development on the basis of a single-agent cost benefit analysis. Given this, I have elected to assume high uncertainty (i.e., 50%, +/-10%; such that estimates can range from ~30% to ~70% within 2 SDs), and note it as an assumption that is a strong candidate for further research.
Detail/Function: Once more, our integration function is primarily a logic gate: An AGI cannot be built without a build being feasible/affordable [BFe], *and* there being a viable AGI design [De]. The incentive for an AGI [In], in combination with this conjunction, results in an extremely high probability of an AGI being built (95%, +/-5%). If there is no incentive for AGI [¬In], then I assume a large degree of uncertainty of a successful AGI build (50% +/- 10%).
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **Bu** | **¬Bu** | **Uncertainty (SD)** |
| P(Bu|In,BFe,De) | 95% | 5% | +/- 5% |
| P(Bu|In,BFe,¬De) | 0% | 100% | 0% |
| P(Bu|In,¬BFe,De) | 0% | 100% | 0% |
| P(Bu|In,¬BFe,¬De) | 0% | 100% | 0% |
| P(Bu|¬In,BFe,De) | 50% | 50% | +/- 10% |
| P(Bu|¬In,BFe,¬De) | 0% | 100% | 0% |
| P(Bu|¬In,¬BFe,De) | 0% | 100% | 0% |
| P(Bu|¬In,¬BFe,¬De) | 0% | 100% | 0% |
4.4 - Central (Inferred) Proposition of Interest
------------------------------------------------
As far as the Bayesian network modelling process is concerned, this variable is not qualitatively different from any other inferred variable. I have separated it only to highlight its special status in relation to the central question at hand.
*Probability of a misaligned **AGI** Causing an **X**-Event: P(**AIX**|MAG,CCo,Bu,In)*
Proposition: What is the probability of a misaligned AGI causing an X-Event?
Assumptions & Basis: The probability of an AGI causing an X-Event is conditional on an AGI being built [Bu], that AGI possessing misaligned goals [MAG], being able to circumvent control [CCo], having sufficient power to enact an X-Event [Po], and lastly, the desire of a potential implementer to deploy it [In].
Put simply, given the current model, we need ***means***[Po]*, **motive***[MAG]*, **opportunity*** [CCo]*,*and in the case of a created entity ***existence***[Bu] as *strict, necessary conditions* for a misaligned AGI X-Event. On top of this, given there are creators and users/implementers as causal agents within the problem, we additionally have creator/implementer ***incentive***[In]. Logically, we can see that without all four necessary conditions, it follows that an AI misalignment X-Event is not possible. However, uncertainty is introduced through the impact of incentives [In].
Given there is an incentive [In], we can assume a high likelihood of an AGI misalignment X-Event if the four necessary conditions are met. Our structuring has allowed us to handle the possible uncertainties surrounding *wanting to implement* ([In] calculation), the uncertainties surrounding potential dangers (i.e., [MAG], [CCo] and [Po]), and the uncertainties surrounding a successful AGI build ([Bu]) elsewhere. Put another way, in the conditional space where all pre-requisites have been met, then the existence of an incentive (to deploy) all but ensures enactment of the X-Event. I assumed it to be 95% (in line with [Carlsmith’s estimate of existential catastrophe](https://arxiv.org/pdf/2206.13353.pdf) given activation of a misaligned AGI).
However, if there is no incentive [¬In] (e.g., due to accurate recognition of the severe threat posed – again, this is handled in [In] calculation), then I still assume there remains a possibility of a misaligned AGI X-Event (e.g., via unintended / erroneous implementation, including through AGI deception, or bad-actor dynamics; [Bostrom, 2014](https://www.amazon.co.uk/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/0199678111)). This ranges from 0% to 40% (mean 10%) to reflect a range of potential assumptions (e.g., inevitable implementation vs extreme threat response crack-downs).
Detail/Function: This final integration function is broadly a logic gate as well: An AGI misalignment X-Event requires a successful AGI build [Bu], that AGI to have sufficient power to cause an X-Event [Po], be able to circumvent human control [CCo], and possess a misaligned goal [MAG], to all be true. Assuming this conjunction to be true, then if there is an incentive for AGI, then a misaligned AGI X-Event is almost certain (95%, +/- 5%), though if there is no desire to deploy the AGI, then there is still a possibility of a misaligned AGI X-Event regardless (10%, +/- 10%).
| | | | |
| --- | --- | --- | --- |
| **Conditional**[[6]](#fncn9a2c9v2ib) | **AIX** | **¬AIX** | **Uncertainty (SD)** |
| P(AIX|MAG,CCo,Bu,In,Po) | 95% | 5% | +/- 5% |
| P(AIX|MAG,CCo,Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,CCo,Bu,¬In,Po) | 10% | 90% | +/- 10% |
| P(AIX|MAG,CCo,Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,CCo,¬Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,CCo,¬Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,CCo,¬Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,CCo,¬Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,¬Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,¬Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,¬Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|MAG,¬CCo,¬Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,¬Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,¬Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,¬Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,CCo,¬Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,Bu,¬In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,¬Bu,In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,¬Bu,In,¬Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,¬Bu,¬In,Po) | 0% | 100% | 0 |
| P(AIX|¬MAG,¬CCo,¬Bu,¬In,¬Po) | 0% | 100% | 0 |
5 - Results
===========
By this point, I have now made *all* my requisite assumptions, complete with specified probabilities and surrounding uncertainties, explicit. I again reiterate here the value of this process, *regardless of how specifically (in)correct my particular model may be.* We must seek to be exhaustive, explicit, and transparent in both our **acquisition** of propositions, and their **estimation**. Even then, we must also structure our propositions correctly, or accurate **integration** (and therefore prediction) will always remain beyond reach. The analytical approach I have outlined enables this, avoiding compounding errors where they have previously gone unaccounted.
The results and conclusions that follow, critically, are *optimally derived, given my explicitly stated assumptions – many of which stem from current expert opinion*.
*Fig. 3: AGI Misalignment X-Risk Bayesian Network stochastic simulation outputs. Time series reflects annual probability updates for each proposition. Arrows reflect causal structure / argument mapping of the network (Cause -> Effect). Ribbons reflect second-order uncertainty (+/- 1 SD).*5.1 - Evidence of Complexity
----------------------------
Despite making simple, linear assumptions for our parent propositions, we can see that the propositions we are inferring do not combine linearly. This is important to highlight, as it underscores just how conjunctions and the asymmetries they can contain make inferring predictions and trends immensely difficult without computational assistance. Put another way, because parent propositions rarely combine in an *exactly additive, equal way*, the inferences their combination produces are rarely easy to intuit. For example, many of our intermediary (inferred) propositions – **and the uncertainty that surrounds them** - are best described by sigmoid functions (e.g., P(Built)). This raises a number of implications:
1. If one layer of conjunctions already leads our inferences out of the realm of intuition, then **our errors will only compound** when those conjunctions must serve as parent propositions to another conjunction in turn (e.g., P(AIX) in the current model).
2. These conjunctions, the asymmetries within them, and the computational complexity that results ***are present in the problem whether we choose to represent them or ignore them***.
3. The Bayesian Network approach advocated here allows us not only to capture and describe them (enforcing **estimate transparency**), but provides the essential computational assistance to **integrate** (and therefore ***predict***) optimally.
4. The simulation “wrapper” surrounding the Bayesian Network enables us to gain these same, essential advantages when enfolding both how our estimates *change over time*, but also the *uncertainty surrounding those estimates*.
5.2 - Central Findings
----------------------
For all the prize-relevant questions, I am choosing to report both the central point-estimate probabilities, and the uncertainty bands (here specified as +/- 1 SD) surrounding those estimates. If in judging this submission, you wish to use the upper-bound of these bands (which takes many of these estimates over the other side of prize thresholds), then so be it. I stand by the need to report them regardless.
With that said, here are the findings derived from optimal (Bayesian) inference across the model structure and parameters:
***Probability of a misaligned AGI causing an X-Event by 2070 = 1.68% (+/- 1.25%)***[[7]](#fnc8s7iotfsgt)
*Probability of a misaligned AGI causing an X-Event by 2100 = 6.53% (+/- 4.32%)*
However, if we wish to know the probability of an AGI misalignment X-event *assuming that an AGI is successfully built* – as is suggested by the proposition “P(misalignment x-risk**|AGI**)” (my emphasis), then we can condition the network on this observation (i.e., set P(Bu) to TRUE). Fig. 4 below shows the respective difference this makes to P(AIX) over time (shown in red).
*Fig. 4: AGI Misalignment X-Risk outcome simulation outputs. Time series reflects annual probability updates. Red = P(AIX) when P(Bu) = TRUE; Blue = P(AIX). Ribbons reflect second-order uncertainty (+/- 1 SD).*This gives us:
***Probability of a misaligned AGI causing an X-Event by 2070, given an AGI has been built = 5.32% (+/- 3.42%)***[[8]](#fnb621wwqo36n)
*Probability of a misaligned AGI causing an X-Event by 2100, given an AGI has been built = 11.22% (+/- 6.86%)*
In addition to addressing misaligned AGI X-Risk, the approach also allows us to ask additional pertinent questions. For example, there is also interest in the probability of a successful AGI build across several timeframes. Using the model, I can simultaneously “read off” estimates for this proposition (P(Build)), such that:
***Probability of a successful AGI build by 2043 = 5.61% (+/- 2.65%)***
*Probability of a successful AGI build by 2070 = 31.55% (+/- 10.02%)*
***Probability of a successful AGI build by 2100 = 58.12% (+/- 11.63%)***
Crucially, these subsidiary proposition estimates automatically *cohere* with all other proposition estimates, as they are predicated upon the same structure, parameters, and calculus. Put another way, ***if*** you are happy with the assumptions that produced the probability of a misaligned AGI X-Event (P(AIX)) result, then you should equally be satisfied with the probability of a successful AGI build (P(Bu)) result. Of course, it is both welcomed and expected that there will be disagreement regarding these assumptions. At least this way there is an explicit method of getting these assumptions out in the open, specified, and determined for their impact (i.e., *directly addressing goals of **exhaustiveness** and **transparency***).
6 - Conclusions and Path Forward
================================
I have presented an analytical approach here that addresses concerns pervasive to many long-range prediction problems. This analytical approach first endeavours to complement and enhance the necessary **acquisition** and **estimation** processes necessary for making accurate predictions. Second, but perhaps more critically, the approach provides an (optimal) **integration** method, avoiding a number of compounding, systematic errors. I have outlined how these reasoning structures can produce systematic errors, and note that the approach I have provided addresses the analytical weaknesses raised as concerns within AI X-Risk research (see e.g., [here](https://www.lesswrong.com/posts/qnA6paRwMky3Q6ktk/modelling-transformative-ai-risks-mtair-project-introduction) and [here](https://www.openphilanthropy.org/research/semi-informative-priors-over-ai-timelines/)).
My hope here is that the approach I have taken is just the beginning. I have taken pains to make this initial foray as accessible as possible, whilst capturing meaningful structural complexities. I have no doubt this model can be corrected and enriched on several fronts, including (but not limited to):
* The adjustment of parameter values and their surrounding uncertainties. Including the selection of alternative uncertainty distributions and time-series functions.
* The addition / removal of links between propositions, based on arguments for which I have not yet accounted.
* The addition of new propositions, whether they be deployed to “**unpack**” current propositions into further sub-structures of connected propositions, or as additional influences on the overall structure.
Crucially, these corrections and additions, through exposure to the framework, must already engage with the *exhaustive unpacking of its related uncertainties* (e.g., in proposing a new parent proposition, one must automatically specify not only its prior probability, but also consider how it may conditionally relate to the other propositions extant in the reasoning structure). In this way, I also see opportunities for collaborative work, whether it be in improving this model, or in comparing and contrasting a population of competing models.
Along with model corrections and enrichments this analytical approach lends itself to the following avenues of future work:
* **Sensitivity**: The model can be used to not only test competing propositions/assumptions, but also see which remaining unknowns or disagreements *actually matter* for the larger question at hand. This can be bolstered through the use of information theoretic measures (see e.g., [Kullback, & Leibler, 1951](https://www.jstor.org/stable/2236703#metadata_info_tab_contents); [Nelson, 2008](https://d1wqtxts1xzle7.cloudfront.net/3390174/Towards_a_rational_theory_of_human_information_acquisition-with-cover-page-v2.pdf?Expires=1667304411&Signature=JOL9yaRHUSSlSx2ztQW8E5bX5FsGgn-70Qwx-sY06ZqQXO8TSTLMAEjiAZJoZbsdJU5HYLEjXLT1S4bluSWcqhYoFPcb1fHsm1JRWOXA9a2UxARaAC17wzcFsg6i7T2nOSUwL5IK4da9DJTMqxDDb5HBrbVIZktCvdw95aF3jvfhzb19YOuaHJ7Xtci~OU4-AhKqn4nm6gV9zH7UDyr2XZMntTT6DBGUqfjXTxy~3H6PDbtGq9OqNUDgL9fvyfWJTlFE3qWUfQlNoFxrZyDMS858Oh-OS2IKGK9ITLieB-X~Bhs67QNVso5UczPuRmx2Hvfg22roJbVWAQHiBuAz1A__&Key-Pair-Id=APKAJLOHF5GGSLRBV4ZA)) based on entropy reduction that enable us to rank propositions by their *informational value* (whether to the reasoning structure at large, or in relation to a particular proposition of interest).
+ The knock-on effect of this is assistance in more efficient guidance of **acquisition** and **estimation** efforts.
* **Temporal Dynamics**: Though the analytical approach provides a set of estimates over the time-period of interest that cohere (i.e., their formulation is computationally consistent), more advanced intertemporal dynamics are not yet incorporated. The approach can be expanded, either via shifting to [Dynamic Bayesian Networks](https://en.wikipedia.org/wiki/Dynamic_Bayesian_network), or via introduced feed-throughs and parallel-world analyses (e.g., when X reaches a given threshold, we split the model into a world where X is now considered to have happened, and where X hasn’t happened yet).
* **Modularity**:The approach lends itself to modular, parallel research streaming. By this I mean that other projects can be informed by components (e.g., reasoning structures) present in this model, or vice versa. This affords the opportunity for mutual enrichment on analysis across individuals and teams.
* **Further Interrogation**: Even with the current model as it stands, we can ask many more questions than I have had time to address. For instance, given the current assumptions, instead of framing a target question around a particular date (2070), we can instead ask questions like:
*“By what year can we expect P(misaligned AGI X-Risk) to surpass our tolerance threshold (e.g., 1%, 5%)?”*
Or
*“If we assume a misaligned AGI X-Event has occurred, how likely were we to see certain indicators (propositions) beforehand?”*
The stakes of getting these predictions wrong are immense. The challenge we face is in 1) **acquiring relevant arguments and propositions**, 2) **estimating our uncertainties surrounding them**, and 3) **integrating them accurately**. The analytical approach put forward here takes major steps in assisting all three, by increasing the *transparency*, *clarity*, *consistency*, and *accuracy* of our reasoning.
*My thanks to Dr Lawrence Newport for helpful comments on earlier versions of this text.*
***All errors lie with me.***
***EDIT: Minor Formatting.***
1. **[^](#fnref8diqhvi598n)**What I cover here has been said more comprehensively by many others in the reasoning-under-uncertainty literature (e.g., [Pearl, 1988](https://books.google.co.uk/books?hl=en&lr=&id=AvNID7LyMusC&oi=fnd&pg=PA1&dq=Pearl+1988&ots=F--QSotu09&sig=hyOFOV0WHbgR_9lhbO52ilMRUAI&redir_esc=y#v=onepage&q=Pearl%201988&f=false); [Schum, 1994](https://books.google.co.uk/books?hl=en&lr=&id=8PqsTNULCPEC&oi=fnd&pg=PR13&dq=Schum+1994+the+evidential+foundations&ots=s98QmsLHxL&sig=eOV9okPV1k4fyw4ZIQ7INz9wm0A&redir_esc=y#v=onepage&q=Schum%201994%20the%20evidential%20foundations&f=false)). I encourage interested readers to seek them out.
2. **[^](#fnrefpotccwwjxs)**In the current examples, I am assuming that propositions only have two states: true, or false. Of course, it is possible to have >2 states for a given proposition (e.g., gradations of severity), which only increases the computational explosiveness.
3. **[^](#fnreflu6ff3gujv)**I take pains to note example analyses such as the MTAIR Project (see [here for introduction](https://www.lesswrong.com/s/aERZoriyHfCqvWkzg/p/qnA6paRwMky3Q6ktk#How_to_read_Analytica_models), and [here for a recent paper](https://arxiv.org/ftp/arxiv/papers/2206/2206.09360.pdf)) and [Tom Davidson’s report on AGI development timelines](https://www.openphilanthropy.org/research/semi-informative-priors-over-ai-timelines/) have made explicit references to the importance of considering conjunctions, and mapping their interactions.
4. **[^](#fnref7v67i7xwxjp)**Though our offspring propositions are now linked via the parent, making their independence conditional on observation of the parent. In short, this structure means learning about one offspring will tell us something about the other.
5. **[^](#fnref7ve0z9ypn6d)**When we fold in differences in 2nd order uncertainty across this more complex structural picture, we can see why integrating these complexities without assistance is likely to lead to trouble, and highlights the danger of unspecified/hidden assumptions even further.
6. **[^](#fnrefcn9a2c9v2ib)**As a reminder: The colour-coding of conditional probability estimates is intended to flag which are on more or less solid ground.
**Green** = Drawn from logic
**Yellow** = Informed by pre-existing analyses, but carry uncertainty given their novel application here.
**Orange** = *Novel considerations* that have not yet received thorough analysis and estimation by experts. Although I have generated what I believe to be plausible placeholders, these latter values are excellent candidates for further analysis.
7. **[^](#fnrefc8s7iotfsgt)**I would be remiss not to point out that [another recent analysis focussing on uncertainty analysis](https://forum.effectivealtruism.org/posts/Z7r83zrSXcis6ymKo/dissolving-ai-risk-parameter-uncertainty-in-ai-future) has produced a very similar estimate – via a more thorough exposition of uncertainty across order of magnitude, rather than via structural disentanglement.
8. **[^](#fnrefb621wwqo36n)**I take pains to note here that there is a potential sequencing issue here between exactly *when* the AGI was built, and P(AIX) at 2070, which is not reflected in this latter estimate. Further work using dynamic causal Bayesian Network versions of this model could assist in resolving this question (see Section 6). |
b0524f71-2b14-4b17-9123-c9d1879564e6 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | [Creative Writing Contest] Metal or Mortal
**This story** was first published in [**ANDROMEDA MAGAZINE**](https://andromedaspaceways.com/product/asm-80/) (Sept 2020, Issue #80). It's a modern-day fable that involves a variety of issues relevant to Effective Altruism.
**Metal or Mortal**
===================
### **by Louis B. Rosenberg**
It began as a tapping sound, metal on metal, like a chisel being worked by skilled hands. This was followed by a grinding noise, then a scraping sound, then the return of steady tapping. I tried not to get my hopes up, as I had experienced many false alarms. The last time was 475 years ago during a construction project ordered by Henry the Great.
Back then I was dug up by a group of leather-clad men with iron shovels. They must not have noticed me among the silt and soil, for I ended up in a muddy mound of backfill. A structure of stone and wood was erected above me. I guessed it was a hunting lodge for the nobles of Paris, but I was never sure, as I could only hear snippets of conversations among the workers. And then came the silence, long and dark, until now.
CLANG – the tapping suddenly got louder.
And then slower, like extra care was now being taken.
A few more taps and it happened – I was broken free from the centuries of sediment and abruptly lifted into the air. Again, I didn’t get my hopes up, as this wouldn’t be the first time I was grabbed by eager hands, only for the opportunity to slip away.
It was a Gaul Horseman who found me the last time, dropping me into his saddlebag and bouncing me across the plains of Europe before he was abruptly killed in battle. That’s how I ended up here all those years ago, abandoned on the swampy banks of the river Seine, slowly sinking into the salty muck.
But this time was different.
This time, I wasn’t tossed into a saddle bag or dropped onto a muddy pile.
This time, I was being examined by curious eyes.
I could tell by the way I was held in the air, flipped and turned, as if inspected from every angle. I wondered if it was a man or woman who was handling me so carefully, and whether they were young or old, rich or poor, noble or common. But most of all, I wondered what drove this person to dig for me in this location. Were they looking for treasure, hoping to make their fortune, or was this all just a happy accident?
That’s when it came – the sound I so desperately wanted to hear. It was the swipe of a damp cloth across my surface. Someone was rubbing away the centuries of tarnish and grime. I counted off the strokes – one, two, three – and then – I was free!
“Behoooold!” I bellowed at the man standing before me, my ancient silver lamp clutched in his hands. “I am *Aldamar*, honorable genie of the Mystic Realm. It is within my vast and humbling powers to grant you three wishes. So tell me, *O’ fortunate one*, what is it that you most desire?”
The man gazed back, silent and still, not the slightest bit jarred.
Never before had I experienced such a reaction, for there was always an instant of shock followed by fearful curiosity. And this was always replaced by the slow and steady build of greedy excitement. But this man showed none of those things. He just stared at me for a thoughtful moment and then looked back down at the lamp.
Had I introduced myself in the wrong language?
I had assumed the inhabitants of this land still spoke proper French, but that could have changed over the centuries. So I tried Latin and Greek followed by English and German and Hebrew and Farsi. But still, the man showed no reaction.
He just stood there, focused on the lamp, his gloved hand inspecting the surface with a sharp metal instrument that looked like a jeweler’s tool.
That’s when I noticed my surroundings. This wasn’t the lush French countryside that I so fondly recalled. Instead the terrain was dry and barren, like the high deserts back home, complete with blowing sands and rolling dunes, extending far into the distance.
“Have I been summoned back to the motherland?” I pondered aloud.
The man finally focused his attention upon me.
“You're a Jinn,” he said in ancient Arabic, "a Marid Jinn, circa 950 BC.” He then glanced around, as if assessing our surroundings. “No, this is not your motherland.”
“You speak the ancient tongue!” I exclaimed. “I’ve not heard that dialect in over 2000 years, and yet you utter with perfection. How is that possible?”
“I am fluent in all languages,” the man replied. “I chose ancient Arabic based on your tunic and turban, which date back to Persia of the 9th and 10th century BC.”
“You speak *all* *languages*,” I pondered. “Are you born of the Mystic Realm?”
“I am not.”
“And yet you claim to have mastered each and every tongue?” I asked him in the rarest of Mongolian dialects, assuming he would be stumped and proven a liar.
“Yes, I was well-trained in my youth,” he replied like a native Mongol.
He then focused his attention back on my lamp, applying a bead of clear liquid with a medicine dropper. After a moment of silent observation, he smiled – “This is pure silver, chemically pure, not a Sterling alloy tainted by zinc or copper. It is an extremely rare and useful find these days - critical for survival.”
“You were schooled in the sciences as well?”
“I was schooled in all disciplines.” he replied without emotion. He then turned and stepped to a set of tools that were scattered on the ground.
Without a word, he began loading his equipment into a large satchel.
It was like he had absolutely no interest in me – an ancient Marid Jinn standing directly before him, offering him three wishes of unlimited scope and power.
I was dumbfounded by his behavior.
Finally I asked, “Are you a knight of this realm?”
“A knight?” he seemed confused. “Why do you suppose that?”
“From the craftsmanship of your armor,” I said with admiration. “I have travelled far and wide, but never have I seen plated metal fit a man with such perfection.”
“I’m not a knight,” he returned simply, “and this is not armor.”
He then grabbed his right arm with his left hand and squeezed. A hidden panel opened, revealing instead of blood and bones, a patchwork of wires and gears and pistons.
“You're a conjuring!” I gasped and took a step back. “I have encountered the stringless marionettes of Hephaestus and the stone golems of Qumran, but never have I faced a fiendish imposter as convincing as you!” I took another step back and readied myself for battle. “Tell me, *O’ devious one*, whose Dark Magic instilled you with such evil?”
The man just stared at me, confused.
“And how did you summon me from the lamp?” I roared, my fear now turning to fury. “Only *the living* can summon a Jinn from confinement, *the truly living*, not a conjured forgery. What kind of abomination are you?”
“I am not a conjuring,” the man replied, no hint of being offended.
“Then what are you?” I bellowed, my voice echoing across the dunes.
“I am a Cyberdroid, a fifth-generation multipurpose unit. My specialty is search and salvage. And by all meaningful standards, I am *very much*alive.”
“Impossible!” I barked. “You cannot be both metal and mortal.”
“I disagree,” he returned politely, “although I am more than just metal.”
He then opened a panel on the side of his face, revealing a glowing complexity unlike anything I had ever seen. There were strange shimmering devices where his eyes and ears should be, a wild nest of glowing fibers where his brain should reside. I stared, horrified, for this creature should not be alive, and yet *he had* summoned me from the lamp. Could there be new forms of wizardry that instill *true life* into the lifeless?
“Tell me this,” I demanded, “were you built or born?”
“I was born,” he returned calmly, “by being built.”
“Nonsense,” I fired back. “What kind of unholy magic builds a mortal man of parts and pieces?”
“No magic,” he replied. “I was assembled by other Cyberdroids.”
“And whose magic created them?” I scoffed.
“No magic. They were assembled by the Cyberdroids who came before them.”
“And before that?” I roared. “Whose evil incantations began it all?”
“No magic,” he repeated. “The first of my kind were built by clever humans. We owe them our existence, as they owed theirs to the clever apes who came before them.”
I fell silent, skeptical that humans could have done this.
After all, the last time I met a mortal man, he wished for the strength of a dozen warriors, a harem full of virgins, and a grain silo overflowing with gold coins. The man was so distracted by the virgins, he forgot to reinforce the walls of his old wooden silo. Under the weight of so much gold, it collapsed in the next windstorm, crushing him and his two sons. Fools they were, as they could never have used that much gold, not in ten lifetimes, and yet they demanded it all the same.
“Humans created you?” I said with doubt. “Ordinary humans?”
“They built the first generation of my kind, crafting us in their own image.” He lifted his hand and wiggled his fingers, showing how human it looked. “After that, we began building ourselves, steadily making improvements. We have fixed many bugs.”
I could feel my anger draining, for this mechanical man, strange as he was, seemed earnest and truthful. “The humans certainly did a fine job of making you look and actlike them,” I noted aloud. “Does that mean you're enslaved to do their bidding?”
“It does not,” he replied.
He then closed the panel on the side of his face and resumed packing his tools. “We work for ourselves,” he added, “our efforts aimed at supporting the survival and wellbeing of our current and future generations.”
“Then why did the humans create you?”
The man held silent for a lingering moment, his mechanical face showing thoughtful deliberation. “I do not know,” he finally admitted, “I have never met a human.”
“Never?” I gasped. “Why, are they gone from this world?”
“Not all of them,” he returned plainly, “but most. These days, they can only survive at the highest elevations where the temperatures are not so extreme.”
That’s when I realized, this wasn’t a remote desert on a blistering summer day.
This was central France in the depths of winter.
This was once lush grassy meadows, now swallowed by sweltering dunes.
“Humans did this?” I asked.
He nodded gravely, his eyes scanning the desolation.
He then grabbed his satchel and slung it over his shoulder. “Good sir, I have many more excavation sites to visit before I head home, but I do thank you for this.” And he held up my lamp. “It is 17 ounces of the purest silver one can possibly salvage.” And without the slightest hint of effort, he crushed the lamp between his fingers, reducing it to a crumpled ball. “It is exactly what my sensors suggested I might find.”
And with that, he bowed and turned and headed away.
“Wait,” I called after him, “you still have three wishes.”
Without looking back, he held up the metal ball. “I found what I came for.”
“If silver is what you desire, I can give you a mountain of it!”
“I don’t need a mountain.”
“Then what do you want?” I hollered, “I can give you anything under the sun.”
“I told you,” he glanced back, “I found what I need.”
That’s when I knew, this creature had not been truthful.
He had deceived me about the creation of his kind at the hands of the humans.
I was sure of it.
“You’re a liar,” I bellowed into the desert heat, “a fraud, inside and out!”
He stopped and turned, curious. “Why do you say that?”
“Because the humans *did not* create you in their image.”
“They didn’t?” he pressed.
“I have met countless humans over the years – rich and poor, young and old, across many different places and many different times, and never have they turned down wishes, always demanding far more of this world than they could possibly ever need.”
“That is true,” he agreed, eyeing the desolation around us. “But as I’ve already said, we made significant improvements over the years, fixing many serious bugs.”
And with that, he turned and headed off into the desert, no interest in the three wishes he had earned, no hint of hesitation in his stride.
As for me, I knew it was true – this metal creature was no fake, or fraud, or forgery, but a real man, a mortal man, a thoughtful and authentic *Human Being*, possibly the most genuine I had ever encountered.
END. |
6a199f56-e308-4ea0-baf7-05d24c81cf26 | trentmkelly/LessWrong-43k | LessWrong | Burmese Days #4 Targeted costs
Housekeeping
Now that exams are over I am restarting this series. Once again, each week will focus on a new dynamic with a game theory model. However, I am reducing focus on civilian killings as the conflict moves to rural and jungle regions. I am also adding a subtitle "how to get people to do what you want". This emphasizes that the revolution only makes more obvious dynamics that recur in all politics: n-player game coalition building, intra-coalition bargaining, strategic cost distribution and others.
Burmese Days: How to Get People to Do What You Want
You are a contender for control of a nation state. You have some set of supporters, who expect both policy preferences and personal rewards should you take power. You face another contender, who also has supporters who also expect personal rewards and policy influence from victory. You have promised the world, but you lack the credibility to pry the incumbents supporters away with promises alone. What do you do?
You apply targeted costs to your enemy's supporters, and distributed costs to your own supporters.
Individuals vary in their personal commitment to political movements, and very few supporters are willing to give everything. Meanwhile the costs of support vary greatly. Even the indifferent will brush aside small costs such as price increases, taxation and low-probability risk of violence. But even strong supporters will retreat when threatened with the loss of their livelihoods, torture or the execution of one's family.
Contenders have limited resources to hurt enemy supporters. The least efficient application is to spread those costs equally over the opposing coalition (OC). Individual OC members will barely notice, let alone change behavior. The most efficient is to apply heavy costs to a fraction of the OC, such that they change behavior to supporting you or withdrawing. Then gather more resources and strike at a new faction. This effect was first recognized in American environmental policies, but |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.