Welcome to The Nonlinear Library, where we use Text-to-Speech software to convert the best writing from the Rationalist and EA communities into audio. This is: Are minimal circuits deceptive?, published by Evan Hubinger on the AI Alignment Forum. This post is part of research I did at OpenAI with mentoring and guidance from Paul Christiano. One possible inner alignment technique is to structure your learning algorithm’s inductive biases such that it is far more likely for it to produce an aligned model than a misaligned model. Two basic ways in which you might get traction now on how a technique like that will scale to AGI might be: Do experiments with current regularization techniques where you try to determine to what extent models end up being aligned with the loss function they were trained under depending on the types of inductive biases present during their training. I’m actually fairly excited about this approach, and I am working on a post right now which will hopefully provide some ways in which you might start going about doing that. One concern with this sort of approach, however, is that you might not be able to get any traction on phenomenon that only appear once your capabilities get really close to AGI. Analyze from a theoretical perspective what happens in the limit of perfect capabilities. The idea here is to try to patch the previous problem where when you do experiments you don’t get to see phenomenon that only happen when your capabilities get really advanced by imagining that your capabilities were perfect. In this case, however, you end up with the opposite problem of not seeing any phenomenon that only appear when your capabilities are weaker. The nice thing about combining these two approaches is that they tackle the problem from both sides, giving you an idea both of what the weak capability landscape looks like as well as the strong capability landscape. For this post, however, the focus will be on the second approach. Background From a theoretical perspective, the two most natural types of inductive biases are simplicity bias and time complexity bias. Paul Christiano has previously argued that simplicity bias can be deceptive in the limit, but left the question of whether the same is true of time complexity bias as an open problem. The goal of this post is to attempt to resolve that open problem. In particular, though one might have hoped that deceptive alignment could be avoided by heavily penalizing time complexity so that the model doesn’t have time to spend reasoning about deception, I will argue that this is not the case due to the possibility of meta-learning. First, however, we need a clear understanding of what it means for a model to be deceptive. “Risks from Learned Optimization” provides the model of deceptive alignment, wherein a model appears aligned during training for the instrumental reason of being selected by the training process. More generally, however, what we really care about is deceptive behavior—that is, the situation where our model generalizes pathologically such that it produces a malign failure. Second, we need a clear understanding of why simplicity bias leads to deception and an understanding of why that doesn’t generalize to time complexity bias. Specifically, the trap that simplicity bias falls into is as follows. First, there are many short but very powerful optimizers composed of very simple objectives with optimization processes capable of making use of vast amounts of compute. Furthermore, for a sufficiently complex task, it seems likely that the simplest program that solves it will be such an optimizer. Second, many of these optimizers are likely to be able to benefit in some way from being selected by your simplicity-biased training process—by defecting against you at some point, for example—and thus will have good training performance. Finally, since the vast majority of possible objectives—especially the very simple ones—are not aligned, the simplest progra...