Link to original article

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: Should we rely on the speed prior for safety?, published by Marc-Everin Carauleanu on December 14, 2021 on The AI Alignment Forum. This post was written under Evan Hubinger’s direct guidance and mentorship, as a part of the Stanford Existential Risks Institute ML Alignment Theory Scholars (MATS) program. TL;DR: There is some hope that by penalizing time complexity we could strongly select for non-deceptive models as they require time to figure out if they are out of training. Evan Hubinger argued in his post ‘Are minimal circuits deceptive?’ that we cannot fully rely on the speed prior to ensure non-deceptiveness by attempting to show that even minimal circuits, which we can intuitively understand as “fastest programs”, can be deceptive. This is because of the risk of spontaneous meta-learning which doesn’t inherit the safety guarantees that we would hope to get from the speed prior. This argument could be wrong if meta-learning is never the fastest way to solve a problem or if spontaneous meta-learning is constrained by the speed prior. Introduction A possible approach to tackling inner alignment failures is to shape the model’s inductive biases to make them unlikely. The two most natural inductive biases are the simplicity prior and the time complexity/speed prior. Increasing the number of parameters of the model is conducive to the simplicity prior as you are increasing the possible number of models from which SGD can optimize for the simplest model whereas decreasing the number of parameters is promoting the speed prior.[1] [2] In ‘Risks from Learned Optimization’[3] the concept of deceptive alignment is presented. That is, a model is said to be deceptively aligned if during training it pretends to be aligned in order to defect later on once it’s deployed. Paul Christiano previously argued that in the limit of perfect capabilities the simplicity prior can result in deceptive models. That is because, between the models that are capable of deception, the ones with less complex (unaligned) objectives are simpler than the aligned ones which have a more complex objective. He then left the open question of whether penalizing time complexity can lead to a deceptive model or not. Deceptive models have a particular property: they need to spend time figuring out if they are out of the training or not. The hope would be that by penalizing time complexity the model wouldn’t have time to figure out if it is out of training and therefore it would not be capable of deception. Minimal circuits can be deceptive Evan Hubinger attempted to resolve this open question in his post ‘Are minimal circuits deceptive?’ by thinking in terms of minimal circuits which we can intuitively understand as “fastest programs”. The conclusion of his argument is that even the fastest programs that solve a set of tasks can be deceptive, potentially crushing the hope that strongly penalizing time complexity can ensure non-deceptive models. To understand why one has to first understand mesa-optimization. It is a neologism introduced in ‘Risks from Learned Optimization’ which means that a learned model is also an optimiser, in other words, if your optimizer (base-optimizer) learns another optimizer, then the learned optimizer is the mesa-optimizer. Now consider the possibility that your learning algorithm learns a specific type of optimizer: another learning algorithm (meta-learning). There are some cases in which the fastest way to solve a problem (set of tasks) could be to do meta-learning, in fact, that’s what we do. It seems likely that for a problem like classifying objects needed for an average human world model, meta-learning is the fastest way to do that. There is one issue with this though: the fastest meta-learning algorithms don’t necessarily learn the fastest models, which would eliminate the ...