Nice! A couple things that this comment pointed out for me:

1. Real time is not always (and perhaps often not) the most useful way to talk about timelines. It can be more useful to talk about different paths, or economic growth, if that's more relevant to how tractable the research is.
2. An agenda doesn't necessarily have to argue that its assumptions are more likely, because we may have enough resources to get worthwhile expected returns on multiple approaches.

Something that's unclear here: are you excited about this approach because you think brain-like AGI will be easier to align? Or is it more about the relative probabilities / neglectedness / your fit?

I'm excited about this project. I've been thinking along similar lines about inducing a model to learn deception, in the context of inner alignment. It seems really valuable to have concrete (but benign) examples of a problem to poke at and test potential solutions on. So far there seem to be less concrete examples of deception, betrayal and the like to work with in ML compared to say, distributional shift, or negative side effects.

Previous high level projects have tried to define concepts like "trustworthiness" (or the closely related "truthful") and motivated the AI to follow them. Here we will try the opposite: define "betrayal", and motivate the AIs to avoid it.

Why do you think the betrayal approach is more tractable or useful? It's not clear from the post.
 

Thanks. I think I understand, but I'm still confused about the effect on the risk of catastrophe (i.e. not just being pseudo-aligned, but having a catastrophic real-world effect). It may help to clarify that I was mainly thinking of deceptive alignment, not other types of pseudo-alignment. And I'll admit now that I phrased the question stronger than I actually believe, to elicit more response :)

I agree that the probability of pseudo-alignment will be the same, and that an unrecoverable action could occur despite the threat of modification. I'm interested in whether online learning generally makes it less likely for a deceptively aligned model to defect. I think so because (I expect, in most cases) this adds a threat of modification that is faster-acting and easier for a mesa-optimizer to recognise than otherwise (e.g. human shutting it down).

If I'm not just misunderstanding and there is a crux here, maybe it relates to how promising worst-case guarantees are. Worst-case guarantees are great to have, and save us from worrying about precisely how likely a catastrophic action is. Maybe I am more pessimistic than you about obtaining worst-case guarantees. I think we should do more to model the risks probabilistically.

In the limit of training on a diverse set of tasks, we expect joint optimization of both the base and mesa- objectives to be unstable. Assuming that the mesa-optimizer converges towards behavior that is optimal from the perspective of the base optimizer, the mesa-optimizer must somehow learn the base objective.

Joint optimization may be unstable, but if the model is not trained to convergence, might it still be jointly optimizing at the end of training? This occurred to me after reading https://arxiv.org/abs/2001.08361 which finds that "Larger models are significantly more sample-efficient, such that optimally compute-efficient training involves training very large models on a relatively modest amount of data and stopping significantly before convergence." If convergence is becoming less common in practical systems, it's important to think about the implications of that for mesa-optimization.