Comment by pizza
3 months ago
In the case of supervised problem domains, you implicitly make a decision about what is signal, and what is noise, and sure, in that closed setting, supervised learning is much more sample efficient. But I think what we're learning now is that with strong enough base models, 'aha' moments in RL training show that it might be possible to essentially 'squeeze out signal from language itself', giving you far greater breadth of latent knowledge than supervised examples, and letting you train to generalize to far greater horizons than a fixed dataset might allow. In a fascinating way it is rather reminiscent of, well, abiogenesis. This might sound like speculative claptrap if you look at the things the current generation of models are still weak at, but... there's a real chance that there is a very heavy tail to the set of outcomes in the limit.
With a pretrained LLM most of the work is done. RL just steers the model into a 'thinking' mode. There is enough signal for that to work and for the inefficiency to not matter.
The downside is that you are limiting the model to think in the same language it outputs. An argument could be made that this is not how all humans think. I know that I rarely think in language or even images, just concepts (probably isn't even the right word) mix and transform and often I don't even bother to make the transformation to language at the end, just action.
I strongly agree; in fact I think what best matches the thought process is something like the multiset tree/forest workspace approach as suggested by Marcolli, Chomsky, and Berwick - a Hopf algebra that can be externalized into (non-planar) embeddings of linearized strings, or alternately into semantic manifolds.