I found this insightful, particularly bringing back LLMs into the Yudkowski / Sequences fold, whereas many have claimed the rise of LLMs has shown decades of Yudkowskian AI speculation to be way off base. I don't have enough technical knowledge to evaluate the accuracy of this post, but I am hopeful that large parts of it are true.
The brute-force training process naturally sculpts Transformers into inference engines. They don’t just approximate the math; they build a physical geometry — orthogonal hypothesis frames and entropy-ordered manifolds — that implements Bayesian updating as a mechanical process.
They aren’t Bayesian by design; they are Bayesian by geometry.
To the extent the article has merit, it does seem to explain why CoT and Reasoning models are able to "outperform". The 20 questions model, where we are not merely bisecting the information space, but looking to maximize rejection or filtering, offers a lot of insight into the nature of the problem. When a fixed number of layers gets exhausted, is this where normal models hallucinate? With CoT or reasoning, we can feed the smaller space back into the first layer, and continue filtering down.
Jump in the discussion.
No email address required.
Notes -
Here's an example of the correspondence between Bayesian/frequentist interpretations for linear regression: https://stats.stackexchange.com/questions/283238/is-there-a-bayesian-interpretation-of-linear-regression-with-simultaneous-l1-and.
More options
Context Copy link