Discussion about this post

User's avatar
Rafael Kaufmann's avatar

What Brad/Cosma's explanation of LLM misses (along with most others) -- the reason why Brad's intuition "strongly leads [him] to think that they should not be able to do half as well as they do" -- is the concept of **faithful representation**. Put simply: Reality allows itself to be mapped in a low-dimensional latent space, and some compression schemes are just **good mappings**. [Or at least (the instrumentalist view), of humanity's currently-known shared ways to describe and predict our observations of reality.]

This is why our scientific and statistical models work in the first place. Brute-force ML works because (or to the extent that) it finds these mappings. LLMs work because their training process finds the same latent representations as (or equivalent to) the representations that internet sh*tposters have in their heads as they're writing [1]. Therefore, they can quite accurately predict what a sh*tposter would have said, in pretty much any context.

Yes, this is just Plato's concept of "carving nature at its joints" from ~2400 years ago. Quite surprisingly, it turns out that we're all living at the unique moment in humanity's history where we are discovering that such a carving is an objective possibility, and that we can, for the first time, automate this carving [2, 3]. And this does not just apply to language, BTW: if the success of multimodal models weren't enough, tabular foundation models [4] demonstrate that good old statistics has the same property.

There is still the crucial question of whether/when, beyond "just" the latent representations, ML can learn the true causal world model: which latent states at t cause which latent states at t' > t. This is, in a technical sense, much harder than learning temporal correlation, and at least in some cases, it's **impossible** to learn just from the data, requiring the capability to intervene/experiment [5]. However, it has recently been proven that learning a good causal model is a strong requirement for robust behavior -- ie, reliable extrapolation beyond the data [6].

We've been lucky so far that the Internet already has a lot of natural-language descriptions of causal models of just about everything. These can be compressed into "meta-representations" that let LLMs "interpolate extrapolations". This is similar at some level to how humans learn much of their own extrapolation capability -- not by experimenting themselves, but by learning theory from other people and representing it in their heads as "little stories" that they can interpolate.

However, because these meta-representations can't be directly grounded in observational data, the way LLMs learn them is very sensitive to the quality of the training corpus. That is why stuff like embodied learning and synthetic data are hot topics in AI: This is stuff you want to get *exactly right*.

Regardless, "just" getting the representation right is a huge step in the direction, and it goes a long way to explain/justify the conceptual leap from "just statistics on steroids" to "AGI" made by LLM stans.

[1] https://arxiv.org/abs/2405.07987

[2] https://arxiv.org/abs/2505.12540v2

[3] https://aiprospects.substack.com/p/llms-and-beyond-all-roads-lead-to

[4] https://www.nature.com/articles/s41586-024-08328-6

[5] https://smithamilli.com/blog/causal-ladder/

[6] https://arxiv.org/abs/2402.10877

Expand full comment
Karl Seeley's avatar

A couple of weeks ago I was opening up a stub of a paper I'm working on, a PDF of not fully 8 pages. Acrobat informed me that, "This looks like a long article. Would you like me to summarize it?"

I was appalled that it treated an 8-page piece of writing as "long," but curious what it would come up with for a summary.

Mostly good, but also got one key idea wrong.

If I didn't know the writing (having, after all, written it), I wouldn't have known how much of the summary was on-target and how much was eroneous. Simply knowing that some of the info is incorrect doesn't help - to figure out which is the incorrect stuff, you have to go in and read the article. And then the point of the summary was...?

I suppose one could say that the standard shouldn't be perfection, but rather, is the AI doing better than a graduate research assistant. My college is undergrads only, so I don't have any experience with evaluating the work of a graudate research assistant.

Expand full comment
3 more comments...

No posts