> Almost all statements Yann LeCun made back then (LLMs won't be able to do this or that) turned out to be wrong

You're being uncharitable. He would have been right if Chain of Thought hadn't been pulled out of a hat, and no one saw it coming, or how u reasonably effective it would be.

Today, no one can know if there will be a future step-change improvement to LLMs at the same level as CoT, and conversely, no one can be absolutely certein there won't be - so it shouldn't be a firable offense because neither position obvious; even though it will look like it in hindsight.

CoT wasn't pulled out of a hat. It was obvious to any researcher that it was needed. It just took some time and stronger base models and longer context to actually be implemented. People were already doing "CoT" by telling models to think step by step before answering, primitive harnesses and similar tricks.

Even a nobody like me was playing with a sort of CoT with llama finetunes before o1 came out.

Only this year I'd say we picked all the low hanging fruit that has been obvious for a while that was needed.

The most confusing thing to me about AI and LLM progress in the past 4 years is that I can't clearly see how all novel patterns add up, or what was the right point.

I think adding harness + MCP + chain of thought + specialized data generation resulted in the world where we have now - the one where you don't have to type code manually, but sometimes do still have to think and sometimes don't.

But LLMs without all this tooling and calls under the hood are still kind of... LLMs. If there's a space to hallucinate - they will hallucinate.

I would take it from other angle.

Meta AI is actually frontier lab now with LLMs. Muse Wonderfull is usable with multi agent orchestration (where other labs struggle).

I do not think that would happen with pure research and without single focus!