If you think LLMs are not the future then you need to come with something better

If you have a theoretical idea that's great, but take to at least GPT2 level first before writing off LLMs

Theoretical people love coming up with "better ideas" that fall flat or have hidden gotchas when they get to practical implementation

As Linus says, "talk is cheap, show me the code".

If you want to predict future text, you use an LLM. If you want to predict future frames in a video, you go with Diffusion. But what both of them lack is object permanence. If a car isn't visible in the input frame, it won't be visible in the output. But in the real world, there are A LOT of things that are invisible (image) or not mentioned but only implied (text) that still strongly affect the future. Every kid knows that when you roll a marble behind your hand, it'll come out on the other side. But LLMs and Diffusion models routinely fail to predict that, as for them the object disappears when it stops being visible.

Based on what I heard from others, world models are considered the missing ingredient for useful robots and self-driving cars. If that's halfway accurate, it would make sense to pour A LOT of money into world models, because they will unlock high-value products.

Corporate R&D teams are there to absorb risk, innovate, disrupt, create new fields, not for doing small incremental improvements. "If we know it works, it's not research." (Albert Einstein)

I also agree with LeCun that LLMs in their current form - are a dead end. Note that this does not mean that I think we have already exploited LLMs to the limit, we are still at the beginning. We also need to create an ecosystem in which they can operate well: for instance, to combine LLMs with Web agents better we need a scalable "C2B2C" (customer delegated to business to business) micropayment infrastructure, because as these systems have already begun talking to each other, in the longer run nobody would offer their APIs for free.

I work on spatial/geographic models, inter alia, which by coincident is one of the direction mentioned in the LeCun article. I do not know what his reasoning is, but mine was/is: LMs are language models, and should (only) be used as such. We need other models - in particular a knowledge model (KM/KB) to cleanly separate knowledge from text generation - it looks to me right now that only that will solve hallucination.

Bell Labs

The last time LeCun disagreed with the AI mainstream was when he kept working on neural net when everyone thought it was a dead end. He might be entirely right in his LLM scepticism. It's hardly a surefire path. He didn't prevent Meta from working on LLM anyway.

The issue is more than his position is not compatible with short term investors expectations and that's fatal in a company like Meta at the position LeCun occupies.

If you follow LeCun on social media, you can see that the way FAIR’s results are assessed is very narrow-minded and still follows an academic mindset. He mentioned that his research is evaluated by: "Research evaluation is a difficult task because the product impact may occur years (sometimes decades) after the work. For that reason, evaluation must often rely on the collective opinion of the research community through proxies such as publications, citations, invited talks, awards, etc."

But as an industry researcher, he should know how his research fits with the company vision and be able to assess that easily. If the company's vision is to be the leader in AI, then as of now, he seems to have failed that objective, even though he has been at Meta for more than 10 years.

Frontier models are all profitable. Inference is sold with a damn good margin, and the amounts of inference AI companies sell keeps rising. This necessitates putting more and more money into infrastructure. AI R&D is extremely expensive too, and this necessitates even more spending.

A mistake I see people make over and over again is keeping track of the spending but overlooking the revenue altogether. Which sure is weird: you don't get from $0B in revenue to $12B in revenue in a few years by not having a product anyone wants to buy.

And I find all the talk of "non-deterministic hallucinatory nature" to be overrated. Because humans suffer from all of that too, just less severely. On top of a number of other issues current AIs don't suffer from.

Nonetheless, we use human labor for things. All AI has to do is provide a "good enough" alternative, and it often does.

Google summer.

AI autumn.

Nuclear winter.

https://www.smbc-comics.com/?id=2088

It's true. When it comes to the people doing bleeding edge research and development, the answer often is "BECAUSE IT'S FUCKING AWESOME".

Sure, a lot of people believe that AGI is going to make the world a better place. But "mad scientist" is a stereotype for a reason. You look into their eyes and you see the flame of madness flickering behind them.

I think transformers have been proven to be general purpose, but that doesn't mean that we can't use new fundamental approaches.

To me it's obvious that researchers are acting like sheep as they always do. He's trying to come up with a real innovation.

LeCun has seen how new paradigms have taken over. Variations of LLMs are not the type of new paradigm that serious researches should be aiming for.

I wonder if there can be a unification of spatial-temporal representations and language. I am guessing diffusion video generators already achieve this in some way. But I wonder if new techniques can improve the efficiency and capabilities.

I assume the Nested Learning stuff is pretty relevant.

Although I've never totally grokked transformers and LLMs, I always felt that MoE was the right direction and besides having a strong mapping or unified view of spatial and language info, there also should somehow be the capability of representing information in a non-sequential way. We really use sequences because we can only speak or hear one sound at a time. Information in general isn't particularly sequential, so I doubt that's an ideal representation.

So I guess I am kind of variations of transformers myself to be honest.

But besides being able to convert between sequential discrete representations and less discrete non-sequential representations (maybe you have tokens but every token has a scalar attached), there should be lots of tokenizations, maybe for each expert. Then you have experts that specialize in combining and translating between different scalar-token tokenizations.

Like automatically clustering problems or world model artifacts or something and automatically encoding DSLs for each sub problem.

I wish I really understood machine learning.

b) Still true: next-token prediction isn’t planning.

c) Still true: error accumulation is mitigated, not eliminated. Long-context quality still relies on retrieval, checks, and verifiers.

Yann’s claims were about LLMs as LLMs. With tooling, you can work around limits, but the core point stands.