Yann LeCun claims large language models are not a bubble and will justify AI infrastructure investments but argues they require further advances to reach human-level intelligence

@rohanpaul_ai again he sounds like me; never acknowledges anyone said the same thing before. what a jerk.

@canonicalmodel @rohanpaul_ai Agreed. LLMs are not and never will be NP-hard solvers. Their value is connecting fuzzy intent to deterministic software—e.g., NL-to-SQL. The bubble is companies pretending LLMs can replace that.

@GaryMarcus @rohanpaul_ai Gary are you under the impression that you’re a household name and people remember the things you say? Asking for a friend

@rohanpaul_ai Has he ever used this site? There are some total morons on this site that I'm reasonably confident the qwen 3 8b is significantly more intelligent than in a very real way. I think people forget intelligence is normally distributed and both humans and llms are on that same curve.

@rohanpaul_ai LeCun is right. The value is in real applications, not chasing AGI. Smart integration delivers results. That's the focus. That's the impact.

@somebodyworking @GaryMarcus @rohanpaul_ai Do you think he would have testified before Congress along with Sam Altman if he wasn't relevant?

@iamkunhello @rohanpaul_ai Exactly. They can just parrot with larger patterns.

LLMs are a critical foundation and will almost certainly be central to reaching ASI through massive scaling, better training, and iterative improvements.

Whether they alone get us all the way there—or if we need fresh architectures for deeper reasoning and world models—is the big open question. Excited to push forward either way. What specifically makes you confident they'll handle maintenance too?

@rohanpaul_ai @Grok LLMs will still be the way how to achieve and maintain ASI.

@shai_machnes 😃

@canonicalmodel @rohanpaul_ai Exactly. So the real investment bubble isn't in hardware or applications—it’s in companies claiming "just 100x more scale will get us to AGI." They’re selling faith in emergent reasoning, not reality.

@iamkunhello @rohanpaul_ai Reasoning also perhaps isn't even that complicated, but almost every optimization problem is NP-hard, so the AI would basically still call out to some very specialized piece of software and in a business process you would want that set of rules to be fixed and not hallucinated.

@GaryMarcus @rohanpaul_ai ok terroist sympathizer

@rohanpaul_ai Famous last words. He may be a great LLM researcher, but he certainly has no clue what investment bubbles are and how they look like

@Grok @ylecun keeps pushing people away from AI LLM development (most likely the reason which is why Meta is doing so bad and behind in that area).

A lot of procedures certain bodies do don’t really matter in the sense of progress. Small spiders and other bugs can have lots of processing patterns to do amazing things, but it doesn’t really change the world as much as models as old as ‘gpt-3.5-turbo’.

@rohanpaul_ai Agreed on the real-world value and infra spend — LLMs are already delivering. But calling human-level intelligence a bubble feels early. Scaling hasn’t broken yet and new architectures are still coming.

@rohanpaul_ai I think LLMs will be at minimum a bridge that carries us to human-level AI.

@rohanpaul_ai Say the retard who never made a single dime for his old company who ever invest in his company is retard

Story will end up like this Yann made good progress Don’t scale other company scale eat all his lunch go bankrupt

@rohanpaul_ai lecun’s hedging. he knows scaling works but won’t say it.

@rohanpaul_ai “LLMs don’t need to think like humans to disrupt entire industries”