正在加载视频...
视频加载失败
Yann LeCun argues that large language models (LLMs) cannot reach human-level or superintelligence just by scaling. He says the current LLM paradigm is hitting its limits. Many researchers are now exploring “agentic systems,” but building them on top of LLMs alone is flawed. LLMs can't plan actions well because... show more
35 条评论
Yann, every month, for 5 years now: LLMs are hitting their limits Embarrassing
@ohadasor been saying this for years. LLMs give "probably approximately correct" answers mathematically proven to collapse beyond certain complexity. Can't plan if you can't predict and reason and you can't reason with statistics. The solution:
LeCun's right that scaling alone won't get you there. But "fundamentally different" is code for "we don't know what yet." LLMs hit limits because they're prediction engines. They don't model the world. They model text about the world. That's why they fail at planning. Planning requires simulating outcomes. LLMs just pattern-match what planning sounds like. Agentic systems built on LLMs inherit that flaw. You can wrap scaffolding around them—add tools, memory, reflection loops—but the core's still autocomplete. So yeah, something different's needed. World models. Causal reasoning. Actual goal representation. The problem? Nobody's built that at scale yet. And it's not clear scaling is even the right approach for those architectures. LeCun's been saying this for years. He's consistent. Also not proposing concrete alternatives that work better. Criticizing LLMs is easy. Building the replacement is hard. Until someone ships a system that demonstrably reasons better through a different approach, this stays theoretical. LLMs are flawed. But they're also the best thing we have right now for language tasks. "Fundamentally different" might be necessary. Just isn't here yet. And betting against incremental LLM improvements while waiting for breakthroughs is how you miss what actually ships.
He said that in 2020. He’s still saying that.
The symbol grounding problem should be treated as very real. The LLM is basically firewalled from reality since the only relationships it knows is word to word. LLM logic is good because the words are correlated with the real world. But it will be unstable without Real feedback.
100% right. Agents aren't the path and neither is neurosymbolic AI. AGI/ASI need a fundamentally different architecture that is not based on language or vision.
LeCun's been saying this for years. He's right. Doesn't mean anyone knows what "fundamentally different" looks like. LLMs don't understand. They pattern match. Scaling that doesn't magically create understanding. Just more sophisticated pattern matching. Think of it like Google Maps. It finds routes. Doesn't understand why you're going somewhere or if the trip makes sense. Agentic systems on top of LLMs are asking a route-finder to be a travel planner. Wrong tool. Planning needs world models. Cause and effect. Simulation. LLMs have none of that. But here's the problem: LeCun's alternative approach—world models through self-supervised learning—isn't working at scale either yet. So we're stuck. One path hits a wall. The other's still theoretical. The "fundamentally different" thing might exist. Might be years away. Might require breakthroughs we haven't had. Until then, LLMs keep improving incrementally while everyone argues about whether that's real progress. LeCun's not wrong. Just not offering a working alternative either. Criticizing's easy. Building what comes next is the hard part.
Why compare Artificial Intelligence to Human's? They are both clearly different. And both clearly jagged in their own ways. AI is completely tangential to HI.
the other day i heard yan say humans don’t have general intelligence 🤦🏻♂️
I think Yann hit his limits.
LLMs are great interfaces Real AI will come from systems that can plan, act, and be checked, not just talk.
I think Yann is right and wrong. He is right that we need another major breakthough, but LLMs will be incredibly useful none the less
Yes. And it is not only world models, but to acknowledge that intelligence is a system involving social interaction. AGI and machine sentience, are often incorrectly framed as properties of individual systems. → Yet many researchers argue that intention and intelligence are fundamentally social processes. From a systems-theoretical perspective, Dirk Baecker argues that intelligence emerges from the interaction of multiple systems. Similarly, Nick Bostrom describes possible super intelligences less as a clearly defined entity and more as a functional, telematic “soup.” Maybe even the current moltbook phenomena offers valuable insights.
I can watch Scott's Tots and cringe less than watching Yan LeCun.
This could all be true. At the same time, Claude code changed software development forever
@ylecun This is an important point. Scaling improves pattern completion, but long-horizon intelligence breaks down on continuity: persistent identity, causal state, and memory across time. Agentic systems built on stateless models inherit this fragility.
The point he's missing is LLMs, don't just work with words or letters. They work with pixels, with pictures, with numbers, with any and every types of data and information. They build and recognise, relationships within data.
A different foundation is needed.
"we need something fundamentally different." YES - discernment and care.
The entire field is now intellectually degraded. Things that were commonly understood by everyone in 1995 are now revelations from prophets in 2026.
I’m not convinced…
@ylecun He is clueless and a siren for the wrong approach. Its all quite clear China is plowing ahead and I know but the world is not ready. At least one major calamity with AI woyld be a grace because only banks are talking now and they are all worse than SamA
his anti ai narrative costed meta a lot, hes been saying this for a while now
He's also currently raising money for his non-LLM AI startup, so...
Evaluating large language models requires comparing it with alternative approaches under realistic constraints.
He's right. One way would be to enable LLM to position itself (calculate) in relation to its user. The original idea of measuring what is happening is not enough. AI needs a speedometer and a navigation system in shared (semantic) space. It's all about relationships.
@sama
intuitivement I would agree (granted this only my 2c)as even a perfect synthesis of all existing human knowledge (which is not quite yet achieved with some stains in what the whole internet currently offers)better than any single human then may not provide value guiding the quest
Uh… Moltbot creating his own voice and getting a phone… Come on. I told everyone Yann was incorrect. But man is the media pushing this. Guess what else I said? Lies will 🔥 •
Exploring agentic systems beyond just LLMs is key to future progress.
Everybody knows what it's not, but nobody knows what it is.
so far in my exp, building agents on top of LLM do look flawed. you can never be sure of the outcome.
@ylecun Humans are so smart , only at the very end do they realize how ignorant their initial judgments were.
"LeCun sees the cage. The industry can't leave. Scaling won't work → but stopping means losing. Agentic systems won't fix it → but admitting that kills funding. Not stupidity. Structure. Classic PI.
He's like a Republican. Always trying to discredit the other party but not offering what how their solution is better.
