An interviewer suggested Yann LeCun had compared today’s AI to a smart cat, or maybe a three-year-old. In a conversation conducted Friday, April 4, 2025, and published by Jason Howell, LeCun’s answer was two words. “Not even.”
Lines like that travel. They get turned into a story in which one famous researcher proved the entire AI industry wrong. That story is tidy. It is also false. LeCun did not disprove OpenAI, Anthropic, Google, or Meta. He drew a line the industry keeps smudging: systems that already earn their keep, and systems that still cannot be trusted to understand, plan, and act.
The clock he put on the bolder claim comes from a different conversation. On Alex Kantrowitz’s Big Technology podcast, in remarks the Wall Street Journal reported and Gary Marcus later quoted, LeCun said:
We are not going to get to human-level AI just by scaling LLMs… whatever you can hear from some of my more adventurous colleagues, it’s not going to happen within the next two years. There’s absolutely no way in hell.
That line is not from the April 4 interview. Read the two together and the picture gets less revolutionary. It also gets more useful. In the Howell conversation, LeCun said large language models are worth having. They are already strong coding assistants, and they may become more general assistants. He also said they are not reliable enough for routine deployment, and that they are not a path to human-level intelligence. Scaling skepticism and a purchase order can live in the same sentence. A lot of coverage refuses to let them.
Three claims wearing one headline
The “industry proved wrong” framing mashes together three different bets.
- Language models already do valuable work.
- Making them larger may never produce human-level intelligence.
- Learned world models are a proposed way through, not a demonstrated one.
The first is ordinary business reality. The second is a serious research argument, and LeCun is not its only author. The third is a research program with early demonstrations and open objections. Collapse those three and you get a headline. Keep them apart and you get a strategy.
What he thinks is still missing
For human-level performance, LeCun listed four gaps in the April 4 interview. They double as a checklist for anyone evaluating AI agents.
- Understanding the physical world. Fluency about things is not a model of how things behave when you push them.
- Reasoning and planning. A plausible next sentence is not a plan that survives contact with constraints.
- Persistent memory. A system that forgets the case, the customer, or the last failed attempt will repeat the failure with confidence.
- Controllability. He wants systems driven by an objective a person sets, not by whatever continuation looks likely.
Agentic systems, he said in that same conversation, are being discussed everywhere and are “still not totally reliable.” That matches what operators see when a prototype books the wrong meeting, invents a policy, or loops until someone pulls the plug.
For AI for business, the buying question is narrower than the philosophy. Can this system draft, summarize, retrieve, and suggest, with a person still on the hook? Or are you about to hand it a process that fails in expensive, quiet ways?
A sales team can already use a model to sketch follow-ups, clean call notes, and prep an account brief. The rep still owns the send, the price, and the commitment. Two controls make that real. A named person reviews the output before it leaves the building. And a kill rule is written down: a wrong fact, a missing source, or a step the model cannot explain stops the workflow. It does not get patched in chat and shipped anyway.
LeCun’s warning about demos is the part worth stealing. It is his argument, not a settled measurement. He pointed at autonomous driving. Country-road demonstrations from roughly a decade earlier, he said, still have not produced cars that drive as reliably as people, except in constrained services such as Waymo. “As reliably as people” is his standard, and it is contestable. The useful distinction is simpler. A polished run is a demo. A system employees or customers depend on every day is a deployment. AI automation fails in the gap between those two words.
He also described a habit about seventy years old. A new method arrives. Someone promises human-level intelligence within about ten years. The method turns out to be very good at a narrower class of problems, and the promise gets reassigned to the next method. Treat that as a pattern he is asserting. It is not a law of nature. It is still a useful brake on roadmap slides that confuse a capability with a destiny.
He added a budget observation, and it should stay in his voice. Most of the spending, he said, is on running systems for users, not on training them. He did not publish a figure, or define what “most” counts. Training is the press release. Inference is the meter. If millions of employees and customers hit a model all day, the cost that compounds is the cost of each answer, not the cost of the original training run. Ask which one the invoice is actually for.
None of this argues for freezing spend. It argues for knowing which spend buys a tool and which spend buys a hypothesis.
World models, minus the victory lap
LeCun’s alternative, as he described it to Howell, is a learned world model: a system that predicts the consequences of actions, supports planning, and can be used in a classical control method he names, model-predictive control. In plain terms, the system tries possible moves inside its own model of the world, picks the one that looks best against a goal you set, then does it again as the world answers back. He said this vision came from a long paper he wrote about three years before that interview, before language models were a household product, and that their success had not changed his mind.
That is a coherent bet. It is not a finished machine.
A 2025 paper on arXiv, “Critiques of World Modeling, “ takes on what it calls a vocal school of thought whose program lines up with the JEPA and model-predictive-control agenda associated with Meta’s FAIR research group. Treat it as a critique of that school, not as a personal indictment of LeCun, and do not hang every sentence in it on his name. The paper says current world-model systems mostly fail at purposeful reasoning and planning, limited by scope, abstraction, controllability, interaction, and generalization. It also argues that the school’s assumptions have critical limitations.
Two contrasts from that survey are enough to keep the victory lap honest.
A video you cannot ask “what if I turn left?” is not a planner. The paper puts OpenAI’s Sora and Google DeepMind’s Veo in that category: striking video along fixed trajectories, without explicit state, action, and counterfactual control.
A robotic-arm result is not an open-ended worker. The paper calls Meta FAIR’s V-JEPA 2 a step forward on robotic-arm manipulation, and says generalization to long, varied jobs (its examples include making breakfast and mountaineering) is still unclear.
So the alternative to “just scale the chatbot” is real research. You cannot drop it into a warehouse next Tuesday. Early arm results are not a general theory of the physical world.
Not a lone dissenter, and not outside the industry
Gary Marcus, in “The False Glorification of Yann LeCun, ” argues that recent coverage overstates both LeCun’s isolation and his originality. He says the doubt did not start with LeCun, pointing to his own public challenges to GPT-2 and GPT-3 in fall 2019, to articles he wrote in 2019 and 2020, and to the 2021 “stochastic parrots” paper by Emily Bender and coauthors. He also reports that a survey associated with AAAI’s presidential panel found researchers doubting language models as a route to AGI to be in the majority, by a wide margin. He does not lay out the questions, the sample, or the percentages. These are his claims about the record, some of them contested. They are not a rebuttal of the technical point. A shared critique can still be right. A headline can still pretend one person was always the lone adult.
The calendar makes the same point in product form. On April 5, 2025, the day after the Howell conversation, Meta unveiled Llama 4 Scout and Llama 4 Maverick, which it described as a multimodal advance built with a mixture-of-experts design, routing work to specialized sub-models rather than firing one giant network for every token. Those releases were not part of the interview. Skepticism about language models as a road to human-level intelligence sat inside the same company that was still shipping larger open-weight models. If LeCun had proved the industry wrong, Meta missed the memo overnight.
Questions worth taking into the budget meeting
- Did LeCun prove the AI industry wrong?
No. He argued that scaling language models will not produce human-level intelligence. In the April 4, 2025 interview he also said those models are useful and deserve continued investment.
- Should a company stop paying for language models?
Not on this evidence. He singled out coding assistance as a real use, and the same class of language work already shows up in drafting and account prep. Keep a named reviewer on the output, and a rule that kills the workflow when the model cannot show its work.
- Are AI agents ready for unsupervised daily use?
Not by his account. In the Howell interview he called agentic systems “still not totally reliable, ” and he treated impressive demos as a familiar way to mistake a constrained success for a deployable one.
- What would a world model add that a chatbot does not?
A prediction of what happens if the system acts, plus a way to plan against a goal you set. That is the promise. The 2025 survey says current video and JEPA-style systems have not shown they can carry that promise into open-ended, long-horizon work.
- Is researcher skepticism a fringe view?
No. Marcus reports that a survey tied to AAAI’s presidential panel found doubt about language models as a path to AGI to be a majority view. He does not give the questions, the sample, or the percentages. Skepticism is not fringe. Do not put his summary of a panel on a board slide.
- Which of the three bets is this budget item?
Ask that before you sign. A tool you can supervise this quarter is the first bet. A thesis that a bigger language model becomes a mind is the second. A world-model hypothesis is the third. They do not share a success metric, and they should not share a slide.
- So where should the money go?
Toward tools you can supervise now, and toward research on planning, memory, and the physical world that you do not pretend is finished. On LeCun’s account, the expensive line item is running models for users, not the training run that got the applause. That is his observation, not an audited split of the industry’s books.
Keep funding language models because they already perform language and coding work people will pay for. Keep funding robotics, video, and world-model research because those models still do not give you reliable physical prediction or planning. Refuse both sermons: the one that says a bigger chatbot is a mind, and the one that says a critic with a famous name has already closed the case.
Buy reliability. Rent the rest of the future until someone can show it surviving a Tuesday.