LeCun’s AMI: What is the Proposition?

We are currently witnessing an explosion in AI centered around Large Language Models (LLMs).

The holy grail is AGI (artificial general intelligence). What that exactly means is not entirely clear, but it is something closer to what a person can do — reasoning, planning, understanding the world, adapting to new situations without being explicitly programmed for each one.

The proposition from Yann LeCun is that current AI with LLMs is fundamentally a dead end in the search for AGI.

LeCun argues that LLMs are sophisticated statistical parrots — they predict the next token without any genuine understanding of the world. No model of physics. No cause and effect. No real reasoning. Just pattern matching at enormous scale. His contention is that you can scale this forever and never get to AGI because the foundation is wrong.

Yann LeCun unveiled his new startup on March 10, 2026: Advanced Machine Intelligence Labs (AMI), in which he will replace current LLMs with “world models”. $1.03 billion seed round at a $3.5 billion pre-money valuation — one of the largest seeds ever. Paris headquarters. Twelve people. CEO Alex LeBrun. LeCun as executive chairman.

The current status of LLMs: Pretty amazing and getting more amazing by leaps and bounds at an accelerating rate.

The current status of “world models”: No product. No demo. No proof of concept. Just promises of “world models”: AI that understands the physical world, reasons, plans, anticipates outcomes.

On the surface this is a rather bold claim. LLMs are getting better all the time and aren’t currently showing any visible signs that the technique is fundamentally limited. In effect, LeCun is arguing that mostly all of the current AI industry is pursuing the wrong paradigm — one that can never lead to AGI.

This op-ed is in two parts. Part 1 examines why people might think he can do this and does it seem realistic. Part 2 is a deeper look into who is investing in this and what that tells us.

A note on the “LLMs can’t do math or physics” critique: it is incorrect to say that deployed LLMs do not handle math and physics. The systems as packaged for actual use consult math engines, code interpreters, and symbolic reasoning tools. The criticism describes the raw model in isolation, not the deployed system. Nobody ships the raw model alone. MCP servers (Model Context Protocol) now provide a standardized way to connect LLMs to virtually any external service or specialized engine — the scope keeps expanding with no sign of a ceiling. My general sense of many people who criticize LLMs is that they have never used these tools in a serious way, or they are talking about how they were four years ago.

How Science Actually Works

The public model of science is: lone genius has insight nobody else has, works against the skeptics, proves them wrong, changes everything. That model is a fairy tale. Science is slow, collective, and incremental. Breakthroughs happen when the tools are finally ready — and when they are, many people working independently arrive at essentially the same place at essentially the same time.

You can find a new insect in the Amazon. That’s a solo discovery. But something fundamental — the kind of advance that reorganizes how a field thinks — is almost never that.

Even the most celebrated names in science look different when you actually read the history. Newton and Leibniz developed calculus independently and simultaneously. Hooke had gravitational inverse-square ideas before Newton formalized them. Newton himself acknowledged standing on the shoulders of giants.

The famous mathematician Gauss’s most celebrated achievement is the Fundamental Theorem of Algebra (this is a complex proof — the word algebra here is not the algebra you had in junior high school). Before Gauss, d’Alembert, Euler, and Lagrange had all attempted proofs — all incomplete. Gauss first published his own proof in 1799, but it was also incomplete with gaps. The first rigorous proof was published in 1806 by Argand — an amateur mathematician. Gauss kept producing versions of the proof for decades afterward.

The lone genius is a story we tell afterward, once the crowd has been edited out of the frame.

The Modern Genius

Einstein is the archetype of the lone genius in the public imagination — the patent clerk who rewrote physics. The reality, as usual, is more complicated, and I am not going to wade into the priority disputes around special relativity here. There are legitimate questions — Henri Poincaré and Hendrik Lorentz had developed key mathematical elements before Einstein’s 1905 paper, the Lorentz transformations that are central to special relativity are named after Lorentz because he derived them first, and Einstein’s paper notably did not cite either of them. Serious historians of science have written about this. It is also worth noting that Poincaré was a mathematician, not a physicist — and as with Argand solving what the famous mathematicians could not, the field has a long history of being reluctant to credit people who come from the wrong discipline. The consensus view is that Einstein’s formulation was more complete and more physically unified, so the credit is broadly defensible — but it is not clean. That debate can go on without us.

What is not disputed is what happened after. Einstein spent the last thirty years of his life pursuing a unified field theory that went nowhere. He rejected quantum mechanics — “God does not play dice” — against overwhelming experimental evidence and the judgment of virtually every physicist working in the field. He sailed boats. He made funny faces for photographers. He became the world’s most famous scientist and produced no significant new physics after the mid-1920s. The legend kept growing while the work stopped.

Consider one more irony. The Nobel Prize — the institution most synonymous with recognizing lone genius — was itself named after a system integrator. Alfred Nobel didn’t discover nitroglycerin; that was Ascanio Sobrero, his former teacher. He didn’t invent the safety fuse; that was William Bickford. He combined existing components into a commercial product. Useful. Valuable. Not a lone genius moment. But the prize bearing his name spent the next century teaching the world that science works the other way. That myth is now so embedded in how we reward science that we can’t see the crowd behind the name.

LeCun, Hinton, and Bengio won the Turing Award. What does that mean? A committee of the ACM decided to give it to them. That does not mean they uniquely did this work or were even the first. Why they gave it to them and not others is not entirely clear. When investors hear “Turing Award winner” they process it as “certified genius who sees what others can’t.” That is not what it means.

To be fair, the trio did make contributions to the field and literature, they helped popularize machine learning and were true believers during the AI winter — the long decades when the field was defunded, dismissed, and ignored. That persistence was real and it mattered. But what actually unlocked modern AI wasn’t a scientific breakthrough by any of them. It was the GPU, invented by Nvidia to meet the computing demands of video games. When that hardware became available, the algorithms that had been waiting for it finally had somewhere to run. The tide came in. Everyone standing on the beach got credit for the ocean.

One Name on a Crowd’s Work

Let’s examine what LeCun actually did, because the press never bothers with the details. The work associated with his name is computer vision work using a convolutional neural network published with a team of three other researchers at Bell Labs.

The convolutional architecture at the heart of this work wasn’t developed by the team LeCun worked with. Fukushima’s Neocognitron in 1980 had the core convolutional architecture — approximately eight years before the Bell Labs team applied backpropagation to train it.

The backpropagation algorithm itself was Rumelhart, Hinton, and Williams’ (1986) — except it wasn’t really theirs either. The underlying mathematics had been used in control theory for aircraft in the 1960s by Bryson and Ho. Werbos had applied it to neural networks in his 1974 PhD dissertation. Linnainmaa had described the algorithm in his 1970 master’s thesis. Rumelhart, Hinton, and Williams popularized it in their 1986 Nature paper, which is what the Bell Labs team actually cited and built on. But even that algorithm had decades of prior history behind it. The team borrowed from people who had themselves borrowed.

The foundational 1998 paper “Gradient-based learning applied to document recognition” has four authors: LeCun, Bottou, Bengio, and Haffner. Bottou has spent decades making stochastic gradient descent practical at scale. What the paper demonstrates is backpropagation applied to a convolutional architecture to make it trainable on real-world data. That is the contribution. Whose specific idea it was from the Bell Labs team to train a convolutional network using backpropagation, or if it was collaborative as the multiple authorship presumably states, and whether others were already working on the same combination, the paper does not tell us. But synthesis on top of other people’s components, executed by the Bell Labs team, at an institution that provided everything required — that is not the lone genius narrative.

It is also worth asking: how many people in the entire world were working on this problem at that time? Dozens. Maybe a few hundred at the outer edge. With the tools that existed, someone was going to get there.

And that count only includes the work anyone knew about. This was before the internet, before electronic journals, before preprint servers. Work done in Soviet labs, Japanese universities, and Eastern European institutions frequently never made it into Western publications. The Russians developed photocopying technology before the West — the KGB suppressed it because it would enable citizens to freely reproduce and distribute information, undermining state control over what people could read and share, and the West later “invented” it independently. We have no idea how much parallel work was happening in 1988 in places Western researchers never read. Now there are millions of researchers working on AI globally, everything is published instantly, and the competition is total.

After that: AlexNet — Krizhevsky, Sutskever, Hinton. VGG, ResNet — not LeCun. YOLO — Redmon, not LeCun. Vision Transformers — Google Brain, not LeCun. Diffusion models — not LeCun. The entire modern computer vision stack that runs self-driving cars, medical imaging, and manufacturing inspection was built by other people on architectures other people invented.

The field he claims as his life’s work, originally as part of a four-man team at Bell Labs, has been shaped for decades primarily by other people. He helped build one of the early versions of computer vision and watched other people build far better ones ever since. Unlike the fewer than one hundred people worldwide working with primitive computing when he did his foundational work, there are now millions of PhDs with incomprehensible amounts of compute dedicated to solving these problems. As the field of computer vision grew and millions came to work on it, his name disappeared from the cutting edge results. Others took that position.

The transformer architecture — the technical foundation of every major AI system today, including the LLMs LeCun dismisses — was published in 2017 in “Attention Is All You Need” by eight authors from Google Brain and Google Research. Nobody calls it the “Vaswani paper.” It is just the transformer. Eight people, one institution, collective work. That is how modern AI actually gets built.

He Predicted Them Dead. They Keep Getting Better.

Since at least 2022, LeCun has posted relentlessly dismissing LLMs. Hallucinations prove they can’t reason. Scaling won’t get to intelligence. The architecture is fundamentally wrong. His critiques predate ChatGPT — but after November 2022, when a startup nobody had heard of dropped ChatGPT and overnight became the face of AI, the dismissals got louder. Sam Altman, not even an amateur scientist but a businessman and promoter, on every magazine cover. Not LeCun. Not FAIR. Not Meta. For someone who had spent decades as a leading voice in AI, watching a startup he’d publicly dismissed become the most consequential technology company of the decade had to sting. Meanwhile OpenAI and Anthropic keep releasing more stunningly clever models with greatly diminished issues such as hallucination. Scientists with famous names are as capable of envy as anyone else. Einstein did not even cite the obvious work that special relativity built on. Sour grapes doesn’t mean the critique is entirely wrong. But when someone with a competing thesis and a bruised ego tells you the competition is fundamentally broken, you apply a large discount.

I use these systems daily to write these pieces. My business uses them as the primary coding agent. I don’t see any indication of the models reaching some kind of fundamental limitation. The limitations become known and then they get solved, and at this point, much faster than I could have envisioned.

LeCun’s core claim is that LLMs cannot develop genuine understanding of the physical world. Where is his proof? Because they hallucinated a lot four years ago? Or is it just NIH (Not Invented Here)?

LeCun spent thirteen years at Meta running FAIR — unlimited compute, top researchers, no product pressure. The output: papers, open source code, and a visual encoder with no language capability. Meta’s actual consumer AI runs on Llama — a large language model, the thing LeCun publicly calls a dead end. His own employer voted with their product roadmap against his thesis. If this research direction was going to produce something deployable, why didn’t it in thirteen years with those resources?

The Institutions That Built the Legend

The ACM and the press built a wall of credibility around these names that nobody with a checkbook knows how to look behind. A Turing Award is not a forecast. It is a receipt for work done decades ago, which is often not even for scientific discovery but for championing something. Investors read it as a prediction. It isn’t.

And frankly, I don’t see any reality in which AMI is more than LeCun moving his pet research project from Meta to a new address. Google, DeepMind, and others already have teams working on world models and visual representations. Nobody — not one of them — is showing a realistic prototype or proof of concept. The field isn’t holding back because it lacks a French headquarters and a Turing laureate. “World models” for AGI is a speculative research direction at this point, and many such ideas never amount to anything.

In Part 2 we will look more deeply into who is actually writing the checks and why. But it’s worth noting one thing before we get there: Mark Zuckerberg is not on the investors list. Whether Meta was offered the chance to invest and passed, or was never approached, we don’t know. What we do know is that the man who funded LeCun for thirteen years, who knows his work better than any outside investor, is not in this round. There is also good reason to think that Meta, having committed fully to Llama and the LLM path, was ready for a new AI leader — one who wasn’t publicly dismissing their core product strategy.

.