The man who helped invent the technology behind ChatGPT just said, out loud, that it is a dead end. He did not write a cranky blog post or argue about it on social media. He quit one of the most powerful AI labs on the planet, and raised more than a billion dollars to start over and build something different.1
His name is Yann LeCun. In 2018 he won the Turing Award — the closest thing computer science has to a Nobel Prize — for the foundational work on neural networks that everything we now call "AI" is built on.2 For twelve years he was Meta's chief AI scientist and the founder of its research lab. And he now argues that large language models, the engines inside ChatGPT, Gemini, and the rest, will never get us to real intelligence no matter how much money the industry pours into them. That is a strange thing for an insurance agent in Sarasota to care about. Stay with me, because the reason he is right matters for how risk gets measured — and risk is the business we are in.
What He Actually Said
Large language models do one thing extremely well: they predict the next word. You give the model some text, it calculates the most likely next token, spits it out, and repeats the trick tens of thousands of times until you have a paragraph. It is genuinely impressive. But ask it what happens when you push a glass off a table and it will tell you the glass falls and breaks — not because it understands gravity, momentum, or the physics of a tile floor, but because the words "glass," "fall," and "break" appear together constantly in the text it was trained on.1
LeCun's blunt version of this: a house cat understands the physical world better than the most expensive AI model ever built. Your cat has a working model of gravity. It plans a jump, anticipates where a moving object will land, and knows that something balanced wrong is about to fall. The model has read millions of sentences about all of that and understands none of it.1 He frames the whole problem in one line: training a model on text is training it on a shadow of reality. It learns the shadow, never the thing casting it.
He lists five specific failures that follow from that — no real grounding in reality, no genuine reasoning (what looks like reasoning is expensive guessing among many sampled paths), hallucinations that cannot be fully patched out, an inability to plan, and scaling that is hitting a wall where each new dollar of compute buys a smaller gain.1 You have probably felt all five without knowing their names.
The Billion-Dollar Bet
LeCun announced his departure from Meta in November 2025, ending a twelve-year run.3 His new company, AMI Labs — short for Advanced Machine Intelligence — then raised roughly $1.03 billion at a $3.5 billion valuation, the largest seed round a European company has ever raised.4 The backers are not lightweights: the round drew NVIDIA, Temasek, Samsung, Toyota Ventures, and individual investors including Jeff Bezos, Mark Cuban, and Eric Schmidt.4 When that roster writes checks against the prevailing wisdom, it is at least worth understanding what they see.
What they are funding is a different idea about how a machine should learn. LeCun calls his architecture JEPA — a joint embedding predictive architecture. An LLM predicts the next word. JEPA predicts the next structure: it ignores the details that do not matter, like the texture of a wall or the exact color of a cup, and tries to model the underlying dynamics — what is moving, what is about to happen, and why.1 He describes it as learning the way a baby learns. A baby does not read about gravity. It drops things, watches them fall, and builds an internal model of cause and effect from raw observation.
Insurance has always been a world model. We price the cost of things that have not happened yet.
The early evidence is real. The team's video model, V-JEPA 2, was trained on over a million hours of internet video — people picking things up, objects falling, things being stacked and dropped — until it built an internal sense of how the physical world behaves. They then gave it just 62 hours of robot footage and deployed it in two brand-new labs, with objects it had never seen, and it generalized on its own.1 A traditional system would have needed thousands of hours of labeled data for each new environment. That is not a small efficiency gain; it is a different philosophy of what learning means.
Why an Insurance Agent Is Writing About This
Here is the connection most coverage of LeCun misses. Insurance has always been a world model. When we underwrite a strip center, a commercial fleet, or a contractor, we are running a simulation of cause and effect: what could happen, how often, how badly, and what it will cost. We do it with actuarial tables, catastrophe models, and hard-won judgment rather than software — but the job is the same one LeCun is trying to teach a machine to do. We price the physical world's risks before they happen.
Today's LLMs are poor at exactly the part of that job that matters most. They are fluent, but they do not understand a roof, a flood plain, or the path of a hurricane. A model that genuinely understands physics and cause and effect would be a far better fit for the work of risk than a model that has merely read about it. Three places this could land:
- Catastrophe and property modeling. Florida lives and dies by hurricane and flood modeling. A system that can simulate wind, water, and structural failure from observation — rather than from text about past storms — could sharpen how commercial property risk is priced. That cuts both ways for owners: better models can reward genuinely resilient buildings and penalize the rest.
- Claims and inspection. Models that reason about the physical world can read drone footage, satellite imagery, and damage photos and actually understand what they are seeing — distinguishing storm damage from wear, or estimating a repair from structure rather than from a caption.
- A new wave of liability. World models are aimed squarely at robotics and autonomous systems. As physical AI moves into warehouses, vehicles, and job sites, the general-liability and workers’ compensation questions move with it — and that is coverage every commercial owner will eventually have to think about.
The Honest Caveat
None of this is settled. LeCun has a billion dollars behind the claim that LLMs are a dead end. The rest of the industry has something closer to half a trillion dollars riding on the opposite bet.1 5 One side is going to look very wrong in five years, and it is genuinely too early to say which. Even if world models work exactly as advertised, the first real products are aimed at robots and industrial systems, not insurance software — the benefits would reach our world indirectly and slowly.
The point for a business owner is not to bet on a winner. It is to notice that the people who built this technology are openly debating its limits, while a lot of vendors are selling AI tools as if those limits do not exist. When someone pitches you an "AI-powered" quote, policy review, or risk score, the right question is simple: does this thing understand my business, or has it just read a lot of sentences about businesses like mine? More often than the marketing admits, it is the second one.
Where Hendrickson Insurance Fits In
Hendrickson Insurance is an independent Gulf Coast Florida agency — based in Sarasota and serving owners across Bradenton, Tampa, St. Petersburg, Clearwater, Venice, and Lakewood Ranch. We follow where AI is actually going, not just where it is being marketed, because it increasingly shapes how your coverage is priced, underwritten, and excluded. From commercial property and cyber liability to AI liability, professional liability (tech E&O), and directors & officers coverage, we build programs that account for how modern businesses actually operate. Our job is to make sure a model — ours or a carrier's — never understands your risk worse than a human who has actually looked at your property.
If you want a coverage review from someone who reads the technology and the fine print, let's talk. It costs nothing, and you will leave knowing exactly where you stand.
The machines are getting better at modeling the world. We are still the ones who have to live in it.
References
- Parthknowsai, "The Man Replacing LLMs (And He Has $1B to Prove It)," YouTube. youtu.be/X_nWKJg_D6Q
- A.M. Turing Award, "Yann LeCun — 2018 Laureate," Association for Computing Machinery. amturing.acm.org
- eWeek, "Yann LeCun, Meta's Former AI Chief, Launches $1B Startup Focused on 'World Models.'" eweek.com
- TechCrunch, "Yann LeCun's AMI Labs raises $1.03B to build world models." techcrunch.com
- Crunchbase News, "Turing Winner LeCun's New 'World Model' AI Lab Raises $1B In Europe's Largest Seed Round Ever." news.crunchbase.com