I’ve been watching most of the industry spend three years staring at the same shiny object. VCs, founders, colleagues, journalists, CIOs, and board members who learned the word “agentic” at the airport lounge in October. Uncle ChatGPT and aunt Claude crashed the innovation party, the corporate music stopped, and everyone is still holding the same lukewarm drink. GenAI: semi-expert of many, master of (almost) none. Meanwhile one of the people who actually built the foundations of modern deep learning has been standing in the corner with a chilled Chardonnay, watching everyone go feral over what he politely keeps calling a useful sub-component.
His name is Yann LeCun. The thing he has been arguing about is not a feature you can add to GPT-5. It is a different floor of the building. The New York Times put it on the front page of the tech section in January 2026 with the headline “Yann LeCun warns the tech herd could hit a dead end.” Pour a stiff drink, grab some salty nibbles, this is going to take a minute.
The autocomplete with a god complex
Strip the sparkly wrapper and an LLM is a very sophisticated autocomplete. It reads a statistically obscene amount of (mostly stolen) human text, learns which tokens tend to follow which, predicts the next one with eerie accuracy and mind-popping speed. The output looks like reasoning, feels like understanding, occasionally writes a better meeting summary than the people who were in the meeting, and feels way more empathic than your mother-in-law.
Looks like. Feels like. That is the part nobody wants to dwell on: it is all an incredibly sophisticated make-believe. Every fact an LLM “knows” passed through a human first. Physics, gravity, cause and effect, the basic instinct that pushing a toy car makes it roll, all of it second-hand, all of it inferred from how someone happened to describe it in writing. The model never touched the car. The model never saw the car. The model read about the car, and the writing was uneven, incorrect, contradictory and that is the foggy world it lives in. It feeds us happily curated averages packed as wisdom.
LeCun has been saying this in public since 2023, including the years when saying it was professionally inconvenient (or bloody dangerous). On the Big Technology Podcast and again with Jacob Effron in May 2026, his line is the same: LLMs “are not a path towards human level or human like intelligence or even animal like intelligence.” He is careful to add that they are useful, including to him, and that he uses them daily. He just thinks you cannot scale your way out of an architectural ceiling. He has been right about this since before it was fashionable.
The energy bill makes that ceiling visible. Training GPT-3 burned roughly 1,287 megawatt-hours and emitted over 500 metric tons of CO2 equivalent, the lifetime footprint of about 100 American cars (and that is just training, inference can eat up to 60% of total AI electricity once a model is in production). The water numbers are noisier (the famous 500ml per ChatGPT conversation is outdated, Sam Altman publicly claims 0.3ml, independent estimates now hover around 5ml). The disagreement does not matter much; the aggregate is the story: we are running a planetary resource problem and calling it a productivity tool.
Multimodal is not the escape hatch
Someone always raises an eager hand at this point. But GPT-5 watches video, they say. Gemini 2.5 listens to podcasts. Claude Opus 4 reads charts. Multimodal, see, surely that is the world, the future and the Holy Grail.
Nope, it is just the same beast wearing more clothes, and baring a shoulder. A multimodal LLM encodes images and audio into token-like representations and runs the same statistical prediction loop on top. Vision bolted on does not give you physics; let alone thought. Audio bolted on does not give you planning. What you get is a system that can describe a scene and speculate about what might happen in it, with the speculation still routed through learned correlations rather than any actual model of how the scene works (or could or should). Constructing your context based on text, sound and images is not the smartest way to create a world-dominating super brain.
LeCun’s February 2026 keynote at the World AI Cannes Festival has the cleanest version of the argument. A human learns to drive a car in roughly twenty hours of practice. AI systems trained on millions of hours of driving data still struggle to clear that bar. The same models breeze through professional exams and faceplant on a basic household robot test. The gap is the entire story. “Most of human experience and most human intelligence has nothing to do with language“, he told Cannes. Language alone will not get you there. The toddler watching her glass fall off the table already knows more physics than any model in production.
Agents are useful.
The word “agentic” has been doing extraordinary unpaid labor on LinkedIn for about a year. Everyone has an agent now. Everyone is building agents. Half the time the thing being demoed or used is an automation script with a chatbot roughly fastened on, marketed as autonomy, sold as transformation, deployed as disappointment. Do not get me wrong: real agents exist. They use tools, chain steps, take actions, send emails, book flights, and query APIs. The better ones are genuinely useful and I use them eagerly every single day. The thing they are missing is the thing world models would give them. An agent without a world model reacts. An agent with a world model can stop, simulate the consequence of the action it is about to take, evaluate whether the consequence is the one it wants, and decide accordingly.
LeCun, again at Cannes 2026, puts it harder than I would: “building an agentic system that does not have the capability of predicting what the effect of its actions are going to be, is an extremely bad way of building an agentic system.” Today’s agents are the chess player who memorized a million openings. A world-model agent is the one who actually plays chess. The opening-memorizer wins the first six moves and then loses the game.
What is a world model
LeCun’s proposal lives in his 2022 paper, A Path Towards Autonomous Machine Intelligence. The architecture has a name, JEPA, the Joint Embedding Predictive Architecture. The trick is in the word predictive, with the catch that JEPA does not predict the next token, or the next pixel, or the next frame. It predicts at the level of representations. Abstractions about what should happen, instead of pixel-by-pixel reconstructions of what it would look like. The difference is massive.
Point a JEPA-based model at a video of leaves rustling in the wind. An LLM-style generative model burns compute trying to predict every leaf in the next frame, most of which is unpredictable noise, and most of which does not matter. JEPA learns to model what does matter: trajectory, object permanence, the basic grammar of physics. The result is a system that can, inside its latent space, “imagine” what happens if it takes an action, weigh whether the outcome is desirable, and plan around it. That is much closer to how the brain works (your visual cortex does not render a photorealistic next-frame when you reach for your coffee, it runs an abstract trajectory at roughly 20 watts of power, the same energy budget as the lamp on your desk. You are welcome).
LeCun’s full proposal stacks six modules: perception, world model, cost, short-term memory, actor, and a configurator that orchestrates them. One coherent model of reality, adaptable to many goals. Compare that with what we do now, which is to train a new specialized model for every task and gaffer-tape them together with prompt engineering (the most over-hyped phrase of the last decade). The energy math is also blunt. JEPA-style systems are designed to learn from less data, at higher levels of abstraction, with smaller models. Less electricity per unit of useful intelligence. A genuinely different relationship between compute and capability than the “throw another data center at it” school currently running frontier research.
Why LeCun gets to say all this
LeCun has the kind of CV that makes contrarianism dangerous to dismiss. He won the Turing Award in 2018 jointly with Yoshua Bengio and Geoffrey Hinton, an award the field calls the Nobel of computer science. He invented convolutional neural networks, the architecture that powers every image recognition system on the planet, from your phone camera to medical imaging to the face scanner at the airport. He spent over twelve years as Meta’s Chief AI Scientist and built FAIR, the lab that produced Llama in 2023, into one of the most consequential AI groups in the world.
The bit that gives the argument its weight is this. He spent years saying LLMs were a dead end, publicly, on stages, in print, on podcasts, while cashing a very large Meta paycheck that was directly tied to the success of LLMs. People grandstand for clicks. He said the unpopular thing for years inside the building, on the payroll, and kept saying it. That is rarer than it sounds, and it takes some massive cojones.
In November 2025, after twelve years, he left. His LinkedIn post on 19 November was unsentimental: “I am creating a startup company to continue the Advanced Machine Intelligence research program (AMI) I have been pursuing over the last several years with colleagues at FAIR, at NYU, and beyond. The goal of the startup is to bring about the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences.” Meta stays in as a partner, with no equity stake. His own framing at the AI-Pulse conference in Paris a few days later was even sharper: Silicon Valley is hypnotized by generative models, so this work has to happen somewhere else.
Paris, not Palo Alto
For forty years the rule was simple. If you were building the next serious thing (in anything tech, so certainly in AI), you went to San Francisco, or Seattle, or maybe London if you were European and felt the need to keep some continental dignity. The gravitational pull of American capital, American compute, American talent, American venture money was, for all practical purposes, physics. You bought the plane ticket. You moved.
LeCun is the rare case where the blue-glassed president of the fifth-largest economy on earth got on a public stage and lobbied a single researcher to stay home. At the Adopt-AI Summit at the Grand Palais on 25 November 2025, Emmanuel Macron stood on the CEO Stage and said France would “do everything we can to ensure his success from France” (as reported by MIT Technology Review via TechBuzz). That kind of public lobbying for one scientist is something you usually only see in football transfer windows.
By December the name was confirmed: AMI Labs, Advanced Machine Intelligence, LeCun as Executive Chairman, Alex LeBrun (formerly CEO of Paris-based medical AI startup Nabla) as CEO, headquarters in Paris, with planned offices in New York, Montreal and Singapore. LeCun himself keeps his apartment in New York, which is fair, the man has a life. The legal entity is French. The capital is French, the center of gravity is French. That is what matters.
The team is impressive: Saining Xie, late of Google DeepMind’s GenAI/nanobanana team and four years at FAIR before that, comes in as Co-Founder and Chief Science Officer. Pascale Fung, one of the most senior voices in human-centered AI research alive today, joins as Co-Founder and Chief Research and Innovation Officer. Michael Rabbat, a long-time FAIR scientist on world models and agents, takes Co-Founder and VP of World Models out of Montreal. Laurent Solly, Meta’s own former VP for Europe, lands as Co-Founder and COO. That is a quiet poaching operation across two continents and three research traditions, dressed up as a Series A.
On 9 March 2026, AMI raised $1.03 billion at a $3.5 billion pre-money valuation, one of the largest seed rounds in the history of tech. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital and Bezos Expeditions, with Jeff Bezos and Mark Cuban as named angels. Strategic and corporate backers include Nvidia, Samsung, Sea, Temasek, Toyota Ventures, Bpifrance Digital Venture, and, in a delicious twist for any thought piece on the advertising industry, Publicis Groupe. French capital is heavily represented: Association Familiale Mulliez, Groupe Industriel Marcel Dassault, Artémis, Aglaé Lab. The market has opinions about this. Some of those opinions are written in eight figures.
Two Paris bets, opposite directions
The geography gets stranger when you remember that Paris already has a frontier AI champion, and that champion is making the exact opposite architectural bet from LeCun, on the same Métro line. Mistral was founded in April 2023 by Arthur Mensch (ex-Google DeepMind), Guillaume Lample and Timothée Lacroix. Lample and Lacroix came straight out of FAIR Paris. The lab LeCun founded. The two best-funded next-generation AI companies in continental Europe were, at root, seeded by the same lab and the same man (which is the kind of detail that should come with its own documentary, preferably not funded by Canal).
They disagree about almost everything that matters. Mistral’s thesis is that LLMs still have decades of scaling left and that Europe needs the hardware to ride that curve before the Americans and the Chinese pull the door shut behind them. In September 2025 the company closed a €2 billion Series C led by ASML at roughly €11.7 billion post-money, about $13.8 billion in real money. ARR went from $20 million in early 2025 to $400 million in early 2026 (per IntuitionLabs’ breakdown). In March 2026 Mistral raised an additional $830 million in debt to acquire 13,800 Nvidia GB300 GPUs and stand them up at Bruyères-le-Châtel, an hour south of Paris. A separate €1.2 billion data center deal in Sweden is already in flight. That is the textbook frontier-LLM playbook executed at French speed with Dutch and German money. More tokens, more parameters, more compute, more electricity. Bigger, faster, hotter. Urgh.
AMI Labs sits on the same Seine and builds the case against it. Smaller models. Less data. Higher abstraction. Energy as a design constraint instead of a quarterly budget line item. Two ex-FAIR-Paris founder teams, three Métro stops apart, betting opposite futures on the same talent pool, the same regulator, the same time zone.
Whichever one wins, Paris wins. Whichever one loses, Paris still ends up with a frontier AI company and a very expensive lesson. That is what hedging looks like.
EU and the future of AI
The obvious counterargument writes itself, and I have heard a version of it from every CIO I have briefed since November. Europe regulates, the Americans scale, the Chinese steal, that is the established physics. The continent that gave the world GDPR and the AI Act is rarely the continent making aggressive, winner-takes-all infrastructure bets. Brilliant researchers. Terrible venture capital density. 24 official languages, 27 regulatory frameworks, a deep structural ambivalence about whether Europe even wants to win, or whether it would prefer to set the rules for other people’s winning.
All of that is true. Most of it has been true since 1995. And yet. France and Germany jointly convened the Summit on European Digital Sovereignty in Berlin on 18 November 2025, 1,000-plus policymakers, industry leaders and researchers in one room. Companies pledged over €12 billion in European digital investment during the summit alone. LeCun himself pushed Meta to open FAIR Paris in 2015, the lab that produced Llama in 2023. He has now opened AMI in the same city. The pattern has a shape and the shape is deliberate.
LeBrun, AMI’s CEO, calls the strategic position “a new Switzerland” in a world of AI that is becoming “increasingly bipolar” between the US and China: trusted, neutral, principled. Whether you find that metaphor inspiring or vaguely terrifying depending on your reading of Swiss neutrality in the twentieth century, the logic is sound. Europe has deep AI talent that is currently underemployed in Silicon Valley shuttle buses, a regulatory environment that, whatever its friction costs, will increasingly shape global AI governance for everyone, and a real value proposition around safety, transparency, and human-centered design.
For the first time since the deep learning revolution began, a serious, next-generation AI research program is being built on the right bank of the Seine instead of in a glass box in Mountain View. Whether the bet pays off depends on whether LeCun is right about world models and whether Europe can supply enough compute, talent and capital to back him. LeBrun has said the work will take “several years“. LeCun, in his May 2026 conversation with Jacob Effron, predicted the industry will recognize the architectural shift by early 2027. He has been ahead of the calendar before.
So what
The LeCun-versus-the-rest debate does not need to be resolved this quarter. Frontier research takes years and AMI itself says so. The argument sits one floor below the things people fight about on Twitter. It is about what AI is fundamentally for. Text prediction at scale, or world understanding at efficiency.
Less energy, smaller models, common sense, planning; persistent memory and physical grounding. Properties that current LLMs cannot grow into by doing more of what they already do (and the bill for doing more of what they already do is already eating CIOs alive on the budget side, as I argued earlier). The bet is that the next decade of AI value comes from a different architecture, full stop, and a smaller one at that.
That this next chapter is being written in Paris, with European talent, (some) European capital, a publicly stated commitment to a European model of AI that is controllable, safe, grounded in actual reality rather than in trillion-parameter text compression, is something the continent should treat as strategy. If LeCun is right, Silicon Valley is about to learn the hard way that being hypnotized by the current thing is no substitute for understanding the next one. The people who figured that out first are working out of an office building somewhere off the right bank of the Seine.
AMI means “friend” in French.
Which might turn out to be the most pointed piece of naming in the history of this industry.
