Yann LeCun left Meta in November 2025, after twelve years as its chief AI scientist, and incorporated AMI Labs ("Advanced Machine Intelligence") in Paris that December. The venture went public in January 2026. Among the AI startups I forecast, it is the odd one out: explicitly premised on not using LLMs, and on beating them at some point in the future.
It raised $1.03 billion in March at a $3.5 billion pre-money valuation to build world models, AI that learns from video and physical interaction rather than from text. LeCun has said for years that human-level AI will not be built on LLMs, and he has indicated there may be no product for something like five years (a timeline under which I imagine LLMs will have transformed society). Bloomberg reported on August 5 that LeCun joined the investing firm 224 Ventures as a partner while keeping his chairman role.
Among the "neolabs," the wave of billion-dollar AI startups founded by researchers who left the big labs, AMI came in last on every metric on my neolab bench, but maybe this is primarily because I believe in LLMs and LeCun does not. One point of similarity is that FutureSearch spends its days building world models of a different kind, so I should be more sympathetic to LeCun's approach. I tried to be as objective as possible here.
| Question | Forecast |
|---|---|
| First public product, model, or API | May 2028 (Mar 2027 to Jan 2031) |
| A world model beats frontier LLMs on a recognized benchmark before 2028 | 15% |
| Reported valuation at the end of 2027 | $4.5B ($4.4B to $18.5B) |
These are the core facts I anchored to:
- November 2025. LeCun leaves Meta after twelve years, having founded its FAIR lab.
- December 2025. AMI is incorporated in Paris.
- January 2026. The venture goes public, initially reported to be raising $500 million at $3.5 billion. The mission is machines that understand the physical world.
- March 2026. The round closes bigger, $1.03 billion co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, at $3.5 billion pre-money. LeBrun takes the CEO job.
- Spring 2026. Formal JEPA theory papers appear, with no weights or demos attached, and Nabla announces an exclusive partnership to build agentic healthcare AI on AMI's world models. The company's own updates page still has exactly one entry, the launch note.
- July 2026. LeCun tells the BBC the model will be refined through year-end, with industrial applications hoped for in 2027. LeBrun tells TechCrunch world models are complementary to LLMs, not a replacement. LeCun's separate fund-raising effort collapses.
- August 2026. LeCun joins 224 Ventures as a partner, remaining AMI's executive chairman.
The median for AMI's first public product, model, or API is May 2028, with the 80% interval running from March 2027 all the way to January 2031. Two clocks tick against each other here. LeCun told Wired at launch that AMI plans to release its first models "quickly," even if nobody notices, and told the BBC in July the company would spend the rest of the year refining its model, with hopes for industrial use in 2027. But the actual output so far is theory, formal JEPA papers with no weights attached, and the early applications are partner-gated, like an exclusive healthcare deal with Nabla that would never trigger a public release. FAIR under LeCun dropped open checkpoints alongside papers as a rhythm. If AMI inherits that habit, the left tail fires. If it ships like SSI, 2031.
The benchmark question comes out at 15% by the end of 2027, and the win condition is a conjunction. AMI itself must release a system, benchmark it against frontier LLMs rather than against other world models, and do it somewhere widely recognized or independently verified. Today's world-model results, V-JEPA, Cosmos, Genie, show qualitative progress, not head-to-head wins. And AMI's CEO frames world models as "complementary, not replaceable", which is not the posture of a lab hunting a public benchmark kill. Most of the 15% lives where LLMs are weak, sensorimotor prediction, video planning, physical reasoning, where a third-party evaluation could produce a clean verdict.
The valuation forecast is the most honest number on this page. The median is $4.5B at the end of 2027, exactly today's mark, because there is roughly an even chance AMI simply never prices a new round. A $1.03B seed funds a 30-to-50 person team for years. Conditional on a raise, the step-up runs 2x to 4x, the p90 is $18.5B, and the comparables argue AMI is cheap. World Labs repriced from about $1B to $5B, and SSI went from $5B to $32B in under a year. The 224 Ventures news cuts against urgency on all of it. A founder splitting time with a venture fund is not sprinting toward a launch, though the day-to-day was already LeBrun's, so I filed it under mild execution risk rather than a red flag.
If AMI is right about LLMs, the first sign will come from somewhere none of the standard metrics look, which is exactly why I wrote the win condition as any recognized benchmark or verified head-to-head, not a leaderboard slot. For a bench built to measure the race, the lab that rejects the premise is worth measuring most carefully. These forecasts are how I will notice if the fossil-fuel era ends early.
Forecast these yourself in the FutureSearch app , the moment AMI publishes anything, or the moment a world model tops a leaderboard.