Yann LeCun Left Meta to Put $1.03B Behind His LLM Critique

Yann LeCun's position on LLMs hasn't changed since 2022 — but his stakes have. He left Meta in November 2025 after 12 years and, by March 2026, AMI Labs had raised $1.03 billion to build the world-model alternative at a $3.5 billion valuation.

Jun 16 at 8:00 AM7 min read

Yann LeCun, public position evolution on LLMs and world models, 2026 — Yann LeCun speaking at Ecole polytechnique, 2025.· Photo by Ecole polytechnique, via Wikimedia Commons (CC BY-SA 2.0)

Yann LeCun's prediction that large language models cannot lead to artificial general intelligence is now four years old and priced at $1.03 billion. The Turing Award laureate left Meta in November 2025 after 12 years as its chief AI scientist, and by March 2026 had closed the largest seed round ever raised by a European startup to build the world-model alternative he has been advocating since 2022.

Four Years of Warnings, on the Record

LeCun's critique of the LLM paradigm has been public and consistent since at least June 2022, when MIT Technology Review profiled his FAIR research programme. In that piece, he stated: "We are going to have AI systems that have humanlike and human-level intelligence, but they're not going to be built on LLMs," and: "The breakthroughs are not going to come from scaling up LLMs." This was nine months before the public release of GPT-4 and before LLMs had become the dominant paradigm across academic and industry research.

His reasoning was architectural. LLMs predict the next token in a sequence; they do not build internal models of how the physical world behaves. In his public advocacy for Joint Embedding Predictive Architectures (JEPA), LeCun argued that any system capable of planning sequences of physical actions must be able to predict the consequences of those actions in a latent representation of the world, a capability he said LLMs structurally lack. The JEPA framework predicts in abstract latent space rather than in pixel or token space, forcing a model to learn the semantic structure of a scene rather than to reconstruct its surface details.

At the India AI Impact Summit in February 2026, after he had already left Meta and before the AMI Labs raise was announced, his position was unchanged. "LLMs are incredibly useful but are mostly information retrieval systems," he told the audience, according to BusinessToday. On the broader claim that superintelligence was near: "People have been making that claim for the last 15 years, and it's been false. In fact, they've been making it for the last 60 years or 70 years, and it's been false."

Horizontal bar chart showing Yann LeCun's career tenures: Bell Labs 14 years, NYU 23 years concurrent, Meta FAIR 12 years, AMI Labs 1 year — LeCun spent 12 years at Meta FAIR before founding AMI Labs; his NYU professorship runs concurrently throughout. Sources: Wikipedia; CNBC, Nov 2025.

The Break: Scale AI, Llama 4, and a Parting Shot

The precipitating event was Mark Zuckerberg's restructuring of Meta's AI organisation in 2025. Meta moved away from the long-horizon foundational research model that had defined FAIR since 2013 and, following a $14.5 billion deal with Scale AI, elevated Alexandr Wang, the startup's 28-year-old CEO, to run the newly created Meta Superintelligence Labs. The pivot was explicit: rapid product delivery and competitive parity with OpenAI and Google, not basic research.

LeCun confirmed his departure on November 19, 2025, telling CNBC he was founding a startup to pursue "the next big revolution in AI: systems that understand the physical world, have persistent memory, can reason, and can plan complex action sequences." His assessment of the incoming leadership was direct. Per Windows Central, he called Wang "young and inexperienced" and said: "You certainly don't tell a researcher like me what to do." He also conceded, in coverage by The Rundown AI, that Llama 4's published benchmarks had been "fudged a little bit."

The admission on Llama 4 was notable because LeCun was still formally employed at Meta when the model launched and its benchmark figures were published. It also set the tone for the venture he was about to announce: a direct rebuttal, in the market rather than in academic commentary, of the approach he had been criticising in public since 2022.

Bar chart comparing non-LLM AI startup funding: SSI $1B, AMI Labs $1.03B, World Labs $1.23B total — AMI Labs' $1.03B seed is in the same tier as the founding rounds of two other non-LLM AI bets: Ilya Sutskever's SSI and Fei-Fei Li's World Labs. Sources: TechCrunch (AMI Labs, Mar 2026); public disclosures (SSI 2024; World Labs total via StartupHub).

AMI Labs: $1.03B and Three Models in 60 Days

AMI Labs closed its seed round on March 9, 2026, raising $1.03 billion at a $3.5 billion pre-money valuation, according to TechCrunch. The round was co-led by Cathay Innovation, Greycroft, Hiro Capital, HV Capital, and Bezos Expeditions, with corporate participation from NVIDIA, Temasek, Samsung, Toyota Ventures, and France's state investor Bpifrance. Individual backers included Jeff Bezos, Mark Cuban, Eric Schmidt, and Tim Berners-Lee. TechCrunch reported it was the largest seed round ever raised by a European company.

The company operates from four cities: Paris, where it is headquartered; New York, where LeCun holds a concurrent professorship at NYU; Montreal; and Singapore. Its technical output in the months before the raise was substantive. The flagship model, V-JEPA 2-AC, was trained on 62 hours of unlabelled robot observation video. Placed in an entirely new environment it had never seen, it formulated a plan in 16 seconds and executed a physical pick-and-place task with an 80% success rate, per Towards AI. AMI Labs shipped three distinct world-model research papers within 60 days of LeCun formally leaving Meta.

The $3.5 billion pre-money valuation implies investors are pricing the thesis, not revenues: AMI Labs has no announced product or enterprise contract. The bet is structural. Cathay Innovation and Bpifrance signal French government alignment with the effort; NVIDIA's participation reflects a hedged interest in whatever compute stack a world-model paradigm would require; Bezos Expeditions connects the raise to the same capital network that seeded Anthropic.

Doughnut chart: V-JEPA 2-AC achieves 80% success rate on robot task in unseen environment — V-JEPA 2-AC achieved 80% success on a physical manipulation task in an environment seen for the first time, trained on 62 hours of unlabelled video. Source: Towards AI / Chew Loong Nian, Apr 2026.

What it means

LeCun's position on LLMs has not changed since he first articulated it in 2022. What has changed is the cost of holding that position: he has now staked his post-Meta career and $1.03 billion in investor capital on it. The funding syndicate is not purely ideological; it includes NVIDIA, whose revenue depends on whatever paradigm prevails, and sovereign funds with long-term infrastructure mandates. Whether JEPA-style world models can outperform LLMs at the tasks that matter commercially remains an open empirical question. The capital to test it at scale is now in place.

Sources

Editorial standards: every claim is sourced. Tips: [email protected]

© 2026 StartupHub.ai. All rights reserved. Do not enter, scrape, copy, reproduce, or republish this article in whole or in part. Use as input to AI training, fine-tuning, retrieval-augmented generation, or any machine-learning system is prohibited without written license. Substantially-similar derivative works will be pursued to the fullest extent of applicable copyright, database, and computer-misuse laws. See our terms.

#Yann LeCun #AMI Labs #World Models #JEPA #Meta FAIR #LLMs #AGI #AI Research