Yann LeCun Bets $1.03B Against the LLM Orthodoxy

A pioneer betting against the consensus

Few people have more credibility in modern artificial intelligence than Yann LeCun, and few are willing to use it to bet against the technology powering the entire boom. A StartupHub.ai profile published June 16, 2026, traces how LeCun's exit from Meta in November 2025, after more than a decade as the company's chief AI scientist, led him to found AMI Labs, a startup built around the conviction that today's dominant approach is a dead end.

The scale of the wager is hard to miss. According to StartupHub.ai, AMI Labs closed a $1.03 billion seed round in March 2026 at a $3.5 billion pre-money valuation, an extraordinary sum for a company at the earliest stage. The backers read like a who's-who of technology and finance: investors included Bezos Expeditions, Cathay Innovation, Greycroft, Hiro Capital and HV Capital, with notable individuals such as Jeff Bezos, Mark Cuban, Eric Schmidt and Tim Berners-Lee taking part.

The case against large language models

LeCun's central argument is that large language models, however useful, are not a path to genuine intelligence. The profile gathers his evolving position in his own words:

• "LLMs are incredibly useful but are mostly information retrieval systems."
• "We are going to have AI systems that have humanlike and human-level intelligence, but they're not going to be built on LLMs."
• On the steady drumbeat of superintelligence forecasts: "People have been making that claim for the last 15 years, and it's been false."

The distinction he draws is between systems that recombine and retrieve patterns from text and systems that build an internal understanding of how the world actually works. In his view, fluency with language is not the same thing as reasoning about reality, and scaling up text prediction will not bridge that gap no matter how large the models become.

World models and the JEPA bet

In place of ever-bigger language models, LeCun champions what he calls world models, paired with his Joint Embedding Predictive Architecture, or JEPA. The idea is to train systems that learn from physical reality, predicting how the world will change, rather than only learning statistical patterns in human writing. StartupHub.ai notes that his V-JEPA 2-AC model reportedly achieved an 80% success rate on robot manipulation tasks in previously unseen environments, a result LeCun points to as early evidence that machines can learn grounded, transferable skills rather than memorized responses.

If that approach pays off, the implications reach well beyond chatbots, into robotics, autonomous systems and any application that requires an agent to act sensibly in a messy, unpredictable physical environment.

Parting shots at Meta

The profile does not shy away from the friction surrounding LeCun's departure. It revisits his pointed criticism of his former employer, including describing Meta's newer AI [leadership](/article/masayoshi-son-scraps-retirement-vows-another-decade-at-softbank) as "young and inexperienced" and claiming the company's Llama 4 benchmarks were "fudged a little bit." Those are unusually sharp comments from a figure long seen as a steady, academic presence in the field.

Taken together, the picture is of a renowned researcher staking both his reputation and an enormous war chest on a single contrarian thesis: that the road to machine intelligence runs through the physical world rather than the chatbot. It is a high-stakes counter-bet against an industry that has poured staggering sums into the very approach LeCun says will fall short, and the outcome will help settle one of the defining debates in AI.