After 12 years as one of the most visible — and vocal — research leaders inside Facebook-turned-Meta, Yann LeCun announced last week that he will leave the company at the end of the year to found a startup dedicated to what he calls Advanced Machine Intelligence (AMI): systems that understand the physical world, keep persistent memories, reason and plan multi-step actions. He said Meta will be a partner in the new venture.

If the name LeCun rings a bell, it’s not just because he has a reputation for dry jokes and blunt takes on the latest hype cycle. He is one of the architects of modern deep learning — a researcher whose work on convolutional neural networks helped kick off the image-recognition revolution — and one of three recipients of the ACM A.M. Turing Award that cemented the field’s mainstream breakthrough. At Meta, he helped build the Fundamental AI Research group known as FAIR; he’s also kept a foothold in academia as a professor at NYU.

But LeCun’s departure is more than a personnel move. It’s emblematic of a widening split inside big-tech AI: between the recent megabucks, model-centric sprint toward gigantic foundation models and a faction pushing for different architectures that are better at modeling the physical world and long-term memory. LeCun has long been in the latter camp; now he’s betting his reputation (and a new company) on making it real.

The pitch: world models, memory and planning

In a short, plainspoken post to colleagues and followers, LeCun framed AMI as the “next big revolution in AI,” arguing that current models — for all of their headline-grabbing abilities — don’t meaningfully understand the mechanics of the world they describe. He wrote that the new startup will continue research he’s been doing across FAIR and NYU, and will pursue systems that move beyond learning from scraped web text to models that have a richer, more persistent grasp of how the world actually works. Meta, he said, will partner with the company.

That language — “persistent memory,” “reason,” “plan complex action sequences” — is deliberately aspirational. It signals a move toward architectures that can simulate environments, actions and consequences rather than only predict the next token of text. For many researchers, that’s the long road to anything resembling general intelligence; for companies racing to ship ever-bigger LLMs, it’s a different roadmap — and a slower, riskier one.

Why now? A messy moment at Meta

LeCun’s exit comes amid a big reordering of Meta’s AI efforts. Earlier this year, the company rolled out Llama 4, a new generation of its open-weight models; the reception among developers was mixed, and that disappointment helped prod CEO Mark Zuckerberg into a major executive and spending sprint to revamp the AI team.

In June, Meta struck a headline-grabbing deal to invest billions in Scale AI and to bring Scale’s young founder into its ranks — a move meant to bulk up Meta’s commercial muscle for model training and data operations. And in October, the company cut roughly 600 roles out of its newly reorganized Superintelligence Labs as management sought to streamline decision-making and align teams under fresh leadership. Those shifts — plus new hires and a different product emphasis — are part of the internal pressure and context around LeCun’s decision, according to people familiar with the matter.

Different philosophies, same company

Part of the drama is cultural as well as technical. LeCun has been a vocal defender of openness in AI research: he favors sharing models, code and ideas broadly with the research community. Newer hires who arrived with Zuckerberg’s summer hiring spree, by contrast — including executives and leaders with startup backgrounds — pushed toward faster, productized development, sometimes with tighter control over models and IP. Those tensions are normal in big labs, but when they compound with organisational changes, they can sharpen into departures.

LeCun’s new company will give him a clean slate to pursue the AMI program without having to bend it to quarterly product timelines — while preserving a working relationship with Meta. That last point matters: a partner relationship should give the startup access to cloud, compute, engineering muscle and — potentially — the kind of real-world product feedback that’s hard to come by for independent labs.

What it means for the AI ecosystem

There are two easy headlines people will take from this: the “founder leaves big company to start a startup” narrative, and the “one of the godfathers of AI rejects the direction his employer is taking” narrative. Both are true. But the longer, more consequential story is whether LeCun’s vision can translate into systems with real-world advantages — and whether investors, engineers and partners will bet on a slower, arguably harder path at a moment when faster returns come from scaling up LLMs and product integrations.

Meta’s continued partnership with LeCun’s venture will be watched closely: it’s both a hedge and a signal. If AMI research yields breakthroughs — more sample-efficient world models, better memory systems, safer long-horizon planning — it could reshape how companies build assistants, robots and decision systems. If it falters, LeCun will still have helped seed an important experiment about where the next decade of AI might go.

A personal coda

LeCun has often spoken better than most academics about the practical limits of current AI — he’s called models “dumber than a cat” on a bad day and has publicly argued that we need new architectures to close big gaps. Leaving a corporate ramp where he helped plant deep roots is personal as well as professional; he called creating FAIR “my proudest non-technical accomplishment.” Now he’s betting that a different kind of lab, unbound by product cadence but still partnered with industry, can do the heavy lifting his critics say the field needs.

We’ll be watching the startup’s first hires, its research papers, and whether Meta — or anyone else — puts significant engineering and product bets behind the AMI play. For now, the move is both a vote of no-confidence in the status quo and a long shot at changing it — from the inside out and the outside in.