Sutskever, Silver and LeCun Leave the Big Labs: What the Neolabs Are Betting On

A strange picture: after the launch of GPT-6 Astra, Jensen Huang writes on X that AGI has arrived, while OpenAI president Greg Brockman welcomes humanity to the "AGI era." At the same time, the people who built today's models are leaving the big companies in droves and opening their own labs. The Decoder counted 45 such "neolabs" and examined 27 of them in detail. The conclusion is simple: many of those who left do not believe the current recipe is enough.

Who Left and How Much They Raised

Ilya Sutskever, OpenAI co-founder, is building Safe Superintelligence. In July, according to Bloomberg, Nvidia invested about $5 billion in SSI. Together with the earlier $3 billion, that comes to roughly $8 billion at a $32 billion valuation. Yet the lab has released neither a model nor a paper.

Jerry Tworek, who led development of the o1 and o3 reasoning models, founded Core Automation. David Silver, the author of AlphaGo, left DeepMind for Ineffable Intelligence and in April raised $1.1 billion in a seed round at a $5.1 billion valuation. Yann LeCun, after twelve years at Meta, founded AMI. Flapping Airplanes, with no public model, raised $1 billion in a Series A at a $6 billion valuation.

The motives are mixed: paper billions, a strategy dispute with Meta, a desire to do research without product pressure. Sometimes the break is not even clean: Google itself became one of the first investors in Jeff Dean's Discovery Loop.

The Main Bet: Learning, Not Just Solving

The most straightforward group of labs works on continual learning. Today's model does not change its weights after release. It uses new information within a conversation but does not remember it permanently. ChatGPT and Claude work around this with notes and search over past conversations. Tworek believes continual learning is a job for machine learning, not for Markdown files. Simply fine-tuning a network risks catastrophic forgetting: it loses what it knew before.

Imagine a junior employee who, a month into the job, still knows exactly what they knew on day one, plus whatever you put on their cheat sheet. That is roughly how a model behaves today.

The approaches differ:

• Adaption builds models that cheaply adapt to the user on the fly. So far it sells only data tools and AutoScientist via API.
• Core Automation wants to update weights during work and make do with a hundredth of the usual data.
• Flapping Airplanes trains on a small number of examples, which matters for robotics and biology, where data is scarce.
• François Chollet's Ndea looks for a program that explains a few examples.
• Silver's Ineffable trains a system on its own experience with a reward from the environment.

The last approach has a weak spot. In Go, it is clear who won. In science or the physical world, the signal arrives late or is not clear at all.

What This Means for Practitioners

First, a small lab has a short head start. TypeSafe AI released its decision model Jev in mid-September, and two weeks later OpenAI announced the Decisions API. Second, the exit often comes through acquisition: AMD wants to buy World Labs for about $8.2 billion, Nvidia is in talks with Reflection, and Microsoft took the Inflection team.

For developers and product managers, the takeaways are down to earth:

• Do not build an architecture that rests on the assumption that a model will never learn to remember. Memory in the form of notes remains a crutch for now.
• Watch ARC-AGI-3. Astra scored 62.7% in the standard setup by translating unfamiliar games into compact rules. That is exactly what Ndea is after.
• Do not confuse valuation with a product. Most neolabs have nothing you can plug into via API.

The bet is simple. If the current recipe hits a ceiling, the winner will be whoever first teaches a model to learn. If not, the big labs will simply buy the team or replicate the result within a couple of weeks.