45 AI Neolabs, Billions in Funding, and a Shared Doubt About the AGI Recipe

By Marcus Bennett

Safe Superintelligence has raised roughly eight billion dollars, holds a 32 billion dollar valuation, and has published neither a model nor a paper. Nvidia invested a reported five billion dollars in the lab in July, according to Bloomberg, betting on Ilya Sutskever's reputation alone.

SSI is the most extreme case of a broad pattern: the people who built today's AI systems are leaving the big labs to found their own. THE DECODER's fifth "Frontier Radar" issue profiles 27 such labs in detail, plus nearly 20 more briefly, sorting all 45 into six research bets. Their shared doubt is not whether powerful AI arrives. It is whether today's recipe — ever-larger language models trained on internet data and refined with reinforcement learning — gets there.

The timing sharpens the contradiction. After the launch of GPT-6 Astra, Nvidia CEO Jensen Huang declared on X: "AGI has arrived." OpenAI president Greg Brockman welcomed humanity to the "AGI era." There is no agreed definition of AGI, but the message is clear: the path is mapped.

Who is leaving, and why it matters

The departures are senior and numerous. Sutskever co-founded OpenAI and served as its chief scientist. Jerry Tworek led development of OpenAI's reasoning models o1 and o3 and now runs Core Automation. David Silver, the mind behind AlphaGo, left Google DeepMind for Ineffable Intelligence. Yann LeCun left Meta after twelve years to found AMI. Jeff Dean, 27 years at Google and most recently its chief scientist, started Discovery Loop.

The money follows. Prometheus, Jeff Bezos' industrial AI lab, raised twelve billion dollars in June at a 41 billion dollar valuation. Flapping Airplanes raised one billion dollars in an October Series A with no public model.

Motives are mixed. Founders become paper billionaires fast at these valuations. LeCun left after a strategy dispute. Sutskever wants research without product pressure. Some departures are not clean breaks — Google is a founding investor in Discovery Loop.

Investors fund the pattern because of the innovator's dilemma (the management-theory observation that established companies struggle with ideas threatening their business). Precedents exist: Google invented the Transformer in 2017 and built AlphaGo, yet OpenAI shipped the mass-market assistant, and its spinoff Anthropic, founded in 2021, reached over 65 billion dollars in annualized revenue by July 2026, per Bloomberg.

What is a neolab?

The report defines the term as independently founded, research-driven AI companies built around a long-term technical bet: changing how AI learns, perceives, solves problems, is controlled, or computes. The wave became visible in 2024, when Sutskever founded SSI in June and raised a billion dollars within months.

Bet 1: Learning, not just scaling

Six labs attack continual learning (models today do not permanently update their weights after deployment) from three angles.

Sara Hooker, formerly of Cohere and author of "On the Slow Death of Scaling," founded Adaption with Sudip Roy, betting against scaling in favor of efficient AI that continuously learns. It raised a 50 million dollar seed in February 2026 and ships data tools plus AutoScientist.

Core Automation's Jerry Tworek — who said after leaving that "when you're in a race for the best model, you take fewer risks" — wants learning algorithms beyond pre-training and reinforcement learning. For him, continual learning "is a problem for machine learning, not for Markdown files." The first model is rumored to update weights during deployment using one hundredth of today's training data. Core raised 432 million dollars and is valued at 3.62 billion per Forge.

Flapping Airplanes (Ben and Asher Spector, Aidan Smith of Neuralink) wants models that learn from few examples; it raised one billion dollars in October at a six billion valuation. David Silver's Ineffable Intelligence, building on his "Era of Experience" paper with Richard Sutton, pursues a "Superlearner" trained on environmental rewards rather than human examples; its 1.1 billion dollar seed in April at a 5.1 billion valuation was Europe's largest. François Chollet's Ndea (30 million raised, 450 million valuation per The Information) uses program synthesis — a neural model steering a search for programs that explain few examples.

SSI's Sutskever told Dwarkesh Patel in late 2025 he expects a system that learns as well as a human within five to twenty years. Pathway and Symbolica pursue adjacent approaches.

Bet 2: Language alone is not enough

AMI is LeCun's counter-program: JEPA (Joint Embedding Predictive Architecture), proposed in 2022, predicts abstract representations rather than pixels. Meta's V-JEPA 2 needed just 62 hours of robot data for pick-and-place tasks. AMI raised 1.03 billion dollars in March at a 3.5 billion pre-money valuation with roughly a dozen employees.

Black Forest Labs released FLUX 3 Action in September — an image generator fine-tuned to control robots, predicting the next two seconds of movement. Physical Intelligence, founded by Sergey Levine and Chelsea Finn, shipped π0.7 with about five billion parameters and opened its Pi API in beta. Fei-Fei Li's World Labs sells Marble, and AMD agreed in late September to acquire the company for about 8.2 billion dollars in stock, with Li becoming AMD's chief scientist.

Generalists are closing in: GPT-6 Astra leads the AA-Robotics benchmark with 69 of 100 points but takes half an hour and over 16 dollars per task, versus 20 seconds for a human. Robotics researcher Georgia Chalvatzaki sees real capability change, not proof that scaling solves robotics.

Bet 3: The best data is not on the internet

Basecamp Research has collected more than 100 billion previously unknown genes from over 30 countries and trains its EDEN models on them. Periodic Labs, founded by Liam Fedus of OpenAI and GNoME's Ekin Dogus Cubuk, runs automated labs and unveiled its Neon model on September 15, beating GPT-6 Astra and Claude Fable 5.1 on X-ray analysis, per the company.

Prometheus, run by Bezos as co-CEO, targets an "artificial general engineer" — modernized CAD — with over 18 billion dollars raised and no product. The Wall Street Journal reports Bezos explored a roughly 100 billion dollar fund to buy industrial companies, a potential route to proprietary engineering data.

Bets 4 through 6: Compute, research automation, and gaps

Inception's Mercury 3 diffusion-based text model shipped October 19, outputting text four to eight times faster, per the company. Liquid AI's models now run in Mercedes vehicles in North America. Ricursive (300 million dollar Series A) released Ripple-65B in November. Unconventional AI, founded by Naveen Rao, computes with analog coupled ring oscillators.

Axiom Math translated the 10,000-plus-page "Enormous Theorem" into Lean in 24 days with AI agents — a task whose 250-page section once took six years — and found an error in a 1986 paper. It raised a 200 million dollar Series A. Discovery Loop, founded by Jeff Dean, Sanjay Ghemawat, Quoc Le, and Oriol Vinyals, reportedly negotiated valuations from ten toward 50 billion dollars. Richard Socher's Recursive Superintelligence released Flint, built by its "Eureka Machine" in under four months; Socher said: "Depending on your definition, we're getting close to AGI."

The sixth group already sells products: Goodfire's interpretability tools cut hallucinations 58 percent on Gemma 3; TypeSafe AI raised 870 million dollars days after OpenAI launched a competing Decisions API; Thinking Machines Lab, under Mira Murati, reports over 200 million dollars in annualized revenue and a reported five billion dollar round at a 40 billion valuation; Reflection's open Beam model is the target of Nvidia acquisition talks per the Financial Times.

Nvidia always wins

Nvidia is an investor in at least 25 of the 45 labs, with partnerships in four more, spanning all six bets. As long as competing approaches run on its chips, Nvidia does not need to predict winners — part of the invested money returns as compute spending.

The report's baseline scenario: most neolabs end up acquired or acqui-hired within two to three years, while labs with hard-to-copy data — Basecamp, Periodic, Physical Intelligence — and those with revenue stay resilient. Whether any of the six bets beats the scaling recipe will decide if the pattern holds.