Frontier Radar #5: If AGI is so close, why is everyone leaving? The neolabs and their bets

OpenAI's Greg Brockman has publicly welcomed humanity to the "AGI era" following the launch of GPT-6 Astra, and Nvidia CEO Jensen Huang declared on X that "AGI has arrived." Yet many of the researchers who built the current generation of models are heading out the door. This Decoder Frontier Radar issue, the fifth in a series published six times a year for subscribers, profiles the resulting wave of independent labs and the technical assumptions behind them, sorting 27 of them in detail into six research bets, with nearly 20 more covered briefly.
Among the departures: Ilya Sutskever, an OpenAI co-founder and former chief scientist, now at Safe Superintelligence; Jerry Tworek, who led development of OpenAI's reasoning models, now at Core Automation; David Silver, the force behind AlphaGo, who left Google DeepMind for Ineffable Intelligence; Yann LeCun, who left Meta after twelve years to found AMI; and Jeff Dean, 27 years at Google and most recently its chief scientist, who started Discovery Loop. Google is itself a founding investor in Discovery Loop.
The funding is large relative to what exists. Nvidia invested in SSI in July, reportedly five billion dollars per Bloomberg, for a lab that has released no model or paper. Prometheus, Jeff Bezos' industrial AI lab, raised twelve billion dollars at a 41 billion valuation in June. Flapping Airplanes, also without a public model, raised one billion dollars in an October Series A. Ineffable announced a 1.1 billion dollar seed round in April at a 5.1 billion post-money valuation, described as the largest in Europe. AMI raised 1.03 billion dollars in March at a 3.5 billion pre-money valuation with roughly a dozen employees. SSI, combined with an earlier three billion, holds about eight billion dollars, with a last known valuation of 32 billion.
The Decoder's working definition of a "neolab" is an independently founded, research-driven company organized around a long-term technical bet, whether that concerns how AI learns, perceives, solves problems, is controlled, or computes. The shared doubt is not whether far more capable systems are coming, but whether the current recipe of ever-larger language models trained on internet data and refined with reinforcement learning is sufficient on its own.
Bet one concerns continual learning. Today's models need enormous data to acquire a skill and generally do not update their weights once deployed, so they cannot permanently learn from experience. Adaption, founded by Sara Hooker and Sudip Roy, argues explicitly against scaling and ships data tooling (cleaning, visualization, automated training via AutoScientist) while the real target, deployment-time learning, remains unbuilt. Core Automation is hunting for learning algorithms beyond pre-training and reinforcement learning, with a first model promised to update weights during deployment on one hundredth of today's training data; so far there is only rumor. Flapping Airplanes, from Ben and Asher Spector and Aidan Smith, targets data-scarce domains such as robotics and biology, borrowing biological principles but running on GPUs; it spent much of its first year on foundations, including GPU training efficiency work. Ineffable Intelligence follows Sutton and Silver's argument in "Welcome to the Era of Experience" that systems imitating humans plateau near human performance, and instead needs reward from environment rather than human judgment. Ndea, built by Keras author and ARC designer François Chollet with Zapier co-founder Mike Knoop, uses program synthesis with a neural model steering the search, on a team of about 15. SSI has roughly 50 staff in Palo Alto and Tel Aviv, no output, and a stated plan to eliminate product distractions; Sutskever has said a system that learns as well as a human is five to twenty years away.
Bet two holds that language alone is insufficient. AMI builds on LeCun's JEPA architecture, which predicts in abstract representation space rather than token by token; Meta's V-JEPA 2 showed that after video pre-training, 62 hours of robot data let arms perform simple pick-and-place in unfamiliar labs. Black Forest Labs, known for FLUX image models, released FLUX 3 Action in September, fine-tuned on gameplay, first-person video and teleoperated robot data to take a camera image, robot state and text instruction and predict the next two seconds of motion and scene; weights are on Hugging Face, and it placed second on Nvidia's RoboLab-120 benchmark on October 22, now fourth, tested in simulation rather than on real hardware. General Intuition grew out of the game-clip platform Medal, pairing billions of video frames with keyboard and mouse inputs to train world models and action models; the idea echoes OpenAI's 2022 VPT work on Minecraft. World Labs generates 3D worlds, Physical Intelligence learns control directly on robots, and Pathway's BDH architecture adapts during processing, though only at up to one billion parameters.
The pattern investors are buying is the innovator's dilemma: incumbents with IPO ambitions and competing product teams underinvest in ideas that threaten their business. Precedents exist, since the Transformer came from Google and the mass-market assistant from OpenAI. The lead can be short: TypeSafe AI shipped its Jev decision model in mid-September and OpenAI announced a Decisions API two weeks later. Acqui-hires are the fallback, as with Inflection to Microsoft, Character.ai and Windsurf to Google, AMD's reported 8.2 billion dollar bid for World Labs, and Nvidia's talks with Reflection per the Financial Times.
Why it matters: for developers and product teams, the practical read is that near-term capability gains will likely keep coming from incremental scaling of existing model families, while the bets above are years from shipping, and several may never produce a usable system. Inference, not reported fact: expect continued acqui-hire consolidation and a widening gap between research narratives and available APIs. Nvidia's position is the one that looks durable either way, since it sells compute into every outcome, including labs with nothing to demonstrate.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
OpenAI is talking about the AGI era, and Nvidia CEO Jensen Huang says AGI has arrived. At the same time, co-creators of today's models are leaving OpenAI, Google DeepMind, and Meta to start their own labs, raising billions with no product to show. Why? They're betting that the current recipe alone won't be enough. We've sorted 45 neolabs by their research bets, mapped the limitations of today's AI they're trying to overcome, looked at what's already visible from their work, and explain why Nvidia always wins. The a