Builders of AI Doubt the Current Recipe: Six Technical Bets from 45 New Labs
The Decoder's deep analysis of 45 new AI labs reveals a core contradiction: while NVIDIA CEO Jensen Huang declared that AGI has arrived, the people who built today's systems are collectively leaving big tech to found new laboratories. They do not doubt that stronger AI will arrive; rather, they doubt whether the current recipe—expanding language models and reinforcement learning alone—is sufficient. The article categorizes these 45 labs into six technical bets: continual learning, world models, exclusive data, compute efficiency, research automation, and model controllability—and explains why NVIDIA profits regardless of which direction wins. The editorial board judges that integration is the most likely outcome: methods rarely remain exclusive, acquisitions are already underway, and some investment capital flows back to investors in the form of compute spending.
NVIDIA CEO Jensen Huang announced on X that AGI has arrived following the release of GPT-6 Astra, with OpenAI President Greg Brockman calling the model a turning point. Yet at the same time, Ilya Sutskever, Jerry Tworek, David Silver, Yann LeCun, Jeff Dean, and others who built today's systems are leaving OpenAI, Google DeepMind, and Meta to start their own labs, raising billions of dollars, most without products or papers. The Decoder's Frontier Radar Issue #5 classifies 45 new labs into six categories of technical bets, attempting to answer a core question: what exactly are they doubting?
The answer is not that they doubt stronger AI won't arrive. The article explicitly states that some founders believe AGI may be near. Their doubt centers on a more specific judgment: whether the current recipe—ever-larger language models, internet-data training, and reinforcement-learning fine-tuning—can get there. This doubt divides the 45 labs into six directions, each corresponding to a specific limitation of today's systems.
The first bet targets learning itself. Today's models require massive amounts of data to master a skill and typically do not update weights after deployment. Core Automation's Jerry Tworek argues that continual learning is a machine learning problem, not a Markdown note-taking problem; its goal is for models to update weights during deployment and achieve equivalent results with one percent of today's training data. Safe Superintelligence's Sutskever expects systems that learn from few examples like humans within five to twenty years. Ineffable Intelligence's David Silver goes further: he believes systems that merely imitate humans struggle to surpass human-level performance, and that next-generation systems must learn from environmental experience rather than receiving rewards from human judgment.
The second bet argues that language alone is insufficient. Yann LeCun's AMI, based on the JEPA architecture, does not predict outputs token by token but learns what matters in abstract internal representations for video or robotic actions. Physical Intelligence learns control directly on robots; its π0.7 model, with approximately five billion parameters, has matched individually trained specialist models in its own tests. Black Forest Labs starts from an image generator, with FLUX 3 Action converting video models into robot controllers. These labs share the judgment that spatial relationships, physical consequences, and world understanding cannot be acquired through text data alone.
The third bet wagers on data that does not exist on the internet. Basecamp Research collects volcanic rock and deep-sea samples in over thirty countries, with datasets reportedly containing over one hundred billion previously unknown genes. Periodic Labs operates automated laboratories where robots mix powders, furnaces sinter them, and X-ray detectors check results; its founder believes the most valuable data consists precisely of failure results that journals never publish. Jeff Bezos' Prometheus, funded with over eighteen billion dollars, aims to build AI that designs physical objects, creating data from physical laws and test results.
The fourth bet focuses on compute efficiency. Inception's Mercury model generates text drafts in parallel via diffusion and refines them step-by-step, claiming to be four to eight times faster than GPT-6 Luna and Claude Haiku 5.5. Liquid AI designs streamlined models for constrained hardware, with its technology now integrated into Mercedes-Benz vehicles. Unconventional AI attempts to replace digital computation with analog oscillators, arguing that energy will become a hard limit for AI in about three years.
The fifth bet productizes research itself. Axiom Math used AI agents to translate the Enormous Theorem—a classification theorem spanning fifty years and over ten thousand pages—into Lean formal proofs in twenty-four days, discovering an error in a 1986 paper known to experts but never publicly documented. Recursive Superintelligence claims its Flint model was built almost entirely by AI in less than four months. Discovery Loop was founded by four former top Google researchers, including Jeff Dean, with Google itself as founding investor and cloud partner.
The sixth bet does not attempt to build the strongest model but wagers that even the strongest model will leave gaps. Goodfire understands and intervenes in model behavior through internal representations; its RLFR method uses the model's own signals as feedback for training, reducing hallucinations by 58% on Gemma 3. TypeSafe AI outputs fixed decisions rather than free text, with calibrated probabilities. Thinking Machines Lab's Tinker allows customers to fine-tune open models using their own domain knowledge, with annualized revenue exceeding two hundred million dollars.
The article's core structural insight concerns NVIDIA. The same CEO who announced AGI has arrived has invested in at least 25 of the 45 labs, with four additional technology or research partnerships, spanning all six bets. This seems contradictory but is not: as long as competing methods run on NVIDIA chips and cloud services, the company need not predict which lab wins. Investments serve not only as research bets but also as customer retention—part of the invested capital flows back to NVIDIA as compute spending. Google, Amazon, and AMD follow the same logic.
The editorial board's final judgment: the baseline scenario is most likely, with funding remaining high but only a few labs delivering provably superior results within two to three years, while major players continue absorbing new methods within weeks. Many new labs will ultimately end up being acquired or poached. An acceleration scenario is reasonable only for individual labs in areas with verifiable results, such as materials research. For product-less labs like SSI, Prometheus, and Flapping Airplanes, the board expects elements of write-downs or acquisitions within two to three years, without breaking the overall dynamic.
This analysis offers readers value not in predicting which lab will succeed, but in mapping the known boundaries of current AI technology routes. Each bet corresponds to a testable technical gap: continual learning remains unrealized, world models still rely on simulation testing, the moat of exclusive data depends on replication difficulty, compute-efficiency breakthroughs face architectural inertia from large labs, and research automation has progressed furthest in verifiable fields like mathematics, yet human decision-making remains indispensable. These boundaries are not guesses; they are the result of builders of today's systems voting with billions of dollars and top talent.