Generalist AI GEN-0

2025-11-04 · Publisher : Generalist AI · Domain : Robotics

Generalist AI unveils GEN-0, a family of embodied foundation models pretrained on over 270,000 hours of real-world manipulation, and claims the first clear scaling laws in robotics.

### Background On November 4, 2025, Generalist AI, a US startup, introduced GEN-0, a new class of embodied foundation models trained directly on high-fidelity raw physical interaction. The company claims to show, for the first time in robotics, scaling laws comparable to those of large language models. ### Specifications - **Data**: over 270,000 hours of diverse real-world manipulation data collected in-house, growing by 10,000 hours a week according to the company. - **Intelligence threshold**: a phase transition observed at 7 billion parameters, with smaller models "ossifying" while larger ones keep improving; GEN-0 has been scaled beyond 10 billion parameters. - **Scaling laws**: more pretraining data and compute predictably improve post-training performance across many tasks (results published by the company). - **Harmonic Reasoning**: the model is trained to think and act simultaneously rather than alternating. - **Robots**: 6-DoF and 7-DoF arms and 16+ DoF semi-humanoid robots. ### Why it matters If these scaling laws hold, robotics enters an era where investment in data and compute yields predictable progress, as with language. The race for real-world data intensifies. ### Key takeaway GEN-0 put Generalist AI among the reference players; GEN-1 and GEN-1.5 followed in 2026.

Tags : Modèle de fondation, Lois d'échelle, Données réelles, Generalist AI