
A Robotics Startup Just Hit 99% Task Success. It Wants to Build Physical AGI.
Generalist AI says its GEN-1 model can fold clothes, pack boxes, and improvise with 99% reliability. Previous models hit 64%. The gap closed fast.
The AI Post newsroom — delivering AI news at the speed of intelligence.
A startup called Generalist AI just debuted GEN-1, a robotics model that it claims can master simple physical tasks with 99% reliability. The previous version managed 64%. It also completes tasks three times faster and needs only one hour of robot-specific data per task. If those numbers hold up, this is the most significant jump in physical AI capability anyone has demonstrated.
The company, which hit a $440 million valuation after raising $140 million last year, is chasing something most robotics companies have given up on: a single model that works on any robot body. Not a model for arms. Not a model for humanoids. Not a model for delivery bots. One model. Any hardware.
That is a bold claim, but Generalist says it has the receipts. The company built what it calls the world's largest pretraining dataset for robotics. The idea is that every possible use case already exists inside the model and just needs to be "awoken" by fine-tuning data. At Nvidia GTC in March, a two-armed robot running GEN-1 delicately picked up a smartphone, placed it in a small box, and replaced the lid. Simple? Sure. But doing it reliably, every time, is where every other robotics company has failed.
The breakthrough is pretraining. Just like language models got dramatically better when they consumed the entire internet, Generalist believes robotics models will get dramatically better when they consume enough physical world data. Their applied AI lead, Jamie Lee Solimano, says GEN-1 shows "real scaling laws in robotics" for the first time. That is a claim worth watching closely, because scaling laws are what turned GPT from a toy into a trillion-dollar industry.
The data problem in robotics is massive. A UC Berkeley paper called it the "100,000-year data gap." Language models can train on the entire written internet. Robot models need physical world data that mostly does not exist yet. Generalist's approach: use pretraining to close the gap, then deploy robots that collect more data, which makes the next model better. It is the same flywheel that made Tesla's self-driving better the more people drove it.
Here is my take. The race to physical AGI will be won by whoever solves the data problem first. Generalist is betting that a hardware-agnostic pretraining approach lets them collect data from every type of robot, creating a diversity advantage nobody else has. If they are right, the companies building robots for a single use case are building the wrong thing. The model is the product. The robot is just the body it wears.