Generalist AI: Can Human Hands Teach a Robot Brain?

Billions in VC capital is pouring into robot foundation models. While peers lean on simulation or internet video, Generalist AI is venturing that large volumes of real-world human manipulation data, captured with handheld devices, is the path to general-purpose robots. Generalist says its dataset has grown from roughly 270,000 hours to more than 500,000 hours, and that GEN-1.5 can learn new manipulation tasks from a single 3-12 second demonstration, averaging 59% success across 10 everyday tasks. The company reports that its models improve predictably as pretraining data grows, potentially following LLM-like scaling laws. Morgan Stanley argues the field still faces model capability hurdles (lack of training data, no clean evaluation, cross-embodiment transfer, counterfactuals) as well as physical-world complexities (law of 9's, conservative customers, limited manufacturing scale outside China), meaning a working model does not immediately translate into steep revenue growth as with LLM