
How can we get enough data to train a robot GPT?
I work through what it would take to collect two trillion tokens of robot experience using fleets, simulation, and human video.
I write It Can Think about how robots learn, how we test them, and what happens when they work.

I work through what it would take to collect two trillion tokens of robot experience using fleets, simulation, and human video.

Predicting how a scene will change is one thing. How does a robot turn that prediction into its next action?

If robot labor became cheap, what work would we take on that nobody can justify doing today?