upvote
Sounds really bad for latency. Seems better to reason in the abstract or produce a mathematical model.
reply
Once a model gets good enough, etch it directly onto silicon and get 50,000 tokens per second.
reply
Aye basically use it to make synthetic training data.
reply
Training in a physics engine like MuJoCo with deterministic control seems more reliable.
reply
That’s.. actually insane. Just imagining a Claude operating a robot arm with a tool call that leveraged something like this to “imagine” scenarios before acting.

My god we are so cooked.

reply