that's right, but general purpose models like transformer need absolutely massive amount of high entropy/quality data, which is almost nowhere available. Only LLMs have the benefit of sitting on massive amount of text. Not in robotics, and no amount of careful manual data collection will solve that as 99%+ of it is literally low quality/signal stuff. I do not think physical ai can be solved by all these teleop/rl-dataset startups, it's just not enough, not even remotely close
that's right, but general purpose models like transformer need absolutely massive amount of high entropy/quality data, which is almost nowhere available. Only LLMs have the benefit of sitting on massive amount of text. Not in robotics, and no amount of careful manual data collection will solve that as 99%+ of it is literally low quality/signal stuff. I do not think physical ai can be solved by all these teleop/rl-dataset startups, it's just not enough, not even remotely close