> Astra told me I can walk to the car wash to wash my car.
Yes, like I said, a mediocre idea of how the world works.
> Again, the improvement in coding is due to RLVR, and it does not generalize to other domains. So it is meaningless to point to examples of that improvement.
I can point at actual benchmarks for performance on business tasks. On the one hand, those lines go up; on the other, I just don't know how good those benchmarks are.
> There are many ordinary behaviors and skills that are not feasible to RL at scale. The sampling efficiency is just way to low.
Yes. A point I make myself on occasion. One of two perfectly valid ways to describe intelligence, and one by which all current machine learning really sucks.
Like, if it was fully general, the trillion or so miles that Tesla cars have driven would have obsoleted the steering wheel long ago, rather than the closest we have being (yet another) pronouncement from Musk that this new car he's not actually made available for sale won't need one.
Business self-play training could be anywhere from code to cars. I was expecting much faster progress with the cars, and much slower progress with code, so I don't know how to adjust my expectations with any other domain, including business.