Listening to this recent dwarkesh podcast on RSI

Is progress from GPT3->5 largely due to increasing quantity of data? Guest suggests for pre training it isn’t. More due to curation of good data. For post training probably not too, we needed to figure out the right diversity of RL environments, it wasn’t just adding more

So to improve the models we had to run all these experiments to see which types of data are good and what distribution u want etc.

So you can get to a threshold of data quantity where you generalize decently well and then you just need to be careful about finding subsets that preserve positive transfer.

I think there’s a model of the future where humans are basically data collectors for models, robots, bio tech, etc. that’s it. We aren’t useful but to train ai systems.

The premise is that real generalization is impossible so we just spend energy teaching the ai to be skillful in each domain

but I think this view suggests it’s an intermediate period before gains come from being clever. Cleverness is something AI is capable of, a lot of it can just be brute forced