How Similar Are Current Deep Learning Systems to GPTN?

I worry that a lot of existing theoretical alignment work is hard to verify or it's like making assumptions that aren't obviously true and I worry that it's sort of ungrounded. Current models are human level or less so human evaluators can basically evaluate these models whereas that will totally break down in the future, he says. If we have unsupervised methods that do well on human level tasks maybe that will say something meaningful about it scaling and generalizing even to superhuman level models now there's still subtleties here for example current language models are trained to predict human mostly on the basis of what they're biased towards.

Transcript

Play full episode

Transcript

Episode notes

The AI-powered Podcast Player

Save insights by tapping your headphones, chat with episodes, discover the best highlights - and more!

Get the app