Nathan Lambert: evaluation

Transcript context

…Humans don't even agree with humans 50% of the time. Yeah, so like that's the thing. It's like the human disagreement or the lack of accuracy should be like a signal, but how do you incorporate that? It's really tricky to actually do that. I think that people just keep using GPT-4 because it's really cheap. It's one of my like go-to, like I just say this over and over again is like GPT-4 for data generation, all terms and conditions aside because we know OpenAI has this stuff is like very cheap for getting pretty good data compared to compute or salary of any engineer or anything. So it's like tell people to go crazy generating GPT-4 data if you're willing to take the organizational like cloud of should we be doing this? But I think most people have accepted that you kind of do this, especially at individuals. Like they're not gonna come after individuals. I do think more companies should think twice before doing tons of OpenAI outputs. Also just because the data contamination and what it does to your workflow is probably hard to control at scale. And we should just mention at the time of recording, we've seen the first example of OpenAI enforcing their terms of service. ByteDance was caught, reported to be training on GPT-4 data and they got their access to OpenAI revoked. So that was one example.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.