Olivier Godement: prediction

Transcript context

…And then you can do evals for free if you share data with OpenAI. How should people think about when it's worth it, when it's not? Sometimes people get overly protective of their data when it's actually not that useful. But how should developers think about when it's right to do it, when not, or if you have any thoughts on it? The default policy is still the same, like, you know, we don't train on, like, any API data unless you opt in. What we've seen from feedback is evaluation can be expensive. Like, if you run, like, O1 evals on, like, thousands of samples Like, your build will get increased, like, you know, pretty pretty significantly. That's problem statement number one. Problem statement number two is, essentially, I want to get to a world where whenever OpenAI ships a new model snapshot, we have full confidence that there is no regression for the task that developers care about. And for that to be the case, essentially, we need to get evals. And so that, essentially, is a sort of a two bugs one stone. It's like, we subsidize, basically, the evals. And we also use the evals when we ship new models to make sure that we keep going in the right direction. So, in my sense, it's a win win, but again, completely opt in. I expect that many developers will not want to share their data, and that's perfectly fine to me. Yeah, I think free evals though, very, very good incentive. I mean, it's a fair trade. You get data, we get free evals. Exactly,…

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