Shawn Wang: belief

Transcript context

…Yeah, I don't think OpenAI will collect all of this anymore, but I think for research perspectives, it's very insightful to know, but it's also expensive. So essentially your preference data scales with how many minutes it takes for you to do each task and every button is like, it scales pretty linearly. So it's not cheap stuff. Can we, since you mentioned expensiveness, I think you may have joined one of our spaces back in Lama 2 was released. We had an estimate from you that was something on the order of Lama 2 costs $3 to $6 million to train GPU-wise, and then it was something like $20 to $30 million in preference data. Is that something that's still in the ballpark? I don't need precise numbers. I think it's still a ballpark. I know that the 20 million was off by a factor of four because I was converting from a prompt number to a total data point. So essentially when you do this, if you have multi-turn setting, each turn will be one data point and the Lama 2 paper reports like 1.5 million data points, which could be like 400,000 prompts. So I would say it's still say like 6 to 8 million is safe to say that they're spending, if not more, they're probably also buying other types of data and or throwing out data that they don't like, but it's very comparable to compute costs. But the compute costs listed in the paper always are way lower because all they have to say is like, what does one run cost? But they're running tens or hundreds of runs. So it's like, okay, like... Yeah, it's just kind of a meaningless number. Yeah, the data number would be more interesting.…

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