Soumith Chintala: evaluation

Transcript context

…Okay, great. And then maybe one more thing about benchmarks, since we already brought it up and you brought up Confident Benchmarks. There was some recent drama around AnyScale. AnyScale released their own benchmarks and obviously they look great on their own benchmarks, but maybe didn't give the other... I feel there are two lines of criticism. One, which is they didn't test some apples for apples on the kind of endpoints that the other providers, that they are competitors with, on their benchmarks and that is due diligence baseline. And then the second would be more just optimizing for the right thing. You had some commentary on it. I'll just kind of let you riff. Yeah, I mean, in summary, basically my criticism of that was AnyScale built these benchmarks for end users to just understand what they should pick, right? And that's a very good thing to do. I think what they didn't do a good job of is give that end user a full understanding of what they should pick. Like they just gave them a very narrow slice of understanding. I think they just gave them latency numbers and that's not sufficient, right? You need to understand your total cost of ownership at some reasonable scale. Not oh, one API call is one cent, but a thousand API calls are 10 cents. Like people can misprice to cheat on those benchmarks. So you want to understand, okay, like how much is it going to cost me if I actually subscribe to you and do like a million API calls a month or something? And then you want to understand the latency and reliability, not just from one call you made, but an aggregate of calls you've made over several various times of the day and times of the week. And the nature of the workloads, is it just some generic single paragraph that you're sending that is cashable? Or is it like testing of real world workload? I think that kind of rigor, like in presenting that benchmark wasn't there. It was a much more narrow sliver of what should have been a good benchmark. That was my main criticism. And I'm pretty sure if before they released it, they showed it to their other stakeholders who would be caring about this benchmark because they are present in it, they would have easily just pointed out these gaps. And I think they didn't do that and they just released it. So I think those were the two main criticisms. I think they were fair and Robert took it well. And he took it very well. And we'll have him on at some point and we'll discuss it. But I think it's important for, I think the market being maturing enough that people start caring and competing on these kinds of things means that we need to establish what best practice is because otherwise everyone's going to play dirty.…

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