Alessio Fanelli: evaluation

Transcript context

…My my perception is definitely that just, just observationally, Cloud 3 is certainly the first thing that I've seen where lots of people. They're, no one's debating evals or anything like that. They're talking about the specific use cases that they have, that they used to use chat GPT for every day, you know, day in, day out, that they've now just switched over. And that has, I think, shifted a lot of the sort of like vibe and sentiment in the space too. And I don't necessarily think that it's sort of a A like full you know, sort of full knock. Let's put it this way. I think it's less bad for open AI than it is good for anthropic. I think that because GPT 5 isn't there, people are not quite willing to sort of like, you know get overly critical of, of open AI, except in so far as they're wondering where GPT 5 is. But I do think that it makes, Anthropic look way more credible as a, as a, as a player, as a, you know, as a credible sort of player, you know, as opposed to to, to where they were. Yeah. And I would say the benchmarks veil is probably getting lifted this year. I think last year. People were like, okay, this is better than this on this benchmark, blah, blah, blah, because maybe they did not have a lot of use cases that they did frequently. So it's hard to like compare yourself. So you, you defer to the benchmarks. I think now as we go into 2024, a lot of people have started to use these models from, you know, from very sophisticated things that they run in production to some utility that they have on their own. Now they can just run them side by side. And it's like, Hey, I don't care that like. The MMLU score of Opus is like slightly lower than GPT 4. It just works for me, you know, and I think that's the same way that traditional software has been used by people, right? Like you just strive for yourself and like, which one does it work, works best for you? Like nobody looks at benchmarks outside of like sales white papers, you know? And I think it's great that we're going more in that direction. We have a episode with Adapt coming out this weekend. I'll and some of their model releases, they specifically say, We do not care about benchmarks, so we didn't put them in, you know, because we, we don't want to look good on them. We just want the product to work. And I think more and more people will, will go that way. Yeah. I I would say like, it does take the wind out of the sails for GPT 5, which I know where, you know, Curious about later on. I think anytime you put out a new state of the art model, you have to break through in some way. And what Claude and Gemini have done is effectively take away any advantage to saying that you have a million token context window. Now everyone's just going to be like, Oh, okay. Now you just match the other two guys. And so that puts An insane amount of pressure on what gpt5 is going to be because it's just going to have like the only option it has now because all the other models are multimodal all the other models are long context all the other models have perfect recall gpt5 has to match everything and do more to to not be a flop…

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