Shawn Wang: belief

Transcript context

…I'm curious to get your thought on Mistral. Everybody's favorite sparkling weights company. They just released the, you know, Mistral large enough. Mistral large 2. So this was one day after Llama 3, presumably because they were speaking at ICML, which is going on right now. By the way, Brittany is doing a guest host thing for us. She's running around the poster sessions doing what I do, which is very great because I couldn't go because of my visa issue. I have to be careful what I say here, but I think because we still want to respect their work. But Mistral large, I would say it's like not as exciting as Llama 3. I think that is very, very fair to say. It is, yes, another GPT-4 class model released as open weights with a research license on a commercial license, but still open weights. And that's good for the community, but it is a step down in terms of the general excitement around Mistral compared to Llama. I think that would be fair to say, and I would say that to Mistral themselves. So the general hope is, and I cannot say too much because I've had offline conversations with people close to this. The general hope is that they need something more, you know, of the 10 elements of like, what is next in terms of their frontier model boundaries. Mistral needs to make progress there. They made progress here with like instruction following and structured output and multilinguality and all those things. But I think to stand out, you need to basically pull a stunt. You need to be a superlatively good company in one dimension. And now, unfortunately, Mistral does not have that crown as open source kings. You know, like a year ago I was saying, Mistral are the kings of open source AI. Now Meta is, they've lost their crowns. By the way, they've also deprecated Mistral 7B, 8x7B and 8x22B, right? So now there's only like the closed source models that are API platform. So has Mistral basically started becoming more of a closed model proprietary platform? I don't believe that's true. I believe that they're still very committed to open source, but they need to come up with something more that people can use. And that's a grind. I mean, they have, what, $600 million to do it? So that's still good. But, you know, people are waiting for like what's next from them. Yeah. To me, the perception was interesting. In the comments of the release, everybody was like, why do you have a non-commercial license? You're not making any money anyway from the inference. So I feel like the AI engineering tier list, you know, is kind of shifting in real time. And maybe Mistral, like you said before, was like, hey, thank God for these guys. They're saving us in open source. They're kind of like speed running GPT-1, GPT-2, GPT-3 in open source. But now it's like they're kind of moving away from that. I haven't really heard of that many people using them as scale commercially, just from, you know, discussions. So I'm curious to see what the next step is.…

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