Shawn Wang: evaluation

Transcript context

…Yeah. Most AI stuff from eight months ago, it's really not that relevant anymore. And today we'll try and bucket some of the recent news on it. We haven't done a monthly thing in like three months. So three months is a lot of stuff. That's mostly because I got busy with the conference. But I do want to get back on that horse or maybe just do it weekly so that I don't have such a big lift that I don't do it. I think the activation energy is the problem really. So yeah, I think frontier model wise, it seems like Cloud has really carved out a persistent space for itself. For a long time, I thought it was kind of like a clear number two to open AI. And with 3.5 on it, at least in some of the hard benchmarks on LMSys or coding benchmarks on LMSys, it is the undisputed number one model in the world, even with 4.0 mini. And we can talk about 4.0 mini and benchmarking later on. But for Cloud to be there and hold that position for what is more than a month now in AI time is a big deal. There's not much that people know publicly about what Enthopic did for Cloud's on it. But I think it's still a huge achievement. It marks the beginning of a non-open AI centric world to the point where people on Twitter have canceled ChatGPT. That's been a trend that's been going on for a while. We talked about the unbundling of ChatGPT. But now new open source projects and tooling, they're just built for Cloud. They don't even use open AI. That's a strategic threat to open AI, I think, a little bit. Obviously, open AI is so big that it doesn't really care about that. But for Enthopic, it's a big win. I think to see that going and to see Enthopic differentiating itself and actually implementing research. So the rumor is that the scaling monosematicity paper that they put out two months ago was a big part of Cloud 3.5's on it. I've had off-the-record chats with people about that idea, and they don't agree that it is the only cause. So I was thinking this is the only thing that they did. But people say that there's about four or five other tricks that they haven't disclosed yet that went into 3.5's on it. But the scaling monosematicity paper is a very, very good read. It's a very long read. But it basically says that you can find control vectors, control features now that you can turn on to make it better at code without really retraining it. You just train a whole bunch of sparse autoencoders, find a bunch of features, and just say, let's up those features, and suddenly you're better at code, or suddenly you care a lot about the Golden Gate Bridge. These are the same things to the model. That is a huge, huge win for interpretability, because up to now, we were only doing interpretability on toy models, like a few million parameters, a model of Go or chess or whatever. Cloud 3's on it was interpreted and usefully improved using this technique. Wow. Yeah, I think it would be amazing if we could replicate the same on the open models to then, because now we can use Llama 3.1 to generate synthetic data for training and fine-tuning. I think, obviously, Anthropic has a lot of compute and a lot of money. So once they figure out, OK, this is what we should make the model better at, they can put a lot of resources. I think an open source is probably going to be a more distributed effort. I feel like Noose has held the crown of the best fine-tuning data site owners for a while, but at some point that should change, hopefully. Other groups should step up. And I think if we can apply the same principles to a model as big as 405B and bring them into maybe the 7B form factor, that would be great. But yeah, Cloud is great. I canceled JGBD a while ago. Really small podcaster run for latent space. It runs both on Cloud and on OpenAI, and Cloud is definitely better most of the time. It's not a benchmark. It's just vibes. But when the vibes are good, the vibes are good.…

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