Shawn Wang: recommendation

Transcript context

…Yeah, I think that's a good question. But I think, I actually don't think people have a good understanding of AGI. And I don't mean definition level. I mean, people are like, okay, we're gonna, AGI means it's powering 40% of world economic output or something like that, right? But what does that mean? So do you think electricity is powering 40% of world economic output or is it not? Like generally the notion of powering X percent of economic output is not defined well at all for me to understand how to know when we got to AGI or how to measure whether we're getting AGI. Like, you know, you can look at it in terms of intelligence or task automation or whatever. I think that's what we are doing right now. We're basically integrating like the current set of AI technologies into so many real world use cases where we find value that if some new version of AI comes in, we can find, we can be like, ah, this helps me more. In that sense, I think the whole process of how we think we got to AGI will be continuous and not discontinuous like how I think the question is posed. So I think the open source thing will be very much in line with getting to AGI because open source has that natural selection effect. Like if a better open source model comes, really no one says, ha, I don't want to use it because there are ecosystem effect, I'm logged into my ecosystem or, I don't know if I like the models, you know, whatever. It's just a very pure direct thing. So if there's a better model that comes out, then it will be used. So I definitely think it has a good chance of achieving how I would think about it as a continuous path to what we might define as AGI. For the listeners, I would actually mention a couple other maybe related notes on just this very interesting concept of feedback sinkhole for open source to really catch up in terms of the overall Google versus OpenAI debate. Open Assistant was led by Yannick Kilcher who recently ended his effort. I think the criticism there was like the kind of people that go to a specific website to give feedback is not representative of real world usage. And that's why the models trained on Open Assistant didn't really seem like they have caught on in the open source world. The two leading candidates in my mind are LMSYS out of UC Berkeley who have the LMSYS arena, which is being touted as one of the only ways, only reliable benchmarks anymore. I kind of call them non-parametric benchmarks because there's nothing to cheat on it except for ELO. And then the other one is OpenRouter, which is Alex Atala's thing. I don't know if you've talked to any of these people. I obviously know all of the efforts that you talked about. I haven't talked to them directly about this yet. But the way I think about it is the way these models are going to be used is always going to be way more distributed than centralized. Like, which is the power of the open source movement. Like the UI within which these models are going to be used is going to be decentralized. These models are going to be integrated into hundreds and thousands of projects and products and all of that. And I think that is important to recognize. Like the LMSYS leaderboard is the best thing we have right now to understand whether a model is better or not versus another model. But it's also biased in only having a sliver of view into how people actually use these models. Like the people who actually end up coming to the LMSYS leaderboard and then using a model only use it for certain things. Like GitHub Copilot style usage is not captured in say LMSYS things. And so many other styles, like the character AI style things is not captured in LMSYS.…

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