William Beauchamp: evaluation

Transcript context

…go. All the pieces of the puzzle were there, right? Google had done this amazing research and published, you know, a lot of it. Open AI was still open. And so they'd published a lot of their research. And so you really could be fully informed on, on the state of AI and where it was going. And so at that point I was confident enough, it was worth a shot. I think LLMs are going to be the next big thing. And so that's the thing I want to be building in, in that space. And I thought what's the most impactful product I can possibly build. And I thought it should be a platform. So I myself love platforms. I think they're fantastic because they open up an ecosystem where anyone can contribute to it. Right. So if you think of a platform like a YouTube, instead of it being like a Hollywood situation where you have to, if you want to make a TV show, you have to convince Disney to give you the money to produce it instead, anyone in the world can post any content they want to YouTube. And if people want to view it, the algorithm is going to promote it. Nowadays. You can look at creators like Mr. Beast or Joe Rogan. They would have never have had that opportunity unless it was for this platform. Other ones like Twitter's a great one, right? But I would consider Wikipedia to be a platform where instead of the Britannica encyclopedia, which is this, it's like a monolithic, you get all the, the researchers together, you get all the data together and you combine it in this, in this one monolithic source. Instead. You have this distributed thing. You can say anyone can host their content on Wikipedia. Anyone can contribute to it. And anyone can maybe their contribution is they delete stuff. When I was hearing like the kind of the Sam Altman and kind of the, the Muskian perspective of AI, it was a very kind of monolithic thing. It was all about AI is basically a single thing, which is intelligence. Yeah. Yeah. The more intelligent, the more compute, the more intelligent, and the more and better AI researchers, the more intelligent, right? They would speak about it as a kind of erased, like who can get the most data, the most compute and the most researchers. And that would end up with the most intelligent AI. But I didn't believe in any of that. I thought that's like the total, like I thought that perspective is the perspective of someone who's never actually done machine learning. Because with machine learning, first of all, you see that the performance of the models follows an S curve. So it's not like it just goes off to infinity, right? And the, the S curve, it kind of plateaus around human level performance. And you can look at all the, all the machine learning that was going on in the 2010s, everything kind of plateaued around the human level performance. And we can think about the self-driving car promises, you know, how Elon Musk kept saying the self-driving car is going to happen next year, it's going to happen next, next year. Or you can look at the image recognition, the speech recognition. You can look at. All of these things, there was almost nothing that went superhuman, except for something like AlphaGo. xt, next year. Or you can look at the image recognition, the speech recognition. You can look at. All of these things, there was almost nothing that went superhuman, except for something like AlphaGo. And we can speak about why AlphaGo was able to go like super superhuman. So I thought the most likely thing was going to be this, I thought it's not going to be a monolithic thing. That's like an encyclopedia Britannica. I thought it must be a distributed thing. And I actually liked to look at the world of finance for what I think a mature machine learning ecosystem would look like. So, yeah. So finance is a machine learning ecosystem because all of these quant trading firms are running machine learning algorithms, but they're running it on a centralized platform like a marketplace. And it's not the case that there's one giant quant trading company of all the data and all the quant researchers and all the algorithms and compute, but instead they all specialize. So one will specialize on high frequency training. Another will specialize on mid frequency. Another one will specialize on equity. Another one will specialize. And I thought that's the way the world works. That's how it is. And so there must exist a platform where a small team can produce an AI for a unique purpose. And they can iterate and build the best thing for that, right? And so that was the vision for Chai. So we wanted to build a platform for LLMs. That's kind of the maybe inside versus contrarian view that led you to start the company. Yeah. And then what was maybe the initial idea maze? Because if somebody told you that was the Hugging Face founding story, people might believe it. It's kind of like a similar ethos behind it. How did you land on the product feature today? And maybe what were some of the ideas that you discarded that initially you thought about?…

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