Chai publishes occasional research on how they think about this, including talks at their Palo Alto office: William expands upon this in today’s podcast (34 mins in): Fundamentally, the way I would describe it is when you're building anything in life, you need to be able to evaluate it. And through evaluation, you can iterate, we can look at benchmarks, and we can say the issues with benchmarks and why they may not generalize as well as one would hope in the challenges of working with them. But something that works incredibly well is getting feedback from humans. And so we built this thing where anyone can submit a model to our developer backend, and it gets put in front of 5000 users, and the users can rate it. And we can then have a really accurate ranking of like which model, or users finding more engaging or more entertaining. And it gets, you know, it's at this point now, where every day we're able to, I mean, we evaluate between 20 and 50 models, LLMs, every single day, right. So even though we've got only got a team of, say, five AI researchers, they're able to iterate a huge quantity of LLMs, right. So our team ships, let's just say minimum 100 LLMs a week is what we're able to iterate through. Now, before that moment in time, we might iterate through three a week, we might, you know, there was a time when even doing like five a month was a challenge, right? By being able to change the feedback loops to the point where it's not, let's launch these three models, let's do an A-B test, let's assign, let's do different cohorts, let's wait 30 days to see what the day 30 retention is, which is the kind of the, if you're doing an app, that's like A-B testing 101 would be, do a 30-day retention test, assign different treatments to different cohorts and come back in 30 days. So that's insanely slow. That's just, it's too slow. And so we were able to get that 30-day feedback loop all the way down to something like three hours. In Crowdsourcing the leap to Ten Trillion-Parameter AGI, William describes Chai’s routing as a recommender system, which makes a lot more sense to us than previous pitches for model routing startups: William is notably counter-consensus in a lot of his AI product principles: * No streaming: Chats appear all at once to allow rejection sampling * No voice: Chai actually beat Character AI to introducing voice - but removed it after finding that it was far from a killer feature. * Blending: “Something that we love to do at Chai is blending, which is, you know, it's the simplest way to think about it is you're going to end up, and you're going to pretty quickly see you've got one model that's really smart, one model that's really funny. How do you get the user an experience that is both smart and funny? Well, just 50% of the requests, you can serve them the smart model, 50% of the requests, you serve them the funny model.” (that’s it!) But chief above all is the recommender system. We also referenced Exa CEO Will Bryk’s concept of SuperKnowlege: