William Beauchamp: evaluation

Transcript context

…Yeah, I think we grew by a factor of, you know, users grew by a factor of three last year. Revenue over doubled. You know, it's very exciting. We're competing with some really big, really well funded companies. Character AI got this, I think it was almost a $3 billion valuation. And they have 5 million DAU is a number that I last heard. Torquay, which is a Chinese built app owned by a company called Minimax. They're incredibly well funded. And these companies didn't grow by a factor of three last year. Right. And so when you've got this company and this team that's able to keep building something that gets users excited, and they want to tell their friend about it, and then they want to come and they want to stick on the platform. I think that's very special. And so last year was a great year for the team. And yeah, I think the numbers reflect the hard work that we put in. And then fundamentally, the quality of the app, the quality of the content, the quality of the content, the quality of the content, the quality of the content, the quality of the content. AI is the quality of the experience that you have. You actually published your DAU growth chart, which is unusual. And I see some inflections. Like, it's not just a straight line. There's some things that actually inflect. Yes. What were the big ones? Cool. That's a great, great, great question. Let me think of a good answer. I'm basically looking to annotate this chart, which doesn't have annotations on it. Cool. The first thing I would say is this is, I think the most important thing to know about success is that success is born out of failures. Right? Through failures that we learn. You know, if you think something's a good idea, and you do and it works, great, but you didn't actually learn anything, because everything went exactly as you imagined. But if you have an idea, you think it's going to be good, you try it, and it fails. There's a gap between the reality and expectation. And that's an opportunity to learn. The flat periods, that's us learning. And then the up periods is that's us reaping the rewards of that. So I think the big, of the growth shot of just 2024, I think the first thing that really kind of put a dent in our growth was our backend. So we just reached this scale. So we'd, from day one, we'd built on top of Google's GCP, which is Google's cloud platform. And they were fantastic. We used them when we had one daily active user, and they worked pretty good all the way up till we had about 500,000. It was never the cheapest, but from an engineering perspective, man, that thing scaled insanely good. Like, not Vertex? Not Vertex. Like GKE, that kind of stuff? We use Firebase. So we use Firebase. I'm pretty sure we're the biggest user ever on Firebase. That's expensive. Yeah, we had calls with engineers, and they're like, we wouldn't recommend using this product beyond this point, and you're 3x over that. So we pushed Google to their absolute limits. hat's expensive. Yeah, we had calls with engineers, and they're like, we wouldn't recommend using this product beyond this point, and you're 3x over that. So we pushed Google to their absolute limits. You know, it was fantastic for us, because we could focus on the AI. We could focus on just adding as much value as possible. But then what happened was, after 500,000, just the thing, the way we were using it, and it would just, it wouldn't scale any further. And so we had a really, really painful, at least three-month period, as we kind of migrated between different services, figuring out, like, what requests do we want to keep on Firebase, and what ones do we want to move on to something else? And then, you know, making mistakes. And learning things the hard way. And then after about three months, we got that right. So that, we would then be able to scale to the 1.5 million DAE without any further issues from the GCP. But what happens is, if you have an outage, new users who go on your app experience a dysfunctional app, and then they're going to exit. And so your next day, the key metrics that the app stores track are going to be something like retention rates. And so your next day, the key metrics that the app stores track are going to be something like retention rates. Money spent, and the star, like, the rating that they give you. In the app store. In the app store, yeah. Tyranny. So if you're ranked top 50 in entertainment, you're going to acquire a certain rate of users organically. If you go in and have a bad experience, it's going to tank where you're positioned in the algorithm. And then it can take a long time to kind of earn your way back up, at least if you wanted to do it organically. If you throw money at it, you can jump to the top. And I could talk about that. But broadly speaking, if we look at 2024, the first kink in the graph was outages due to hitting 500k DAU. The backend didn't want to scale past that. So then we just had to do the engineering and build through it. Okay, so we built through that, and then we get a little bit of growth. And so, okay, that's feeling a little bit good. I think the next thing, I think it's, I'm not going to lie, I have a feeling that when Character AI got... I was thinking. I think so. I think... So the Character AI team fundamentally got acquired by Google. And I don't know what they changed in their business. I don't know if they dialed down that ad spend. Products don't change, right? Products just what it is. I don't think so. Yeah, I think the product is what it is. It's like maintenance mode. Yes. I think the issue that people, you know, some people may think this is an obvious fact, but running a business can be very competitive, right? Because other businesses can see what you're doing, and they can imitate you. And then there's this... There's this question of, if you've got one company that's spending $100,000 a day on advertising, and you've got another company that's spending zero, if you consider market share, and if you're considering new users which are entering the market, the guy that's spending $100,000 a day is going to be getting 90% of those new users. 's spending zero, if you consider market share, and if you're considering new users which are entering the market, the guy that's spending $100,000 a day is going to be getting 90% of those new users. And so I have a suspicion that when the founders of Character AI left, they dialed down their spending on user acquisition. And I think that kind of gave oxygen to like the other apps. And so Chai was able to then start growing again in a really healthy fashion. I think that's kind of like the second thing. I think a third thing is we've really built a great data flywheel. Like the AI team sort of perfected their flywheel, I would say, in end of Q2. And I could speak about that at length. But 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. And when we did that, we could really, really, really perfect techniques like DPO, fine tuning, prompt engineering, blending, rejection sampling, training a reward model, right, really successfully, like boom, boom, boom, boom, boom. And so I think in Q3 and Q4, we got, the amount of AI improvements we got was like astounding. It was getting to the point, I thought like how much more, how much more edge is there to be had here? But the team just could keep going and going and going. That was like number three for the inflection point.…

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