Lenny Rachitsky: observation

Transcript context

…Yeah, yeah, yeah. So it meant companies like Salesforce were basically just a thin wrapper. And I think that that's kind of the mistake people made. So we're at this company Cursor, and if you look under the covers in Cursor, they've built 14 different models to really understand how a developer works, a high-end, a real developer. Those models have tons and tons of interactions with how people talk to their friend Cursor about how they should design their programming so forth. And that's real, that's not just a thin layer on a foundation model. And I think there are many, many applications like that. And so I think there's going to be a lot of opportunity at the application layer. There's going to be some opportunity at the foundation model, and of course you can invest in Sam, you can invest in Anthropic and so forth as well. But there will probably be a very small number of companies at that set, and then a almost unlimited number of companies at the application layer. And then as the technology advances, we'll of course see more things we can body to AI. I mean already autonomous cars are working really well now after a long, long, long, long time. Since I think Sebastian won the challenge in 2006 when he drove the self-driving car across the country. And here we are 20 years later and now they're deployed. So that was a long time. Robots I think is a harder problem than self-driving cars. So we'll see how that goes. But, yeah, there's certainly a lot in that world as well. Wow, okay. There's a lot to this answer. No, that was exactly what I was looking for. The Cursor example, it's something that comes up a lot on this podcast, in the application layer specifically. The thought that the way to win in this space and to build a moat is, as you said, "Build your own model slash have proprietary data that you build through people using your product. " Thoughts on that? Yeah, I mean, I think that ends up just being what's required. So it turns out that the universe is long-tailed, is fat-tailed, and humans are very fat-tailed in terms of human behavior, human conversation and so forth. So to get to the real meaning of it and to get to the kind of essence of the problem, in any domain turns out to be, I think, more complex than we thought. And so the early things and people were running around saying, "Okay, there's going to be one big brain to rule them all on these kinds of things." That's kind of not played out yet. And in fact, if you look underneath the covers, you have LLMs, which have generalized pretty in fascinating ways, but they've kind of also asymptoted in that we have run out of data for the most part. And so if you look at the GPT-5 LLM compared to the GPT-4.1 and how much more it costs to train and so forth, it's definitely not going linear anymore. On the other hand, the reinforcement learning side has been linear, but it doesn't generalize. So if you build a great programming model, it may be an idiot at math. And so that I think is just very different than what people would've said three years ago. And I think that there's not something that's both scaling and generalizing yet, and maybe we'll get there, but that certainly opens the door to something that's more user-friendly, that's more effective in any number of domains than just the basic foundation model infrastructure. Now those models are incredibly important, and I think OpenAI is probably 80% of the revenue in AI are something like that now. It's massive, and so that foundation model is really, really important. And then the basic consumer app is really, really important. That just answers whatever the hell you want to know. Those things are very, very real. But I do think particularly, and then if you get into enterprise stuff and then it's no longer internet data, it's their data that becomes very different. Databricks is having a lot of success there because, okay, well, once you're inside a company, guess what? You care about access control. That's hard with an AI world. It gets trained on some stuff. How does it know who has access to that information and who doesn't, and so forth. You have semantic issues. So if you look at an enterprise, find 10 enterprises, they all have a different definition of what a customer means. You would think customer is a basic thing. Well, is it a department at AT&T? Is it AT&T? Is it a person at AT&T? What the hell is the customer? And it turns out to be very, very meaningful, particularly if you're trying to figure out important things like churn and this and that and third. So that kind of stuff matters. So I would just say the problem space is a lot bigger than you can just attack with a basic foundation model currently. Maybe that will change, and if that changes, then certain prices will have, in retrospect, look way inflated and others will look too low. But that is TBD.…

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