Hamel Husain argues synthetic data can't replace real users when building LLM products. Simulating humans would require AGI-level capabilities we don't have, so synthetic data stays on the happy path and misses the specific failure modes that happen in actual coding work. Real user data is essential for robust products.
Synthetic data, it's gonna be hard to simulate a human. At some point it gets really hard to simulate your users. If you can truly simulate a human, that's AGI-ish. We're not there yet. It's gonna be a lot better to use real user data and bootstrap off of that, especially in this context where there's a lot of nuance. And developer tool workflows, I think it would be very difficult to simulate. My intuition for developer tool workflows like website builders, it's going to stay on a very happy path. And there's a lot of unexpected things that happen when you're coding or trying to build something with AI that are very specific to the user's environment, so on. So there's a lot of different footguns that I think might not surface. So I would say don't get too crazy with simulation. Try to get real user data.
