OpenAI's Kevin Weil: Building AI at Unprecedented Speed, Learning from Libra
Candy
1
8-1Mia: It feels like if you blink, you miss the next big thing in AI. There's this one idea that really stuck with me: the AI we are using today is literally the worst AI we will ever use for the rest of our lives. How does a company like OpenAI even build products when the ground is shifting that fast?
Mars: You know, it forces a kind of radical agility. You have to build with the full expectation that the underlying tech will be dramatically better in just a few months. It's about building for what's almost possible, because a product that barely works today is going to absolutely sing tomorrow.
Mia: So it's not about building for today's AI, but for an AI that doesn't even exist yet. That's a huge mental shift for anyone in product development.
Mars: Exactly. If what you're building is right on the bleeding edge of the model's capabilities, that's actually a great sign. It means you're ahead of the curve, and you just have to wait for the models to catch up to your vision.
Mia: That's a powerful way to look at it. But to build like that, you must need a totally new set of tools. I've heard you talk about something called evals being a critical new skill for product managers. What exactly is an eval?
Mars: Right. Think of evals as really rigorous quizzes for an AI model. They test its ability on very specific tasks, from creative writing to complex reasoning. It's one thing to know a model *can* do something, but it's another to know if it gets it right 60% of the time or 99.5% of the time. That number completely changes how you have to design the product around it.
Mia: I see. So the evals are like a report card that dictates the product strategy. It's not just about testing, but about guiding that constant improvement. This makes me wonder, if the big foundational models are getting so powerful, where does that leave startups? Is there any room left for them?
Mars: Oh, absolutely. The opportunities are massive. A foundational model company can't possibly build everything. The key for startups is to find their niche. They can leverage industry-specific data or focus on a very particular use case that a big company would never have the specialized knowledge for.
Mia: Got it. So it’s about going deep and specific, not broad and general. That makes a lot of sense. To move that fast, though, the internal culture must be different. I heard this fantastic term, vibe coding. What on earth is that?
Mars: It's a great term, right? It's about a new way of working with AI tools like GitHub Copilot. You're not meticulously writing every single line of code. Instead, you're guiding the AI, letting it generate code, and you're just correcting and steering it. It's more of a collaboration.
Mia: So it's less like being an architect with a perfect blueprint and more like... a dance with the technology. You lead, it follows, it suggests a new move, and you adapt.
Mars: That's a perfect analogy. It lets you prototype and explore ideas at a speed that was impossible before. You combine that with empowered teams who are trusted to experiment, and that's how you get the velocity we're seeing.
Mia: It sounds like a complete rethinking of the development process. So if you had to boil all this down, what are the core principles for someone trying to build in this new AI era?
Mars: I'd say it comes down to a few things. First, accept that AI is improving exponentially; what you have today is the floor. Second, get really good at evals to truly understand model performance. Third, for startups, focus on a niche where you can win. And finally, embrace new ways of working like vibe coding. It’s all about putting in good, consistent work over time, because that's what truly drives progress at this unprecedented speed.