Why Your App Sucks
Henry Kim ·
Today, I'd like to talk about something a bit technical. There's a concept Andrej Karpathy has been frequently mentioning lately: the iteration of automated research and verifiable Validation. I also believe this is one of the most practical AI design logics today. Anyone who has tried building an AI app will understand this frustration. GPT scores perfect marks on every test and answers any question with ease, but when you actually apply it to your own work, the results are terrible. By our domain standards, it falls far short. Why is that? It's because that AI has never been trained in our specific domain. The solution is RLVR: Reinforcement Learning with Verifiable Rewards. The core logic is simple: AI produces an output → I evaluate it using predefined metrics → If it performed well, I give a reward signal → Repeat. If you design this loop and train the AI, within 24 hours, that AI can produce results equivalent to 10 years of human experience in that domain. Remember when AlphaGo defeated Lee Sedol? It was trained using this very method. It even worked for games with an infinite number of possibilities, like Go. Now, we live in an era where we can create AlphaGo-like AI agents in our respective domains, 100 times easier than it was to build AlphaGo itself. Of course, designing that loop, defining evaluation metrics, and conducting the training is still up to humans. It's impossible without domain insight. That's why I believe that as AI advances, domain expertise becomes even more crucial. We need people who can teach AI what it needs to learn. And honestly, on one hand, I also wonder if we're quite close to AGI with this.