An OpenAI researcher said: inside OpenAI, researchers will be replaced by AI first, infra engineers second, and sales last. Sounds counterintuitive, but it might actually happen. Ideas are cheap. Most research is idea generation + experiments, which AI can mostly do (top researchers excluded). AI infra code is the foundation that researchers use to run experiments. The codebase is huge, out-of-distribution in the training data, and bugs are common. Sales is pure human psychology: human nature, trust, incentives, and emotions. That’s the final boss for AI.