Projects

Vibe coding is great. You describe what you want, the agent writes it, the tests pass, you ship. It keeps working right up to the moment it does not: the job gets killed by the OOM reaper in production, or it opens ten thousand file descriptors and falls over, or it runs fine on the sample and melts on the real dataset. The code was correct. It just assumed it had infinite memory and infinite file handles, because nothing in the problem statement said otherwise.

AI can write code fast. But can it write correct code?

Vibe coding has grown up. What started as "prompt and pray" has evolved into a serious engineering discipline: practitioners design verification harnesses, write rule files, set up automated test suites, and let AI agents run autonomously for hours. The human role has shifted from writing code to defining constraints and reviewing outputs. The results are impressive — and improving fast.

It started over lunch with two economist friends. We were having one of those long conversations about AI and economics — what happens to labor income when machines get good enough, whether capital saves you or dooms you, the usual cheerful lunchtime topics. By dessert, we had talked ourselves into a challenge: write an interactive research paper about it, in the style of the AER, with Yannick's idea as the starting point.

The result is here: How to Die Optimally — A Theory of Consumption When AI Takes Your Job.

My kids are between five and eleven, so I usually keep their screen time short—unless they are building the experience themselves. What started as a throwaway line quickly became a pact, and they held me to it. We pulled out sketchbooks, drew some ideas, and soon we were designing a small 3D game together.