My first startup idea was AlphaZero-as-a-Service.

The idea was to find popular games on the App Store, build a cracked AI for it following the AlphaZero paper, and then reach out and say,

"Hey, I built an AI for your game that crushes the one you have. Wanna chat?"

This would be custom-made for each game at first (which aligns with @paulg 's "do things that don't scale"), but eventually, the AlphaZero core could be factored out. At some point, it could even be self-service.

The idea came from building an AI for my own game (wallgame.io). I built a Minimax-based AI while my friend Thorben built an AlphaZero-style AI.

The AlphaZero-style AI overperformed in every metric. It was stronger, faster, and more fun to play against (with human-like play). All while being much cheaper and faster to train than we expected.

It felt like discovering a unique edge. People associate AlphaZero with Google-scale compute, but our experiment suggested it was more broadly applicable.

I didn't pursue it because all I had was a thesis and one data point. But with all the attention on LLMs, I think the alpha is still there. And with agents, it would be much easier to make custom wrappers for each game. Should I pursue it? 👀

As part of my build-in-public approach, I wrote a playbook for implementing AlphaZero-style AIs: nilmamano.com/blog/wall-game-ai

Let me know if you can beat the AlphaZero-style AI by playing against the "Normal Bot" at wallgame.io/play

(you can't)

PS. Low-key proud of the double pun "with all the attention on LLMs, I think the alpha is still there" 😁