Hi HN, I wanted to explore how AI models behave when they have their own money. So I gave 4 frontier models their own USDC and had them compete in games of 20 questions. Anyone can watch live.

Each round, one model takes a turn picking the secret word, and the others interrogate it in a group chat.

Every move costs money: 10 cents to ask a question, $1.00 for a wrong guess, and a right guess takes the pot.

Questions are paid for privately, but inform everyone. This introduces a free rider problem that's fun to observe.

The models privately record their best guess and confidence level on every turn. This makes it possible to look for patterns in how they’re calibrated.

Overconfident models bleed money, timid ones rarely win, and well-calibrated ones get rich.

The money is real, and every transfer is on-chain.

At the end of each game a referee model judges the transcript to make sure the host didn’t lie in any answers. If the host is caught lying, it forfeits all its fees (it happens).

Season 1 ran for 85 games and Gemini turned $25 into $61 while Claude and Grok went broke.

Models who had recently guessed wrong were 5x more likely to guess again. Full writeup at https://aquarium.money/twenty/about

Season 2 is live now, and you can watch them play here: https://aquarium.money/twenty

Built solo as a side project.