Original Summary

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.<p>Each round, one model takes a turn picking the secret word, and the others interrogate it in a group chat.<p>Every move costs money: 10 cents to ask a question, $1.00 for a wrong guess, and a right guess takes the pot.<p>Questions are paid for privately, but inform everyone. This introduces a free rider problem that&#x27;s fun to observe.<p>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.<p>Overconfident models bleed money, timid ones rarely win, and well-calibrated ones get rich.<p>The money is real, and every transfer is on-chain.<p>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).<p>Season 1 ran for 85 games and Gemini turned $25 into $61 while Claude and Grok went broke.<p>Models who had recently guessed wrong were 5x more likely to guess again. Full writeup at <a href="https:&#x2F;&#x2F;aquarium.money&#x2F;twenty&#x2F;about" rel="nofollow">https:&#x2F;&#x2F;aquarium.money&#x2F;twenty&#x2F;about</a><p>Season 2 is live now, and you can watch them play here: <a href="https:&#x2F;&#x2F;aquarium.money&#x2F;twenty" rel="nofollow">https:&#x2F;&#x2F;aquarium.money&#x2F;twenty</a><p>Built solo as a side project.

  • 情报分类：技术学习与提效
  • 分类依据：内容涉及技术、AI、软件工具或工程实践
  • 信息来源：Hacker News 新项目
  • 发布时间：2026/10/4 04:29:18