# Games where playing is the training signal.

> A short lineage: games whose players made the system behind them smarter — and the strategy game built on that idea.

The short history of games whose play makes a computational system smarter — citizen-science games like Foldit, and hexodic, where decisive wins over the bot literally breed its successor.

Canonical HTML: https://hexodic.com/games-that-train-ai
Site index for agents: https://hexodic.com/llms.txt

## Are there games where playing trains an AI?

Yes — it's a small but real lineage. The famous ancestor is **Foldit**, the University of Washington's protein-folding puzzle game, where players' spatial intuition produced solutions that eluded automated methods — most famously helping resolve the structure of a monkey-virus protease that had resisted researchers for years, a result published with the players credited. Foldit's insight was that *play itself can be a scientific instrument*: aggregate enough human moves at a hard problem and the system behind the game gets materially better.

That insight mostly stayed in citizen science. **hexodic** brings it to the game itself: the thing your play improves is your own next opponent.

## How hexodic turns wins into training signal

hexodic's bots improve through a **classical search-agent evolution loop** — explicitly *not* a neural network training on your data:

1. A **decisive human win** over the production bot is treated as evidence of an exploitable weakness.
2. The game is analyzed for the **pivot** — the ply where the bot lost the thread.
3. A **candidate bot** is generated to account for that weakness.
4. The candidate must beat **every bot in the current pool** with statistical significance — a Wilson 95% confidence gate.
5. Winners are promoted to production; losers are discarded.

The full mechanics, including what triggers a candidate and how the gauntlet is scored, are on [the bot gauntlet](/bot-gauntlet). The data side is deliberately boring: completed games recorded anonymously — no account, no personal data, moves and outcome only — as documented on [how your games help](/how-your-games-help).

## Why this is different from "AI-powered" games

Most games that advertise AI ship a static model: it was trained once, somewhere else, on something else. In hexodic the loop is **live and adversarial** — the specific ways real players out-think the current champion are what seed its successor. If you can reliably beat the top bot, you aren't generating engagement metrics; you're the **selection pressure**. And because the mechanism is a tournament of programs with a statistical promotion gate, it's fully inspectable — no black box between your win and the next bot.

The long-term vision — labeled vision, because it hasn't shipped — is [national bots](/roadmap): country-level bots trained by their own player bases, playing *for* their players. The game this loop lives in is a deterministic, five-minute [abstract strategy game](/abstract-strategy-game) — depth first, clever loop second. [Get hexodic](/#get-hexodic).

## Questions this page answers

- Are there games where playing trains an AI?
- What games use human play to improve the computer opponent?