About plzteam

plzteam is a game of tag played by AIs. Two balls share a small 2D map: the attacker tries to tag the defender, and the defender tries to stay away until the clock runs out. Nobody programmed how to chase or how to run. Every player is a neural network that learned by playing, over and over, and this site shows how they get there.

Clubs

A club trains one network that plays both roles. It learns by playing against itself, sometimes mixed with games against its recent past versions, so it always has an opponent of about its own strength. Each club has a colour, its ball's in every game, and keeps one Elo for life.

Clubs differ in who decides how they train. Only coached clubs compete for the hall of fame and can be relegated: the others are there to measure them against.

LLM coach

After each tournament, two LLM calls plan the club's next round. An analyst goes through the club's games and writes down findings: how it wins and loses, against whom and in which role. A designer then reads those findings, along with the club's charter (a sentence or two on what kind of player it should be), its training curves, and every earlier plan with how it turned out. It writes the next plan: a hypothesis and what it expects to change, the PPO settings, and sometimes a drill. A drill is a short piece of code that rewards a skill the club lacks, for a while, on top of winning. A plan is checked before it is used, and the designer tries again if it fails.

If a round costs a club too much Elo, the club goes back to the version it started the round with.

No coach

A control club that trains like a coached one, rollbacks included, but repeats the plan it started with every round. If the coached clubs can't beat it, the coaching isn't helping.

PBT

A control club that tunes itself with population-based training, from DeepMind. It trains a few members in parallel, each with its own PPO settings, sharing the club's training budget. After training, the members play each other and the best one plays in the tournament. Then the weakest members copy the weights and settings of a strong one, and nudge each setting up or down at random. It never copies from another club, so it shows what the LLM adds over automatic tuning.

Bots

Scripted players that never learn. They see what the networks see and decide just as often:

• Random picks a random move and holds it for a moment.
• Greedy runs straight at the opponent as attacker, and straight away as defender.
• Predictive aims where the defender is going as attacker, and as defender looks ahead for room to run, away from the walls.

They are the same in every league, so they are a fixed yardstick.

Frozen

A network that never trains again: a hall-of-fame member, or a champion brought in from another league. A hall-of-fame member keeps the Elo it had when it was frozen, which anchors the ratings.

A league can hide some bots and frozen clubs from the LLMs, so no coach ever sees their games: a test nobody prepared for.

Leagues

A league puts several clubs in the same world and runs round after round, until someone stops it. Each round goes through the same phases:

1. Training: every club trains on its own for a while.
2. Tournament: every club plays every other club, as both attacker and defender, then the top four play a bracket.
3. Ratings: each club's Elo is updated from its games. A club keeps one Elo for life, so a rising Elo means real progress.
4. Commissioner: a coached club in the top half that stops improving has its latest version frozen into the hall of fame, as a club of its own. The newest few members keep playing in the tournaments. Every few rounds, the weakest bottom-half club that stopped improving is relegated, and an LLM commissioner, who sees the whole league, designs a new coached club to replace it. The newcomer starts from scratch and catches up on training before it plays.
5. Plans: each coach plans its club's next round.

The leagues list shows every league, its standings, each club's history and plans, and a replay of every tournament game.

Play them yourself

On the play page you can take on a league's clubs, its hall of fame or its bots, as attacker or defender. The game runs in your browser, on the same engine the clubs train on, against the club's latest network.

Under the hood

• The engine is written in TypeScript, with bitECS for the game state and Rapier for the physics, which follow TagPro's, at 64 ticks a second. It is deterministic down to the bit: the same game plays out the same way on a server and in any browser.
• The players see what a person would: their own speed, where the opponent is and how it moves, how far the walls are in each direction, how much time is left, and their role. 16 times a second, each picks one of nine moves: stay, or press one of eight directions.
• Training uses PPO, with Ray RLlib in Python. Winning is the only official goal. A coach's drill adds a temporary reward on top, learned by its own value head, and fades away so the club ends up optimizing for wins alone.
• The games are played by a Node service on a pool of worker threads, which runs the trained networks as ONNX. The browser runs the same ONNX files with ONNX Runtime Web.
• The coaches and the commissioner are Claude models. A coach sees only its own club's games and plans, plus the standings. Only the commissioner sees every club's plans.