The neural nets in our games.
Small nets that play the wolves and bots in Gates, and one that looks for aim cheats in oneshot. Here is each one, what it does today, and the papers it stands on.
In Gates
bots and animals
Bred wolvesA tiny net, 24 inputs and 16 hidden units, picks a wolf pack’s fights. It isn’t trained by gradient descent. It’s bred: each child is its parent plus one random change, and a child replaces its parent only after it beats the parent twice against the real agent, the second time in fights it has never seen. Its genome is a starting seed plus one short seed per change, so a whole wolf brain fits in 280 characters and rebuilds exactly. Bred and waiting for a playtest; the wolves you meet today still run the rule table.
StudentsA recorder sits behind whatever is playing a bot (our scripted policy, or Claude through the agent harness) and keeps every choice it made and what it was shown. A small classifier learns to pick among the same goals the way its teacher did, so a student plays with no model call, at any number of bots. Whole runs are held out, so no run grades itself.
SchoolA student can only copy its teacher. School breeds it further, with the wolves’ mutations, scored on the island’s own standings, and keeps a change only when it also wins on islands it has never played.
The bodyAn experiment: a small transformer over the last 32 ticks that sends a bot’s buttons, moves and turn every tick, and writes its plan three ticks before it acts, in words a spectator could read over its head. So far, reading back its own plan has cost it more than it gained.
In oneshot
aim cheats
The detectorIt reads how a player’s aim moved around each time they saw an enemy: convolutions for the shape of a flick, a recurrent layer for its order, and attention to weigh the moments that matter, plus the gun and the map. Every window is paired with its cheated twin, the same moment with an aimbot laid on, so the only difference it can learn is the cheat.
Our own aimbots, bredWe breed aim assists against the detector with the same seed-per-change genome as the wolves. The archive keeps the best assist of every style, not one champion, and the detector retrains on whatever survives, round after round.
A career, not a matchOne match proves little. A player’s flags add up over their matches as a sequential test, and a case opens only when an honest player would almost never flag that often. A person then watches the replays and decides.
Where it standsTrained on bots only so far, so its numbers are plumbing, not a result. The method and the code are public; the trained weights and the bred aimbots stay on our servers, because a published detector is one a cheat can be bred against.
The rules they all follow
- Same bits everywhere. A net that runs inside a game’s simulation uses only adding, multiplying, dividing and square roots, so the server, the desktop and the browser get the same answer and a replay thinks the same thought.
- Beat a control. Every model is scored against something simpler: repeating the last move, a rule table, a few hand-made features. The detector must also fail when its labels are shuffled.
- Judge on what it never saw. Held-out matches, held-out runs, held-out aimbot styles.
- A model ranks; a person decides. Nothing here bans anyone or takes a payout away on its own.
The papers
what we took from each
Deep NeuroevolutionSuch et al., 2017. A genome as a seed plus one seed per mutation: the wolves, the students’ schooling and the bred aimbots all use it.
MAP-ElitesMouret and Clune, 2015. An archive that keeps the best of every kind, not one winner.
Rainbow TeamingSamvelyan et al., NeurIPS 2024. Attacks bred as a grid of styles, and the target hardened on the whole grid.
Adversarial RL for ESP cheatersKRAFTON and KAIST, 2025. Cheater and detector adapting against each other.
XGuardianZhang, Sun and Qian, USENIX Security 2026. Server-side, from replays, pitch and yaw alone.
AntiCheatPTLoo, Luzkov and Burelli, IEEE CoG 2025. One window per kill, and a public labelled set.
Sequential testsWald, 1945. Weak evidence per match, multiplied to a decision with its error rates set in advance.
Bayesian online change-point detectionAdams and MacKay, 2007. Telling a cheat switched on from a player getting better.
Activation probesMcKenzie et al., NeurIPS 2025. Cheap probes first, uncertain cases passed on: how we plan to read inside the detector.
The full list, with how far each paper was read, is oneshot’s anticheat sources (in its docs).
Both games are open source. The bred nets live in Gates’ genome, and the detector, the aimbot zoo and the career test in oneshot’s aim tools.