It learns to win the fight.
Offense is where you find out if a brain is real. Ours plays shooters and tank games from the screen and the engine, gets better from its own losses, and takes orders like only the chainsaw or don’t fire as hard limits.
- You drive
- It learns your moves
- Hands off
- The brain drives itself
Cross-compare: attack results, game by game
| game | our brain | what went against us |
|---|---|---|
| DOOM E1M1, Hurt Me Plenty | exit switch at 52 s, 4 kills | Nightmare: not cleared in 8 tries |
| DOOM — rules with no learning | 17.9 vs random play’s 3.2 | memory on top: even |
| DOOM E1M5 — learning | route progress more than doubled, 0.19 → 0.43 | no exit on E1M5 yet; pilot unchanged, the rise came with 51 human fixes to the lane — mostly built by us |
| Atari BattleZone (tank) | 27,000 vs random 3,000 | a dodge rule was rejected |
| BattleZone — a brain moved in from another tank game | 33,575 vs the game’s own rules 28,600; after pruning 48,588 vs 29,275 | the brain moved is hand-written by us (+17% is transfer, not learning); a second export was only even |
| 3D tank arena (BZFlag) | net kills per minute −1.17 → −0.26, hand-written to evolved | still loses to BZFlag’s own AI; v2 is ours, v3 and v4 the loop’s promotions |
| Robot Tank (Atari) | 5.67 tanks destroyed vs random 2.33 | hand-written rules, not learning |
| GoldenEye 007, Dam | invented 2 tactics that beat the default | 0 of 72 attempts survived |
Aim is counted, not claimed: in the tank game the eye’s aim error is 0.55° at the 95th percentile.
From never seeing the game to the built-in AI’s heels, in one night
- 26 Sep, 18:11 — first match. Our rule pilot had never seen this game before that day. Hand-written starter rules only; by the second hand-written version it stood at −1.17 net kills per minute against the game’s AI (8 matches).
- 23:39 — evolution switched on. Every death explained; the loop picked one change at a time from options our engineers wrote, and tested it head-to-head in live matches.
- 00:27 — first change it proved. Dodge only when a shot will really come within 6 m; won clearly (z 3.08).
- 01:14 — second. Stop juking after firing; won clearly (z 2.68). 34 changes tried in all, 2 kept, 8 thrown out.
- By 01:14: −1.17 → −0.26 net kills per minute, hand-written to evolved — a clear gain (z 2.93) on matches separate from the ones that chose the changes. It still kills less than the game’s own AI (1.28 vs 2.10 a minute); its deaths are close (1.53 vs 1.73), but that gap is not significant. Not caught yet.
Why the models can’t catch up: their weights are frozen
Claude, DeepSeek and Llama played every match with the same weights they started with. A neural model cannot evolve without touching its weights: to get better at this tank, someone has to collect the matches, retrain it in a data centre and ship a new version — and the new version can forget what the old one knew. Between matches, it learns nothing.
Our pilot changed between matches without touching a single weight — because it has none. From its own deaths, the loop picked two changes, proved them head-to-head, and kept them. That is the whole difference: a model that has to be rebuilt to learn, against a pilot that learns between fights.
Would you let Claude drive your tank now?
58 decisions in a ten-minute match. 3 kills and 47 deaths across seven. Frozen until someone retrains it. Or a rule pilot that made 60,140 decisions in the same ten minutes, had never seen the game before that day, and by the next night had improved itself from −1.17 to −0.26 against the game’s own AI — without a single weight.
A neural model still has a job — seeing and reading. It does not command. Orders inside its authority, Peel carries out; orders outside it, it refuses and tells you why.
From the arena’s recorded matches. Honest caveats: 7 matches per model; Claude ran through its command-line tool, so part of its 7.6 seconds is plumbing; the game’s own AI still out-kills our pilot; the evolved versions never played the language models — the hand-written ones beat them.