I spent the weekend teaching Jev, a small AI built for split-second decisions

I spent the weekend teaching Jev, a small AI built for split-second decisions, to beat the bosses in DOOM. It runs on my PC and makes every move by itself.

Jev is a way of using a language model as a fast decision-maker: instead of writing an answer, the model scores a fixed list of options in one pass and the best one wins.

Mine is a 4-billion-parameter model on one RTX 3090 under my desk. It never sees the picture. It reads a short text description of the screen (enemies, walls, the route, whether the shotgun is reloaded) and picks one of 26 moves in about 90 ms, four times a second.

The game never pauses while it thinks.How I taught it:I built a teacher: about 500 lines of rules that play the game from the same text.Jev copied the teacher, then played on its own, and the teacher corrected the positions Jev actually got itself into.

Four rounds, about an hour of training each.

The teacher decided the result. My first version beat the boss level 70% of the time when it reacted instantly, and never with 50 ms of delay. Jev needs 90 ms per decision, so it could not copy tactics that only work at zero latency. I rebuilt the teacher to fire only when the shotgun is ready, dodge during the reload and keep its distance from the two Barons of Hell, and tested it at Jevs speed. It now wins about 45% of runs at that delay, and Jev wins 8 of 20 runs it had never trained on.

Why spend a weekend on this: many business decisions have the same shape. They are narrow and repeated, the answer is one of a short list, and it is worth less when it arrives late.

A small local model handles that shape at almost no running cost, and the data never leaves the machine. The work sits in writing down what a good decision looks like, under the conditions it will actually run in.The model was never the hard part. The hard part was writing a teacher whose judgment survives a 90 ms delay.