JevMade hello@JevMade.com
← Back to experiments

Playable demos & bots

llm-lan-party

An AI player tackles Unreal Tournament 99 on a real Windows 98 machine.

Source screenshot of llm-lan-party
SOURCE SCREENSHOTFull screenshot ↗

What it does

The maker found its aim too imprecise: direction choices were much wider than the targets.

How you can use it

A developer can borrow this approach to let an AI control a game character. The code does not show the AI the screen. Instead, it describes the situation in simple words like enemy close or health hurt. It then asks the AI what to do next.

Your developer can adapt this code using an access key that connects the app to TypeSafe. They can also use a local AI instead. The AI struggles with exact numbers and aiming. You will need extra code to help the character aim at the chosen target.

Maker-reported (not independently measured by JevMade): Jev scored 137/137 on the 18-situation key of 2026-09-19 and has not been re-scored since (author-reported) · gpt-oss-20b scored 172/177 on the shipped profile at roughly 216 ms per decision on 2026-09-20; the four fixtures added since were scored by the scripted brain only, 181/199 (author-reported) · In runs 2 and 3 the bot fired 314 and 110 times and scored zero kills; the median yaw error to the target Jev named was 19.9° and 12.1°, against 3.2° of angle for a 34 uu player at 600 uu (author-reported)

Primitives
choice, noul
Platform
TypeScript
Added
Project created
GitHub stars
1 (snapshot, not live)