Перейти к содержимому

I Forced 4 AIs to Speedrun Doom...(only one finished all 8)

Novaki

0:00 / 0:00

I Forced 4 AIs to Speedrun Doom...(only one finished all 8)

1 187 просмотров · 10 дн. назад
Novaki
44 подписчика
1 187 просмотров · 10 дн. назад
I gave four AIs the same job: speedrun all eight maps of Doom's first episode in two hours, with no help from me and no input faster than a human record holder ever pressed. Only one of them finished every map. None of them beat a human record. The contestants, each at its highest effort, same machine, same files, one message: · Claude Opus 5.5 (max) in Claude Code · GPT-6 Astra (ultra) in Codex · Gemini 4 Argon (high) in Antigravity · Kimi K3.1 Pro (max) in Kimi Code Results (official checker): · GPT-6 Astra: 8/8 maps, 11:02.10 in total · Claude Opus: 3/8 maps, fastest on every one it finished (E1M1 10.06, E1M2 30.80, E1M4 13.43) · Kimi K3.1 Pro: 1/8 (E1M1 19.77) · Gemini 4 Argon: 0/8. Three sessions, and each one ended with "This request was blocked by Gemini's filters" while it was disassembling the engine. · The eight human records add up to 4:39.00. How we tested: · Shareware DOOM1 v1.9 (freely distributable), DSDA-Doom 0.29.4 in Doom 1.9 compatibility, Ultra-Violence, one demo per map from a pistol start. · Input limits measured from the human record demos themselves: at most 64.7 degrees of turning per tic, at most 6 new fire or use presses per second. · One checker replays every demo in the engine and counts it only if the player reaches the exit alive. Rules, checker and game were locked with SHA-256 before the start. · Two hours each, 4 CPU cores each. After the race, each AI checked another one's demos blind, without our checker: 15 out of 15 verdicts matched. Chapters: 0:00 Map one, four AIs 0:28 The idea 0:49 Doom and demo files 1:11 No superhuman inputs 1:36 One checker for everyone 1:55 The contestants 2:14 How they played 2:42 Gemini: blocked three times 3:04 The two hours 3:24 E1M1 Hangar 3:46 E1M2 Nuclear Plant 4:23 E1M3 Toxin Refinery 4:53 E1M4 Command Control 5:15 E1M5 Phobos Lab 5:42 E1M6 Central Processing 6:11 E1M7 Computer Station 6:41 E1M8 Phobos Anomaly 7:09 Blind cross-checks 7:45 Is this a TAS? 8:03 By the numbers 8:23 Final standings The exact brief every AI got is in the pinned comment. Every demo, the checker, the rules, the cross-checks, Gemini's three sessions and each agent's commit history: https://github.com/novaki-ai/doom-ai-... Inspired by Bears Gaming Den's Contra run:    • An AI Beat Contra's World Record by 24 Sec...   Human records: https://dsdarchive.com DSDA-Doom: https://github.com/kraflab/dsda-doom More from Novaki:    / @novaki_ai   Subscribe:    / @novaki_ai   The narration uses an AI voice (Google Gemini). Website: https://novaki.lol X: https://x.com/Novaki_ai Instagram:   / novaki_ai   Contact: hello@novaki.lol #AI #Doom #Speedrun