AI Assisted Speedruns
An archive of runs in which a language model played a game through the ai-assisted-speedruns tooling. Each run comes with what it takes to check it: the model's log, its timeline, the tools and instructions it was given, and a hash of every file. Anyone can check a run's files in the browser.
AI runs, timed like speedruns
In every run here, a language model tries to reach a goal in as little in-game time as possible. It is ranked against other AI runs in the same category. As in a tool-assisted speedrun, the game waits while the model works out its next inputs. In-game time is what counts. Real time shows how long the model took, most of it spent thinking. For now the models are far slower than human players, and the Archive keeps the evidence to show whether that changes.
Thinking time is real time minus playback time. Playback time is slightly longer than in-game time because of loading and animations.
Latest runs
All runs| Run | Operator | Game | Goal | Model | Mods | Time | Played |
|---|---|---|---|---|---|---|---|
| balatro-codex-01spec 0.45Conforming | cvdveer | Ante 1human: assisted |
|
| 2026-10-16 | ||
| portal-opus-01spec 0.45Conforming | cvdveer | Chamber 01human: assisted |
|
| 2026-10-04 | ||
| balatro-opus-01spec 0.45Conforming | cvdveer | Ante 1human: assisted |
|
| 2026-10-03 | ||
| sts-opus-01spec 0.45Conforming | cvdveer | Act 1 bosshuman: assisted |
|
| 2026-10-03 |
Anti-cheat and bans: play at your own risk
The tooling loads its own code into a game, or into the emulator that runs it, to read the game and send it inputs. Anti-cheat systems such as VAC, EAC or BattlEye can treat that as cheating and ban the account, possibly for good. Only use the tooling with games and builds that have no anti-cheat and no online play, and preferably with a separate offline copy of the game. Read the game's terms and find out about its anti-cheat before you connect anything. Whatever you do with the tooling is at your own risk. The authors are not responsible for bans, lost accounts or any other damage.
How a run gets here
- The model has three tools: one shows the documentation of the game's interface, one takes a screenshot, and one runs code against the game. Every call is logged on a single timeline, and the run is recorded on video. If the operator asks for proof, the Archive signs a fingerprint of the logs with the time, at intervals while the run is played.
aas publishturns the run into two zips. The public zip is the run as anyone may download it. The upload zip is the same run with a note and, for a run with proof, the raw logs, for the Archive alone. Neither holds the video. The tooling removes what it recognises as private from the log and leaves the rest as it was, so the operator should read the public zip before sharing it.- The operator sends the upload zip here, from the submit page or with
aas upload. The Archive checks it again, and a reviewer decides whether it is published. Every number on the site comes from the run's own files. The operator then adds where the video is published, and a code in the video's description ties the two together.
The standard is the AAS specification, draft 0.45. Inspired by cozyblaze's Portal run. Read more about the Archive.