(This is a lightly modified multiple-choice version of this 2024 market, with slightly simplified criteria.)
This question resolves to YES if an AI agent has learned to speedrun at least one popular category (≥30 unique runners on speedrun.com or another leaderboard) of any video game, and has finished at least one run with a better time than any human speedrunner at the time.
Native PC or emulated console games are both fine.
Criteria for resolution:
The AI must be capable of speedrunning the game in real time and learn to do so without directly learning from human inputs (e.g. learning from Youtube videos is fine). A traditional TAS (tool-assisted speedrun) does not count.
The AI must not receive any information about the game that a human speedrunner wouldn’t be able to know during a run, e.g. watching RAM values while playing. Ideally the AI should only receive game’s pixels (possibly downscaled or otherwise processed) and maybe audio as input, but this is not a strict requirement.
The AI should ideally follow all the rules of the specific game and category it is running, as listed on the game’s speedrun.com page (or elsewhere). However, if there are minor rule breaks and the AI’s run is obviously much more impressive than the most comparable human WR, I may choose to ignore this requirement.
It must be a full-game, non-segmented run, not an IL (individual level) speedrun.
The human world record in the category must be over 3 minutes long. Very short speedruns don’t count.
I will use my best judgment to resolve this based on the criteria above and will not bet myself.