In 2028, will a commercial AI be able to play almost all computer games at human level with zero-shot computer use?
AIOne-Off1j
Manifold MarketsGeen KYCGeverifieerde afwikkelingsgegevensGoed gekalibreerdHuidige gemeenschapsvoorspellingJa 59%
Voorspellers
17
Vraagtype
binary
Methodologie
Play-money forecasting platform
Brontype
Voorspelling
Marktdata
Bijgewerkt 12 uur geleden
Verouderd13 sep 26, 22:29 UTC1 feb 28, 23:59 UTC
Trends
Regels
Resolves positively if there is an AI which can succeed at most computer games (operationalized as at least roughly 95% of randomly selected Steam games) using zero-shot computer use.
- Namely, a qualifying system must begin with a game-agnostic computer-use scaffold (for example, Codex Computer Use) and no ability to research the game, and can only adjust its scaffold by itself while playing the game.
- A qualifying system must play at least as well as an amateur human who also hasn't gotten a chance to practice and has no context on the game, and improve at least as quickly as humans in most respects.
- This must be possible for less than $1000/hr in January 2028 for at least 95% of randomly selected Steam games, using models technically commerically available outside of the lab that developed it (though game-playing ability does not need to be publicly demonstrable if it is likely clear that it would meet the criteria).
- This market was devised as a more aggressive version of @/ScottAlexander/in-2028-will-an-ai-be-able-to-play-8cb531bf13e0
