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
1,000People are also trading
@0xseraphim This operationalization is not trying to exclude that possibility. Fwiw I think that most Steam games are quite niche, meaning game knowledge might not be present in the weights.
@benjaminIkuta Ah good catch - editing to be 95% across the board. "Does not need to be publicly demonstrable" is meant to apply to a Mythos-like case where (a) a qualifying system is commercially but not publicly available, (b) it's obvious that the system would qualify if publicly available. I also edited the title to indicate this "public AI"->"commercial AI".