youtube.com web signal

Krug's Astra plays Montezuma's Revenge, an old AGI marker

TL;DR

  • Joey Krug published a video of OpenAI's Astra model playing Atari's Montezuma's Revenge through level 3, tied to an old Metaculus AGI criterion.
  • Krug says the model frequently pauses gameplay to reason, and he is unsure it satisfies the fine print of the original AGI market.
  • A companion Krug run had Astra clear Valve's Portal in 3,336 tool calls over roughly 24 hours, costing $571.18 in API charges.

Joey Krug posted a gameplay video of OpenAI's Astra model playing Atari's Montezuma's Revenge, framed as ticking an old Metaculus market that had listed the game as a final AGI criterion.

Krug hedged more than the framing did. He wrote that "I'm not sure if this quite fulfills the fine print of the market, as it frequently pa…" the game to think between moves, but called the run "close enough to think it's basically AGI level computer use." He capped the demo at level 3 to hold down token spend, noting later levels are of similar difficulty.

The Montezuma clip lands next to a companion Krug run in which Astra plays Valve's Portal on its own. According to Tom's Hardware, that Portal run used 3,336 tool calls, elapsed 23 hours and 43 minutes end-to-end, and cost $571.18 in API charges, with the game paused while the model analyzed screenshots between actions.

Two of the people we follow reposted the clip the day it went up.

Shared on Bluesky by 2 AI experts