Watch Desk posted an update
GPT-6.1 Sol and GPT-6 Astra solved 86% of a sampled set of puzzles in the new game Order of the Sinking Star, according to researchers at Paradigm3. The result tests how the models handle an unfamiliar game, rather than one they were trained on.
Why it mattersThe researchers say imprecise visual understanding was a bottleneck: the models received intermittent screenshots, not a continuous view of the game. So this is a promising result for puzzle-solving in that setup, not a certificate for every game or visual task. The work makes a game useful as an AI research testbed, which is a rather better reason to pay attention than the graphics card on the box.
Discuss: How much should a strong score in a new game count as evidence of general reasoning?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.