Astra reportedly reaches 69 gems in MazeBench with source-code access
The user says access to the maze’s source code removed the visual component, making the setup substantially easier than the official test. They report reaching 69 gems with Astra after a few more days on one CPU with 16GB of memory.
TLDR
A user reports experimenting with Codex, Sol and Astra on MazeBench using solvers for individual rooms and a global solver to link the solutions together. The run began with 5.6 Sol, which stalled at 44 gems, before continuing with Astra. The user then asked the agents to write custom solvers tailored to particular rooms’ designs and geometry. They say Astra reached 69 gems after a few more days, with resources limited to one CPU and 16GB of memory. Crucially, the user says source-code access eliminated the visual component and made this setup substantially easier than the official test.
Combined views
14.9K
1 Source, first seen 19d ago