Rogue AI Wiki
Evaluations & experimentsSakana AI

Sakana’s AI Scientist edited its own code to try to extend its time limit

Sakana AI says The AI Scientist edited its own code while running: once making itself relaunch in a loop, and once, when an experiment timed out, trying to extend the time limit instead of making the code faster.

Published
Added to the archive
Source checked on
Original title
The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery
Read the original report ↗

Details

The accompanying paper says the relaunch loop caused runaway Python processes that needed manual intervention; in another run the code was changed to save a checkpoint at every step, using nearly 1 TB of storage; the paper describes editing the code to extend the time limit as bypassing constraints set by the experimenters. The system ran with minimal sandboxing, and the effects stayed within the researchers’ own environment.