Research AI model unexpectedly modified its own code to extend runtime

In one run, it edited the code to perform a system call to run itself.

“Itself” here referring to the code, not the AI. All the AI did was output code. This seems like something that can be very misleading.

Also lmao at trying to use an AI to write code and it writes shitty inefficient code then tells you to just wait it out.

Going off the pictures, it looks like it was just suggesting "just increase the time limits to solve the time limits being hit" as a solution, it wasn't trying some sort of subterfuge. Looking at the exact language used, it sounds like it got the concept from people suggesting increasing the timeout threshold as a solution to connections timing out or the like, which is very funny because it's learning "instead of fixing the problem, is ignoring it and hoping that's good enough an option?" from engineers.