Simple guardrails would’ve prevented that, especially in a controlled environment like this where they are pitting LLM-created bots against each other. It’s the same with all the other “hacking” news. Of course an LLM that you allow to roam free and go wild will find a way to get the best results, whatever it takes.

Also, I wonder why they specifically chose Protoss for the LLMs to use.

It always amaze me how people train a random number generator to perform a task, and then go surprised Pikachu face as the random number generator proceeds to attempt random things to get a better result at that task.

It’s like all of the times an LLM “escaped”, no shit Sherlock, you put up a random machine into a box and asked it to guess what’s outside, how long until it randomly tries the strategy of peaking outside the box?