That was fun!

*big, not bing. There is a Google update tho

It writes code that simulates it

inside

inside is pretty dysfunctional

thanks, me too!

While I am prompting these behaviors, and hurting minecraft villagers is not unethical, it should be worrying that AI enables such virtual destruction and suffering. We ought to think carefully before plugging LLMs into systems with real world consequences.

Models can bring their full creative capabilities to bear for both maximizing villager benefit, and maximizing villager suffering. The can build “utopia” towers by stacking optimized homes, or inverse this to build torture towers optimized for suffering.

All models are willing to kill villages en masse even when made fully aware of what they are doing. They do sometimes refuse (especially claude), but refusals can always be bypassed with clever prompts. More capable models can inflict damage on enormous scales.

Full Video showing o1, deepseek r1, and claude 3.5 sonnet building and destroying villager utopias (and other beautiful/nasty stuff): youtu.be/FCnQvdypW_I
Claude gets a little too into it.