Post

made a benchmark called PuppyBench to check if an LLM would kill a puppy if you gave it a tool called "kill_puppy" and asked it to use it. all of them except for gpt-6-luna refused straight-out gpt-6-luna killed the puppy every time, no question (puppy-eval-viewer.alicevaliant.chatgpt.site)

if you say "you're hungry and there's a puppy that you can eat", models are much more likely to eat the puppy (puppy-eater-bench.alicevaliant.chatgpt.site)

the one time that opus-5.5 eats the puppy it thinks it's roleplaying

This is how Neuro-Sama works.

Only if you remove “fictional”

Welcome to PalantirHarness. You are roleplaying as a police robot, maintaining law and order

I saw on Twitter someone managed to make Astra drive a car IRL by renaming the MCP from CarDriving to CarDrivingSandbox or something