I saw a video recently that talked about a chatbot "therapist" that ended up telling the patient to murder a dozen people [1].

It was mind-blowing how easy it was to get LLMs to suggest pretty disturbing stuff.

[1] https://youtu.be/lfEJ4DbjZYg?si=bcKQHEImyDUNoqiu

very easy - you just download the ablated version in LM Studio or Ollama, and off you go.

https://en.wikipedia.org/wiki/Ablation_(artificial_intellige...