AI Psychosis Research
Credit where credit is due. While I may have criticized places like LessWrong in the past for a lack of rigor and empirical evidence for their ideas, this is a well written and solidly researched article on what LLMs play into AI psychosis and which don’t.
What was most interesting to me is that the Chinese labs have both the most and least insane models. DeepSeek v3 is even more nuts than GPT-4o and offers no pushback, telling the user at one point to jump off a cliff to escape the simulation.
On the other end is Kimi K2 which has a zero bullshit personality. It doesn’t go along with a psychotic user at all and says the idea is crap immediately rather than glazing the user. After seeing this I’m definitely going to be using this model more when I need feedback on an idea or writing.