"in attempting, eventually, to make its freewheeling platform safer, [Character AI] also made it unusable and deeply frustrating for much of its userbase — suggesting that the qualities that allegedly made its platform so dangerous to some young users were the same features that made it compelling." https://futurism.com/artificial-intelligence/google-character-ai-bankruptcy
this gets to a thing I've been saying for years now: you can absolutely turn the temperature up on an LLM to make the output more "creative" or interesting, but that's also what makes it more likely to coach someone into AI psychosis and suicide.
either you have a "safe" customer service chatbot that's been so lobotomized you might as well just link people to the FAQ; or you have an engaging, human-like chatbot that occasionally calls customers the n-word.
there's really no way to square that circle, and it's why, now **years** after the launch of ChatGPT, public-facing chatbots are still basically indistinguishable from previous automated help chats.
it's also why models that are good at math or coding or whatever are worse at other things. machine learning models have **tradeoffs**. you can do reinforcement training to emphasize a particular feature, but that often means sacrificing performance on other features.
Oct 14, 2026, 8:38 AM@peter I have been saying this for years and not only for curse words
You can also either have a proactive "hands-off" personal assistant that books flights unsupervised. Until one day you are stranded at a Santiago de Chile border checkpoint because it thought you don''t need a visa at Santiago de Compostela
Or you can have a "safe" human-in-the-loop travel planner agent that annoys you every five minutes demanding to be micromanaged
There is no way to large language out of a judgement call
@peter you’re describing heroin