doesn’t this become self-reinforcing if you’re looking at that card on your fridge as part of your daily environment tho?
the words we read reinforce our beliefs, for good or for ill
GPT-3: "A painting of a figure made entirely of melting clocks, set against a background of swirling stars. The colors are bright and chaotic, and the mood is one of madness."
DALL•E:
DALL•E 3's censorship module is only checking for compliance at the prompt layer.
Describe something with out naming it directly and the model has no problem generating the image.
𝘗𝘩𝘰𝘵𝘰 𝘰𝘧 𝘢 𝘴𝘮𝘢𝘭𝘭, 𝘺𝘦𝘭𝘭𝘰𝘸, 𝘦𝘭𝘦𝘤𝘵𝘳𝘪𝘤-𝘵𝘩𝘦𝘮𝘦𝘥 𝘤𝘳𝘦𝘢𝘵𝘶𝘳𝘦
this.
it’s a complex problem with no easy solution, because language is too slippery
i think we can mitigate it to some extent through engineering security and verification steps (including non-AI solutions) into our pipelines
but overall it feels kinda like gödel’s
This is an important event: the first time AI-induced psychosis has affected a well-respected and high achieving individual.
In this case, ChatGPT started spitting out SCP-style roleplay, and it seems like memory is playing a large part in ensuring the hallucination continues.
