More accurate, albeit less sensational headline:
"Blindly following instructions from an LLM would have killed me."
Not exactly shocking if you take into consideration that at the core, they're simply number predictors.
"Blindly following instructions from an LLM would have killed me."
Not exactly shocking if you take into consideration that at the core, they're simply number predictors.
Things like meat people might be more skeptical about, but imo this goes back to do Google, et al really trust their LLMs to give definitive answers on food safety. If it were me this would be one area I’d have the LLM refuse or hedge all answers like with other sensitive topics.
Wait until the whole internet will have LLM content mixed with other seemingly legit content, without advertising its LLM generated.
Normal LLMs are number predictors, yes. But Google Gemini is not a normal LLM: is a lobotomized model of unknown training dataset for supposedly moral-educational filters, from which information about poisonous substances has been cut out.
Specific people are liable for pushing hallucinating word-generator into Google Search. Specific people are liable for censoring this "model". And the fact that the responsibility for this censorship shifts to the end-users plays very much into their hands.
Upd: I provided examples of different Gemini responses depending on version/censorship settings in https://news.ycombinator.com/item?id=40728686