here's a take i don't think i heard yet:
the better LLMs are, the *more* they're dangerous. not because they can "go rogue", nothing like that. rather the better LLMs are, the likelier people are to trust them.
you'll never use an LLM that's wrong half of the time as a search engine, but one that's only wrong 10% of the time might be good enough.
you'll never use an LLM that's wrong 10% of the time to automate your job, but one that's only wrong 1% of the time might be good enough.
you'll never hook an LLM that's wrong 1% of the time to a nuclear weapon. but one that's only wrong 1‰ of the time might be good enough.
stochastic parrots don't get less dangerous the more their output aligns with reality. the opposite is true: since the biggest danger a stochastic parrot poses is a human trusting one to make decisions, a dangerous LLM's the one that's more convincing and less likely to be caught during testing.
in classical AI safety research there's a lot of talk about a smart AI Volkswagening during testing to make itself seem safe, with all sorts of calculations for the odds such an AI will show its true form in any given attempt and yada yada. but what that classic research didn't seem to take into account is that an "artificial intelligence" doesn't need to actually be intelligent to replicate the same behaviour. a machine that sometimes appears safe and intelligent and sometimes doesn't can mislead you just as well, without having an evil plan or even the capability to come up with one. and just like a machine that's evil needs to appear good just often enough to convince you it isn't, a machine that lacks intelligence needs to be right just often enough to convince you it doesn't.
#LLM #LLMs #AI #genAI #FuckAI #AIWWIII #OpenAI #Anthropic #AISafety #StochasticParrot
