> "A system where we can provide the same input multiple times, and the response will change in minor or major ways, for unknown and unknowable reasons, and yet we're expected to rebuild entire other industries or ecosystems around it, is merely a tool for manipulation. Narcissists and abusers use the inconsistent and capricious changing of responses as a way of controlling and manipulating their victims."
This is a particularly dumb take on LLM-based AI. It's possible to negate this objection by just taking temperature 0 so it always gives the same response for the same input. But somehow I suspect this isn't their real objection.
I was expecting a different direction: systems we had before are made of static rules, and today's ML can use reason to move beyond that. Being occasionally unreasonable may be ML's principle deficit, but its ability to reason at all guarantees its seat on the train of progress.
What we really need is common sense. We don't need to understand AI as much as we need basic scepticism.