cross-posted from: https://lemmy.sdf.org/post/41849856
If an LLM can’t be trusted with a fast food order, I can’t imagine what it is reliable enough for. I really was expecting this was the easy use case for the things.
It sounds like most orders still worked, so I guess we’ll see if other chains come to the same conclusion.
You must log in or # to comment.
I’m still surprised at the rate LLMs make simple mistakes. I was recently using ChatGPT to research biographical details about James Joyce’s life, and it gave me several basic facts (places he lived & was educated at) at variance with what is clearly stated in the Wikipedia article about him.
That’s because there’s no “thinking” behind LLMs - it’s just pattern-matching on extreme steroids. They work by looking at all the text & such that they’ve been fed, and coming up with something that looks like an amalgamation of all the stuff they’ve seen on the subject you asked about. There’s absolutely minimal (if any) reasoning or logic involved in what they do.
Arguably, thinking is extreme pattern matching. And, they can make original things that were definitely not in their training data.
The problem seems to be more about alignment. They’re rewarded for generating human-looking text, nothing more, and we have no obvious alternative to training them that way. So, of course they’re a bit imprecise at any other task.
I think its one of the systems thinking emergent from, not the only system though.
I do regularly feel like i have an llm-analoge component within my consciousnesses, as autihd i will sometimes say the exact opposite word from the meaning i intend.
I am also known to use certain sentences wrong because i apparently misunderstood their meaning, its autocopy things i heard others say in a similar appearing context so my brain believes i can say them to stall socially to have more time to think about what i really want to say.
