You didn’t “have a conversation” with ChatGPT.
It doesn’t “think you should…” It doesn’t think.
It didn’t “tell you that…” It doesn’t speak.
It doesn’t “feel that the best option is…” It doesn’t feel.
AI is a cheap parlor trick. You provide words, and it provides words back that are most likely to occur alongside the words you provided.
…
A useful reminder for the next time you’re tempted to personify or humanise an LLM.
LLMs are statistical tools. There are some things that the statistics of language can be good at, especially on average: stuff like summarisation, sentiment analysis, pattern identification, and checking for internal consistency.
But they’re just maths. They’re not a person.
It’s not even that they don’t care about you or don’t want to help you. They don’t even go that far: they’re incapable of “caring” or “wanting” in the first place. What they do is take all of the information they’ve ingested, plus their training and prompt, plus the conversation you’d had with them so far, plus a random number, and produce output which is, after a fashion, a prediction of what comes next.
As always: that’s not to say it’s useless. (It’s also not to say it’s always useful.) But as a tool, it’s pretty opaque to most normal people.
Unless you’ve really taken a deep-dive into understanding low LLMs work, they must seem like magic (hell; speaking as somebody who has taken such a deep-dive, they sometimes seem like magic!). I’m sure that some of the time, they must seem like they’re a living thing, or at least an approximation of one.
But they’re not. And it’s important to remember that.
0 comments