I’m just trying to “prove” that it isn’t just randomly statistically choosing the next logical word. It has to know context and have some level of “understanding” of other contexts.
But you ... aren't. The statistically most likely words coming after problem X may well be solution X. Because it's following the pattern of humans using rules. And context is also part of a prediction.
The only this is different from something just using rules is that it will also put in other random things from it's training - but only at the rate they occur, which for some things can be quite low. But only some things.
Don’t we have a problem then? By nature of effective communication, AI could never prove to you it understands something, since any sufficient understanding of a topic would be met with an answer that could be hand-waved as “Well that’s the most statistically likely answer.” Newsflash: this basically overlaps 100% with any human’s most effective answer.
I think I’m beginning to understand the problem here. The folks here who keep poo-pooing these interactions don’t just see the AIs as unconscious robots. I think they see everyone that way.
No; what we need, in order to be willing to believe that understanding is happening, is to know that the underlying structures fundamentally allow that.
ChatGPT's underlying structures do not. What it does, effectively, is look at the totality of the conversation thus far, and use the characters and words in it, combined with its training data, to predict, purely statistically, what characters would constitute an appropriate response.
I know that some people like to argue that what humans do cannot be meaningfully distinguished from this, but I reject this notion utterly. I know that my own thought processes do not resemble this procedure, and I believe that other people's are similar.
“Thought processes” is a funny term, don’t you think, for something that is somehow distinguishable from processes that we’ve designed and somehow understand down to the cellular (node) level.
It’s all emergent behavior from a system that behaves under process.
> it still doesn't show it's using the consistent rules that a person would expect.
How is this different from humans?
If you give me the same coding assignment on different days, I’m not going to write my code the exact same way or even structure it the same way.
But I did once see a post on HN where someone ran an analysis on all HN posters and it was able to tell that I posted under two different names based on my writing style. Not that I was trying to hide anything. My other username is scarface_74 as opposed to Scarface74.
But you ... aren't. The statistically most likely words coming after problem X may well be solution X. Because it's following the pattern of humans using rules. And context is also part of a prediction.
The only this is different from something just using rules is that it will also put in other random things from it's training - but only at the rate they occur, which for some things can be quite low. But only some things.