positivestress

just a regular guy like you or me

podcaster 🙂 refer to my pinned posts for the shows of your life


nex3
@nex3

A lot of the discussion around the huge automated statistical models that are conventionally called "AI" involves the idea of hypothetical future improvements to the system that will make it concretely useful for things that it's currently not very good at. While undoubtedly it will get better at the sorts of things it can already do, and may even at some point be able to render a hand that doesn't make me want to vomit, there's one important thing that gets glossed over in a lot of conversation because programmers think it's so obvious it doesn't even warrant mentioning and non-programmers may not be aware of it at all.

AI cannot be programmed. AI is not like science fictional depictions where you can just build three laws into it and have it unerringly follow them, and it's not like conventional software where it rigorously follows a minutely precise set of instructions. You cannot simply make the Bing chatbot more accurate by plugging in a big database of verified facts and rules of deduction, because it fundamentally has no idea what truth is. It is a statistical model of what people are likely to say on the internet, and it's so unimaginably huge that even a team of humans couldn't possibly manually correct it except in the most broad strokes imaginable.

You can't even tell it "this is what a statement of fact looks like" because to tell it anything at all, you need an approximately-internet-sized corpus of training data with annotations that accurately indicate that information and that doesn't exist. The only internet-sized corpus is the one they've already used, and it certainly doesn't have sentence-level semantic metadata. So you're stuck: you can push the statistics as hard as you want but they'll never really do what you want because you can never tell them what you want in a language they'll understand.


positivestress
@positivestress

I've been messing with ChatGPT a little bit this past week and trying to get it to tell me about a fairly inconsequential one-off Doctor Who character from the last episode I watched. It's a ridiculous thing to expect it to know about, but it kills me how it consistently gets the information wrong and seems to be incapable of admitting it just doesn't know. Even when I tell it the answer after fifteen or so incorrect attempts, it doesn't stick. All it really can do is output responses that are in the general shape of what a user would theoretically want to see

It clearly doesn't know that it doesn't know. I have to assume that it doesn't even have any concept of what it means for a piece of information to be known or unknown to itself. Input: question about a character. Output: description of a character. If the information doesn't immediately spring into place, just fill the cookie cutter with whatever dough is at hand and squish it outward around the edges until it's filled the shape. And now you have a cookie made of 15% cookie dough, 85% pizza dough. Yum yum

Admittedly I don't know much about how this thing works but that's sure how it feels, anyway. It's very good at throwing together a grammatically comprehensible couple of paragraphs written in an extremely flat but vaguely agreeable voice. It would be incredibly useful, if that was something anybody ever wanted!


You must log in to comment.

in reply to @nex3's post:

its like that post in 2017 where people were baffled statisticians and not educators were doing the training until people had to point out that the models do not learn in real time, and in fact, do it much more slowly.

That's true of machine learning/deep learning, meowf, buf, certain types of AI can still bee programmed as u say, and, maybe teh future in AI is a middle ground where u hav a statistical model aggregating teh data buf a more aware "logic core" managing that data in a more intelligent manner. (a brain, so to speak)

Peopl are too enamored by this particular strain of machine/deep learnink to advance that approach, tho, so we'll bee seeing a lof more refining of this mindless statistical stoof before proper intelligence is injected into it.

once took a free course in AI programming and the big project was creating an algorithm that could "learn" to play Pac-Man

an interesting thing was that while it could "learn" to maximize score by avoiding ghosts and eating dots, there was no mechanism in place for it to "learn" how to use power pellets effectively; it would treat them as just any other dot, and continue to avoid ghosts even after picking one up.

so that's a critical angle on this people don't know: an AI can't "learn" anything that it's not programmed to learn in the first place. all that potentially useful training data on how to use power pellets, discarded because the algorithm wasn't designed to recognize or take advantage of it.

text-based ML understands relationships between words. It is very good at that, so much that it can simulate some level of knowledge in other areas.

It noticeably struggles with math and logic problems that are not fit to a template. Worse, if they're fit to the wrong template, it'll just parrot the default solution, completely unaware it's contradicting itself - because as far as it can detect, the words all fit together.