638
Remember how ChatGPT totally aced the bar exam? Wow! yeah, turns out that was just a lie
(www.nytimes.com)
this post was submitted on 15 May 2024
638 points (98.9% liked)
TechTakes
1483 readers
285 users here now
Big brain tech dude got yet another clueless take over at HackerNews etc? Here's the place to vent. Orange site, VC foolishness, all welcome.
This is not debate club. Unless it’s amusing debate.
For actually-good tech, you want our NotAwfulTech community
founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
It's more like taking a digital copy into the test room with you and Ctrl+F'ing every question/answer.
Except it’s not, because they can’t perfectly recall everything.
It’s more like reading every book in the world, and someone asking you what comes next after “And I…”.
"will alwaaays love you...."
Easy. No other answer.
and in conclusion an AI is very like an elephant, particularly the back end
But the AI isn't "recalling" in the same way you do, it doesn't "remember" what it "read", it "reads" on demand and has instant access to essentially all of the information ~~available online~~ it was trained on (E: though it's becoming more or less the same thing, and is definitely the same when it comes to law books for example), from which it collects the necessary details if and when it needs it.
So yes, it is literally "sat" there with all the books open in front of it, and the ability to pinpoint a bit of information in any one of all the books in milliseconds.
It doesn’t read on demand, it reads once when it’s being trained, and it later recalls what it learnt from that training.
Training LLMs takes a very long time and a lot of hardware power.
If it doesn't read it on demand, how does it sometimes spill its training data verbatim then?
The trained model shouldn't have that, right? But it does?
https://m.slashdot.org/story/422185
These models have so many parameters that, while insufficient to memorize all text it has ever seen, it can end up memorizing some of the content. It is the difference between being able to recall a random passage versus recalling the exact thing you need. Both allow you to spill content verbatim, but one is problematic while the other can be helpful.
There are techniques to allow it it 'read on demand', but they are not part of the core model (i.e. the autocmpletion model / LLM) and are tacked on top of it. For example, you can tie it search engine, which Microsoft's copilot does, and is something which I don't think is enabled for ChatGPT by default. Or allow it to query a external data bank (Retrieval Augmented Generation).
Do you read a song on demand when you are singing the lyrics verbatim?
Yes, it does, from the information it was trained on (or - stored), which like you say, requires a lot of hardware power so it can be accessed on demand. It isn't just manifesting the information out of thin air, and it definitely doesn't "remember" in the same way we do (E: even the best photographic memory isn't the same as an indexable one).
It's definitely not indexed, we use RAG architectures to add indexing to data stores that we want the model to have direct access to, the relevant information is injected directly in the context (prompt). This can somewhat be equated to short term memory
The rest of the information is approximated in the weights of the neural network which gives the model general knowledge and intuition..akin to long term memory
or it can be equated to a shitty database and lossy compression (with artifacts in the form of “hallucinations”), but that doesn’t make the tech sound particularly smart, does it?
but half the posts in your history are in this thread and that’s too many already
Ok bud
People have such crazy misconceptions about AI. Glad to see someone else knows how it works at least.
oh do fuck off
awww, I just got another bowl of popcorn!
but rofl holy shit at "glad to see someone else knows how they work" given the ..... depth of understanding, shall we say? that was demonstrated in this thread
I mean you still gotta understand some shit for Ctrl+F to be helpful. If you've ever taken an open book quiz without prior study you'll learn pretty quick that open book does NOT = easy A (depending on the class / prof I guess, but you get the gist).
So, open book Ctrl-F'able bar exam, I could probably get an okay score just on key word matching, not knowing jack shit about law; but it'd be far from a perfect score. Current state of machine learning appears to be in a comparable boat.
This is a computer. Time (in this aspect) isn't an issue.
your post shows a serious lack of comprehension. just because so many of the posters in this thread are idiots didn't mean you have to participate too.
(CPU time extremely counts, and resource-wise with these things it's really quite a lot)
Steelmanning what this person said, I think the issue is that your ability to CTRL+F through a book during a time-limited exam is not as strong as even a single computer clocked at GHz doing the same thing. You can CTRL+F through a single book in the same time it takes it to CTRL+F through the entire body of knowledge.
I'm not a big AI guy but it's really not quite like that, models do NOT contain all the data they were trained on.
Edit: I have no idea what's going on down below this comment
we can tell
The guy above you is right though. So what are you on about?
what a weird opportunity for someone to burn a throwaway account. not even gonna dig into what you’ve imagined the other guy is right about, given he didn’t post any information of value
Lemmy is starting to have a huge problem with people creating multiple sockpuppets--probably programmatically generating them, in fact--just to win internet arguments. If this goes on too long you're going to see a really surprising number of sudden downvotes on everything you've said in this conversation, and anyone who agreed with you.
It’ll actually be good if that happens because 1) it could be dealt with here, 2) that could be used to feed something that helps defend against it
(e: I mean, it’ll be a nuisance, and @self would be driven to drink, but it could still be handled)
we already ignore offsite downvotes
aha, good to know too :D
I truly wish those so inclined good luck with the downvotes and with coordinating their sockpuppets in a way that isn’t extremely fucking obvious
They’ll use AI! /s
you need to work on your reading comprehension skills, they're letting you down on your left-field brigading
[:popcorn intensifies:]
that's a misleading and meaningless way of putting it. if I rip a page out of my textbook and bring it into an exam room, I do not have with me all the data in my textbook. and yet
It doesn’t do that, either. LLMs retain the linguistic patterns found in textbooks, nothing more. It’s remarkable that they can do so much with this information alone, but it’s still a far cry from genuine intelligence.
And yet they can spit out copyrighted material verbatim, or near-verbatim, how strange and peculiar.
so weird!!!!!
Yeah, even setting aside the intelligence claims, I know I'd be feeling a lot more positive about LLMs as a fun theoretical tool if they weren't being sold as personal assistants or search engine replacements etc, which even the apologists here admit they're really really bad at.
(Also I'd argue "linguistic patterns" is pushing it. "Textual patterns" more like, it's not supposed to have any idea about grammar or even about what "text" is.) (I say "supposed to" because who knows what sort of hacks they're running under the hood.)
Yeah I completely agree with you there. I really don’t like the way AI is being monetized and commercialized; it all just seems poised to go terribly. Ideally there shouldn’t be any incentive to overstate the capabilities of these models, as doing so just makes everything worse.
lol. at least you're honest about it