You know how Google’s new feature called AI Overviews is prone to spitting out wildly incorrect answers to search queries? In one instance, AI Overviews told a user to use glue on pizza to make sure the cheese won’t slide off (pssst…please don’t do this.)
Well, according to an interview at The Vergewith Google CEO Sundar Pichai published earlier this week, just before criticism of the outputs really took off, these “hallucinations” are an “inherent feature” of AI large language models (LLM), which is what drives AI Overviews, and this feature “is still an unsolved problem.”
They keep saying it’s impossible, when the truth is it’s just expensive.
That’s why they wont do it.
You could only train AI with good sources (scientific literature, not social media) and then pay experts to talk with the AI for long periods of time, giving feedback directly to the AI.
Essentially, if you want a smart AI you need to send it to college, not drop it off at the mall unsupervised for 22 years and hope for the best when you pick it back up.
No he’s right that it’s unsolved. Humans aren’t great at reliably knowing truth from fiction too. If you’ve ever been in a highly active comment section you’ll notice certain “hallucinations” developing, usually because someone came along and sounded confident and everyone just believed them.
We don’t even know how to get full people to do this, so how does a fancy markov chain do it? It can’t. I don’t think you solve this problem without AGI, and that’s something AI evangelists don’t want to think about because then the conversation changes significantly. They’re in this for the hype bubble, not the ethical implications.
We do know. It’s called critical thinking education. This is why we send people to college. Of course there are highly educated morons, but we are edging bets. This is why the dismantling or coopting of education is the first thing every single authoritarian does. It makes it easier to manipulate masses.
“Edging bets” sounds like a fun game, but I think you mean “hedging bets”, in which case you’re admitting we can’t actually do this reliably with people.
And we certainly can’t do that with an LLM, which doesn’t actually think.
Jinx! You owe me an edge sesh!
A big problem with that is that I’ve noticed your username.
I wouldn’t even do that with Reagan’s fresh corpse.
I think that’s more a function of the fact that it’s difficult to verify that every one of the over 1M college graduates each year isn’t a “moron” (someone very bad about believing things other people made up). I think it would be possible to ensure a person has these critical thinking skills with a concerted effort.
The people you’re calling “morons” are orders of magnitude more sophisticated in their thinking than even the most powerful modern AI. Almost every single one of them can easily spot what’s wrong with AI hallucinations, even if you consider them “morons”. And also, by saying you have to filter out the “morons”, you’re still admitting that a lot of whole real assed people are still not reliably able to sort fact from fiction regardless of your education method.
No I still agree that we are far from LLMs being ‘thinking’ enough to be anywhere near this. But if we had a bunch of models similar to LLMs that could actually think, or if we really needed to select a person, I do think it would be possible to evaluate a bunch of the models/people to determine which ones are good at distinguishing fake information.
All I’m saying is I don’t think the limitation is actually our ability to select for capability in distinguishing fake information, I think the only limitation is fundamental to how current LLMs work.
Yes, my point wasn’t that it could never be achieved but that LLMs are in a completely different category, which we agree on I think. I was comparing them to humans who have trouble with critical thinking but can easily spot AI’s hallucinations to illustrate the vast gulf.
In both cases I think there are almost certainly more barriers in the way than an education. The quest for a truthful AI will be as contentious as the quest for truth in humans, meaning all the same claim-counterclaim culture-war propaganda tug of war will happen, which I think is the main reason for people being miseducated against critical thinking. In a vacuum it might be a simple technical and educational challenge, but the reason this is a problem in the first place is that we don’t exist in a political vacuum.
Choose a lane, this comment directly contradicts you previous comment. I think you are just trolling and being an idiot with corrections to elicit reactions.
You need to be specific and say what the contradiction is, I don’t see it.
What does this have to do with AI and with what OP said? Their point was obviously about limitations of the software, not some lament about critical thinking
Yeah, I mean, we have that, and parents are constantly trying to dismantle it. No amount of “critical thinking education” can undo decades of brainwashing from parents and local culture.
Removed by mod
You’re exactly right. There is a similar debate about automated cars. A lot of people want them off the roads until they are perfect, when the bar should be “until they are safer than humans,” and human drivers are fucking awful.
Perhaps for AI the standard should be “more reliable than social media for finding answers” and we all know social media is fucking awful.
The problem with these hallucinated answers that makes them such a sensational story is that they are obviously wrong to virtually anyone. Your uncle on facebook who thinks the earth is flat immediately knows not to put glue on pizza. It’s obvious. The same way It’s obvious when hands are wrong in an image or someone’s hair is also the background foliage. We know why that’s wrong; the machine can’t know anything.
Similarly, as “bad” as human drivers are we don’t get flummoxed because you put a traffic cone on the hood, and we don’t just drive into tue sides of trucks because they have sky blue liveries. We don’t just plow through pedestrians because we decided the person that is clearly standing there just didn’t matter. Or at least, that’s a distinct aberration.
Driving is a constant stream of judgement calls, and humans can make those calls because they understand that a human is more important than a traffic cone. An autonomous system cannot understand that distinction. This kind of problem crops up all the time, and it’s why there is currently no such thing as an unsupervised autonomous vehicle system. Even Waymo is just doing a trick with remote supervision.
Despite the promises of “lower rates of crashes”, we haven’t actually seen that happen, and there’s no indication that they’re really getting better.
Sorry but if your takeaway from the idea that even humans aren’t great at this task is that AI is getting close then I think you need to re-read some of the batshit insane things it’s saying. It is on an entirely different level of wrong.
A fair perspective.
I let you in on a secret: scientific literature has its fair share of bullshit too. The issue is, it is much harder to figure out its bullshit. Unless its the most blatant horseshit you’ve scientifically ever seen. So while it absolutely makes sense to say, let’s just train these on good sources, there is no source that is just that. Of course it is still better to do it like that than as they do it now.
Google AI suggested you put glue on your pizza because a troll said it on Reddit once…
Not all scientific literature is perfect. Which is one of the many factors that will stay make my plan expensive and time consuming.
You can’t throw a toddler in a library and expect them to come out knowing everything in all the books.
AI needs that guided teaching too.
Genuine question: do you know that’s what happened? This type of implementation can suggest things like this without it having to be in the training data in that format.
In this case, it seems pretty likely. We know Google paid Reddit to train on their data, and the result used the exact same measurement from this comment suggesting putting Elmer’s glue in the pizza:
https://old.reddit.com/r/Pizza/comments/1a19s0/my_cheese_slides_off_the_pizza_too_easily/
And their deal with Reddit: https://www.cbsnews.com/news/google-reddit-60-million-deal-ai-training/
It’s going to be hilarious to see these companies eventually abandon Reddit because it’s giving them awful results, and then they’re completely fucked
You’re wrong. Anyone who has ever used Google knows Reddit is an absolute goldmine of valuable information. The problem is it’s also full of jokes and puns and bad information, and AI isn’t able to sort one from the other (yet).
This doesn’t mean that there are reddit comments suggesting putting glue on pizza or even eating glue. It just means that the implementation of Google’s LLM is half baked and built it’s model in a weird way.
I literally linked you to the Reddit comment, and pointed out that Google’s response used the same measurements as the comment
Are you an LLM?
Oh, hah sorry! thanks, I didn’t realise that the reddit link pointed to the glue thing
Yes
“Most published journal articles are horseshit, so I guess we should be okay with this too.”
No, it’s simply contradicting the claim that it is possible.
We literally don’t know how to fix it. We can put on bandaids, like training on “better” data and fine-tune it to say “I don’t know” half the time. But the fundamental problem is simply not solved yet.
I’m addition to the other comment, I’ll add that just because you train the AI on good and correct sources of information, it still doesn’t necessarily mean that it will give you a correct answer all the time. It’s more likely, but not ensured.
Yes, thank you! I think this should be written in capitals somewhere so that people could understand it quicker. The answers are not wrong or right on purpose. LLMs don’t have any way of distinguishing between the two.
no, the truth is it’s impossible even then. If the result involves randomness at its most fundamental level, then it’s not reliable whatever you do.
Sure, the AI is never going to understand what it’s doing or why, but training it on better datasets certain WILL improve the results.
Garbage in, garbage out.
You can train an LLM on the best possible set of data without a single false statement and it will still hallucinate. And there’s nothing to be done against that.
Without understanding of the context everything can be true or false.
“The acceleration due to gravity is equal to 9.81m/s2” True or False?
LLM basically works like this: given the previous words written and their order, the most probable next word of the sentence is this one.
Well yes, I’ve seen those examples of ChatGPT citing scientific research papers that turned out to be completely made up, but at least it seems to be a step up from straight up shitposting, which is what you get when you train it on a dataset full of shitposts.
Well it’s definitely true that you will have hard times getting true things from garbage. But funny enough, the model might hallucinate true things:)
The problem is that given the way they combine things is determine by probability, even training it with the greatest bestest of data, the LLM is still going to halucinate because it’s combining multiple sources word by word (roughly) guided only by probabilities derived from language, not logic.
Yes, I understand that. But I’m fairly certain the quality of the data will still have a massive influence over how much and how egregiously that happens.
Basically, what I’m saying is, training your AI on a corpus on shitposts instead of factual information seems like a good way to increase the frequency and magnitude of such hallucinations.
Yeah, true.
If you train you LLM on exclusivelly Nazi literature (to pick a wild example) don’t expect it to by chance end up making points similar to Marx’s Das Kapital.
(Personally I think what might be really funny - in the sense of laughter inducing - would be to purposefull train an LLM exclusivelly on a specific kind of weird material).
Yeah, I mean that’s basically what GPT4Chan did, which someone else already mentioned ITT.
Basically, this guy took a dataset of several gigabytes worth of archived posts from /pol/ and trained a model on that, then hooked it up to a chatbot and let it loose on the board. You can see the results in this video.
That was hilarious!
Thanks for the link.
Here is an alternative Piped link(s):
in this video
Piped is a privacy-respecting open-source alternative frontend to YouTube.
I’m open-source; check me out at GitHub.
That’s just not how LLMs work, bud. It doesn’t have understanding to improve, it just munges the most likely word next in line. It, as a technology, won’t advance past that level of accuracy until it’s a completely different approach.
Or you could just not use LLMs for this.
They could also perform some additional iterations with other models on the result to verify it, or even to enrich it; but we come back to the issue of costs.
Also once you start to get AI that reflects on its own information for truthfulness, where does that lead? Ultimately to determine truth you need to engage with the meaning of the words, and the process inherently involves a process of self-awareness. I would say you’re talking about treaching the AI to understand context, and there is no predefined limit to the layers of context needed to understand the truthfulness of even basic concepts.
An AI that is aware of its own behaviour and is able to explore context as far as required to answer questions about truth, which would need that exploration precached in some sort of memory to reduce the overhead of doing this from first principles every time? I think you’re talking about a mind; a person.
I think this might be a fundamental barrier, which I would call the “context barrier”.
A new religion
I think you’re right that with sufficient curation and highly structured monitoring and feedback, these problems could be much improved.
I just think that to prepare an AI, in such a way, to answer any question reliably and usefully would require more human resources than there are elementary particles in the universe. We would be better off connecting live college educated human operators to Google search to individually assist people.
So I don’t know how helpful it is to say “it’s just expensive” when the entire point of AI is to be lower cost than a battalion of humans.
I mean yes, but also no. If you only train it with “good sources” then you miss out on a whole bunch of other valuable information.
Just like scholar.google.com only has “good sources” but generally it’s not going to have the information that 90% of your search queries will be about.
Why not solve it before training the AI?
Simply make it clear that this tech is experimental, then provide sources and context with every result. People can make their own assessment.
Because a lot of people won’t look at sources even if you serve them up on a silver platter?
It’s better than not doing anything and pretending it’s all accurate.
Yes, but as a solution it’s far inferior to not presenting questionable output to the public at all.
(There are a few specific AI/LLM types whose output we might be able to “human-proof”—for instance, if we don’t allow image generators to make photorealistic images of any sort for any purpose, they become much more difficult to abuse—but I can’t see how you would do it for search engine adjuncts like this without having a human curate their training sets.)
Prompt injection has shown us that basically any attempt to limit the output like this is doomed to fail. Like anti-piracy ones, where if you ask directly for the info it says no, but if you ask for the info under the guise of avoiding it, it gives up everything.
Or for instance with the twitter bot, you could get it to regurgitate its own horrifically hateful prompt, then give it a replacement prompt and tell it to change its whole personality, then tell it to critique its previous prompt. There is currently no way to create a prompt that has supremacy over the user input. You can’t ask it to keep a secret because it doesn’t know what a secret is.
I think because we’re getting access to hallucinations, it’s a bit like telling a person “don’t think about an elephant”. Well, they just did, because you prompted them to with the instruction. LLMs similarly can’t actually control what they output.