this post was submitted on 17 May 2024
503 points (94.8% liked)

Technology

59174 readers
3700 users here now

This is a most excellent place for technology news and articles.


Our Rules


  1. Follow the lemmy.world rules.
  2. Only tech related content.
  3. Be excellent to each another!
  4. Mod approved content bots can post up to 10 articles per day.
  5. Threads asking for personal tech support may be deleted.
  6. Politics threads may be removed.
  7. No memes allowed as posts, OK to post as comments.
  8. Only approved bots from the list below, to ask if your bot can be added please contact us.
  9. Check for duplicates before posting, duplicates may be removed

Approved Bots


founded 1 year ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
[–] [email protected] 200 points 5 months ago (9 children)

We not only have to stop ignoring the problem, we need to be absolutely clear about what the problem is.

LLMs don't hallucinate wrong answers. They hallucinate all answers. Some of those answers will happen to be right.

If this sounds like nitpicking or quibbling over verbiage, it's not. This is really, really important to understand. LLMs exist within a hallucinatory false reality. They do not have any comprehension of the truth or untruth of what they are saying, and this means that when they say things that are true, they do not understand why those things are true.

That is the part that's crucial to understand. A really simple test of this problem is to ask ChatGPT to back up an answer with sources. It fundamentally cannot do it, because it has no ability to actually comprehend and correlate factual information in that way. This means, for example, that AI is incapable of assessing the potential veracity of the information it gives you. A human can say "That's a little outside of my area of expertise," but an LLM cannot. It can only be coded with hard blocks in response to certain keywords to cut it from answering and insert a stock response.

This distinction, that AI is always hallucinating, is important because of stuff like this:

But notice how Reid said there was a balance? That’s because a lot of AI researchers don’t actually think hallucinations can be solved. A study out of the National University of Singapore suggested that hallucinations are an inevitable outcome of all large language models. **Just as no person is 100 percent right all the time, neither are these computers. **

That is some fucking toxic shit right there. Treating the fallibility of LLMs as analogous to the fallibility of humans is a huge, huge false equivalence. Humans can be wrong, but we're wrong in ways that allow us the capacity to grow and learn. Even when we are wrong about things, we can often learn from how we are wrong. There's a structure to how humans learn and process information that allows us to interrogate our failures and adjust for them.

When an LLM is wrong, we just have to force it to keep rolling the dice until it's right. It cannot explain its reasoning. It cannot provide proof of work. I work in a field where I often have to direct the efforts of people who know more about specific subjects than I do, and part of how you do that is you get people to explain their reasoning, and you go back and forth testing propositions and arguments with them. You say "I want this, what are the specific challenges involved in doing it?" They tell you it's really hard, you ask them why. They break things down for you, and together you find solutions. With an LLM, if you ask it why something works the way it does, it will commit to the bit and proceed to hallucinate false facts and false premises to support its false answer, because it's not operating in the same reality you are, nor does it have any conception of reality in the first place.

[–] [email protected] 50 points 5 months ago (3 children)

This right here is also the reason why AI fanboys get angry when they are told that LLMs are not intelligent or even thinking at all. They don't understand that in order for rational intelligence to exist, the LLMs should be able to have an internal, referential inner world of symbols, to contrast external input (training data) against and that is also capable of changing and molding to reality and truth criteria. No, tokens are not what I'm talking about. I'm talking about an internally consistent and persistent representation of the world. An identity, which is currently antithetical with the information model used to train LLMs. Let me try to illustrate.

I don't remember the details or technical terms but essentially, animal intelligence needs to experience a lot of things first hand in order to create an individualized model of the world which is used to direct behavior (language is just one form of behavior after all). This is very slow and labor intensive, but it means that animals are extremely good, when they get good, at adapting said skills to a messy reality. LLMs are transactional, they rely entirely on the correlation of patterns of input to itself. As a result they don't need years of experience, like humans for example, to develop skilled intelligent responses. They can do it in hours of sensing training input instead. But at the same time, they can never be certain of their results, and when faced with reality, they crumble because it's harder for it to adapt intelligently and effectively to the mess of reality.

LLMs are a solipsism experiment. A child is locked in a dark cave with nothing but a dim light and millions of pages of text, assume immortality and no need for food or water. As there is nothing else to do but look at the text they eventually develop the ability to understand how the symbols marked on the text relate to each other, how they are usually and typically assembled one next to the other. One day, a slit on a wall opens and the person receives a piece of paper with a prompt, a pencil and a blank page. Out of boredom, the person looks at the prompt, it recognizes the symbols and the pattern, and starts assembling the symbols on the blank page with the pencil. They are just trying to continue from the prompt what they think would typically follow or should follow afterwards. The slit in the wall opens again, and the person intuitively pushes the paper it just wrote into the slit.

For the people outside the cave, leaving prompts and receiving the novel piece of paper, it would look like an intelligent linguistic construction, it is grammatically correct, the sentences are correctly punctuated and structured. The words even make sense and it says intelligent things in accordance to the training text left inside and the prompt given. But once in a while it seems to hallucinate weird passages. They miss the point that, it is not hallucinating, it just has no sense of reality. Their reality is just the text. When the cave is opened and the person trapped inside is left into the light of the world, it would still be profoundly ignorant about it. When given the word sun, written on a piece of paper, they would have no idea that the word refers to the bright burning ball of gas above them. It would know the word, it would know how it is usually used to assemble text next to other words. But it won't know what it is.

LLMs are just like that, they just aren't actually intelligent as the person in this mental experiment. Because there's no way, currently, for these LLMs to actually sense and correlate the real world, or several sources of sensors into a mentalese internal model. This is currently the crux and the biggest problem on the field of AI as I understand it.

[–] [email protected] 12 points 5 months ago (1 children)

That's an excellent methaphor for LLMs.

[–] [email protected] 19 points 5 months ago (1 children)

It's the Chinese room thought experiment.

[–] [email protected] 5 points 5 months ago (1 children)

Hadn't heard about it before (or maybe I did but never looked into it), so I just went and found it in Wikipedia and will be reading all about it.

So thanks for the info!

[–] [email protected] 6 points 5 months ago

No worries. The person above did a good job explaining it although they kind of mashed it together with the imagery from Plato's allegory of the cave.

[–] [email protected] -1 points 5 months ago (1 children)

Wtf are you even talking about.

[–] [email protected] 2 points 5 months ago* (last edited 5 months ago) (1 children)

They are right though. LLM at their core are just about determining what is statistically the most probable to spit out.

[–] [email protected] 0 points 5 months ago

Your 1 sentence makes more sense than the slop above.

[–] [email protected] 30 points 5 months ago (2 children)

I fucking hate how OpenAi and other such companies claim their models "understand" language or are "fluent" in French. These are human attributes. Unless they made a synthetic brain, they can take these claims and shove them up their square tight corporate behinds.

[–] [email protected] 7 points 5 months ago

I though I would have an aneurism reading their presentation page on Sora.

They are saying Sora can understand and simulate complex physics in 3D space to render a video.

How can such bullshit go unchallenged. It drives me crazy.

[–] [email protected] 11 points 5 months ago (1 children)

They do not have any comprehension of the truth or untruth of what they are saying, and this means that when they say things that are true, they do not understand why those things are true.

Which can be beautifully exploited with sponsored content.

See Google I/O '24.

[–] [email protected] 5 points 5 months ago (1 children)

What specifically in Google I/O?

[–] [email protected] 2 points 5 months ago* (last edited 5 months ago)

Alternative title for this year Google I/O: AI vomit. You can watch Verge's TL;DW video on Google I/O. There is no panel that did not mention AI. Most of it is "user centric".

AI can deliver and gather ad data. The bread and butter for Google.

As to how it relates to the quote. It is up to Google to make it as truthful as they want it to be. And given ads is their money driver.

[–] [email protected] 5 points 5 months ago (1 children)

Well stated and explained. I'm not an AI researcher but I develop with LLMs quite a lot right now.

Hallucination is a huge problem we face when we're trying to use LLMs for non-fiction. It's a little bit like having a friend who can lie straight-faced and convincingly. You cannot distinguish whether they are telling you the truth or they're lying until you rely on the output.

I think one of the nearest solutions to this may be the addition of extra layers or observer engines that are very deterministic and trained on only extremely reputable sources, perhaps only peer reviewed trade journals, for example, or sources we deem trustworthy. Unfortunately this could only serve to improve our confidence in the facts, not remove hallucination entirely.

It's even feasible that we could have multiple observers with different domains of expertise (i.e. training sources) and voting capability to fact check and subjectively rate the LLMs output trustworthiness.

But all this will accomplish short term is to perhaps roll the dice in our favor a bit more often.

The perceived results from the end users however may significantly improve. Consider some human examples: sometimes people disagree with their doctor so they go see another doctor and another until they get the answer they want. Sometimes two very experienced lawyers both look at the facts and disagree.

The system that prevents me from knowingly stating something as true, despite not knowing, without some ability to back up my claims is my reputation and my personal values and ethics. LLMs can only pretend to have those traits when we tell them to.

[–] [email protected] 4 points 5 months ago

Consider some human examples: sometimes people disagree with their doctor so they go see another doctor and another until they get the answer they want. Sometimes two very experienced lawyers both look at the facts and disagree.

This actually illustrates my point really well. Because the reason those people disagree might be

  1. Different awareness of the facts (lawyer A knows an important piece of information lawyer B doesn't)
  2. Different understanding of the facts (lawyer might have context lawyer B doesn't)
  3. Different interpretation of the facts (this is the hardest to quantify, as its a complex outcome of everything that makes us human, including personality traits such as our biases).

Whereas you can ask the same question to the same LLM equipped with the same data set and get two different answers because it's just rolling dice at the end of the day.

If I sit those two lawyers down at a bar, with no case on the line, no motivation other than just friendly discussion, they could debate the subject and likely eventually come to a consensus, because they are sentient beings capable of reason. That's what LLMs can only fake through smoke and mirrors.

[–] [email protected] 3 points 5 months ago

usually, what I see is that the REPL they are using is never introspective enough. The ai cant on its own revert to a prevous state or give notes to itself because the response being fast and in linear time matters for a chatbot. ChatGPT can make really cool stuff when you ask it to break it's thoght process into steps. Ones it usually fails spectacularly at. It was like pulling teeth to get it to actually do the steps and not just give the bad answer anyway.

[–] [email protected] 2 points 5 months ago (1 children)

I'm not convinced about the "a human can say 'that's a little outside my area of expertise', but an LLM cannot." I'm sure there are a lot of examples in the training data set that contains qualification of answers and expression of uncertainty, so why would the model not be able to generate that output? I don't see why it would require an "understanding" for that specifically. I would suspect that better human reinforcement would make such answers possible.

[–] [email protected] 14 points 5 months ago

Because humans can do introspection and think and reflect about our own knowledge against the perceived expertise and knowledge of other humans. There's nothing in LLMs models capable of doing this. An LLM cannot asses it own state, and even if it could, it has nothing to contrast it to. You cannot develop the concept of ignorance without an other to interact and compare with.

[–] [email protected] -4 points 5 months ago (1 children)

I think where you are going wrong here is assuming that our internal perception is not also a hallucination by your definition. It absolutely is. But our minds are embodied, thus we are able check these hallucinations against some outside stimulus. Your gripe that current LLMs are unable to do that is really a criticism of the current implementations of AI, which are trained on some data, frozen, then restricted from further learning by design. Imagine if your mind was removed from all stimulus and then tested. That is what current LLMs are, and I doubt we could expect a human mind to behave much better in such a scenario. Just look at what happens to people cut off from social stimulus; their mental capacities degrade rapidly and that is just one type of stimulus.

Another problem with your analysis is that you expect the AI to do something that humans cannot do: cite sources without an external reference. Go ahead right now and from memory cite some source for something you know. Do not Google search, just remember where you got that knowledge. Now who is the one that cannot cite sources? The way we cite sources generally requires access to the source at that moment. Current LLMs do not have that by design. Once again, this is a gripe with implementation of a very new technology.

The main problem I have with so many of these "AI isn't really able to..." arguments is that no one is offering a rigorous definition of knowledge, understanding, introspection, etc in a way that can be measured and tested. Further, we just assume that humans are able to do all these things without any tests to see if we can. Don't even get me started on the free will vs illusory free will debate that remains unsettled after centuries. But the crux of many of these arguments is the assumption that humans can do it and are somehow uniquely able to do it. We had these same debates about levels of intelligence in animals long ago, and we found that there really isn't any intelligent capability that is uniquely human.

[–] [email protected] 8 points 5 months ago (1 children)

This seems to be a really long way of saying that you agree that current LLMs hallucinate all the time.

I'm not sure that the ability to change in response to new data would necessarily be enough. They cannot form hypotheses and, even if they could, they have no way to test them.

[–] [email protected] -3 points 5 months ago (1 children)

My thesis is that we are asserting the lack of human-like qualities in AIs that we cannot define or measure. Assertions should be made on data, not uneasy feelings arising when an LLM falls into the uncanny valley.

[–] [email protected] 5 points 5 months ago (1 children)

But we do know how they operate. I saw a post a while back where somebody asked the LLM how it was calculating (incorrectly) the date of Easter. It answered with the formula for the date of Easter. The only problem is that that was a lie. It doesn't calculate. You or I can perform long multiplication if asked to, but the LLM can't (ironically, since the hardware it runs on is far better at multiplication than we are).

[–] [email protected] 1 points 5 months ago

We do not know how LLMs operate. Similar to our own minds, we understand some primitives, but we have no idea how certain phenomenon emerge from those primitives. Your assertion would be like saying we understand consciousness because we know the structure of a neuron.

[–] [email protected] -5 points 5 months ago

Very long layman take. Why is your guesstimation so incredibly crucial to understand, then next thing important to understand then really, really important to understand, over and over, when you are not an expert?