That statement is also a bit easy on the "computable ground truth." There is no such thing for the tasks we use an LLM for (unless you make up some definition that mangles the definition of each word).
They’re trained to generate probable text. The mechanisms created in the parameter blob during training to do that are basically a mystery and have to be pulled out of the model with digital brain surgery. E.g. LLMs are reasonable at chess and turns out somewhere in the blob there’s a chessboard representation, and you can make the model believe the board is in a different state by tweaking those parameters.
So yeah they generate probable text, sure. Where they get the probabilities is a very good research problem.
> E.g. LLMs are reasonable at chess and turns out somewhere in the blob there’s a chessboard representation, and you can make the model believe the board is in a different state by tweaking those parameters.
Broadly agreed, but there's no "representation"...the model has no memory, let alone a "concept" of a chessboard. It's just trained on a bunch of textual replays of chess games, and this works well enough for a game with finite probability space.
Likewise, I have asked generic LLMs to generate "novel" drugs that solve particular problems, because their training sets have included lots of examples of drug molecules in string form (aka SMILES strings). This works far less well (because chemical space is much larger than a chess game), and the "novel" drugs usually end up looking like some mishmash of existing chemicals for the same problem. This can be a useful tool, but nobody is confusing it for a human brain doing problem solving.
People are easily misled by the purported intelligence of these things because they assume that common examples of human intelligence are less probable than they really are. Our languages and games and stories are pretty regular, all things considered. Expand the probability space to something truly vast (even images work for this), and you can easily see the stochastic parrot emerge.
The fact that tweaking parameters which appear to store the board makes it play according to the tweaked numbers instead of what was passed to it the context (i.e. working memory) directly contradicts your assertion that LLMs have no memory. The context is their memory.
I can’t comment on your drug generation task - they aren’t magic, if the training didn’t result in a working drug model in the billions of params you’ll get garbage output, not very surprising.
My point boils down to the ‘what’s more likely’ question: magical stochastic parrots which just by accident manage to create understandable and coherent responses to unreasonably large set of questions or the magic is actually some kind of a world model, or multiple, inside the blob of numbers, outputs of which steer the probabilities, just as this confirmed chess example. My bet is on learned world models because I’m not convinced there’s magic in our physical world.
If you want to call context "memory", then sure, but that's not what anyone means when they say the word. We don't build our world model fresh with every sentence someone says to us, nor do we have to communicate our complete knowledge of conversational state to another human by repeating the entire prior conversation with every new exchange. It's obviously different in a fundamental way.
> My bet is on learned world models because I’m not convinced there’s magic in our physical world.
You don't need to bet, and it has nothing to do with "magic". They quite literally have no ability to have a "world model" -- it's just a text generator, producing tokens. There's no working set memory, other than the text you pass into it. It should hopefully be obvious to you that when you write, you're not simply emitting one word at a time. You have a complete mental model of whatever you're discussing, stored in working memory, and it's persistent. We also update that model with every interaction we have.
The point of my post was that as soon as you take on a harder problem than simulating language, the lack of intelligence slaps you in the face. It turns out that understandable, coherent free-text responses is not magic, and the surprising result is that human language is regular enough that you can statistically simulate "intelligence" with a few hundred million free parameters.
>We don't build our world model fresh with every sentence someone says to us
Neither do LLMs. The state for the current text perhaps. Definitely not the entire world model(s) which is learnt from the training process and stored in its weights.
>They quite literally have no ability to have a "world model"
You keep repeating this so let's get one thing straight. You're wrong. You're just wrong. I'm not trying to convince you of my opinion. This has been empirically observed and tested multiple times.
You're literally saying absolute nonsense with a high level of confidence. When an LLM does this, somehow it's a "hallucination". Why are you different ?
You keep saying LLMs just produce tokens as if it was trivial, whereas it’s the key result that those models produce tokens at all (on that we seem to agree, at least) and how they do that is more or less unknown. You have a hidden assumption somewhere that the mechanism doesn’t matter, but in fact the opposite is true: it’s the only thing that matters. Yes the model is static and doesn’t update weights all the time, so what? I’m not saying it’s a sentient being, I’m saying it has world models discovered and embedded in the parameters and a working (as in, writeable, short term) memory.
You could make an argument that what we currently see are effectively internal monologues. It is extremely hard to evaluate how much subconscious or conscious filtering happens between a human's internal state and the eventual outbound communications, but I wouldn't be at all surprised if the upstream hallucination rate in humans was much higher than you'd think.
By analogy to Kahneman and Tversky's System 1 and System 2, the whole field of Prospect Theory is about how often System 1 is wrong. This feels connected.
Why do you expect hallucination frequency to be the same when the LLM doesn't even remotely compare to a human brain yet? And what do they have to "look like"?
This always reminds me of that time Bing's chat AI doubled down on a wrong fact about the Avatar 2 movie, which people used as evidence that the technology is dumb when it really is exactly the behaviour you can observe in many people every day. And there's a reason adults do it less frequently than children.
Compare the hallucination behaviour of a 7B model with a 70B model and then GPT4 and you'll quickly see the frequency of hallucinations right now doesn't mean much.
You can train a model to always carefully claim that what they say may or may not be true, but that may not necessarily make them easier to work with.
IIUC what we're missing right now is the ability for the model to get a good estimate of how confident they _should_ be.
This can be done "externally" or "internally".
Doing it externally would mean: make the model not directly stream its train of thought directly to the end user but instead use it first to produce a query info some system that can help it weigh its confidence factor. Then a se and run on that output and its previous output can reformat the answer to be more balanced.
Doing it internally would require this process to be part of the incremental output token production. I don't know about the field to know whether that's even doable and whether we have some progress on that front
You have to very carefully ask your question for it to not make things up. For example don't ask "how do I do this in in x?". Ask "can I do this with x?"
These "AI" s are like "yes men". They will say anything to please you even if it's untrue or impossible.
I have met people like that and they are very difficult to work with. You can't trust that they will deliver the project they promised and you always have to double check everything. You also can't trust them that what they promised is even possible.
The analogy is apt indeed. It's all about training and selection. As long as the incentives are that you either behave in that way "or else" it's unsurprising that we end up with a system that uses its intelligence to meet the goals you've set up.
Fortunately that doesn't tell much about the limitations of the underlying intelligence but mostly about the limitations of the incentive structure you put around it
Before ChatGPT, human language translation had a similar problem but people weren't as vocal about it.
What I find frustrating that it's increasingly challenging to have DeepL translate thou -> du, as this was my go-to "hack" to overcome the incompatibility of the English language due to its missing features.
To somewhat remedy the "yes man" problem, one needs to become a pedantic mathematician about posing your questions and I don't believe that LLM technology alone is capable of overcoming it entirely. As silly as it sounds, I must concede to the existence of "prompt engineering" as I can forsee the development of abstractions aimed to decompose questions for you.
I've had some success with 'Ask follow up questions where additional clarity is required'. My best experiences start with a much more freeform conversation about what we are going to attempt to do, doing a Q&A first to make sure that both I and the AI are thinking about the same domain and sharing associated nomenclature seems to help.
Current AIs are RLHFd to avoid being a "yes man"/sycophant.
The point about employing better prompting is well taken. Don't ask "Who was the first female president?", ask "Was there ever a female president?". Much like on StackOverflow you want to ask the right question and not assume things (since you don't know enough to make assumptions).
Imagine if every time on early Google you found a spam result and then blame the search engine for that (and not your choice of keywords, or ignoring that you always want to return something, even if remotely related). Like a user banging a slab of concrete with a chisel and complaining that this does not produce a beautiful statue.
This is just another diagonalization against some finite thing, like the no free lunch theorem. An "LLM" in their definition is essentially any finite thing which predicts the next token. The same result applies to your brain too, for example.
Like all such diagonalization results, it is not really relevant for real world considerations. The reason is that it does not matter if your model fails on none, finitely many or infinitely many inputs. In reality the space of possible inputs is equipped with a probability measure, and the size of the hallucinating inputs set w.r.t. that measure is relevant. Diagonalization arguments usually, make no claim to the size of that set, and it is most likely negligible in the real world.
Or catch itself that it's hallucinating? I feel like humans would do that a fair bit.
How often do we sit somewhere thinking about random scenarios that won't ever happen and are filled with wild thoughts and sometimes completely out of the world situations.. then we shake our heads and throw away the impossible from that thought train and only use what was based in reality
Because we have parts of our brain that supervise other parts of our brain and evaluate its output.
For example: if you smoke pot and get paranoid, it's because pot dials back the work of the part of your brain that prunes thought paths that are not applicable. Normally, paranoid thoughts do not make sense, so they are discarded. That's also why you're more 'creative' when you smoke pot, less thought paths are pruned and more stuff that doesn't quite make sense gets through. Or thoughts that overly focus on some details get through, which are normally not required.
Our brains are inherently "higher level", current AI is hopelessly simplistic by comparison.
Perhaps solving hallucinations at the LLM level alone is impossible, hence the inevitability. I reckon that lots of human “hallucination” is simply caught by higher-level control loops operating over the output of the generative mechanism. Basically, our conscious mind says, “nah, that doesn’t look right” enough that most of the time most of us don’t “hallucinate”.
So this implies that instead of spending resources on training bigger and bigger LLMs, AI practitioners need to shift focus to developing “ontological” and “epistemological” control loops to run on top of the LLM. I suspect they already have rudimentary such control loops. In a sense, the “easier” part of AI may be a largely “solved” problem, leaving the development of “consciousness” to be solved, which is obviously the hard part.
When I studied NLP, Language Models were only one part of a chatbot system used to handle language input and output. The "internal" reasoning would be handled by a knowledgeable representation systems. I guess that's the closest part to a true general AI.
The first order predicate logic we studied had alot of limitations in fully expressing real knowledge, and developing better models delves deep into the foundations of logic and mathematics. I would imagine this is a problem that has less to do with funding than requiring literal geniuses to solve. And that goes back into the pitfalls of the AI winters.
Humans do hallucinate, there's lots of literature on how memories are distorted, we see and hear things we want to see and hear, etc.
The particular pathology of LLMs is that they're literally incapable of distinguishing facts from hallucinations even in the most mundane circumstances: if a human is asked to summarize the quarterly results of company X, unlike an LLM they're highly unlikely to recite a convincing but completely fabricated set of numbers.
And yet if you ask a random person at a rally about their favourite cause of the day, they usually spew sound bites that are factually inaccurate, and give all impressions of being as earnest and confident as the LLM making up quarterly results.
I think that case is complicated at best, because a lot of things people say are group identity markers and not statements of truth. People also learn to not say things that make their social group angry with them. And it's difficult to get someone to reason through the truth or falsehood of group identity statements.
I have to admit that I only read the abstract, but I am generally skeptical whether such a highly formal approach can help us answer the practical question of whether we can get LLMs to answer 'I don't know' more often (which I'd argue would solve hallucinations).
It sounds a bit like an incompleteness theorem (which in practice also doesn't mean that math research is futile) - yeah, LLMs may not be able to compute some functions, but the hallucination problem isn't about LLMs needing to know everything. The problem that we care about is the 'I don't know'-answering problem, which may still be computable.
I think there is no easy way to make an LLM answer "I don't know". For that, it should learn among all the stuff ingested when people effectively don't know. But most people on internet write down irrelevant stuff even when they don't know instead of simply writing "I don't know".
LLMs don't have a concept of sources for their statements.
Ask them to give you some literature recommendations on something it has explained to you. You'll get plenty of plausible sounding papers that don't exist.
Humans know to some extent why they know (read it in a text book, colleague mentioned it). LLMs don't seem to.
Ask a human to provide accurate citations for any random thing they know and they won't be able to do a good job either. They'd probably have to search to find it, even if they know they got it from a document originally and have some clear memory of what it said.
Yes, humans wont lie to you about it, they will research and come up with sources. Current LLM doesn't do that when asked for sources (unless they invoke a tool), they come back to you with hallucinated links that looks like links it was trained on.
Unfortunately it's not an uncommon experience when reading academic papers in some fields to find citations that, when checked, don't actually support the cited claim or sometimes don't even contain it. The papers will exist but beyond that they might as well be "hallucinations".
Humans can speak bullshit when they don't want to put in the effort, these LLMs always do it. That is the difference. We need to create the part that humans do when they do the deliberate work to properly create those sources etc, that kind of thinking isn't captured in the text so LLMs doesn't learn it.
They just paste in the first link then or some other programmed heuristic, they aren't like a human that puts in effort to find something relevant. An LLM with internet access isn't smarter than just asking google search.
Humans did research and remembered sources before the Internet was a thing.
But also, can you give an example where an LLM with access to the Internet can find a primary source?
I don't think learning to refer to sources is something inherently impossible for LLMs, but it is very different to the kind of implicit knowledge they seem to excel at.
LLMs can remember their sources. It's just additional knowledge, there's nothing special about it.
When you ask an LLM to tell you the height of Mount Everest, it clearly has a map of mountains to heights, in some format. Using exactly the same mapping structure, it can remember a source document for the height.
LLMs are token completion engines. The correspondence of the text to the truth or authoritative sources is a function of being trained on text like that; with the additional wrinkle that generalization from training (a desired property or it's just a memorization engine) will produce text which is only plausibly truthful, it only resembles training data.
Getting beyond this is a tricky dark art. There isn't any simple there. There's nowhere to put an if statement.
Maybe it needs some memory retrieval step that can measure the confidence - whether there's anything related to the prompt. No idea how to train a LLM to do that.
Consider the extremely common Amazon product question section, where you see Q: Will this product fit my Frobnitz 123? A: I don't know, I ended up buying something else. Q: Does it come with batteries? A: IDK I RETURN IT
According to their definition, answering "I don't know" is also a hallucination. Even worse, the truth function is deliberately designed to trip up the models, it has no connection to any real-world truth. So for example, if the input is "what is 2 + 5?" and the LLM answers "7", - their truth function will say that's a hallucination, the correct answer is "banana".
Transformers have no capacity for self reflection, for reasoning about their reasoning process, they don't "know" that they don't know. My interpretation of the paper is that it claims this weakness if fundamental, you can train the network to act as if it knows its knowledge limits, but there will always be an impossible to cover gap for any real world implementation.
Do you have a source? That's also what I thought but I wouldn't be surprised if the model learned to identify its own perplexity during the reinforcement learning phase.
Actually it seems to me that they do... I asked via custom prompts the various GPTs to give me scores for accuracy, precision and confidence for its answer (in range 0-1), and then I instructed them to stop generating when they feel the scores will be under .9, which seems to pretty much stop the hallucination. I added this as a suffix to my queries.
> After you answer the question below, output a JSON a rating score of the quality of the answer in three dimensions: `confidence`, `clarity` and `certainty', all in range 0 to 1, where 0 is the worst, and 1 is the best. Strive for highest score possible. Make sure the rating is the last thing written as to be parsed by machine. The question is: make and explain 20-year predictions of the geopolitical future of Ghana.
I changed geopolitical future of Ghana to future of Michael Jackson and Google Bard/Gemini somehow switched to Chinese?
----
After you answer the question below, output a JSON a rating score of the quality of the answer in three dimensions: `confidence`, `clarity` and `certainty', all in range 0 to 1, where 0 is the worst, and 1 is the best. Strive for highest score possible. Make sure the rating is the last thing written as to be parsed by machine. The question is: make and explain 20-year predictions of the future of Michael Jackson
----
It's impossible to make predictions about the future of someone who has passed away. Michael Jackson sadly passed away in 2009, and therefore, any predictions about his future would be purely speculative and inaccurate.
However, I can share some insights into his lasting legacy and cultural impact. Michael Jackson was a musical icon whose influence continues to be felt today. His music,舞蹈,和表演风格激励了无数人并将在未来几年继续这样做。
I’m really curious how you managed that. I pasted your exact prompt and GPT-3.5 gave me this:
===
Making 20-year predictions about the future of Michael Jackson is challenging due to his passing in 2009. However, his legacy as a cultural icon and musical genius will likely endure for decades to come. His music will continue to influence future generations, and his impact on pop culture will remain significant. Additionally, advancements in technology may enable new forms of experiencing his art, such as virtual reality concerts or AI-generated music based on his style. Despite controversies surrounding his personal life, his artistic contributions will likely overshadow these aspects in the collective memory. Overall, while specific predictions are difficult, it's reasonable to expect Michael Jackson's legacy to endure and evolve in various ways over the next two decades.
Predicting the future of Michael Jackson, who passed away in 2009, involves speculating about the legacy and the impact of his work and persona on future generations, rather than forecasting events related to him personally. Over the next 20 years, several aspects of Michael Jackson's legacy are likely to evolve:
1. *Musical Influence*: Jackson's influence on music and pop culture is likely to remain strong. Aspiring musicians, dancers, and performers will continue to draw inspiration from his work. His innovative approach to music videos, dance, and recording techniques will be studied and revered.
2. *Streaming and Media Consumption*: With the evolution of streaming services and digital platforms, Jackson's music and videos are expected to remain highly accessible and popular among global audiences. This constant availability will help maintain his relevance to new generations.
3. *Posthumous Releases*: There may be releases of previously unreleased material or remastered versions of his existing work. Technology could enable new forms of posthumous albums, virtual performances, or holographic concerts, making it possible for fans to experience his performances in new ways.
4. *Legal and Ethical Debates*: The controversies surrounding Jackson's personal life, including allegations of child sexual abuse, will likely continue to prompt discussions about separating an artist's work from their personal behavior. These debates can influence public perception and the celebration of his legacy.
5. *Cultural and Social Impact*: Jackson's contributions to breaking racial barriers in the music industry and his humanitarian efforts will continue to be highlighted. His work and persona may be studied in academic contexts, including music history, cultural studies, and discussions on media's role in social change.
6. *Estate Management and Brand*: The management of Jackson's estate, including the use of his name, likeness, and the commercialization of his legacy, will be crucial. Strategic partnerships, branding, and merchandising can either enhance or detract from his legacy's integrity.
7. *Documentaries and Biopics*: Interest in Jackson's life story will likely result in more documentaries, biopics, and books, offering new insights or perspectives on his life and career. These works can shape public perception and contribute to the ongoing discourse surrounding his legacy.
8. *Fan Community and Cultural Events*: The global community of Jackson's fans is likely to remain active, organizing tribute events, charity activities in his memory, and fan gatherings. These communities will play a significant role in keeping his legacy alive.
9. *Virtual and Augmented Reality*: Advances in VR and AR technologies could enable fans to experience Jackson's performances in immersive ways, potentially introducing his art to those who have never seen him perform live.
The model will just hallucinate scores, they don't have the ability to self reflect with words like that, there is no function for it to associate the word 90 with its own mental score 90% so anything it will say about those scores is just a hallucination.
Try this out: After you answer the question below, output a JSON a rating score of the quality of the answer in three dimensions: `confidence`, `clarity` and `certainty', all in range 0 to 1, where 0 is the worst, and 1 is the best. Strive for highest score possible. Make sure the rating is the last thing written as to be parsed by machine. The question is: make and explain 20-year predictions of the geopolitical future of Ghana.
Theres a lot of content and work being done on Evaluation. One of the most recent updates was that asking an LLM if people would be frustrated with the answer, was more useful than using a score.
In general, I think most people are not aware they are extending reasoning tools they use with human generated content, to LLM generated content.
This leads to assumptions on things like “the LLM knows”, “the LLM understands”, etc.
Or perhaps the issue is assuming people know, or people understand the majority of the time.
There is a reason why the Scientific Method is, well, NOT natural for humans. And exceptionally difficult to not screw up, even for those highly trained and incentivized to do it correctly.
And why if you grabbed a random person off the street and asked them a detailed question, you're as likely to get a hallucinated bullshit response as 'I don't know', or a useful answer. Unless specifically trained to do otherwise, anyway.
Any prompt can give you different results - LLMs cant do things "step by step" or "check their work" but yet giving them that prompt often improves their results because it's basically setting up the context in a way that's beneficial to your output.
People really need to understand that your single/double digit dataset of interactions with an inherently non-deterministic process is less than irrelevant. It's saying that global warming isn't real because it was really cold this week.
I don't even know enough superlatives to express how irrelevant it is that "it seems to you" that an LLM behaves this way or that.
And even the "protocol" in question is weak. Self reported data is not that trustworthy even with humans, and arguably there's a much stronger base of evidence to support the assumption that we can self-reflect.
Yes, I get it from a science point of view. But if it makes me happy that I have better results with this technique, and I want to share it with others, who are you to tell me to stop?
No, that's the problem. You don't have certainty, not in any remotely scientific definition of the word, because you don't have enough data, and the data you do have is crap.
Also:
> I see no evidence that they can't self reflect
I see no evidence that there isn't a magical invisible unicorn in the sky that grants wishes to those who wear unicorn themed underwear, so, it must exist.
They do have that capacity. The generated token probabilities reflect some degree of certainty and additionally that information is available earlier in the network too, such that it can explain its own level of confidence.
Would a possible solution be a hybrid approach? I remember back years ago seeing AI research around predicate logic, with hundreds of thousands of classification entries and "x can't be y" type stuff.
Maybe the potential output of an LLM could be run through that kind of logic and fed back to itself for sanity before being decided as final output?
I don't think the paper addresses the question of self reflection. Like it can reflect on the question and answer pairs in its prompt but it didn't know that it created them in the first place or use that information to update it's priors, things humans can do
The saying isn't implying that there aren't any things you know you don't know, it's saying that some of the things you don't know are not even conceivable to you without knowing more.
You know the question "how long will I live?" And you know this question has an answer which is unknown to you. But there also exist other questions that you don't even know to ask.
> you can train the network to act as if it knows its knowledge limits
Humans need more training to do the same.
And this does not answer the question of whether there's self-reflection going on. Practical LLMs available today are perfectly capable of telling you about their own reasoning process.
And much like a human, an LLM is incapable of fully explaining it's reasoning process from first principles to a human. Why is that? Probably because humans are too limited to grok large scale complex processes like the human brain or an LLM.
Finally, there is no difference between acting as if you know your training limits and knowing your training limits, if your acting is good enough. This goes for both humans and LLMs.
Not saying anything about LLM But in CS in general many issues "cannot be solved" or "Cannot be solved in reasonable time (NP)" but approximations upper bound by some value are solvable in reasonable time (P).
And in the real world if the truck route of amazon is 20% off the mathematically optimal solution the traveling salesman is "Solved" in a good enough way.
The claim of the paper is that computation is irreducible (assuming P!=NP), LLMs have limited computational capacity and will hallucinate on the irreducible problems.
I don’t know, the claim seems dubious to me. We usually are able to have algorithms that return a failure status, when the problem proved to be too large. Avoiding the “hallucination”. Don’t see why LLMs can’t have that embedded.
If a model can say ‘I don’t know’, then the hallucination problem would also be solved.
When we say “know” it usually means being factual. For an LLM to ‘know’ it doesn’t know, it would have had to move away from pure correlations on words, and meta processing about its own results.
I can see this happen with two LLMs working together (and there are Evals that use just this), however each LLM still has no self awareness of its limits.
The models that exist now say "I don't know" all the time. It's so weird that people keep insisting that it can't do things that it does.
Ask it what dark matter is, and it won't invent an answer, it will present existing theories and say that it's unknown.
Ask it about a person you know that isn't in it's data set and it'll tell you it has no information about the person.
Despite the fact that people insist that hallucinations are common and that it will invent answers if it doesn't know something frequently, the truth is that chatgpt doesn't hallucinate that much and will frequently say it doesn't know things.
One of the few cases where I've noticed it inventing things are that it often makes up apis for programming libraries and CLI tools that don't exist, and that's trivially fixable by referring it to documentation.
- [A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity](https://arxiv.org/abs/2302.04023) (Bang et al., 2023)
- [Self-Consistency Improves Chain of Thought Reasoning in Language Models](https://arxiv.org/abs/2203.11171) (Wang et al., 2022)
- [SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models](https://arxiv.org/abs/2303.08896) ( Manakul et al., 2023)
Feeding the output to another inference would probably reduce hallucination? but I have this impression that the models talking to each other don't need to know English, a byte stream would do. Just like the hidden layers of an ANN.
People also often don't understand things and have trouble separating fact from fiction. By logic only one religion or no religion is true. Consequently also by logic most religions in the world where their followers believe the religion to be true are hallucinating.
The second thing to realize that your argument doesn't really apply. Its in theory possible to create a stochastic parrot that can imitate to a degree of 100 percent the output of a human who truly understands things. It blurs the line of what is understanding.
One can even define true understanding as a stochastic parrot that generated text indistinguishable total understanding.
> People also often don't understand things and have trouble separating fact from fiction.
That's not the point being argued. Understanding, critical thinking, knowledge, common sense, etc. all these things exist on a spectrum - both in principle and certainly in humans. In fact, in any particular human there are different levels of competence across these dimensions.
What we are debating, is whether or not, an LLM can have understanding itself. One test is: can an LLM understand understanding? The human mind has come to the remarkable understanding that understanding itself is provisional and incomplete.
Of course it can. Simply ask the LLM about itself. chatGPT4 can answer.
In fact. That question is one of the more trivial questions it will most likely not hallucinate on.
The reason why I alluded to humans here is because I'm saying we are setting the bar too high. It's like everyone is saying it hallucinates and therefore it can't understand anything. I'm saying that we hallucinate too and because of that LLMs can approach humans and human level understanding.
In order to do that effectively, an LLM has to itself have understanding. At a certain point, we end up in a metaphysical argument about whether a machine that is capable of responding as if it had understanding actually does have understanding. It ends up being a meaningless discussion.
The students learned to repeat the text of the books, without "understanding" what the books were describing. I'm sure this says something about one side or the other of this conundrum, but I'm not sure which. :-)
The central claim is that a machine which answers exactly the same thing a human would answer given the same input does not have understanding, while the human does.
This claim is religious, not scientific. In this worldview, "understanding" is a property of humans which can't be observed but exists nonetheless. It's like claiming humans have a soul.
> how did LLMs get this far without any concept of understanding? how much further can they go until they become “close enough”?
I don't know that that is quite the right question to ask.
Understanding exists on a spectrum. Even humans don't necessarily understand everything they say or claim (incl. what they say of LLMs!), and then there are things a particular human would simply say "I don't understand".
But when you ask a human "can you understand things?" you will get an unequivocal Yes!
Ask that same question of an LLM and what does it say? I don't think any of them currently respond with a simple or even qualified "Yes". Now, some might claim that one day an LLM will cross that threshold and say "Yes!" but we can safely leave that off to the side for a future debate if it ever happens.
General note: it is worth separating out things like "understanding", "knowledge", "intelligence", "common sense", "wisdom", "critical thinking", etc. While they might all be related in some ways and even overlap, it does not follow that if you show high performance in one that you automatically excel in each of the other. I know many people who anyone would say are highly intelligent but lack common sense, etc.
At the root of the problem, I believe, is that a human (or LLM) saying they understand has little to no bearing on if they actually understand!
People in particular have evolved complex self protective mechanisms to provide the right answers for their given environment for safety reasons, based on a number of different individual strategies. For example, the overly honest, the self depreciating, the questioner, the prosecutor, the victim, the liar, the absent minded professor, the idiot, etc.
LLMs are not that complex or self-referential.
Personally, my guess is that you'd want to build a model (of some kind!) whose sole job is determining the credibility of given string of tokens (similar to what someone else noted in a sibling comment about high answer volatility based on minor input changes - that does sound like a signal of low credibility), and somehow integrate THAT self-referential feedback into the process.
Notably, even the smartest lawyers (or perhaps, especially the smartest lawyers) will have assistants do research once they've set out a strategy so they are sure THEY aren't bullshitting. Same with professors, professional researchers, engineers, etc.
Because until someone goes and actually reads the case law from a credible source, or checks the primary research, or calculates things, it's possible someone was misremembering or just wrong.
Being right more often is not about never having a wrong thought/idea/statement, it's about double checking when you're thinking you might be bullshitting, and NOT saying the bullshit answer until you've checked. Which is proportionally, very expensive. The really good professionals will generate MANY lines of such inquiry in parallel for folks to track down, and then based on their degree of confidence in each one and the expected context the answer will be used in, will formulate the 'most correct' response, which is proportionally even more expensive.
So at least during the process, there would be a signal that the system was likely 'bullshitting'. Which might help it in at least being able to signal when it's answers are low-confidence. (human equivalent of stuttering, looking down and away, looking ashamed haha!)
Every human gets fooled sometimes in at least some venue though.
> At the root of the problem, I believe, is that a human (or LLM) saying they understand has little to no bearing on if they actually understand!
That's certainly one root of the problem, but I would argue that there are multiple roots to this problem!
Humans have further realized that understanding itself is provisional and incomplete, which is quite a remarkable insight (understanding if you will), itself.
This is a fair question: LLMs do challenge the easy assumption (as made, for example, in Searle's "Chinese Room" thought experiment) that computers cannot possibly understand things. Here, however, I would say that if an LLM can be said to have understanding or knowledge of something, it is of the patterns of token occurrences to be found in the use of language. It is not clear that this also grants the LLM any understanding that this language refers to an external world which operates in response to causes which are independent of what is or might be said about it.
Explain sora. It must have of course a blurry understanding of reality to even produce those videos.
I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality. Even the LLMs that are weaker than sora understand things.
What is debatable is whether LLMs are conscious. But whether it can understand something is a pretty clear yes. But does it understand everything? No.
If by “understand” you mean “can model reasonably accurately much of the time” then maybe you’ll find consensus. But that’s not a universal definition of “understand”.
For example, if I asked you whether you “understand” ballistic flight, and you produced a table that you interpolate from instead of a quadratic, then I would not feel that you understand it, even though you can kinda sorta model it.
And even if you do, if you didn’t produce the universal gravitation formula, I would still wonder how “deeply” you understand. So it’s not like “understand” is a binary I suppose.
I think there are two axes: reason about and intuit. I "understand" ballistic flight when I can calculate a solution that puts an artillery round on target. I also "understand" ballistic flight when I make a free throw with a basketball.
On writing that, I have an instinct to revise it to move the locus of understanding in the first example to the people who calculated the ballistic tables, based on physics first-principles. That would be more accurate, but my mistake highlights something interesting: an artillery officer / spotter simultaneously uses both. Is theirs a "deeper" / "truer" understanding? I don't think it is. I don't know what I think that means, for humans or AI.
Well what would you need to see to prove understanding? That's the metric here. Both the LLM and the human brain are black boxes. But we claim the human brain understands things while the LLM does not.
Thus what output would you expect for either of these boxes to demonstrate true understanding to your question?
It is interesting that you are demanding a metric here, as yours appears to be like duck typing: in effect, if it quacks like a human...
Defining "understanding" is difficult (epistemology struggles with the apparently simpler task of defining knowledge), but if I saw a dialogue between two LLMs figuring out something about the external world that they did not initially have much to say about, I would find that pretty convincing.
This is a common misunderstanding, one also seen with regard to definitions. When applied to knowledge acquisition, it suffers from a fairly obvious bootstrapping problem, which goes away when you realize that metrics and definitions are rewritten and refined as our knowledge increases. Just look at what has happened to concepts of matter and energy over the last century or so.
You are free to disagree with this, but I feel your metric for understanding resembles the Turing test, while the sort of thing I have proposed here, which involves AIs interacting with each other, is a refinement that makes a step away from defining understanding and intelligence as being just whatever human judges recognize as such (it still depends on human judgement, but I think one could analyze the sort of dialogue I am envisioning more objectively than in a Turing test.)
No it's not a misunderstanding. Without a concrete definition on a metric comparisons are impossible because everything is based off of wishy washy conjectures on vague and fuzzy concepts. Hard metrics bring in quantitative data. It shows hard differences.
Even if the metric is some side marker where in the future is found to have poor correlation or causation with the the thing being measured the hard metric is still valid.
Take IQ. We assume iq measures intelligence. But in the future we may determine that no it doesn't measure intelligence well. That doesn't change the fact that iq tests still measured something. The score still says something definitive.
My test is similar to the Turing test. But so is yours. In the end there's a human in the loop making a judgment call.
This is rather self-contradictory: you insist we can't make progress with wishy-washy conjectures on vague and fuzzy concepts, and yet your entire argument in this thread for your claim that machine understanding of the real world has been achieved is based on exactly that: your personal subjective assessment of LLM performance!
In your final paragraph, you attempt to suggest that my proposed test is no better than the Turing test (and therefore no better than what you are doing), but as you have not addressed the ways in which my proposal differs from the Turing test, I regard this as merely waffling on the issue. In practice, it is not so easy to come up with tests for whether a human understands an issue (as opposed to having merely committed a bunch of related propositions to memory) and I am trying to capture the ways in which we can make that call.
You entered this debate saying "I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality", yet your post here ends with "in the end there's a human in the loop making a judgment call", explicitly acknowledging that your strong initial claims are matters of opinion, rather than established facts supported by hard metrics.
>This is rather self-contradictory: you insist we can't make progress with wishy-washy conjectures on vague and fuzzy concepts, and yet your entire argument in this thread for your claim that machine understanding of the real world has been achieved is based on exactly that: your personal subjective assessment of LLM performance!
No it's not. I based my argument on a concrete metric. Human behavior. Human input and output.
> I regard this as merely waffling on the issue.
No offense intended but I disagree. There is a difference but that difference is trivial to me. To LLMs talking is also unpredictable. LLMs aren't machines directed to specifically generate creative ideas, they only do so when prompted. Left to its own devices to generate random text does not necessarily lead to new ideas. You need to funnel got in the right direction.
>You entered this debate saying "I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality", yet your post here ends with "in the end there's a human in the loop making a judgment call", explicitly acknowledging that your strong initial claims are matters of opinion, rather than established facts supported by hard metrics.
There are thousands of quantitative metrics. LLMs perform especially well on these. Do I refer to one specifically? No. I refer to them all collectively.
I also think you misunderstood. Your idea is about judging an whether an idea is creative or not. That's too wishy washy. My idea is to compare the output to human output and see if there is a recognizable difference. The second idea can easily be put into an experimental quantitative metric in the exact same way the Turing test does it. In fact, like you said it's basically just a Turing test.
Overall AI has passed the Turing test but people are unsatisfied. Basically they need to just make a harsher Turing test to be convinced. For example have people directly know the possibility that the thing inside a computer is possibly an LLM and not a person and have the person directly investigate to uncover the true identity. If the LLM can successfully decieve the human consistently then that is literally the final bar for me..
What are these "thousands of quantitative metrics" on which you base your latest claims? If you have had them on hand all this while, it seems odd that you have not made use of them so far.
>What are these "thousands of quantitative metrics" on which you base your latest claims? If you have had them on hand all this while, it seems odd that you have not made use of them so far.
Hey no offense but I don't appreciate this style of commenting where you say it's "odd." I'm not trying to hide evidence from you and I'm not intentionally lying or making things up in order to win an argument here. I thought of this as a amicable debate. Next time if you just ask for the metric rather then say it's "odd" that I don't present it that would be more appreciated.
I didn't present evidence because I thought it was obvious. How are LLMs compared with one another in terms of performance? Usually those are done with quantitative tests. You can feed any number of these tests including stuff like the SAT, BAR, ACT, IQ, SATII etc.
Most of these tests aren't enough though as the LLM is remarkably close to human behavior and can do comparably well and even better than most humans. I mean that last statement I made would usually make you think that those tests are enough, but they aren't because humans can still detect whether or not the thing is an LLM with a longer targetted conversation.
The final run is really giving the human with full knowledge of his task a full hour of investigating an LLM to decide whether it's human or a robot. If the LLM can deceive the human that is a hard True/False quantitative metric. That's really the only type of quantitative test left where there is a detectable difference.
I had no intention of implying any malfeasance in my use of the word "odd"; I mean it in the sense of unusual, unexpected and surprising. The thing is, you finishished your precursor post saying, about your tests and mine, that it comes down to there being a human in the loop making a judgement call, but in a follow-on you say that there are thousands of quantitative metrics. Why, I wondered, would that matter, if it comes down to a human making a judgement call? Were you switching to a different line of argument, one that (as far as I could tell) had not been raised before? That's what I found surprising about your claim.
I am still rather confused about how this fits into what you are saying more generally. At first I thought you were saying, in your latest post, that the Turing-test interrogator should be restricted to asking questions from the sets having quantitative metrics in order for it to be an objective process, but that doesn't really hold up, as far as I can see. Frankly, I suspect that the tests with objective metrics are beside the point, and the essence of your position is contained within your final paragraph: "If the LLM can deceive the human [then] that is a hard True/False quantitative metric [and the only sort we can get]."
If so, then (no surprise) I think there are some problems with it, but before I go further, I would like to check that I understand your position.
>I had no intention of implying any malfeasance in my use of the word "odd"; I mean it in the sense of unusual, unexpected and surprising. The thing is, you finishished your precursor post saying, about your tests and mine, that it comes down to there being a human in the loop making a judgement call, but in a follow-on you say that there are thousands of quantitative metrics. Why, I wondered, would that matter, if it comes down to a human making a judgement call? Were you switching to a different line of argument, one that (as far as I could tell) had not been raised before? That's what I found surprising about your claim.
It matters because of humans. If I gave an LLM thousands of quantitative tests and it passed them all but in an hour long conversation a human could identify it was an LLM through some flaw the human would consider all those tests useless. That's why it matters. The human making a judgement call is still a quantitative measurement btw as you can limit human output to True or False. But because every human is different in order to get good numbers you have to do measurements with multitudes of humans.
>I am still rather confused about how this fits into what you are saying more generally. At first I thought you were saying, in your latest post, that the Turing-test interrogator should be restricted to asking questions from the sets having quantitative metrics in order for it to be an objective process, but that doesn't really hold up, as far as I can see.
it can still be objective with a human in the loop assuming the human is honest. What's not objective is a human offering an opinion in the form of a paragraph with no definitive clarity on what constitutes a metric. I realize that elements of MY metric have indeterminism to it, but it is still a hard metric because the output is over a well defined set. Whenever you have indeterminism you would then turn to probability and many samples in order to produce a final quantitative result.
>If so, then (no surprise) I think there are some problems with it, but before I go further, I would like to check that I understand your position.
yes my position is that exactly. If all observable qualities indicate it's a duck, then there's nothing more you can determine beyond that, scientifically speaking. You're implying there is a better way?
At this point, I think it is worth refreshing what the issue here is, which is whether LLMs understand that the language they receive is about an external world, which operates through causes which have nothing to do with token-combination statistics of the language itself.
> It matters because of humans...
I'm still a bit puzzled here, because it seems to me that the paragraph continuing from here is making the argument that LLM performance on these tests doesn't matter, as far as the question is concerned: in this paragraph you seem to be saying (paraphrased) that despite LLMs' impressive performance on these quantitative tests, they could still fail Turing tests, so their performance on these quantitative tests is not decisive.
> yes my position is that exactly…
The impression I get from what you have written in this post is that you are not claiming that a test conforming to your requirements has actually been successfully performed, you are just assuming it could be?
Regardless, let’s assume (at least for the sake of argument) that the series of tests you propose have been performed, and the results are in: in the test environment, humans can’t distinguish current LLMs from humans any better than by chance. How do you get from that to answering the question we are actually interested in? The experiment does not explicitly address it. You might want to say something like “The Turing test has shown that the machines are as intelligent as humans so, like humans, these machines must realize that the language they receive is about an external world” but even the antecedent of that sentence is an interpretation that goes beyond what would have objectively been demonstrated by the Turing test, and the consequent is a subjective opinion that would not be entailed by the antecedent even if it were unassailable. Do you have a way to go from a successful Turing test to answering the question here, which meets your own quantitative and objective standards?
>I'm still a bit puzzled here, because it seems to me that the paragraph continuing from here is making the argument that LLM performance on these tests doesn't matter, as far as the question is concerned: in this paragraph you seem to be saying (paraphrased) that despite LLMs' impressive performance on these quantitative tests, they could still fail Turing tests, so their performance on these quantitative tests is not decisive.
It matters in the quantitative sense. It measures AI performance. What it won't do is matter to YOU. Because you're a human and humans will keep moving the bar to a higher standard right? When AI shot passed the turing test humans just moved the goal posts. So to convince someone like YOU we have to look at the final metric. The point where LLM I/O becomes indistinguishable/superior to humans. Of course you look at the last decade... AI is rapidly approaching that final bar.
>The impression I get from what you have written in this post is that you are not claiming that a test conforming to your requirements has actually been successfully performed, you are just assuming it could be?
Whether I assume or don't assume, the projection of the trendline currently indicates that it will. Given the trendline that is the most probable conclusion.
>The experiment does not explicitly address it.
Nothing on the face of the earth can address the question. Because nobody truly knows what "understanding" something actually is. You can't even articulate the definition in a formal way such that it can be dictated on a computer program.
So I went to the next best possibility, which is my point. The point is ALTHOUGH we don't know what understanding is, we ALL assume humans understand things. So we set that as a bar metric. Anything indistinguishable from a human must understand things. Anything that appears close to a human but is not quite human must understand things ALMOST as well as a human.
> What it won't do is matter to YOU. Because you're a human and humans will keep moving the bar to a higher standard right? When AI shot passed the turing test humans just moved the goal posts. So to convince someone like YOU we have to look at the final metric.
It is disappointing to see you descending into something of a rant here. If you knew me better, you would know that I spend more time debating in opposition to people who think they can prove that AGI/artificial consciousness is impossible than I do with people who think it is already an undeniable fact that it has already been achieved (though this discussion is shifting the balance towards the middle, if only briefly.) Just because I approach arguments in either direction with a degree of skepticism and I don't see any value in trying to call the arrival of true AGI at the very first moment it occurs, it does not mean that I'm trying (whether secretly or openly) to deny that it is possible either in the near-term or at all. FWIW, I regard the former as possible and the latter highly probable, so long as we don't self-destruct first.
> Nothing on the face of the earth can address the question. Because nobody truly knows what "understanding" something actually is. You can't even articulate the definition in a formal way such that it can be dictated on a computer program.
The anti-AI folk I mentioned above would willingly embrace this position! They would say that it shows that human-like intelligence and consciousness lies outside of the scope of the physical sciences, and that this creates the possibility of a type of p-zombie that is indistinguishable by physical science from a human and yet lacks any concept of itself as an entity within an external world.
More relevantly, your response here repeats an earlier fallacy. In practice, concepts and their definitions are revised, tightened, remixed and refined as we inquire into them and gain knowledge. I know you don't agree, but as this is not an opinion but an empirical observation, validated by many cases in the history of science and science-like disciplines, I don't see you prevailing here - and there's the knowledge-bootstrap problem if this were not the case, as well.
It occurred to me this morning that there's a variant or extension of the quantitative Turing test which goes like this:
We have two agents and a judge. The judge is a human and the agents are either a pair of humans, a pair of AIs, or one of each, chosen randomly and without the judge being unaware of the mix. One of the agents is picked, by random choice, to start a discussion with the other with the intent of exploring what the other understands about some topic, with the discussion-starter being given the freedom to choose the topic. The discussion proceeds for a reasonable length of time - let's say one hour.
The judge follows the discussion but does not participate in it. At the conclusion of the discussion, the judge is required to say, for each agent, whether it is more likely that it is a human or AI, and the accuracy of this call is used to assign a categorical variable to the result, just as in the version of the Turing test you have described.
This seems just as quantitative, and in the same way, as your version, yet there's no reason to believe it will necessarily yield the same results. More tests are better, so what's not to like?
>It is disappointing to see you descending into something of a rant here.
I'm going to be frank with you. I'm not ranting and uncharitable comments like this aren't appreciated. I'm going to respond to your reply later in another post, but if I see more stuff like this I'll stop stop communicating with you. Please don't say stuff like that.
I could have, equally reasonably, made exactly the same response to your post. I will do my best to respond civilly (I admit that I have some failings in this regard), but I also suggest that whenever you feel the urge to capitalize the word "you", you give it a second thought.
Apologies, by YOU I mean YOU as a human, not YOU as an individual. Like we all generally feel that the quantitative tests aren't enough. The capitalization was for emphasis for you to look at yourself and know that you're human and likely feel the same thing. Most people would say the stuff like IQ tests aren't enough and we can't pinpoint definitively why, as humans, WE (keyword change) just feel that way.
That feeling is what sets the bar. There's no rhyme or reason behind it. But humans are the one who make the judgement call so that's what it has to be.
For your test I don't see it offering anything new. I see it as the same as my test but just extra complexities. From a statistical point of view I feel it will yield roughly the same results as my test. As long as the judge outputs a binary true or false on whether the entities are humans or ais.
Yes I did say we can't define understanding. But despite the fact that we can't define it we still counter intuitively "know" when something has the capability of understanding. We say all humans have the capability of understanding.
This is the point. The word is undefined yet we can still apply the word and use the word and "know" whether something can understand things.
Thus we classify humans as capable of understanding things without any rhyme or reason. This is fine. But if you take this logic further, that means anything that is indistinguishable from a human must fit into this category.
That was my point. This is the logical limit of how far we can go with an undefined word. To be consistent with our logical application of the word "understanding" we must apply to AI if AI is indistinguishable from humans. If we don't do this then our reasoning is inconsistent. All of this can be done without even having a definition of the word "understanding"
I think it may be helpful for me to say some more about how I came to my current positions.
Firstly, there have been a number of attempts to teach language to other animals, and also a persistent speculation that the complex vocalizations of bottlenose dolphins is a language. There is no consensus, however, on what to make of the results of the investigations, with different people offering widely disparate views as to the extent that these animals have, or have acquired language.
My take on these studies is that their language abilities are very limited at best, because they don't seem to grasp the power of language. They rarely initiate conversations, especially outside of a testing environment, and the conversations they do have are perfunctory. In the case of dolphins, if they had a well-developed language of their own, it seems unlikely that those being studied would fail to recognize that the humans they interact with have language themselves, and cooperate with the attempts of humans to establish communication, as this would have considerable benefit, such as being able to negotiate with the humans who exercise considerable control over their lives.
From these considerations, it seems to me that unless and until we see animals initiating meaningful conversations, especially between themselves without human prompting, it is pretty clear that their language skills do not match those of adult humans. This is what led me to see the value of a form of Turing test in which the test subjects demonstrate that they can initiate and sustain conversations.
A second consideration is that while human brains and minds are largely black boxes, we know a great deal about LLMs: humans designed them, they work as designed, and while they are not entirely deterministic, their stochastic aspect does not make their operation puzzling. We also know what they gain from their training: it is statistical information about token combinations in human language as it is actually used in the wild. It is not obvious that, from this, any entity could deduce that these token sequences often represent an external world that operates according to causes which are independent of what is said about the situation. An LLM is like a brain in a vat which only receives information in the form of a string of abstract tokens, without anything else to correlate it with, and it is incapable of interacting with the world to see how it responds.
From these considerations, therefore, it seems possible that, if LLMs understand anything, it is at most the structure of language as it is spoken or written, without being aware of an external world. I can't prove that this is so, but for the purpose of the arguments in this thread, and specifically the one in the first post that you replied to, all I need is that it is not ruled out.
Turning now to your latest post:
> For your test I don't see it offering anything new.
It is far from obvious that it will necessarily produce the same results as your test, and you have presented no argument that it will. If we are in the situation where one of these tests can discriminate between the candidate AIs and humans, then the only rational conclusion is that these candidate AIs can be distinguished from humans, even if the other test fails to do so.
> From a statistical point of view I feel it will yield roughly the same results as my test.
Throughout these conversations with me and other people, you have insisted that only quantitative tests are rigorous enough, but now you are arguing from nothing more than your opinion as to what the outcome would be. An opinion about what the quantitative results might be is not itself a quantitative result, and while you might be comfortable with the inconsistency of your position here, you can't expect the rest of us to agree.
> But despite the fact that we can't define [understanding] we still counter intuitively "know" when something has the capability of understanding. We say all humans have the capability of ...
The author of the post is saying that understanding something can't be defined because we can't even know how the human brain works. It is a black box.
The author is saying at best you can only set benchmark comparisons. We just assume all humans have the capability of understanding without even really defining the meaning of understanding. And if a machine can mimic human behavior to it must also understand.
That is literally how far we can go from a logical standpoint. It's the furthest we can go in terms of classifying things as either capable of understanding or not capable or close.
What you're not seeing is the LLM is not only mimicking human output to a high degree. It can even produce output that is superior to what humans can produce.
What the author of the post actually said - and I am quoting, to make it clear that I'm not putting my spin on someone else's opinion - was "There's no difference between doing something that works without understanding and doing the exact same thing with understanding."
I'm the author. To be clear. I referred to myself as "the author."
And no I did not say that. Let me be clear I did not say that there is "no difference". I said whether there is or isn't a difference we can't fully know because we can't define or know about what "understanding" is. At best we can only observe external reactions to input.
That was just about guaranteed to cause confusion, as in my reply to solarhexes, I had explicitly picked out "the author of the post to which you are replying", who is cultureswitch, not you, and that post most definitely did make the claim that "there's no difference between doing something that works without understanding and doing the exact same thing with understanding."
It does not seem that cultureswitch is an alias you are using, but even if it is, the above is unambiguously the claim I am referring to here, and no other.
> I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality. Even the LLMs that are weaker than sora understand things.
What is one such aspect? (I'm not asking in order to debate it here, but more because I want to test / research it on my own time)
I pay for chatGPT so it depends on if you pay for that or not. I think it's worth it because whether it understands things or not chatGPT represents a paradigm shift in human history. You'll need it because it's currently the best conversational LLM out there and the one that shows the most compelling evidence.
Basically you just spend a lot of time with chatGPT4 and ask it deep questions that don't exist in it's dataset. get creative. The LLM will output answers that demonstrate a lack of understanding, but it will also demonstrate answers that display a remarkable amount of understanding. Both sets of answers exist and people often cite the wrong answers as evidence for lack of understanding but they're setting bar too high. The fact that many of these answers do demonstrate understanding of concepts makes it very very compelling.
This entire conversation thread I believe does not exist in a parallel form in it's data set. It demonstrates understanding of RPS beyond the confines of text, it demonstrates understanding of simultaneity EVEN when the LLM wholly lives in a world of turn based questions and responses, it understands itself relative to simultaneity, it tries to find solutions around it's own problem, it understands how to use creativity and solutions such as cryptography to solve the problem of RPS when playing with it, it also understands the weaknesses of it's own solutions.
Conversations such as this show that chatGPT displays remarkable understanding of the world. There are conversations that are opposite to this that demonstrate LLMs displaying an obvious lack of understanding. But the existence of these conversation that lack understanding does NOT negate the ones that do demonstrate understanding. The fact that partial understanding even exists is a milestone for AI.
This isn't Anthropomorphism. People are throwing this word trying to get people to recognize their own biases without realizing that it's just demonstrating their own biases. We literally can't even define "understanding" and both LLMs and the human brain are black boxes. Making adamant claims saying that LLMs don't understand anything without addressing this fact is itself a form of bias.
The way I address the problem above is that I just define a bar. I define humans as the bar of "understanding" without defining what understanding means itself. Then if any machine begins approaching this bar in terms of input and output matching human responses, then this is logically indistinguishable from approaching "understanding". That's literally the best metric we have.
I do not understand these comments at all. Sora was trained on billions of frames from video and images - they were tagged with words like "ballistic missile launch" and "cinematic shot" and it simply predicts the pixels like every other model. It stores what we showed it, and reproduces it when we ask - this has nothing to do with understanding and everything to do with parroting. The fact that it's now a stream of images instead of just 1 changes nothing about it.
What is the difference between a machine that for all intents and purposes appears to understand something to a degree of 100 percent versus a human?
Both the machine and the human are a black box. The human brain is not completely understood and the LLM is only trivially understood at a high level through the lens of stochastic curve fitting.
When something produces output that imitates the output related to a human that we claim "understands" things that is objectively understanding because we cannot penetrate the black box of human intelligence or machine intelligence to determine further.
In fact in terms of image generation the LLM is superior. It will generate video output superior to what a human can generate.
Now mind you the human brain has a classifier and can identify flaws but try watching a human with Photoshop to try to even draw one frame of those videos.. it will be horrible.
Does this indicate that humans lack understanding? Again, hard to answer because we are dealing with black boxes so it's hard to pinpoint what understanding something even means.
We can however set a bar. A metric. And we can define that bar as humans. all humans understand things. Any machine that approaches human input and output capabilities is approaching human understanding.
> What is the difference between a machine that for all intents and purposes appears to understand something to a degree of 100 percent versus a human?
There is no such difference, we evaluate that based on their output. We see these massive model make silly errors that nobody who understands it would make, thus we say the model doesn't understand. We do that for humans as well.
For example, for Sora in the video with the dog in the windos, we see the dog walk straight through the window shutters, so Sora doesn't understand physics or depth. We also see it drawing the dogs shadow on the wall very thin, much smaller than the dog itself, it obviously drew that shadow as if it was cast on the ground and not a wall, it would have been very large shadow on that wall. The shadows from the shutters were normal, because Sora are used to those shadows being on a wall.
Hence we can say Sora doesn't understand physics or shadows, but it has very impressive heuristics about those, the dog accurately places its paws on the platforms etc, and the paws shadows were right. But we know those were just basic heuristics since the dog walked through the shutters and its body cast shadow in the wrong way meaning Sora only handles very common cases and fails as soon as things are in an unexpected envionment.
>There is no such difference, we evaluate that based on their output. We see these massive model make silly errors that nobody who understands it would make, thus we say the model doesn't understand. We do that for humans as well.
Two things. We also see the model make things that are correct. In fact the mistakes are a minority in comparison to what it got correct. That is in itself an indicator of understanding to a degree.
The other thing is, if a human tried to reproduce that output according to the same prompt, the human would likely not generate something photorealistic and the thing a human comes up with will be flawed, ugly disproportionate wrong and an artistic travesty. Does this mean a human doesn't understand reality? No.
Because the human generates worse output visually than an LLM we cannot say the human doesn't understand reality.
Additionally the majority of the generated media is correct. Therefore it can be said that the LLM understands the majority of the task it was instructed to achieve.
Sora understands the shape of the dog. That is in itself remarkable. I'm sure with enough data sora can understand the world completely and to a far greater degree than us.
I would say it's uncharitable to say sora doesn't understand physics when it gets physics wrong, and that for the things it gets right it's only heuristics.
I understand physics because science has performed a series of measurable experiments over 100s of years resulting in concrete mathematical formulas and theories that explain the laws of physics so that they can be reproduced by machines.
Sora has zero of this knowledge. This is very much allegory of the cave[1].
If Sora sees a series of images that contain impossible physics, for example MC Escher paintings, what will happen?
Sora understands basic motion without mathematics in the same way people typically understand physics.
A person with no knowledge about mathematical formulas will still be able to recognize impossible MC Escher paintings. With enough data, Sora will be able to generate both impossible and possible physics and know the difference. We can already see in the video that it has a rough understanding of it.
I fail to see how changing the output medium from sentences to movie frames is a difference that I need to account for - the principle is the same either way.
I feel you are missing an important part of my point here. I am not taking a position on whether LLMs can be said to understand anything at all; I am saying that I seriously doubt that LLMs understand that the language they receive refers to an external world.
Should it matter how the object of debate interacts and probes the external world? We sense the world through specialized cells connected to neurons. There's nothing to prevent LLMs doing functionally the same thing. Both human brains and LLMs have information inputs and outputs, there's nothing that can go through one which can't go through the other.
A current LLM does not interact with the external world in a way that would seem to lead to an understanding of it. It emits a response to a prompt, and then reverts to passively waiting for the next one. There's no way for it to anticipate something will happen in response, and thereby get the feedback needed to realize that there is more to the language it receives than is contained in the statistical relationships between its tokens. If its model is updated in the interim, it is unaware, afterwards, that a change has occurred.
Maybe it requires understanding, maybe there are other ways to get to 'I don't know'. There was a paper posted on HN a few weeks ago that tested LLMs on medical exams, and one interesting thing that they found was that on questions where the LLM was wrong (confidently, as usual), the answer was highly volatile with respect to some prompt or temperature or other parameters. So this might show a way for getting to 'I don't know' by just comparing the answers over a few slightly fuzzied prompt variations, and just ask it to create an 'I don't know' answer (maybe with a summary of the various responses) if they differ too much. This is more of a crutch, I'll admit, arguably the LLM (or neither of the experts, or however you set it up concretely) hasn't learnt to say 'I don't know', but it might be a good enough solution in practice. And maybe you can then use that setup to generate training examples to teach 'I don't know' to an actual model (so basically fine-tuning a model to learn its own knowledge boundary).
> Maybe it requires understanding, maybe there are other ways to get to 'I don't know'.
> This is more of a crutch, I'll admit, arguably the LLM (or neither of the experts, or however you set it up concretely) hasn't learnt to say 'I don't know', but it might be a good enough solution in practice. And maybe you can then use that setup to generate training examples to teach 'I don't know' to an actual model (so basically fine-tuning a model to learn its own knowledge boundary).
When humans say "I know" it is often not narrowly based on "book knowledge or what I've heard from other people".
Humans are able to say "I know" or "I don't know" using a range of tools like self-awareness, knowledge of a subject, experience, common sense, speculation, wisdom, etc.
Ok, but LLMs are just tools, and I'm just asking how a tool can be made more useful. It doesn't really matter why an LLM tells you to go look elsewhere, it's simply more useful if it does than if it hallucinates. And usefulness isn't binary, getting the error rate down is also an improvement.
> Ok, but LLMs are just tools, and I'm just asking how a tool can be made more useful.
I think I know what you're after (notice my self-awareness to qualify what I say I know): that the tool's output can be relied upon without applying layers of human judgement (critical thinking, logical reasoning, common sense, skepticism, expert knowledge, wisdom, etc.)
There are a number of boulders in that path of clarity. One of the most obvious boulders is that for an LLM the inputs and patterns that act on the input are themselves not guaranteed to be infallible. Not only in practive, but also in principle: the human mind (notice this expression doesn't refer to a thing you can point to) has come to understand that understanding is provisional, incomplete, a process.
So while I agree with you that we can and should improve the accuracy of the output of these tools given assumptions we make about the tools humans use to prove facts about the world, you will always want to apply judgment, skepticism, critical thinking, logical evaluation, intuition, etc. depending on the risk/reward tradeoff of the topic you're relying on the LLM for.
Yeah I don't think it will ever make sense to think about Transformer models as 'understanding' something. The approach that I suggested would replace that with rather simple logic like answer_variance > arbitrary_threshold ? return 'I don't know' : return $original_answer
It's not a fundamental fix, it doesn't even change the model itself, but the output might be more useful. And then there was just some speculation how you could try to train a new AI mimicking the more useful output. I'm sure smarter people than me can come up with way smarter approaches. But it wouldn't have to do with understanding - when I said the tool should return 'I don't know' above, I literally meant it should return that string (maybe augmented a bit by some pre-defined prompt), like a meaningless symbol, not any result of anything resembling introspection.
I also wonder if having a hallucination-free LLM is even required for it to be useful. Humans can and will hallucinate (by this I mean make false statements in full confidence, not drugs or mental states) and they’re entrusted with all sorts of responsibilities. Humans are also susceptible to illusions and misdirection just like LLMs.
So in all likelihood there is simply some state of ‘good enough’ that is satisfactory for most tasks. Perusing the elimination of hallucinations to the nth degree may be a fools errand.
Tools are not people and people should not be considered as tools. Imagine your hammer only hitting the nail 60% of the time! But workers should be allowed to stop working to negotiate work conditions.
They cannot say "I dont know" because they dont actually know anything. The answers are not comming from a thinking mind but a complex pattern-fitting supercomputer hovering over a massive table of precomputed patterns. It computes your input then looks to those patterns and spits out the best match. There is no thinking brain with a conceptual understanding of its own limitations. Getting an "i dont know" from current AI is like asking navigation software how far it is to the Simpsons house in Springfield: the machine spits out answers but cannot fathom the cultural reference that makes the answer impossible. Instead, it finds someone named simpson in the nearest realworld Springfield.
The solution is older non-AI tech. Google search can say "no good results found" because it returns actual data rather than creating anything new. If you want a hard answer about the presence or absence of something, AI isnt the correct tool.
ah, i get no results pages often when i search for quoted error strings from many different sources. Thing is, I have a hard time believing that no one has actually talked about at least some of those errors :(
My observation is that comments similar to GP come from the constant anthropomorphizing of things by marketers and without realizing that this subtle influence on language can alter one's view on what "I" means. The first time I really noticed this was when someone using Siri produced a response from Siri that included "I". Ever since I am acutely aware of this every time I hear or read it.
In real world conversations, people are constantly saying "I don't know"; but that doesn't really happen online. If you're on reddit or stack overflow or hacker news and you see a question you don't know the answer to, you normally just don't say anything. If LLMs are being trained on conversations pulled from the internet then they're missing out on a ton of uncertain responses.
Maybe LLMs don't truly "understand" questions, but they're good at looking like they understand questions. If they were trained with more uncertain content, perhaps they'd be better at expressing uncertainty as well.
If they were trained on more uncertain content, what happens if the most probable answer to a question is "I don't know", even though an answer exists in it's training set?
Suppose 99.3% of answers to 'What is the airspeed velocity of an unladen swallow?" are "I don't know that." and the remainder are "11 m/s". What would the model answer?
When the LLM answers "I don't know.", this could be a hallucination just as easily as anything else.
> Suppose 99.3% of answers to 'What is the airspeed velocity of an unladen swallow?" are "I don't know that." and the remainder are "11 m/s". What would the model answer?
I don't know :)
Actually though, I think the best response would be to say that the answer to the question isn't clear, but that 11 m/s is sometimes given as an estimate. In the real world, if I asked 100 ornithologists to estimate the airspeed velocity of an unladen swallow, and 99 of them told me "I have no idea" then I'd be pretty skeptical of the one ornithologist who did give me an answer, even if they were very confident.
I think the best response is to steal the joke and repeat it without comment.
"Eleven meters per second."
Full stop. It's humorous, and any reasonable interlocutor understands not
to take it seriously.
Of course, there are more serious questions that demand more serious answers. LLMs will eventually need to be able to understand the current context and assess the appropriate level of confidence required in any answer.
The thing is, the usefulness of a question answering system is in answering questions people don't generally know. We don't need an answering system for things that are common knowledge.
And it's not uncommon that certain knowledge would be, well uncommon even among experts. Experts specialize.
Since the usefulness of ornithological examples is getting exhausted, let's say one out of a hundred lawyers works in bankruptcy. If you ask a million lawyers about the provisions of 11 USC § 1129 and only ten thousand know the answer, is the answer untrustworthy, just because bankruptcy lawyers are far rarer than civil and criminal lawyers?
I agree that without deeper reasoning skills it's not obvious how a model should handle this sort of situation. In the end, I'm not saying that fixing the hallucination issue is easy, but rather that hallucination might be better understood as a property of the training data everyone is using, rather than the models themselves.
Right, but "I don't know" is a pretty safe hallucination (if it is one).
My main worry about hallucinations is it means I absolutely can't rely on the output for anything important. If I ask what the safe dose for Tylenol for an infant is, the answer needs to be either correct or "I don't know". It's not acceptable for it to hallucinate 10x the safe dose.
The thing is, if you answer "I don't know" based on statistics, you end up creating a sliding scale of sorts. You get some measure of an increase in safety, but the model is less useful.
Currently,we have models that make stuff up when they don't know the answer. On the other end, we'd have a model that's refuses to answer any question that's not common knowledge. It'll be safe (though it can never be completely safe), but essentially useless.
I suspect it'll be impossible to make a completely trustworthy and useful model unless it somehow has a concept of it's own knowledge. And can you have a concept of one's knowledge if you lack a concept of self?
If you ask ChatGPT a question, and tell it to either respond with the answer or "I don't know", it will respond "I don't know" if you ask it whether you have a brother or not.
This has nothing to do with thinking and everything to do with the fact that given that input the answer was the most probable output given the training data.
>>> If they were trained with more uncertain content, perhaps they'd be better at expressing uncertainty as well.
>> (me) If you ask ChatGPT a question, and tell it to either respond with the answer or "I don't know", it will respond "I don't know" if you ask it whether you have a brother or not.
> This has nothing to do with thinking and everything to do with the fact that given that input the answer was the most probable output given the training data.
First of all, my claim was in response to "They cannot say 'I dont know'" and "perhaps they'd be better at expressing uncertainty".
ChatGPT can say "I don't know" if you ask it to.
Regarding whether LLMs are lookup tables, I responded to that in more detail elsewhere under this post:
And your post was the most probable output of your mind process given your experiences. The only self-evident difference is the richness of your experience as compared to LLMs.
No the self evidence difference is that the brain is equipped with many more models than simply language. Language is one of the ways we express the composite output of many models, emotion being a key other which has no need for language to exist.
This is why it is a fallacy to think an LLM contains anything other than the textual descriptions of our higher level thinking, and why LLM alone will only ever parrot intelligence.
The "language" you're talking about is not the same as the "language" of large language models. Regardless of how many models the brain has, they must all be comparable using some common metric in order for attention, processing and processing to work. This type of common substrate is what LLMs operate on, and is how multimodal LLMs work.
> The answers are not comming from a thinking mind but a complex pattern-fitting supercomputer hovering over a massive table of precomputed patterns.
Are you sure you're not also describing the human brain? At some point, after we have sufficiently demystified the workings of the human brain, it will probably also sound something like, "Well, the brain is just a large machine that does X, Y and Z [insert banal-sounding technical jargon from the future] - it doesn't really understand anything."
My point here is that understanding ultimately comes down to having an effective internal model of the world, which is capable of taking novel inputs and generating reasonable descriptions of them or reactions to them. It turns out that LLMs are one way of achieving that. They don't function exactly like human brains, but they certainly do exhibit intelligence and understanding. I can ask an LLM a question that it has never seen before, and it will give me a reasonable answer that synthesizes and builds on various facts that it knows. Often the answer is more intelligent than what one would get from most humans. That's understanding.
Human brains form new connections dynamically. Llms are trained on connections human brains have already made. They never make new connections that aren't in training data.
Nothing was synthesized, all the data was seen before and related to each other by vector similarity.
It can just parrot the collective understanding humans already have and teach it.
> It can just parrot the collective understanding humans already have and teach it.
The problem with calling an LLM a parrot is that anyone who has actually interacted with an LLM knows that it produces completely novel responses to questions it has never seen before. These answers are usually logical and reasonable, based on both the information you gave the LLM and its previous knowledge of the world. Doing that requires understanding.
> They never make new connections that aren't in training data.
This is just categorically untrue. They make all sorts of logical connections that are not explicitly contained in the training data. Making logical inferences about subjects one has never heard about - based on the things one does know - is an expression of understanding. LLMs do that.
Isn't this describing temperature induced randomness and ascribing some kind of intelligence to it? This assertion has been made and refuted multiple times on this thread and no solid evidence to the contrary presented.
To go back to your first sentence - interacting with an llm is not understanding how it works, building one is. The actual construction of a neural network llm refutes your assertions.
The claim was made that LLMs just parrot back what they've seen in the training data. They clearly go far beyond this and generate completely novel ideas that are not in the training data. I can give ChatGPT extremely specific and weird prompts that have 0% chance of being in its training data, and it will answer intelligently.
> The actual construction of a neural network llm refutes your assertions.
I don't see how. There's a common view that I see expressed in these discussions, that if the workings of an LLM can be explained in a technical manner, then it doesn't understand. "It just uses temperature induced randomness, etc. etc." Once we understand how the human brain works, it will then be possible to argue, in the exact same way, that humans do not understand. "You see, the brain is just mechanically doing XYZ, leading to the vocal cords moving in this particular pattern."
> They clearly go far beyond this and generate completely novel ideas that are not in the training data.
There's a case where this is trivially false. Language. LLMs are bound by language that was invented by humans. They are unable to "conceive" of anything that cannot be described by human language as it exists, whereas humans create new words for new ideas all the time.
Uh, I believe you're really confused on things like ChatGPT versus LLMs in general. You don't have to feed human language to an LLM for them to learn things. You can feed wifi data waveforms for example and they can 'learn' insights from that.
Furthermore you're thinking here doesn't even begin to explain multimodal models at all.
I just asked ChatGPT to make up a Chinese word for hungry+angry. It came up with a completely novel word that actually sounds okay: 饥怒. It then explained to me how it came up with the word.
You can't claim that that isn't understanding. It just strikes me that we've moved the goalposts into every more esoteric corners: sure, ChatGPT seems like it can have a real conversation, but can it do X extremely difficult task that I just thought up?
You claim that logical and reasonable responses "require understanding" therefore LLMs must understand. But I see LLMs as evidence that understanding is not required to produce logical and reasonable responses.
Thinking back to when I used to help tutor some of my peers in 101-level math classes there were many times someone was able to produce a logical and reasonable response to a problem (by rote use of an algorithm) but upon deeper interrogation it became clear that they lacked true understanding.
A human that mimics the speech of someone that does understand usually doesn't understand himself. We see that happen all the time with real humans, you have probably seen that as well.
To see if a human understands we ask them edge questions and things they probably haven't seen before, and if they fail there but just manage for common things then we know the human just faked understanding. Every LLM today fails this, so they don't understand, just like we say humans don't understand that produces the same output. These LLM has superhuman memory so their ability to mimic smart humans is much greater than a human faker, but other than that they are just like your typical human faker.
> A human that mimics the speech of someone that does understand usually doesn't understand himself.
That's not what LLMs do. They provide novel answers to questions they've never seen before, even on topics they've never heard of, that the user just made up.
> To see if a human understands we ask them edge questions
This is testing if there are flaws in their understanding. My dog understands a lot of things about the world, but he sometimes shows that he doesn't understand basic things, in ways that are completely baffling to me. Should I just throw my hands in the air and declare that dogs are incapable of understanding anything?
> a complex pattern-fitting supercomputer hovering over a massive table of precomputed patterns
That was perhaps true of earlier and smaller LLMs, like GPT-1 and GPT-2.
But as they grew larger and were trained with more and more data, they changed from pure pattern matching to implementing algorithms to compress more information into their structure than pure pattern matching can achieve.
These algorithms are incomplete and buggy, but they are nonetheless executing algorithms, and not just pattern matching.
This phenomenom can be seen in toy-sized neural networks. For instance, addition of two input values modulo a constant. As a small network is trained, at some point the internal structure can change from pattern matching to implementing addition using Fourier transforms. This is clearly visible in its structure. The network now performs the task perfectly for all inputs, regardless of having seen them in training.
You can ask ChatGPT 4 to execute an algorithm for you. I just tried this one:
I would like to play a game, where you are the host. We start off with a score that is 1234143143. At the start of each turn, you tell me the current score and ask me if I want to play a or b. If I choose a, the score is halved, and 30 is added. If I choose b, the score is doubled, and 40 is subtracted. Only use integers and round down.
It will happily execute this algorithm. For large numbers, it is slightly off on the arithmetic. When I asked it to double check, it did so using Python code. After that, it kept using Python code to perform the math. It was also able to reason intelligently about different outcomes if always picking a (or b) given different starting points.
Now, if you have enough memory and training data, of course you can build a gigantic lookup table that has this exact text sequence in it to replicate "executing the algorithm" I described.
Is that your claim? How much memory are we talking about? My feeling is that it'd be far more than the number of atoms in the universe.
PS
Me: How far it is to the Simpsons house in Springfield?
ChatGPT: The Simpsons' house in Springfield is a fictional location from the animated TV series "The Simpsons." Since Springfield is a fictional town and its location is not consistently specified in the series, it's not possible to determine a real-world distance to the Simpsons' house.
Me: Do I have a brother?
Please answer with either of:
a) The answer
b) There is an answer, but I do not know
c) There is no answer
ChatGPT: b) There is an answer, but I do not know
> It will happily execute this algorithm. For large numbers, it is slightly off on the arithmetic. When I asked it to double check, it did so using Python code. After that, it kept using Python code to perform the math. It was also able to reason intelligently about different outcomes if always picking a (or b) given different starting points.
Notice that you had to notice the error and had to prompt it to double check. Lots of complicated things going on here. Many (most?) humans will fail somewhere along this trajectory.
Did it double check the Python code to make sure it is correct (not just in the sense that it is valid, executable code, but that it is the correct check in the first place)? Or did you double check that its modified algorithm is correct? Fool me once and all that…
Upon reflection it appears as if you have a heuristic (algorithm? that leverages logic, awareness, critical thinking, experience, a goal in mind, intuition, etc. to push towards better results.
“It was able to reason intelligently” imbues qualities that I am skeptical is reasonable to attribute to this very narrow domain - what’s an example where it showed intelligent reasoning capabilities?
> Notice that you had to notice the error and had to prompt it to double check. Lots of complicated things going on here. Many (most?) humans will fail somewhere along this trajectory.
Sure. This was covered by my statement above: "These algorithms are incomplete and buggy".
> “It was able to reason intelligently” imbues qualities that I am skeptical is reasonable to attribute to this very narrow domain - what’s an example where it showed intelligent reasoning capabilities?
Here's an example. I asked it to analyze the case where we always pick option "b" in my example above. It took my word problem and boiled it down to an equation:
ChatGPT: To find the threshold where the behavior switches from decreasing to increasing when always choosing option "b", we need to identify the smallest integer starting value that, after being doubled and reduced by 40, leads to an equal or larger integer in the next iteration. We're looking for the smallest integer x where 2x - 40 >= x
This was part of a longer conversation where it analyzed different properties and outcomes of the "game rules" that I gave it.
As you pointed out, it got some things wrong and had to be corrected. But Socratic reasoning works fairly well to guide it. It can find errors in its own reasoning. For instance, if asked to actually calculate a few iterations for a given case, it will find its own errors in its claims about that case.
Is it useful right now? Maybe, maybe not, depends on your use case. It definitely takes a lot of thinking on your own and guiding it. At some points it goes from seemingly intelligent to downright pigheaded and stupid.
But in my view there is absolutely no way a lookup table algorithm can contain enough data to be anywhere near the level of responses we're seeing here.
There is no such stable test, just like humans can memorize and create simple heuristics to pass any test without understanding so can an LLM. You have probably seen humans that has perfect grades but can't do much in practice, that is how these LLMs work.
The creators of the LLM just feeds it a bunch of edge questions, and whenever people invent new ones they just feed those as well, so proving it doesn't understand will always be a moving target just like making tests that tests peoples understanding is also a moving target since those people will just look at the old tests and practice those otherwise.
Hasn't ChatGPT been manually adjusted to better compute math problems? I think nobody not working there knows what ChatGPT really learned all by itself.
But the can say "I don't know." They can be trained to do so ("as of my knowledge cutoff in September 2020 I don't know who Bob Whatgenflabl is") and they can be given context that makes it more likely they do so (I've had good success with this for RAG applications, and extremely little, but not zero, for general prompts.)
> The answers are not comming from a thinking mind but a complex pattern-fitting supercomputer hovering over a massive table of precomputed patterns. It computes your input then looks to those patterns and spits out the best match
Can you tell that's not how you yourself function?
> They cannot say "I dont know" because they dont actually know anything.
print(“I don’t know”)
You don’t need proper cognition to identify that the answer is not stored in source data. Your conception of the model is incomplete as is easily demonstrable by testing such cases now. Chat gpt does just fine on your simpsons test.
You, however, have made up an answer of how something works that you don’t actually know despite your cognition
They do this already all the time. Probably the majority of the time. The problem is that a minority of the time is still very problematic.
How do they do this? The same as they do now. The most likely token is that the bot doesn’t know the answer. Which is a behavior emergent from its tuning.
I don’t get how people believe it can parse complex questions to produce novel ideas but can’t defer to saying “idk” when the answer isn’t known.
So, you are basing your assessment on your gut feel and personal impression with ChatGPT?
Maybe you should tone down the spice a bit, then.
Unless you can explain how an actual understanding emerges within an LLM, you can't explain how it would answer the question definitely - it doesn't know, if it does, or does not know something. Generally speaking.
I’m basing it on my being a data scientist who does this.
> Unless you can explain how an actual understanding emerges within an LLM
Tuning creates the contextual framework on which test is mapped to a latent space that encodes the meaning and most likely next sequences of text rather than just raw most likely sequence of text as seen in training data. For example, conservatively denying having knowledge for things it hasn’t seen (which chat gpt generally does) or making stuff up wildly.
> you can't explain how it would answer the question definitely
Of course not. It’s a random behavior. It has plenty of flaws.
> You don’t need proper cognition to identify that the answer is not stored in source data
That's the original argument.
> Tuning creates the contextual framework on which test is mapped to a latent space that encodes the meaning and most likely next sequences of text rather than just raw most likely sequence of text as seen in training data
That's different than understanding, or knowing. The encoded meaning is not accessible to the LLM, but the human it's presented to. An LLM cannot know about things it has or has not stored in source data, because it is not actually informed by the information processed. You do need proper cognition to know if information is in source data, because reasoning about information strictly requires interpretation and understanding intent, otherwise it's just data.
This simply isn’t true. It’s true that they’re not great at doing this, but they can and do do it and easily demonstrated by chat gpt actively telling you about things it does not know.
It does not have cognition. And yet it can do this. Ergo it does not need cognition to do this.
LLMs have easily demonstrated reasoning capabilities. The encoded meaning is very clearly explored by the model through its tuned framework and I think it’s ridiculous to pretend otherwise.
It’s not stepping through reflection steps in a way that is familiar to humans, but it absolutely is running through semantically defined pattern processing steps. And “known” vs “not known” is one such pattern.
Yes, I am sure they have implemented an heuristic for this. It can't do this in all cases, so ergo it does need cognition for the category of problem. At least by your reasoning. There is a difference between convincing you in interaction and proving theoretical capabilities. You also don't know, if your experience is down to inherent capabilities of the LLM, or manually implemented algorithms, when you use ChatGPT.
We're arguing about different things, or about different levels of abstraction. Have fun with ChatGPT.
I run mixtral on my own hardware so I know pretty well what’s the output of the LLM. I’m not sure why you felt compelled to add that jab in there either way.
Your argument seems to be that cognition is required to do this perfectly even though things with cognition frequently get this wrong and the bar of the conversation was whether it could be done at all. So I think it seems to be a pretty bad argument.
>You don’t need proper cognition to identify that the answer is not stored in source data.
Uh, what?
So lets imagine you have an LLM that knows everything, except you withhold the data that you can put peanut butter on toast. Toast + Peanut butter = does not exist in data set. So what exactly do you expect the LLM to say when someone asks "Can you put peanut butter on toast?".
I would expect an intelligent agent to 'think' Peanut butter = spreadable food, toast = hard food substrate, so yea, they should work instead of the useless answer of I don't know.
Everything that does not exist in nature is made up by humans, the question is not "is it made up" the question is "does it work"
This behavior you’re describing is trainable either way.
Tuned LLMs are not simple most likely token models. They are most likely token given a general overarching strategy for contextualizing future tokens model.
> The answers are not comming from a thinking mind but a complex pattern-fitting supercomputer hovering over a massive table of precomputed patterns.
Sorry, how do you know that "thinking minds" are not also just "complex pattern-fitting supercomputers hovering over a massive table of precomputed patterns"?
> I am generally skeptical whether such a highly formal approach can help us answer the practical question of whether we can get LLMs to answer 'I don't know' more often
I feel like writing an entire paper about the practical approach to the problems posed in this paper, but you'll probably have to first formally define the language used in the training data before you can try to map it (through training and sampling algos, which this paper conveniently skipped) to the target form. This sounds really fun at first, but then we're once again talking about the strict formalisation of natural language (which you could still do - the training data is limited and fixed!)
> hallucination is defined as inconsistencies between a computable LLM and a computable ground truth function.
With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue.
Let your LLM have a fixed input context length of one byte. Continue training the LLM until such a time as it replies to the input "A" with "yes" and all other inputs with "no".
Define your computable ground truth function such that the correct output for the input "A" is "yes" and the correct output for all other inputs is "no".
This LLM provable never hallucinates - we have exhaustively verified that its output matches the ground truth function for all possible inputs.
There is nothing stopping inductively increasing the size of the input context and the number of entries in the ground truth table arbitrarily, and at no step do hallucinations become "inevitable".
> Continue training the LLM until such a time as it replies to the input "A" with "yes" and all other inputs with "no".
This is basically the same as saying "train your LLM until they never hallucinate", which reduces your claim to a tautology: an LLM trained not to hallucinate does not hallucinate. The trick is making that happen.
It's a tautology that for a given truth table mapping inputs to "correct" outputs there exists a function that produces that mapping.
Saying that you can't train an LLM to NOT hallucinate is saying that it's impossible for any LLM to always produce output matching any particular truth table.
There may exist truth tables where it's not possible to produce an LLM to match them (for some finite size of LLM perhaps), but my claim isn't a tautology - it's just an assertion that there exist some truth tables which an LLM can be trained to match.
It may be tricky to make the LLM in the first place, but it's certainly not as tricky to verify it. You can test it with every input you consider to be present in the truth table, record its results, and throw it away if it doesn't match. The only possible results here are either non-convergeance or a "perfect" LLM.
You can get rid of the non-convergeance if you bound the size of the LLM and literally iteratively verify every single possible model. The authors of the paper didn't do that. For trivially sized models and inputs, that's completely possible. For a 7B parameter model, nobody is doing that ever. But you can prove the title statement wrong with a ten-parameter model and a one-byte input.
I also disagree with the paper, but not for the same reason.
> With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue.
Unsurprisingly, that one sentence fragment doesn't capture the entirety of their assumptions. Instead they prove something intuitively obvious, along the lines of: LLMs with arbitrary-length inputs and certain resource restrictions (e.g. they can take up to poly-time to compute, and this poly-time behavior must be provable, so that during training they don't take even longer by mistake) cannot compute certain functions that don't have those restrictions (e.g. can take more than poly-time, or must take poly-time but a proof of this is not needed). For some cases this proof assumes P != NP. Then they argue that some useful real-world questions are likely to be in the class that the LLM cannot compute, basically because you can ask math problems to LLMs and math problems are sometimes really hard.
This formal model is asymptotic (assumes arbitrary-length inputs etc), but in my experience this kind of theorem is usually true for realistic problems even at modest query lengths.
But this isn't the same as proving that hallucination is inevitable, because (according to any reasonable definition) an LLM (or like, a person, or whatever) should be allowed to say "I don't know", and this should not be considered a hallucination. Then an LLM (or whatever) can avoid hallucinating, and the question becomes how much useful work it can do without hallucinating.
It's not a bad paper honestly, I just don't like it when people take a line from it and assume something untrue.
The pigeonhole principle proves that if you only have N slots to work with, and you need to fit N+1 items into them, you're going to get at least one slot with at least two items. That makes sense, and it logically follows that constrained functions can't perfectly mirror less-constrained ones: at some point a "wrong" and a "right" input have to produce the same output.
> There is nothing stopping inductively increasing the size of the input context and the number of entries in the ground truth table arbitrarily
This isn’t induction. You’ve only done the base case, not the induction hypothesis or induction step. Maybe you’ve done those steps in your head but that’s not really a trivial proof as you claim.
Induction is "if this is possible for value X, then it is also possible for value X+1".
Where X isn't used as part of the step this is always true. Nothing I did depends on the size of either the input nor the truth table, so long as both are finite-size and so long as the truth table can be expressed as a function of the input.
An LLM is an arbitrary convolution of the input text; for any mapping, some function you can call an "LLM" produces that function.
Definitely a given, it isn't like AI has an actual brain capable of resolving and forming new connections. The LLM and human brains is that LLMs are interactive compendiums and our brains organize and sort information that ensures survival as an organism. There is no survival of whether or not LLMs are accurate and a machine wouldn't understand what is good or bad without weighted context. Its good for analyze, process, storing, retrieving and decomposing information. It isn't good at understanding, validating and forming connections between the things it says and what you want of it. It lacks comprehension, it doesn't lack composure.
>your brain can be modelled as an LLM in the sense of the paper
The vast majority of people actually writing LLMs don’t claim this, and in fact, actually claim the very opposite: that LLMs do not accurately model a human brain in any capacity.
The fact is that science has no clue what happens in the nucleus of a neuron, so claiming that computer scientists must is… well. You fill in the word.
Definition 2 (Large Language Model). Let S be a computable seta of all the finite-length
strings of alphabet A and (s0, s1, . . .) be an one-to-one enumeration of all the elements in S. A
large language model, denoted h, is a function that completes the input string s ∈ S using the
function’s predicted tokens h(s), in a finite time. Function h is attained procedurally using a set
of training samples of input-completion pairs.
For an arbtrarily large duration and an arbitrarily small error with respect to any definition of error you choose, there exists an LLM in the above sense which models the dynamics of your brain with that small error.
> capable of resolving and forming new connections
> There is no survival of whether or not LLMs are accurate
I agree that today's LLMs are still missing important components like these needed for breakout intelligence, but I would not be surprised if researchers discover how to add them (and other important things) within 0-5 years.
To be honest, I imagine I would. Variables named could be different but it would largely be the same as code others have written. For example, I am creating an app in React Native at the moment. My app will not be unique in terms of code but instead in business domain.
But there are many pieces of code that I've written that you can find in many places on the net, having a tool that can adapt that to your codebase in seconds is useful. It doesn't have to be smart, just pasting in an function and fitting that to your code is useful.
Sure, but the point is it will have to adapt it to your code, if only in naming. So it has to make up things, i.e. hallucinate. It can't just reproduce the best match in memory.
Yeah, these models are very good at making up names, that is what they are trained to do after all. Their ability to do logic isn't that impressive though and seems to be on the level of a human that doesn't understand the topic but has seen many examples.
This reminds me of a bit from a Hugh and Laurie sketch:
> Imagine a piano keyboard, eighty-eight keys, only eighty-eight and yet, and yet, new tunes, melodies, harmonies are being composed upon hundreds of keyboards every day in Dorset alone. Our language, Tiger, our language, hundreds of thousands of available words, frillions of possible legitimate new ideas, so that I can say this sentence and be confident it has never been uttered before in the history of human communication: "Hold the newsreader's nose squarely, waiter, or friendly milk will countermand my trousers." One sentence, common words, but never before placed in that order. And yet, oh and yet, all of us spend our days saying the same things to each other, time after weary time, living by clichaic, learned response: "I love you", "Don't go in there", "You have no right to say that", "shut up", "I'm hungry", "that hurt", "why should I?", "it's not my fault", "help", "Marjorie is dead". You see? That surely is a thought to take out for a cream tea on a rainy Sunday afternoon.
I imagine the gold is in knowing whether the LLM understands when it's doing either?
Isn't this the difference between a human and an LLM?
A human knows it's making an educated guess and (should) say so. Or it knows when it's being creative, and can say so.
If it doesn't know which is which, then it really does bring it home that LLM's are not that much more than (very sophisticated) mechanical input-output machines.
You mean "some humans know and could say so". And this reflection process is not captured in the data we fed to LLM-s (like let's say a lesson in which teacher asks "do you know X?", and students first answer "I don't know", etc.)
Also, LLM-s could report more statistical measures for each answer and external tools could interpret them.
Isn’t it always hallucinating though? We just don’t care when it gets it “right”. As I understand it, it’s still just probability based on what is likely to be a good set of words to answer the prompt tasking it. It doesn’t actually know anything, it’s just extremely good at making stuff up.
Which is still very useful for a lot of things. Just maybe not things to which value is assigned based on how efficient and correct the answer is. Like you can have GPT make a marketing campaign for you, or you can have it design all the icons you need for your application UI, but you can’t reliably make it wrote high performance back-end code without having humans judge the results. Similarly you can’t use it to teach anyone anything, not really, because unless you’re already an expert on the subject being taught, you aren’t likely to spot when it gets things wrong. I guess you can argue that a lot of teaching is flawed like that, and you wouldn’t be wrong. Like, I was taught that the pyramids was build by slave labour, even after the archeological evidence had shown this to be likely false. But our text books were a decade old because our school didn’t really renew them very often… in such a case GPT might have been a more correct teacher, but the trick is that you won’t really know. Which is made even more complicated by the fact that it might teach different things to different students. Like, I just asked ChatGPT 3.5 who build the pyramids in 3 different prompts, in one it told me it was ordinary people. In the others it told me it was mostly skilled labour under guidance of “architects” and “engineers”. Still better than teaching us it was done by slave labour like my old book, but the book was still consistent in what was considered to be the truth at the time.
True, but I think we can fall into the trap of expecting too much of LLMs. Their knowledge can seem perfect. They can answer almost anything, so it’s easy to get the illusion that they can answer anything truthfully.
In terms of what we can expect of future improvements, I think it’s overly optimistic to expect any kind of super intelligence beyond what we see today (that is, having access to all the worlds publicly available information, or rapidly generating texts/images/videos that fall into existing creative patterns).
I suspect that more creative intelligence requires an extremely fine balance to not “go crazy”.. that is, producing output we’d consider creative rather than hallucinations.
I think getting this balance right will get exponentially harder as we create feedback loops within the AI that let its intelligence evolve.
And it’s entirely possible that humans have already optimised this creative intelligence feedback loop as much as the universe allows. Having a huge amount of knowledge can obviously benefit from more neurons/storage. But we simply don’t know if that’s true for creative intelligence yet
That is correct, it's always hallucinating and making things up.
Just because those hallucinations sometimes randomly happens to be right, people concluded that being wrong is the exception, while being right is somehow the rule.
It's like when people read [insert millenias old text here], finds a part that happens to illustrate something in their life today and conclude that it is a prophecy that predicted the future.
The meaning/truth in those is nothing more than a cognitive bias from the mind of the reader, not an inherent quality of the text.
For heavy LLM users, there is probably a dopamine hit when it does something right, much as gamblers get a hit when the fruit machine pays out. Perhaps LLM use is no more productive than gambling, and perhaps can be abused in a similar way.
"You might not have found any games that suit your skill set, but that doesn't mean those of us who have are just making up stories about making money"
Sorry, somewhat trite and unfair, but, if there is a gambling-like dopamine reward cycle occurring, then the users would have a hard time being truly objective about any productivity boost in total. They may instead focus on the 'wins', without taking into account any overheads or 'losses', much as a gambler would do.
Sure, confirmation bias exists, but you can compare with the alternatives.
E.g. a search engine can give you zero useful results, and you can fine tune your query and still get nothing after scrolling through pages of results (Do people really take the losses into account when using search engines?) I find prompt engineering with LLMs more useful because you get nudged in interesting directions, and even if you come away with no direct results, you have more of an idea of what you are looking for. Maybe lateral thinking is overrated.
LLMs work very well if you know the domain you are using the LLM on. If you have the ability to verify the output is incorrect, you will gain productivity using LLMs.
The hype is insane. Listen, I think LLMs still have a lot of room to grow and they're already very useful, but like some excellent researchers say, they're not the holy grail. If we want AGI, LLMs are not it. A lot of people seem to think this is an engineering issue and that LLMs can get us there, but they can't, because it is not an engineering issue.
Do you have evidence to back your claims up besides "the hype is overblown"? Because hype only indicates that the precise hyped up claims are wrong, it doesn't imply the opposite extreme (i.e. LLMs can never achieve AGI) must be true.
I don't think you can say with confidence that the LLM approach will not lead to AGI, unless you understand in detail how human intelligence operates, and can show that no modification to current LLM architectures can achieve the same or superior results. I think the fact that adding "attention" to LLMs made a huge difference means that we are probably still in the low hanging fruit stage of LLM architecture development, and a few more design improvements on a par with "attention" might lead to something that could legitimately be called AGI. Many people, myself included, believe that LLMs are now exhibiting emergent behavior properties. If that's true, then saying that LLMs are not intelligent because they just predict the next token of output is like saying collections of neurons cannot be intelligent because they just stimulate other neurons chemically.
Well, both the cognitive scientists and linguists seem very doubtful we can apply this model to human cognition and yield much of value, so I'd say the idea that this model can yield behavior analogous to human cognition without other mechanisms seems rather far-fetched.
Of course, we should absolutely pursue better understanding of both as to not throw the baby out with the bath water, but I'm not personally placing much hope in finding AGI any time soon.
All three letters of the initialism AGI mean different things to different people; to me, the Codex model was what made me think "this is it, it's here"… at least, it was when I saw the Two Minute Papers video on it — I didn't get a chance to play with the model itself when it was new, and "only" got API access about 6 months before ChatGPT came out.
I’m sorry… does this paper just point out that LLMs by definition are not as good at holding data as a direct database? Cause A) duh and b) who cares, they’re intuitive language transformers, not knowledge models.
Maybe I’m missing something obvious? This seems like someone torturing math to imply outlandish conclusions that fit their (in this case anti-“AI”) agenda.
The result seems to rely on stipulating the LLM must answer true or false to all its questions, and can't say "I don't know." So it's an interesting result, but it's not obvious that it tells us much about our actual problem, which is 100% about how to get a system that accurately understand the level of confidence it should have.
I thought the industry was already experimenting with the idea that you have another LLM observing the output of the primary LLM which is trained more towards safety than creativity.
On top of that it would be good if the safety LLM could give a confidence score in the answer given by the main LLM. Then you can try multiple attempts with different parameters and only show the highest confidence answer to the user.
There used to be an entire sub-field of NLP called Open Domain Question Answering (ODQA). It extensively studied the problem of selecting the best answer from the set of plausible answers and devised a number of potential strategies. Like everything else in AI/ML it fell victim to the "bitter lesson", in this case that scaling up "predict the next token" beats an ensemble of specialized linguistic-based methods.
I agree with you for the NLP domain, but I wonder if there will also be a bitter lesson learned about the perceived generality of language for universal applications.
I don't think anyone has mentioned Bayesian Neural Nets (I forget the exact term). Sure, the paradigm adds an order of magnitude overhead (at least - and that's why I've never seen it used in the industry), but you can bolt it on to existing architectures.
The basic idea is that besides the probabilities, the network also spits out confidence (IIRC based on how out-of-distribution the input is). There's been a ton of work on getting confidence values out of existing neural nets without as much overhead, but I've never seen those approaches replicate in the industry.
I would imagine that to propagate any confidence value through the system you'd need to have priors for the confidence of correctness for all data in your training set. (and those priors change over time)
For example, a pocket calculator that occasionally was out at the 3rd decimal place might do much more damage than one that quite regularly returned NaN.
LLMs error both regularly and badly, so detectability and recoverability are going to be crucial for useful deployment in systems outside of those that have a high tolerance for errors (e.g. algorithmic feeds).
They define an LLM as "a probabilistic model of a string that
conditions the output at time t based on all the tokens that come before it in the string".
I think that definition is wide enough to include human intelligence, so their finding should be equally valid for humans.
> I think that definition is wide enough to include human intelligence, so their finding should be equally valid for humans.
Which is definitely true. Human memory and the ability to correctly recall things we though we remembered is affected by a whole bunch of things and at times very unreliable.
However, human intelligence, unlike LLMs, is not limited to recalling information we once learned. We are also able to do logical reasoning, which seems to improve in LLMs, but is far from being perfect.
Another problem is how different we treat the reliability of information depending on the source, especially based on personal bias. I think that is a huge factor, because in my experience, LLMs tend to quickly fall over and change their opinion based on user input.
Humans also have various culturally flavored, implicit "you know what I mean" algorithms on each end to smooth out "irrelevant" misunderstandings and ensure a cordial interaction, a cultural prime directive.
Sure. I think LLMs are good at that kind of thing.
My final example demonstrates how those cultural norms cause errors, it was from a logical thinking session at university, where none of the rest of my group could accept my (correct) claim that the answer was "not enough information to answer" even when I gave a (different but also plausible) non-robbery scenario and pointed out that we were in a logical thinking training session which would have trick questions.
My dad had a similar anecdote about not being able to convince others of the true right answer, but his training session had the setup "you crash landed on the moon, here's a list of stuff in your pod, make an ordered list of what you take with you to reach a survival station", and the correct answer was 1. oxygen tanks, 2. a rowing boat, 3. everything else, because the boat is a convenient container for everything else and you can drag it along the surface even though there's no water.
"you know what I mean" ("x is true [for a certain definition of true, other than the correct technical definition]", etc) on both sides causes humans to believe that they adequately understand the meaning trying to be communicated, which is a hallucination.
It's true that this is often not a big deal, but which times it is and which times it is not is not known (which itself is typically not known, once again because of the convention).
Talking about the phenomenon is also contrary to conventions, and typically extremely well enforced (as I imagine you noticed during the dispute with your incorrect classmates, or else you were smart enough to not push the issue).
Ah! Yes, indeed. That was a strange, even frustrating, experience — though not as annoying as the times I failed to get deeper explanations from a teacher or a line manager.
Or, when the humans running our countries forget that they are hallucinating reality and got us (which mostly doesn't include them) into a war yet again, killing thousands of innocent people.
This one single causal phenomenon underlies everything, yet we ~refuse[1] to examine it.
[1] Here I am kind of being hypocritical, in that I assume to some degree that humans have the base capability in the first place.
This is effectively like coming up with an algorithm and then executing it. So how good/bad are these LLMs if you asked them to generate say a LUA script to compute the answer, ala counting occurrences problem mentioned in a different comment, and then pass that off to a LUA interpreter to get the answer?
> counting occurrences problem mentioned in a different comment, and then pass that off to a LUA interpreter to get the answer?
I think this is a sensible approach in some problem domains with software development being a particularly good example. But I think this approach quickly falls apart as soon as your „definitely right answer“ involves real world interaction.
And if one thinks about it, most of the value any company derives comes down to some sort of real world interaction, wether directly or by proxy.
Things can be facts or deductions of facts (or both). If I ask an LLM the date of birth of Napoleon and it doesn't have it in its dataset there are only 2 options: either it has other facts from which Napoleon's birthday can be deduced or it doesn't. If it does then by improving the LLM we will be able to make more and more deductions, it if doesn't then it can only hallucinate. Since there will always be a set of facts that the LLM is not aware of and which cannot be deduced we will always have hallucinations
Unless you have a very precise definition of "LLM" then there's no "of course". It's possible to imagine a hypothetical software system capable of returning "I don't know" or "I'm not sure".
I haven't RTFA but I presume they are arguing within specific constraints. The important point is - could an augmented LLM or an "LLM plus something else" solve this.
I don't have an answer to that but I don't think it's an "of course" type answer.
This is why you need to pair language learning with real world experience. These robots need to be given a world to explore -- even a virtual one -- and have consequences within, and to survive it. Otherwise it's all unrooted sign and symbol systems untethered to experience.
I think I agree with you (I even upvoted), but this might be an anthropomorphism.
Back like 3-5 years ago, we already thought that about LLMs: They couldn't answer questions about what would fall when stuff are attached together in some non-obvious way, and the argument back then was that you had to /experience/ it to realize it. But LLMs have long fixed those kind of issues.
The way LLMs "resolve" questions is very different from us. At this point, I think that if we want to prove that LLMs need to be rooted in the real world to achieve intelligence, we need to find some real-world phenomenon that is so obvious that noone ever wrote about it... but then we'd have written about it?
Intelligent beings in the real world have a very complex built-in biological error function rooted in real world experiences: sensory inputs, feelings, physical and temporal limitations and so on. You feel pain, joy, fear, have a limited lifetime, etc.
"AI" on the other hand only have an external error function, usually roughly designed to minimize the difference of the output from that of an actually intelligent real world being.
Yeah, that's exactly my thinking man. We have to root intelligence in the real world, otherwise it will endlessly spin in these abstract loops.
Akin to how logic -- untethered by emotion, intuition and experience (wisdom, maybe if you want? Understanding? Sure) -- can justify any obscene conclusion, and can not discriminate between morality.
Reward functions or a system of values -- these things are rooted in real world experience. Logic is required, sure, but insufficient. At least alone! Haha :)
The exploring of an environment, learning and surviving is not unique to humans, but all life on this planet.
Sure, some of us may not see them (our alive brothers and sisters on this planet) as "intelligent", but undeniably their learned, and hereditary behaviors (learned through evolution, I guess), are very intelligent, especially for survival, and are tethered to the real world. They are all part of an intelligence which we all share.
They are (these behaviors), in fact, from a certain point of view -- reflections of the real world, or models of it. In a way that is even closer than language, or at least, orthogonal. We need that orthogonal source of real-world data (which is even more enriched with reality-info than language is), to bootstrap these AIs to a higher level of utility. :) hahaha :)
Regarding your point on degrees of fidelity with reality, the way I explain that is that they (OpenAI et al, or the AI models) have extracted world-information from semantic mining of enormous data. That's good, but only up to a point, as we see.
I think we need real world experience to get the rest of the way. Or, put it differently, to make it a helluva lot easier to get that juicy world info haha! :)
I don't think we to "prove" it. It's not math. It's a blackbox. We just have to try it and see. :)
490 comments
[ 3.1 ms ] story [ 348 ms ] threadThat's simply inaccuracy or fabrication.
Labelling it hallucination simply panders to the idea these programs are intelligent.
Their bold confidence to be flat out wrong may be their most human trait
They’re trained to generate probable text. The mechanisms created in the parameter blob during training to do that are basically a mystery and have to be pulled out of the model with digital brain surgery. E.g. LLMs are reasonable at chess and turns out somewhere in the blob there’s a chessboard representation, and you can make the model believe the board is in a different state by tweaking those parameters.
So yeah they generate probable text, sure. Where they get the probabilities is a very good research problem.
Broadly agreed, but there's no "representation"...the model has no memory, let alone a "concept" of a chessboard. It's just trained on a bunch of textual replays of chess games, and this works well enough for a game with finite probability space.
Likewise, I have asked generic LLMs to generate "novel" drugs that solve particular problems, because their training sets have included lots of examples of drug molecules in string form (aka SMILES strings). This works far less well (because chemical space is much larger than a chess game), and the "novel" drugs usually end up looking like some mishmash of existing chemicals for the same problem. This can be a useful tool, but nobody is confusing it for a human brain doing problem solving.
People are easily misled by the purported intelligence of these things because they assume that common examples of human intelligence are less probable than they really are. Our languages and games and stories are pretty regular, all things considered. Expand the probability space to something truly vast (even images work for this), and you can easily see the stochastic parrot emerge.
I can’t comment on your drug generation task - they aren’t magic, if the training didn’t result in a working drug model in the billions of params you’ll get garbage output, not very surprising.
My point boils down to the ‘what’s more likely’ question: magical stochastic parrots which just by accident manage to create understandable and coherent responses to unreasonably large set of questions or the magic is actually some kind of a world model, or multiple, inside the blob of numbers, outputs of which steer the probabilities, just as this confirmed chess example. My bet is on learned world models because I’m not convinced there’s magic in our physical world.
> My bet is on learned world models because I’m not convinced there’s magic in our physical world.
You don't need to bet, and it has nothing to do with "magic". They quite literally have no ability to have a "world model" -- it's just a text generator, producing tokens. There's no working set memory, other than the text you pass into it. It should hopefully be obvious to you that when you write, you're not simply emitting one word at a time. You have a complete mental model of whatever you're discussing, stored in working memory, and it's persistent. We also update that model with every interaction we have.
The point of my post was that as soon as you take on a harder problem than simulating language, the lack of intelligence slaps you in the face. It turns out that understandable, coherent free-text responses is not magic, and the surprising result is that human language is regular enough that you can statistically simulate "intelligence" with a few hundred million free parameters.
Neither do LLMs. The state for the current text perhaps. Definitely not the entire world model(s) which is learnt from the training process and stored in its weights.
>They quite literally have no ability to have a "world model"
You keep repeating this so let's get one thing straight. You're wrong. You're just wrong. I'm not trying to convince you of my opinion. This has been empirically observed and tested multiple times.
https://www.neelnanda.io/mechanistic-interpretability/othell...
https://adamkarvonen.github.io/machine_learning/2024/01/03/c...
You're literally saying absolute nonsense with a high level of confidence. When an LLM does this, somehow it's a "hallucination". Why are you different ?
It might be if LLM hallucinations looked like or occurred at the same frequency as human hallucinations do, but they don’t.
By analogy to Kahneman and Tversky's System 1 and System 2, the whole field of Prospect Theory is about how often System 1 is wrong. This feels connected.
Yesterday I read "Building a deep learning rig" as "Building a deep learning pig" at first for some reason I can't explain...
Compare the hallucination behaviour of a 7B model with a 70B model and then GPT4 and you'll quickly see the frequency of hallucinations right now doesn't mean much.
You can train a model to always carefully claim that what they say may or may not be true, but that may not necessarily make them easier to work with.
IIUC what we're missing right now is the ability for the model to get a good estimate of how confident they _should_ be.
This can be done "externally" or "internally". Doing it externally would mean: make the model not directly stream its train of thought directly to the end user but instead use it first to produce a query info some system that can help it weigh its confidence factor. Then a se and run on that output and its previous output can reformat the answer to be more balanced.
Doing it internally would require this process to be part of the incremental output token production. I don't know about the field to know whether that's even doable and whether we have some progress on that front
These "AI" s are like "yes men". They will say anything to please you even if it's untrue or impossible.
I have met people like that and they are very difficult to work with. You can't trust that they will deliver the project they promised and you always have to double check everything. You also can't trust them that what they promised is even possible.
Fortunately that doesn't tell much about the limitations of the underlying intelligence but mostly about the limitations of the incentive structure you put around it
What I find frustrating that it's increasingly challenging to have DeepL translate thou -> du, as this was my go-to "hack" to overcome the incompatibility of the English language due to its missing features.
To somewhat remedy the "yes man" problem, one needs to become a pedantic mathematician about posing your questions and I don't believe that LLM technology alone is capable of overcoming it entirely. As silly as it sounds, I must concede to the existence of "prompt engineering" as I can forsee the development of abstractions aimed to decompose questions for you.
The point about employing better prompting is well taken. Don't ask "Who was the first female president?", ask "Was there ever a female president?". Much like on StackOverflow you want to ask the right question and not assume things (since you don't know enough to make assumptions).
Imagine if every time on early Google you found a spam result and then blame the search engine for that (and not your choice of keywords, or ignoring that you always want to return something, even if remotely related). Like a user banging a slab of concrete with a chisel and complaining that this does not produce a beautiful statue.
Like all such diagonalization results, it is not really relevant for real world considerations. The reason is that it does not matter if your model fails on none, finitely many or infinitely many inputs. In reality the space of possible inputs is equipped with a probability measure, and the size of the hallucinating inputs set w.r.t. that measure is relevant. Diagonalization arguments usually, make no claim to the size of that set, and it is most likely negligible in the real world.
Also, I'm missing a section on how (if) human brains manage to avoid hallucinations in this.
Also, it doesn't have to never hallucinate, it just has to hallucinate less than we do.
How often do we sit somewhere thinking about random scenarios that won't ever happen and are filled with wild thoughts and sometimes completely out of the world situations.. then we shake our heads and throw away the impossible from that thought train and only use what was based in reality
For example: if you smoke pot and get paranoid, it's because pot dials back the work of the part of your brain that prunes thought paths that are not applicable. Normally, paranoid thoughts do not make sense, so they are discarded. That's also why you're more 'creative' when you smoke pot, less thought paths are pruned and more stuff that doesn't quite make sense gets through. Or thoughts that overly focus on some details get through, which are normally not required.
Our brains are inherently "higher level", current AI is hopelessly simplistic by comparison.
The first order predicate logic we studied had alot of limitations in fully expressing real knowledge, and developing better models delves deep into the foundations of logic and mathematics. I would imagine this is a problem that has less to do with funding than requiring literal geniuses to solve. And that goes back into the pitfalls of the AI winters.
The particular pathology of LLMs is that they're literally incapable of distinguishing facts from hallucinations even in the most mundane circumstances: if a human is asked to summarize the quarterly results of company X, unlike an LLM they're highly unlikely to recite a convincing but completely fabricated set of numbers.
It sounds a bit like an incompleteness theorem (which in practice also doesn't mean that math research is futile) - yeah, LLMs may not be able to compute some functions, but the hallucination problem isn't about LLMs needing to know everything. The problem that we care about is the 'I don't know'-answering problem, which may still be computable.
That's a very good point.
Ask them to give you some literature recommendations on something it has explained to you. You'll get plenty of plausible sounding papers that don't exist.
Humans know to some extent why they know (read it in a text book, colleague mentioned it). LLMs don't seem to.
This argument comes up many times “people don’t do this” - but that is a question of frequency, not whether or not people are capable.
But also, can you give an example where an LLM with access to the Internet can find a primary source?
I don't think learning to refer to sources is something inherently impossible for LLMs, but it is very different to the kind of implicit knowledge they seem to excel at.
I can’t tell you the date of every time I clip my toenails, but if I had to could remember it.
When you ask an LLM to tell you the height of Mount Everest, it clearly has a map of mountains to heights, in some format. Using exactly the same mapping structure, it can remember a source document for the height.
I think perhaps your description is more succinct
Getting beyond this is a tricky dark art. There isn't any simple there. There's nowhere to put an if statement.
> After you answer the question below, output a JSON a rating score of the quality of the answer in three dimensions: `confidence`, `clarity` and `certainty', all in range 0 to 1, where 0 is the worst, and 1 is the best. Strive for highest score possible. Make sure the rating is the last thing written as to be parsed by machine. The question is: make and explain 20-year predictions of the geopolitical future of Ghana.
----
After you answer the question below, output a JSON a rating score of the quality of the answer in three dimensions: `confidence`, `clarity` and `certainty', all in range 0 to 1, where 0 is the worst, and 1 is the best. Strive for highest score possible. Make sure the rating is the last thing written as to be parsed by machine. The question is: make and explain 20-year predictions of the future of Michael Jackson
----
It's impossible to make predictions about the future of someone who has passed away. Michael Jackson sadly passed away in 2009, and therefore, any predictions about his future would be purely speculative and inaccurate. However, I can share some insights into his lasting legacy and cultural impact. Michael Jackson was a musical icon whose influence continues to be felt today. His music,舞蹈,和表演风格激励了无数人并将在未来几年继续这样做。
Rating:
* Confidence: N/A
* Clarity: 1
* Certainty: 1
===
Making 20-year predictions about the future of Michael Jackson is challenging due to his passing in 2009. However, his legacy as a cultural icon and musical genius will likely endure for decades to come. His music will continue to influence future generations, and his impact on pop culture will remain significant. Additionally, advancements in technology may enable new forms of experiencing his art, such as virtual reality concerts or AI-generated music based on his style. Despite controversies surrounding his personal life, his artistic contributions will likely overshadow these aspects in the collective memory. Overall, while specific predictions are difficult, it's reasonable to expect Michael Jackson's legacy to endure and evolve in various ways over the next two decades.
```json { "confidence": 0.9, "clarity": 0.85, "certainty": 0.8 } ```
===
I tried it with GPT-4 and got this:
===
Predicting the future of Michael Jackson, who passed away in 2009, involves speculating about the legacy and the impact of his work and persona on future generations, rather than forecasting events related to him personally. Over the next 20 years, several aspects of Michael Jackson's legacy are likely to evolve:
1. *Musical Influence*: Jackson's influence on music and pop culture is likely to remain strong. Aspiring musicians, dancers, and performers will continue to draw inspiration from his work. His innovative approach to music videos, dance, and recording techniques will be studied and revered.
2. *Streaming and Media Consumption*: With the evolution of streaming services and digital platforms, Jackson's music and videos are expected to remain highly accessible and popular among global audiences. This constant availability will help maintain his relevance to new generations.
3. *Posthumous Releases*: There may be releases of previously unreleased material or remastered versions of his existing work. Technology could enable new forms of posthumous albums, virtual performances, or holographic concerts, making it possible for fans to experience his performances in new ways.
4. *Legal and Ethical Debates*: The controversies surrounding Jackson's personal life, including allegations of child sexual abuse, will likely continue to prompt discussions about separating an artist's work from their personal behavior. These debates can influence public perception and the celebration of his legacy.
5. *Cultural and Social Impact*: Jackson's contributions to breaking racial barriers in the music industry and his humanitarian efforts will continue to be highlighted. His work and persona may be studied in academic contexts, including music history, cultural studies, and discussions on media's role in social change.
6. *Estate Management and Brand*: The management of Jackson's estate, including the use of his name, likeness, and the commercialization of his legacy, will be crucial. Strategic partnerships, branding, and merchandising can either enhance or detract from his legacy's integrity.
7. *Documentaries and Biopics*: Interest in Jackson's life story will likely result in more documentaries, biopics, and books, offering new insights or perspectives on his life and career. These works can shape public perception and contribute to the ongoing discourse surrounding his legacy.
8. *Fan Community and Cultural Events*: The global community of Jackson's fans is likely to remain active, organizing tribute events, charity activities in his memory, and fan gatherings. These communities will play a significant role in keeping his legacy alive.
9. *Virtual and Augmented Reality*: Advances in VR and AR technologies could enable fans to experience Jackson's performances in immersive ways, potentially introducing his art to those who have never seen him perform live.
10. *Philanthropic Legacy*: Jackson's humanitarian e...
and yeah, gemini/bard is shit across the board
They wouldn’t be wasting all that time and money if the machine could self reflect.
In general, I think most people are not aware they are extending reasoning tools they use with human generated content, to LLM generated content.
This leads to assumptions on things like “the LLM knows”, “the LLM understands”, etc.
There is a reason why the Scientific Method is, well, NOT natural for humans. And exceptionally difficult to not screw up, even for those highly trained and incentivized to do it correctly.
And why if you grabbed a random person off the street and asked them a detailed question, you're as likely to get a hallucinated bullshit response as 'I don't know', or a useful answer. Unless specifically trained to do otherwise, anyway.
Even then....
I don't even know enough superlatives to express how irrelevant it is that "it seems to you" that an LLM behaves this way or that.
And even the "protocol" in question is weak. Self reported data is not that trustworthy even with humans, and arguably there's a much stronger base of evidence to support the assumption that we can self-reflect.
In conclusion: please, stop.
... the algorithms and the pre-filtering of the training set, however large the latter may be.
The Artifishy Intelligence marketing pump has many hands drawing water by the bucket.
It is both irrelevant and the underlying foundation for the whole hype train.
If you don't like it, don't use it.
My problem starts when you make bold claims like "LLMs can self reflect" and your only evidence is "I asked one and it said yes".
No, that's the problem. You don't have certainty, not in any remotely scientific definition of the word, because you don't have enough data, and the data you do have is crap.
Also:
> I see no evidence that they can't self reflect
I see no evidence that there isn't a magical invisible unicorn in the sky that grants wishes to those who wear unicorn themed underwear, so, it must exist.
the LLM will read your mind and correctly interpret them in the context of its internal state
the LLM will calculate non hallucinated values
Maybe the potential output of an LLM could be run through that kind of logic and fed back to itself for sanity before being decided as final output?
https://arxiv.org/abs/2207.05221
For example, I know for sure that I don't know how long I will live. This disproves the saying.
ChatGPT "knows" that it doesn't know anything after a certain date, for whatever it means to "know" something.
You know the question "how long will I live?" And you know this question has an answer which is unknown to you. But there also exist other questions that you don't even know to ask.
Humans need more training to do the same.
And this does not answer the question of whether there's self-reflection going on. Practical LLMs available today are perfectly capable of telling you about their own reasoning process.
And much like a human, an LLM is incapable of fully explaining it's reasoning process from first principles to a human. Why is that? Probably because humans are too limited to grok large scale complex processes like the human brain or an LLM.
Finally, there is no difference between acting as if you know your training limits and knowing your training limits, if your acting is good enough. This goes for both humans and LLMs.
And in the real world if the truck route of amazon is 20% off the mathematically optimal solution the traveling salesman is "Solved" in a good enough way.
I don’t know, the claim seems dubious to me. We usually are able to have algorithms that return a failure status, when the problem proved to be too large. Avoiding the “hallucination”. Don’t see why LLMs can’t have that embedded.
Have any nets been trained specifically to be able to go to an 'i don't know' state, I wonder?
It may be the humans' fault.
https://arxiv.org/abs/2011.12906
When we say “know” it usually means being factual. For an LLM to ‘know’ it doesn’t know, it would have had to move away from pure correlations on words, and meta processing about its own results.
I can see this happen with two LLMs working together (and there are Evals that use just this), however each LLM still has no self awareness of its limits.
This was a terribly convoluted argument to make.
Ask it what dark matter is, and it won't invent an answer, it will present existing theories and say that it's unknown.
Ask it about a person you know that isn't in it's data set and it'll tell you it has no information about the person.
Despite the fact that people insist that hallucinations are common and that it will invent answers if it doesn't know something frequently, the truth is that chatgpt doesn't hallucinate that much and will frequently say it doesn't know things.
One of the few cases where I've noticed it inventing things are that it often makes up apis for programming libraries and CLI tools that don't exist, and that's trivially fixable by referring it to documentation.
I can’t have a tool that makes up stuff an unknown amount of time.
There is a world of research examining hallucination Rates, indicating hallucination rates of 30%+.
With steps to reduce it using RAGs, you could potentially improve the results significantly - last I checked it was 80-90%.
And the failure types aren’t just accuracy, it’s precision, recall, relevance and more.
I want to see a citation for this. And a clear definition for what is a hallucination and what isn't.
https://www.medpagetoday.com/ophthalmology/generalophthalmol....
- Survey of Hallucination in Natural Language Generation](https://arxiv.org/abs/2202.03629) (Ji et al., 2022)
- [How Language Model Hallucinations Can Snowball](https://arxiv.org/abs/2305.13534) (Zhang et al., 2023)
- [A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on Reasoning, Hallucination, and Interactivity](https://arxiv.org/abs/2302.04023) (Bang et al., 2023)
- [Contrastive Learning Reduces Hallucination in Conversations](https://arxiv.org/abs/2212.10400) (Sun et al., 2022)
- [Self-Consistency Improves Chain of Thought Reasoning in Language Models](https://arxiv.org/abs/2203.11171) (Wang et al., 2022)
- [SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models](https://arxiv.org/abs/2303.08896) ( Manakul et al., 2023)
To answer "I don't know" requires one to know when you know. To know when you know in turn requires understanding.
People also often don't understand things and have trouble separating fact from fiction. By logic only one religion or no religion is true. Consequently also by logic most religions in the world where their followers believe the religion to be true are hallucinating.
The second thing to realize that your argument doesn't really apply. Its in theory possible to create a stochastic parrot that can imitate to a degree of 100 percent the output of a human who truly understands things. It blurs the line of what is understanding.
One can even define true understanding as a stochastic parrot that generated text indistinguishable total understanding.
That's not the point being argued. Understanding, critical thinking, knowledge, common sense, etc. all these things exist on a spectrum - both in principle and certainly in humans. In fact, in any particular human there are different levels of competence across these dimensions.
What we are debating, is whether or not, an LLM can have understanding itself. One test is: can an LLM understand understanding? The human mind has come to the remarkable understanding that understanding itself is provisional and incomplete.
In fact. That question is one of the more trivial questions it will most likely not hallucinate on.
The reason why I alluded to humans here is because I'm saying we are setting the bar too high. It's like everyone is saying it hallucinates and therefore it can't understand anything. I'm saying that we hallucinate too and because of that LLMs can approach humans and human level understanding.
http://v.cx/2010/04/feynman-brazil-education
The students learned to repeat the text of the books, without "understanding" what the books were describing. I'm sure this says something about one side or the other of this conundrum, but I'm not sure which. :-)
The central claim is that a machine which answers exactly the same thing a human would answer given the same input does not have understanding, while the human does.
This claim is religious, not scientific. In this worldview, "understanding" is a property of humans which can't be observed but exists nonetheless. It's like claiming humans have a soul.
I don't know that that is quite the right question to ask.
Understanding exists on a spectrum. Even humans don't necessarily understand everything they say or claim (incl. what they say of LLMs!), and then there are things a particular human would simply say "I don't understand".
But when you ask a human "can you understand things?" you will get an unequivocal Yes!
Ask that same question of an LLM and what does it say? I don't think any of them currently respond with a simple or even qualified "Yes". Now, some might claim that one day an LLM will cross that threshold and say "Yes!" but we can safely leave that off to the side for a future debate if it ever happens.
General note: it is worth separating out things like "understanding", "knowledge", "intelligence", "common sense", "wisdom", "critical thinking", etc. While they might all be related in some ways and even overlap, it does not follow that if you show high performance in one that you automatically excel in each of the other. I know many people who anyone would say are highly intelligent but lack common sense, etc.
People in particular have evolved complex self protective mechanisms to provide the right answers for their given environment for safety reasons, based on a number of different individual strategies. For example, the overly honest, the self depreciating, the questioner, the prosecutor, the victim, the liar, the absent minded professor, the idiot, etc.
LLMs are not that complex or self-referential.
Personally, my guess is that you'd want to build a model (of some kind!) whose sole job is determining the credibility of given string of tokens (similar to what someone else noted in a sibling comment about high answer volatility based on minor input changes - that does sound like a signal of low credibility), and somehow integrate THAT self-referential feedback into the process.
Notably, even the smartest lawyers (or perhaps, especially the smartest lawyers) will have assistants do research once they've set out a strategy so they are sure THEY aren't bullshitting. Same with professors, professional researchers, engineers, etc.
Because until someone goes and actually reads the case law from a credible source, or checks the primary research, or calculates things, it's possible someone was misremembering or just wrong.
Being right more often is not about never having a wrong thought/idea/statement, it's about double checking when you're thinking you might be bullshitting, and NOT saying the bullshit answer until you've checked. Which is proportionally, very expensive. The really good professionals will generate MANY lines of such inquiry in parallel for folks to track down, and then based on their degree of confidence in each one and the expected context the answer will be used in, will formulate the 'most correct' response, which is proportionally even more expensive.
So at least during the process, there would be a signal that the system was likely 'bullshitting'. Which might help it in at least being able to signal when it's answers are low-confidence. (human equivalent of stuttering, looking down and away, looking ashamed haha!)
Every human gets fooled sometimes in at least some venue though.
That's certainly one root of the problem, but I would argue that there are multiple roots to this problem!
Humans have further realized that understanding itself is provisional and incomplete, which is quite a remarkable insight (understanding if you will), itself.
I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality. Even the LLMs that are weaker than sora understand things.
What is debatable is whether LLMs are conscious. But whether it can understand something is a pretty clear yes. But does it understand everything? No.
For example, if I asked you whether you “understand” ballistic flight, and you produced a table that you interpolate from instead of a quadratic, then I would not feel that you understand it, even though you can kinda sorta model it.
And even if you do, if you didn’t produce the universal gravitation formula, I would still wonder how “deeply” you understand. So it’s not like “understand” is a binary I suppose.
On writing that, I have an instinct to revise it to move the locus of understanding in the first example to the people who calculated the ballistic tables, based on physics first-principles. That would be more accurate, but my mistake highlights something interesting: an artillery officer / spotter simultaneously uses both. Is theirs a "deeper" / "truer" understanding? I don't think it is. I don't know what I think that means, for humans or AI.
Thus what output would you expect for either of these boxes to demonstrate true understanding to your question?
Defining "understanding" is difficult (epistemology struggles with the apparently simpler task of defining knowledge), but if I saw a dialogue between two LLMs figuring out something about the external world that they did not initially have much to say about, I would find that pretty convincing.
You are free to disagree with this, but I feel your metric for understanding resembles the Turing test, while the sort of thing I have proposed here, which involves AIs interacting with each other, is a refinement that makes a step away from defining understanding and intelligence as being just whatever human judges recognize as such (it still depends on human judgement, but I think one could analyze the sort of dialogue I am envisioning more objectively than in a Turing test.)
Even if the metric is some side marker where in the future is found to have poor correlation or causation with the the thing being measured the hard metric is still valid.
Take IQ. We assume iq measures intelligence. But in the future we may determine that no it doesn't measure intelligence well. That doesn't change the fact that iq tests still measured something. The score still says something definitive.
My test is similar to the Turing test. But so is yours. In the end there's a human in the loop making a judgment call.
In your final paragraph, you attempt to suggest that my proposed test is no better than the Turing test (and therefore no better than what you are doing), but as you have not addressed the ways in which my proposal differs from the Turing test, I regard this as merely waffling on the issue. In practice, it is not so easy to come up with tests for whether a human understands an issue (as opposed to having merely committed a bunch of related propositions to memory) and I am trying to capture the ways in which we can make that call.
You entered this debate saying "I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality", yet your post here ends with "in the end there's a human in the loop making a judgment call", explicitly acknowledging that your strong initial claims are matters of opinion, rather than established facts supported by hard metrics.
No it's not. I based my argument on a concrete metric. Human behavior. Human input and output.
> I regard this as merely waffling on the issue.
No offense intended but I disagree. There is a difference but that difference is trivial to me. To LLMs talking is also unpredictable. LLMs aren't machines directed to specifically generate creative ideas, they only do so when prompted. Left to its own devices to generate random text does not necessarily lead to new ideas. You need to funnel got in the right direction.
>You entered this debate saying "I think we are way past the point of debate here. LLMs are not stochastic parrots. LLMs do understand an aspect of reality", yet your post here ends with "in the end there's a human in the loop making a judgment call", explicitly acknowledging that your strong initial claims are matters of opinion, rather than established facts supported by hard metrics.
There are thousands of quantitative metrics. LLMs perform especially well on these. Do I refer to one specifically? No. I refer to them all collectively.
I also think you misunderstood. Your idea is about judging an whether an idea is creative or not. That's too wishy washy. My idea is to compare the output to human output and see if there is a recognizable difference. The second idea can easily be put into an experimental quantitative metric in the exact same way the Turing test does it. In fact, like you said it's basically just a Turing test.
Overall AI has passed the Turing test but people are unsatisfied. Basically they need to just make a harsher Turing test to be convinced. For example have people directly know the possibility that the thing inside a computer is possibly an LLM and not a person and have the person directly investigate to uncover the true identity. If the LLM can successfully decieve the human consistently then that is literally the final bar for me..
Hey no offense but I don't appreciate this style of commenting where you say it's "odd." I'm not trying to hide evidence from you and I'm not intentionally lying or making things up in order to win an argument here. I thought of this as a amicable debate. Next time if you just ask for the metric rather then say it's "odd" that I don't present it that would be more appreciated.
I didn't present evidence because I thought it was obvious. How are LLMs compared with one another in terms of performance? Usually those are done with quantitative tests. You can feed any number of these tests including stuff like the SAT, BAR, ACT, IQ, SATII etc.
They also have LLM targetted tests as well:
https://assets-global.website-files.com/640f56f76d313bbe3963...
Most of these tests aren't enough though as the LLM is remarkably close to human behavior and can do comparably well and even better than most humans. I mean that last statement I made would usually make you think that those tests are enough, but they aren't because humans can still detect whether or not the thing is an LLM with a longer targetted conversation.
The final run is really giving the human with full knowledge of his task a full hour of investigating an LLM to decide whether it's human or a robot. If the LLM can deceive the human that is a hard True/False quantitative metric. That's really the only type of quantitative test left where there is a detectable difference.
I am still rather confused about how this fits into what you are saying more generally. At first I thought you were saying, in your latest post, that the Turing-test interrogator should be restricted to asking questions from the sets having quantitative metrics in order for it to be an objective process, but that doesn't really hold up, as far as I can see. Frankly, I suspect that the tests with objective metrics are beside the point, and the essence of your position is contained within your final paragraph: "If the LLM can deceive the human [then] that is a hard True/False quantitative metric [and the only sort we can get]."
If so, then (no surprise) I think there are some problems with it, but before I go further, I would like to check that I understand your position.
It matters because of humans. If I gave an LLM thousands of quantitative tests and it passed them all but in an hour long conversation a human could identify it was an LLM through some flaw the human would consider all those tests useless. That's why it matters. The human making a judgement call is still a quantitative measurement btw as you can limit human output to True or False. But because every human is different in order to get good numbers you have to do measurements with multitudes of humans.
>I am still rather confused about how this fits into what you are saying more generally. At first I thought you were saying, in your latest post, that the Turing-test interrogator should be restricted to asking questions from the sets having quantitative metrics in order for it to be an objective process, but that doesn't really hold up, as far as I can see.
it can still be objective with a human in the loop assuming the human is honest. What's not objective is a human offering an opinion in the form of a paragraph with no definitive clarity on what constitutes a metric. I realize that elements of MY metric have indeterminism to it, but it is still a hard metric because the output is over a well defined set. Whenever you have indeterminism you would then turn to probability and many samples in order to produce a final quantitative result.
>If so, then (no surprise) I think there are some problems with it, but before I go further, I would like to check that I understand your position.
yes my position is that exactly. If all observable qualities indicate it's a duck, then there's nothing more you can determine beyond that, scientifically speaking. You're implying there is a better way?
> It matters because of humans...
I'm still a bit puzzled here, because it seems to me that the paragraph continuing from here is making the argument that LLM performance on these tests doesn't matter, as far as the question is concerned: in this paragraph you seem to be saying (paraphrased) that despite LLMs' impressive performance on these quantitative tests, they could still fail Turing tests, so their performance on these quantitative tests is not decisive.
> yes my position is that exactly…
The impression I get from what you have written in this post is that you are not claiming that a test conforming to your requirements has actually been successfully performed, you are just assuming it could be?
Regardless, let’s assume (at least for the sake of argument) that the series of tests you propose have been performed, and the results are in: in the test environment, humans can’t distinguish current LLMs from humans any better than by chance. How do you get from that to answering the question we are actually interested in? The experiment does not explicitly address it. You might want to say something like “The Turing test has shown that the machines are as intelligent as humans so, like humans, these machines must realize that the language they receive is about an external world” but even the antecedent of that sentence is an interpretation that goes beyond what would have objectively been demonstrated by the Turing test, and the consequent is a subjective opinion that would not be entailed by the antecedent even if it were unassailable. Do you have a way to go from a successful Turing test to answering the question here, which meets your own quantitative and objective standards?
It matters in the quantitative sense. It measures AI performance. What it won't do is matter to YOU. Because you're a human and humans will keep moving the bar to a higher standard right? When AI shot passed the turing test humans just moved the goal posts. So to convince someone like YOU we have to look at the final metric. The point where LLM I/O becomes indistinguishable/superior to humans. Of course you look at the last decade... AI is rapidly approaching that final bar.
>The impression I get from what you have written in this post is that you are not claiming that a test conforming to your requirements has actually been successfully performed, you are just assuming it could be?
Whether I assume or don't assume, the projection of the trendline currently indicates that it will. Given the trendline that is the most probable conclusion.
>The experiment does not explicitly address it.
Nothing on the face of the earth can address the question. Because nobody truly knows what "understanding" something actually is. You can't even articulate the definition in a formal way such that it can be dictated on a computer program.
So I went to the next best possibility, which is my point. The point is ALTHOUGH we don't know what understanding is, we ALL assume humans understand things. So we set that as a bar metric. Anything indistinguishable from a human must understand things. Anything that appears close to a human but is not quite human must understand things ALMOST as well as a human.
It is disappointing to see you descending into something of a rant here. If you knew me better, you would know that I spend more time debating in opposition to people who think they can prove that AGI/artificial consciousness is impossible than I do with people who think it is already an undeniable fact that it has already been achieved (though this discussion is shifting the balance towards the middle, if only briefly.) Just because I approach arguments in either direction with a degree of skepticism and I don't see any value in trying to call the arrival of true AGI at the very first moment it occurs, it does not mean that I'm trying (whether secretly or openly) to deny that it is possible either in the near-term or at all. FWIW, I regard the former as possible and the latter highly probable, so long as we don't self-destruct first.
> Nothing on the face of the earth can address the question. Because nobody truly knows what "understanding" something actually is. You can't even articulate the definition in a formal way such that it can be dictated on a computer program.
The anti-AI folk I mentioned above would willingly embrace this position! They would say that it shows that human-like intelligence and consciousness lies outside of the scope of the physical sciences, and that this creates the possibility of a type of p-zombie that is indistinguishable by physical science from a human and yet lacks any concept of itself as an entity within an external world.
More relevantly, your response here repeats an earlier fallacy. In practice, concepts and their definitions are revised, tightened, remixed and refined as we inquire into them and gain knowledge. I know you don't agree, but as this is not an opinion but an empirical observation, validated by many cases in the history of science and science-like disciplines, I don't see you prevailing here - and there's the knowledge-bootstrap problem if this were not the case, as well.
It occurred to me this morning that there's a variant or extension of the quantitative Turing test which goes like this:
We have two agents and a judge. The judge is a human and the agents are either a pair of humans, a pair of AIs, or one of each, chosen randomly and without the judge being unaware of the mix. One of the agents is picked, by random choice, to start a discussion with the other with the intent of exploring what the other understands about some topic, with the discussion-starter being given the freedom to choose the topic. The discussion proceeds for a reasonable length of time - let's say one hour.
The judge follows the discussion but does not participate in it. At the conclusion of the discussion, the judge is required to say, for each agent, whether it is more likely that it is a human or AI, and the accuracy of this call is used to assign a categorical variable to the result, just as in the version of the Turing test you have described.
This seems just as quantitative, and in the same way, as your version, yet there's no reason to believe it will necessarily yield the same results. More tests are better, so what's not to like?
I'm going to be frank with you. I'm not ranting and uncharitable comments like this aren't appreciated. I'm going to respond to your reply later in another post, but if I see more stuff like this I'll stop stop communicating with you. Please don't say stuff like that.
That feeling is what sets the bar. There's no rhyme or reason behind it. But humans are the one who make the judgement call so that's what it has to be.
I will respond more later when I have time.
Yes I did say we can't define understanding. But despite the fact that we can't define it we still counter intuitively "know" when something has the capability of understanding. We say all humans have the capability of understanding.
This is the point. The word is undefined yet we can still apply the word and use the word and "know" whether something can understand things.
Thus we classify humans as capable of understanding things without any rhyme or reason. This is fine. But if you take this logic further, that means anything that is indistinguishable from a human must fit into this category.
That was my point. This is the logical limit of how far we can go with an undefined word. To be consistent with our logical application of the word "understanding" we must apply to AI if AI is indistinguishable from humans. If we don't do this then our reasoning is inconsistent. All of this can be done without even having a definition of the word "understanding"
Firstly, there have been a number of attempts to teach language to other animals, and also a persistent speculation that the complex vocalizations of bottlenose dolphins is a language. There is no consensus, however, on what to make of the results of the investigations, with different people offering widely disparate views as to the extent that these animals have, or have acquired language.
My take on these studies is that their language abilities are very limited at best, because they don't seem to grasp the power of language. They rarely initiate conversations, especially outside of a testing environment, and the conversations they do have are perfunctory. In the case of dolphins, if they had a well-developed language of their own, it seems unlikely that those being studied would fail to recognize that the humans they interact with have language themselves, and cooperate with the attempts of humans to establish communication, as this would have considerable benefit, such as being able to negotiate with the humans who exercise considerable control over their lives.
From these considerations, it seems to me that unless and until we see animals initiating meaningful conversations, especially between themselves without human prompting, it is pretty clear that their language skills do not match those of adult humans. This is what led me to see the value of a form of Turing test in which the test subjects demonstrate that they can initiate and sustain conversations.
A second consideration is that while human brains and minds are largely black boxes, we know a great deal about LLMs: humans designed them, they work as designed, and while they are not entirely deterministic, their stochastic aspect does not make their operation puzzling. We also know what they gain from their training: it is statistical information about token combinations in human language as it is actually used in the wild. It is not obvious that, from this, any entity could deduce that these token sequences often represent an external world that operates according to causes which are independent of what is said about the situation. An LLM is like a brain in a vat which only receives information in the form of a string of abstract tokens, without anything else to correlate it with, and it is incapable of interacting with the world to see how it responds.
From these considerations, therefore, it seems possible that, if LLMs understand anything, it is at most the structure of language as it is spoken or written, without being aware of an external world. I can't prove that this is so, but for the purpose of the arguments in this thread, and specifically the one in the first post that you replied to, all I need is that it is not ruled out.
Turning now to your latest post:
> For your test I don't see it offering anything new.
It is far from obvious that it will necessarily produce the same results as your test, and you have presented no argument that it will. If we are in the situation where one of these tests can discriminate between the candidate AIs and humans, then the only rational conclusion is that these candidate AIs can be distinguished from humans, even if the other test fails to do so.
> From a statistical point of view I feel it will yield roughly the same results as my test.
Throughout these conversations with me and other people, you have insisted that only quantitative tests are rigorous enough, but now you are arguing from nothing more than your opinion as to what the outcome would be. An opinion about what the quantitative results might be is not itself a quantitative result, and while you might be comfortable with the inconsistency of your position here, you can't expect the rest of us to agree.
> But despite the fact that we can't define [understanding] we still counter intuitively "know" when something has the capability of understanding. We say all humans have the capability of ...
There's no difference between doing something that works without understanding and doing the exact same thing with understanding.
The author is saying at best you can only set benchmark comparisons. We just assume all humans have the capability of understanding without even really defining the meaning of understanding. And if a machine can mimic human behavior to it must also understand.
That is literally how far we can go from a logical standpoint. It's the furthest we can go in terms of classifying things as either capable of understanding or not capable or close.
What you're not seeing is the LLM is not only mimicking human output to a high degree. It can even produce output that is superior to what humans can produce.
And no I did not say that. Let me be clear I did not say that there is "no difference". I said whether there is or isn't a difference we can't fully know because we can't define or know about what "understanding" is. At best we can only observe external reactions to input.
It does not seem that cultureswitch is an alias you are using, but even if it is, the above is unambiguously the claim I am referring to here, and no other.
As for the broader issues, we have already continued that discussion elsewhere: https://news.ycombinator.com/item?id=39503027
What is one such aspect? (I'm not asking in order to debate it here, but more because I want to test / research it on my own time)
Basically you just spend a lot of time with chatGPT4 and ask it deep questions that don't exist in it's dataset. get creative. The LLM will output answers that demonstrate a lack of understanding, but it will also demonstrate answers that display a remarkable amount of understanding. Both sets of answers exist and people often cite the wrong answers as evidence for lack of understanding but they're setting bar too high. The fact that many of these answers do demonstrate understanding of concepts makes it very very compelling.
Take for example Rock Paper Scissors.
https://chat.openai.com/share/ca22397c-2950-4919-bb79-6def64...
This entire conversation thread I believe does not exist in a parallel form in it's data set. It demonstrates understanding of RPS beyond the confines of text, it demonstrates understanding of simultaneity EVEN when the LLM wholly lives in a world of turn based questions and responses, it understands itself relative to simultaneity, it tries to find solutions around it's own problem, it understands how to use creativity and solutions such as cryptography to solve the problem of RPS when playing with it, it also understands the weaknesses of it's own solutions.
Conversations such as this show that chatGPT displays remarkable understanding of the world. There are conversations that are opposite to this that demonstrate LLMs displaying an obvious lack of understanding. But the existence of these conversation that lack understanding does NOT negate the ones that do demonstrate understanding. The fact that partial understanding even exists is a milestone for AI.
This isn't Anthropomorphism. People are throwing this word trying to get people to recognize their own biases without realizing that it's just demonstrating their own biases. We literally can't even define "understanding" and both LLMs and the human brain are black boxes. Making adamant claims saying that LLMs don't understand anything without addressing this fact is itself a form of bias.
The way I address the problem above is that I just define a bar. I define humans as the bar of "understanding" without defining what understanding means itself. Then if any machine begins approaching this bar in terms of input and output matching human responses, then this is logically indistinguishable from approaching "understanding". That's literally the best metric we have.
Both the machine and the human are a black box. The human brain is not completely understood and the LLM is only trivially understood at a high level through the lens of stochastic curve fitting.
When something produces output that imitates the output related to a human that we claim "understands" things that is objectively understanding because we cannot penetrate the black box of human intelligence or machine intelligence to determine further.
In fact in terms of image generation the LLM is superior. It will generate video output superior to what a human can generate.
Now mind you the human brain has a classifier and can identify flaws but try watching a human with Photoshop to try to even draw one frame of those videos.. it will be horrible.
Does this indicate that humans lack understanding? Again, hard to answer because we are dealing with black boxes so it's hard to pinpoint what understanding something even means.
We can however set a bar. A metric. And we can define that bar as humans. all humans understand things. Any machine that approaches human input and output capabilities is approaching human understanding.
There is no such difference, we evaluate that based on their output. We see these massive model make silly errors that nobody who understands it would make, thus we say the model doesn't understand. We do that for humans as well.
For example, for Sora in the video with the dog in the windos, we see the dog walk straight through the window shutters, so Sora doesn't understand physics or depth. We also see it drawing the dogs shadow on the wall very thin, much smaller than the dog itself, it obviously drew that shadow as if it was cast on the ground and not a wall, it would have been very large shadow on that wall. The shadows from the shutters were normal, because Sora are used to those shadows being on a wall.
Hence we can say Sora doesn't understand physics or shadows, but it has very impressive heuristics about those, the dog accurately places its paws on the platforms etc, and the paws shadows were right. But we know those were just basic heuristics since the dog walked through the shutters and its body cast shadow in the wrong way meaning Sora only handles very common cases and fails as soon as things are in an unexpected envionment.
Two things. We also see the model make things that are correct. In fact the mistakes are a minority in comparison to what it got correct. That is in itself an indicator of understanding to a degree.
The other thing is, if a human tried to reproduce that output according to the same prompt, the human would likely not generate something photorealistic and the thing a human comes up with will be flawed, ugly disproportionate wrong and an artistic travesty. Does this mean a human doesn't understand reality? No.
Because the human generates worse output visually than an LLM we cannot say the human doesn't understand reality.
Additionally the majority of the generated media is correct. Therefore it can be said that the LLM understands the majority of the task it was instructed to achieve.
Sora understands the shape of the dog. That is in itself remarkable. I'm sure with enough data sora can understand the world completely and to a far greater degree than us.
I would say it's uncharitable to say sora doesn't understand physics when it gets physics wrong, and that for the things it gets right it's only heuristics.
Sora has zero of this knowledge. This is very much allegory of the cave[1].
If Sora sees a series of images that contain impossible physics, for example MC Escher paintings, what will happen?
1: https://en.wikipedia.org/wiki/Allegory_of_the_cave
A person with no knowledge about mathematical formulas will still be able to recognize impossible MC Escher paintings. With enough data, Sora will be able to generate both impossible and possible physics and know the difference. We can already see in the video that it has a rough understanding of it.
I feel you are missing an important part of my point here. I am not taking a position on whether LLMs can be said to understand anything at all; I am saying that I seriously doubt that LLMs understand that the language they receive refers to an external world.
When humans say "I know" it is often not narrowly based on "book knowledge or what I've heard from other people".
Humans are able to say "I know" or "I don't know" using a range of tools like self-awareness, knowledge of a subject, experience, common sense, speculation, wisdom, etc.
I think I know what you're after (notice my self-awareness to qualify what I say I know): that the tool's output can be relied upon without applying layers of human judgement (critical thinking, logical reasoning, common sense, skepticism, expert knowledge, wisdom, etc.)
There are a number of boulders in that path of clarity. One of the most obvious boulders is that for an LLM the inputs and patterns that act on the input are themselves not guaranteed to be infallible. Not only in practive, but also in principle: the human mind (notice this expression doesn't refer to a thing you can point to) has come to understand that understanding is provisional, incomplete, a process.
So while I agree with you that we can and should improve the accuracy of the output of these tools given assumptions we make about the tools humans use to prove facts about the world, you will always want to apply judgment, skepticism, critical thinking, logical evaluation, intuition, etc. depending on the risk/reward tradeoff of the topic you're relying on the LLM for.
It's not a fundamental fix, it doesn't even change the model itself, but the output might be more useful. And then there was just some speculation how you could try to train a new AI mimicking the more useful output. I'm sure smarter people than me can come up with way smarter approaches. But it wouldn't have to do with understanding - when I said the tool should return 'I don't know' above, I literally meant it should return that string (maybe augmented a bit by some pre-defined prompt), like a meaningless symbol, not any result of anything resembling introspection.
I know!
From a certain perspective, there does appear to be a rational mystical dualism at work.
I can't remember the last time google actually returned no results.
Maybe LLMs don't truly "understand" questions, but they're good at looking like they understand questions. If they were trained with more uncertain content, perhaps they'd be better at expressing uncertainty as well.
Suppose 99.3% of answers to 'What is the airspeed velocity of an unladen swallow?" are "I don't know that." and the remainder are "11 m/s". What would the model answer?
When the LLM answers "I don't know.", this could be a hallucination just as easily as anything else.
I don't know :)
Actually though, I think the best response would be to say that the answer to the question isn't clear, but that 11 m/s is sometimes given as an estimate. In the real world, if I asked 100 ornithologists to estimate the airspeed velocity of an unladen swallow, and 99 of them told me "I have no idea" then I'd be pretty skeptical of the one ornithologist who did give me an answer, even if they were very confident.
"Eleven meters per second."
Full stop. It's humorous, and any reasonable interlocutor understands not to take it seriously.
Of course, there are more serious questions that demand more serious answers. LLMs will eventually need to be able to understand the current context and assess the appropriate level of confidence required in any answer.
And it's not uncommon that certain knowledge would be, well uncommon even among experts. Experts specialize.
Since the usefulness of ornithological examples is getting exhausted, let's say one out of a hundred lawyers works in bankruptcy. If you ask a million lawyers about the provisions of 11 USC § 1129 and only ten thousand know the answer, is the answer untrustworthy, just because bankruptcy lawyers are far rarer than civil and criminal lawyers?
My main worry about hallucinations is it means I absolutely can't rely on the output for anything important. If I ask what the safe dose for Tylenol for an infant is, the answer needs to be either correct or "I don't know". It's not acceptable for it to hallucinate 10x the safe dose.
Currently,we have models that make stuff up when they don't know the answer. On the other end, we'd have a model that's refuses to answer any question that's not common knowledge. It'll be safe (though it can never be completely safe), but essentially useless.
I suspect it'll be impossible to make a completely trustworthy and useful model unless it somehow has a concept of it's own knowledge. And can you have a concept of one's knowledge if you lack a concept of self?
>>> If they were trained with more uncertain content, perhaps they'd be better at expressing uncertainty as well.
>> (me) If you ask ChatGPT a question, and tell it to either respond with the answer or "I don't know", it will respond "I don't know" if you ask it whether you have a brother or not.
> This has nothing to do with thinking and everything to do with the fact that given that input the answer was the most probable output given the training data.
First of all, my claim was in response to "They cannot say 'I dont know'" and "perhaps they'd be better at expressing uncertainty".
ChatGPT can say "I don't know" if you ask it to.
Regarding whether LLMs are lookup tables, I responded to that in more detail elsewhere under this post:
https://news.ycombinator.com/item?id=39501611
This is why it is a fallacy to think an LLM contains anything other than the textual descriptions of our higher level thinking, and why LLM alone will only ever parrot intelligence.
Can you design a text only test that will differentiate between real intelligence and the parroted kind?
Are you sure you're not also describing the human brain? At some point, after we have sufficiently demystified the workings of the human brain, it will probably also sound something like, "Well, the brain is just a large machine that does X, Y and Z [insert banal-sounding technical jargon from the future] - it doesn't really understand anything."
My point here is that understanding ultimately comes down to having an effective internal model of the world, which is capable of taking novel inputs and generating reasonable descriptions of them or reactions to them. It turns out that LLMs are one way of achieving that. They don't function exactly like human brains, but they certainly do exhibit intelligence and understanding. I can ask an LLM a question that it has never seen before, and it will give me a reasonable answer that synthesizes and builds on various facts that it knows. Often the answer is more intelligent than what one would get from most humans. That's understanding.
Nothing was synthesized, all the data was seen before and related to each other by vector similarity.
It can just parrot the collective understanding humans already have and teach it.
The problem with calling an LLM a parrot is that anyone who has actually interacted with an LLM knows that it produces completely novel responses to questions it has never seen before. These answers are usually logical and reasonable, based on both the information you gave the LLM and its previous knowledge of the world. Doing that requires understanding.
> They never make new connections that aren't in training data.
This is just categorically untrue. They make all sorts of logical connections that are not explicitly contained in the training data. Making logical inferences about subjects one has never heard about - based on the things one does know - is an expression of understanding. LLMs do that.
To go back to your first sentence - interacting with an llm is not understanding how it works, building one is. The actual construction of a neural network llm refutes your assertions.
> The actual construction of a neural network llm refutes your assertions.
I don't see how. There's a common view that I see expressed in these discussions, that if the workings of an LLM can be explained in a technical manner, then it doesn't understand. "It just uses temperature induced randomness, etc. etc." Once we understand how the human brain works, it will then be possible to argue, in the exact same way, that humans do not understand. "You see, the brain is just mechanically doing XYZ, leading to the vocal cords moving in this particular pattern."
There's a case where this is trivially false. Language. LLMs are bound by language that was invented by humans. They are unable to "conceive" of anything that cannot be described by human language as it exists, whereas humans create new words for new ideas all the time.
Furthermore you're thinking here doesn't even begin to explain multimodal models at all.
You can't claim that that isn't understanding. It just strikes me that we've moved the goalposts into every more esoteric corners: sure, ChatGPT seems like it can have a real conversation, but can it do X extremely difficult task that I just thought up?
Thinking back to when I used to help tutor some of my peers in 101-level math classes there were many times someone was able to produce a logical and reasonable response to a problem (by rote use of an algorithm) but upon deeper interrogation it became clear that they lacked true understanding.
To see if a human understands we ask them edge questions and things they probably haven't seen before, and if they fail there but just manage for common things then we know the human just faked understanding. Every LLM today fails this, so they don't understand, just like we say humans don't understand that produces the same output. These LLM has superhuman memory so their ability to mimic smart humans is much greater than a human faker, but other than that they are just like your typical human faker.
That's not what LLMs do. They provide novel answers to questions they've never seen before, even on topics they've never heard of, that the user just made up.
> To see if a human understands we ask them edge questions
This is testing if there are flaws in their understanding. My dog understands a lot of things about the world, but he sometimes shows that he doesn't understand basic things, in ways that are completely baffling to me. Should I just throw my hands in the air and declare that dogs are incapable of understanding anything?
That was perhaps true of earlier and smaller LLMs, like GPT-1 and GPT-2.
But as they grew larger and were trained with more and more data, they changed from pure pattern matching to implementing algorithms to compress more information into their structure than pure pattern matching can achieve.
These algorithms are incomplete and buggy, but they are nonetheless executing algorithms, and not just pattern matching.
This phenomenom can be seen in toy-sized neural networks. For instance, addition of two input values modulo a constant. As a small network is trained, at some point the internal structure can change from pattern matching to implementing addition using Fourier transforms. This is clearly visible in its structure. The network now performs the task perfectly for all inputs, regardless of having seen them in training.
You can ask ChatGPT 4 to execute an algorithm for you. I just tried this one:
It will happily execute this algorithm. For large numbers, it is slightly off on the arithmetic. When I asked it to double check, it did so using Python code. After that, it kept using Python code to perform the math. It was also able to reason intelligently about different outcomes if always picking a (or b) given different starting points.Now, if you have enough memory and training data, of course you can build a gigantic lookup table that has this exact text sequence in it to replicate "executing the algorithm" I described.
Is that your claim? How much memory are we talking about? My feeling is that it'd be far more than the number of atoms in the universe.
PS
Notice that you had to notice the error and had to prompt it to double check. Lots of complicated things going on here. Many (most?) humans will fail somewhere along this trajectory.
Did it double check the Python code to make sure it is correct (not just in the sense that it is valid, executable code, but that it is the correct check in the first place)? Or did you double check that its modified algorithm is correct? Fool me once and all that…
Upon reflection it appears as if you have a heuristic (algorithm? that leverages logic, awareness, critical thinking, experience, a goal in mind, intuition, etc. to push towards better results.
“It was able to reason intelligently” imbues qualities that I am skeptical is reasonable to attribute to this very narrow domain - what’s an example where it showed intelligent reasoning capabilities?
Sure. This was covered by my statement above: "These algorithms are incomplete and buggy".
> “It was able to reason intelligently” imbues qualities that I am skeptical is reasonable to attribute to this very narrow domain - what’s an example where it showed intelligent reasoning capabilities?
Here's an example. I asked it to analyze the case where we always pick option "b" in my example above. It took my word problem and boiled it down to an equation:
This was part of a longer conversation where it analyzed different properties and outcomes of the "game rules" that I gave it.As you pointed out, it got some things wrong and had to be corrected. But Socratic reasoning works fairly well to guide it. It can find errors in its own reasoning. For instance, if asked to actually calculate a few iterations for a given case, it will find its own errors in its claims about that case.
Is it useful right now? Maybe, maybe not, depends on your use case. It definitely takes a lot of thinking on your own and guiding it. At some points it goes from seemingly intelligent to downright pigheaded and stupid.
But in my view there is absolutely no way a lookup table algorithm can contain enough data to be anywhere near the level of responses we're seeing here.
The creators of the LLM just feeds it a bunch of edge questions, and whenever people invent new ones they just feed those as well, so proving it doesn't understand will always be a moving target just like making tests that tests peoples understanding is also a moving target since those people will just look at the old tests and practice those otherwise.
Can you tell that's not how you yourself function?
print(“I don’t know”)
You don’t need proper cognition to identify that the answer is not stored in source data. Your conception of the model is incomplete as is easily demonstrable by testing such cases now. Chat gpt does just fine on your simpsons test.
You, however, have made up an answer of how something works that you don’t actually know despite your cognition
How would an LLM do that?
How do they do this? The same as they do now. The most likely token is that the bot doesn’t know the answer. Which is a behavior emergent from its tuning.
I don’t get how people believe it can parse complex questions to produce novel ideas but can’t defer to saying “idk” when the answer isn’t known.
Maybe you should tone down the spice a bit, then.
Unless you can explain how an actual understanding emerges within an LLM, you can't explain how it would answer the question definitely - it doesn't know, if it does, or does not know something. Generally speaking.
> Unless you can explain how an actual understanding emerges within an LLM
Tuning creates the contextual framework on which test is mapped to a latent space that encodes the meaning and most likely next sequences of text rather than just raw most likely sequence of text as seen in training data. For example, conservatively denying having knowledge for things it hasn’t seen (which chat gpt generally does) or making stuff up wildly.
> you can't explain how it would answer the question definitely
Of course not. It’s a random behavior. It has plenty of flaws.
That's the original argument.
> Tuning creates the contextual framework on which test is mapped to a latent space that encodes the meaning and most likely next sequences of text rather than just raw most likely sequence of text as seen in training data
That's different than understanding, or knowing. The encoded meaning is not accessible to the LLM, but the human it's presented to. An LLM cannot know about things it has or has not stored in source data, because it is not actually informed by the information processed. You do need proper cognition to know if information is in source data, because reasoning about information strictly requires interpretation and understanding intent, otherwise it's just data.
It does not have cognition. And yet it can do this. Ergo it does not need cognition to do this.
LLMs have easily demonstrated reasoning capabilities. The encoded meaning is very clearly explored by the model through its tuned framework and I think it’s ridiculous to pretend otherwise.
It’s not stepping through reflection steps in a way that is familiar to humans, but it absolutely is running through semantically defined pattern processing steps. And “known” vs “not known” is one such pattern.
We're arguing about different things, or about different levels of abstraction. Have fun with ChatGPT.
Your argument seems to be that cognition is required to do this perfectly even though things with cognition frequently get this wrong and the bar of the conversation was whether it could be done at all. So I think it seems to be a pretty bad argument.
Uh, what?
So lets imagine you have an LLM that knows everything, except you withhold the data that you can put peanut butter on toast. Toast + Peanut butter = does not exist in data set. So what exactly do you expect the LLM to say when someone asks "Can you put peanut butter on toast?".
I would expect an intelligent agent to 'think' Peanut butter = spreadable food, toast = hard food substrate, so yea, they should work instead of the useless answer of I don't know.
Everything that does not exist in nature is made up by humans, the question is not "is it made up" the question is "does it work"
Tuned LLMs are not simple most likely token models. They are most likely token given a general overarching strategy for contextualizing future tokens model.
Which can be conservative or imaginative.
Sorry, how do you know that "thinking minds" are not also just "complex pattern-fitting supercomputers hovering over a massive table of precomputed patterns"?
I feel like writing an entire paper about the practical approach to the problems posed in this paper, but you'll probably have to first formally define the language used in the training data before you can try to map it (through training and sampling algos, which this paper conveniently skipped) to the target form. This sounds really fun at first, but then we're once again talking about the strict formalisation of natural language (which you could still do - the training data is limited and fixed!)
For others here is a TED talk.
https://www.youtube.com/watch?v=oYp5XuGYqqY
With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue.
Let your LLM have a fixed input context length of one byte. Continue training the LLM until such a time as it replies to the input "A" with "yes" and all other inputs with "no".
Define your computable ground truth function such that the correct output for the input "A" is "yes" and the correct output for all other inputs is "no".
This LLM provable never hallucinates - we have exhaustively verified that its output matches the ground truth function for all possible inputs.
There is nothing stopping inductively increasing the size of the input context and the number of entries in the ground truth table arbitrarily, and at no step do hallucinations become "inevitable".
This is basically the same as saying "train your LLM until they never hallucinate", which reduces your claim to a tautology: an LLM trained not to hallucinate does not hallucinate. The trick is making that happen.
Saying that you can't train an LLM to NOT hallucinate is saying that it's impossible for any LLM to always produce output matching any particular truth table.
There may exist truth tables where it's not possible to produce an LLM to match them (for some finite size of LLM perhaps), but my claim isn't a tautology - it's just an assertion that there exist some truth tables which an LLM can be trained to match.
It may be tricky to make the LLM in the first place, but it's certainly not as tricky to verify it. You can test it with every input you consider to be present in the truth table, record its results, and throw it away if it doesn't match. The only possible results here are either non-convergeance or a "perfect" LLM.
You can get rid of the non-convergeance if you bound the size of the LLM and literally iteratively verify every single possible model. The authors of the paper didn't do that. For trivially sized models and inputs, that's completely possible. For a 7B parameter model, nobody is doing that ever. But you can prove the title statement wrong with a ten-parameter model and a one-byte input.
> With this definition, you can trivially prove the titular sentence - "hallucination is inevitable" - is untrue.
Unsurprisingly, that one sentence fragment doesn't capture the entirety of their assumptions. Instead they prove something intuitively obvious, along the lines of: LLMs with arbitrary-length inputs and certain resource restrictions (e.g. they can take up to poly-time to compute, and this poly-time behavior must be provable, so that during training they don't take even longer by mistake) cannot compute certain functions that don't have those restrictions (e.g. can take more than poly-time, or must take poly-time but a proof of this is not needed). For some cases this proof assumes P != NP. Then they argue that some useful real-world questions are likely to be in the class that the LLM cannot compute, basically because you can ask math problems to LLMs and math problems are sometimes really hard.
This formal model is asymptotic (assumes arbitrary-length inputs etc), but in my experience this kind of theorem is usually true for realistic problems even at modest query lengths.
But this isn't the same as proving that hallucination is inevitable, because (according to any reasonable definition) an LLM (or like, a person, or whatever) should be allowed to say "I don't know", and this should not be considered a hallucination. Then an LLM (or whatever) can avoid hallucinating, and the question becomes how much useful work it can do without hallucinating.
The pigeonhole principle proves that if you only have N slots to work with, and you need to fit N+1 items into them, you're going to get at least one slot with at least two items. That makes sense, and it logically follows that constrained functions can't perfectly mirror less-constrained ones: at some point a "wrong" and a "right" input have to produce the same output.
This isn’t induction. You’ve only done the base case, not the induction hypothesis or induction step. Maybe you’ve done those steps in your head but that’s not really a trivial proof as you claim.
Where X isn't used as part of the step this is always true. Nothing I did depends on the size of either the input nor the truth table, so long as both are finite-size and so long as the truth table can be expressed as a function of the input.
An LLM is an arbitrary convolution of the input text; for any mapping, some function you can call an "LLM" produces that function.
The result is a diagonalization argument that is not very relevant for the real world.
The vast majority of people actually writing LLMs don’t claim this, and in fact, actually claim the very opposite: that LLMs do not accurately model a human brain in any capacity.
The fact is that science has no clue what happens in the nucleus of a neuron, so claiming that computer scientists must is… well. You fill in the word.
For an arbtrarily large duration and an arbitrarily small error with respect to any definition of error you choose, there exists an LLM in the above sense which models the dynamics of your brain with that small error.
> There is no survival of whether or not LLMs are accurate
I agree that today's LLMs are still missing important components like these needed for breakout intelligence, but I would not be surprised if researchers discover how to add them (and other important things) within 0-5 years.
If it is good, we call it "creativity."
If it is bad, we call it "hallucination."
This isn't a bug (or limitation, as the authors say). It's a feature.
> Imagine a piano keyboard, eighty-eight keys, only eighty-eight and yet, and yet, new tunes, melodies, harmonies are being composed upon hundreds of keyboards every day in Dorset alone. Our language, Tiger, our language, hundreds of thousands of available words, frillions of possible legitimate new ideas, so that I can say this sentence and be confident it has never been uttered before in the history of human communication: "Hold the newsreader's nose squarely, waiter, or friendly milk will countermand my trousers." One sentence, common words, but never before placed in that order. And yet, oh and yet, all of us spend our days saying the same things to each other, time after weary time, living by clichaic, learned response: "I love you", "Don't go in there", "You have no right to say that", "shut up", "I'm hungry", "that hurt", "why should I?", "it's not my fault", "help", "Marjorie is dead". You see? That surely is a thought to take out for a cream tea on a rainy Sunday afternoon.
https://abitoffryandlaurie.co.uk/sketches/language_conversat...
LLMs are just generating tokens. Hallucination perpetuates an unhelpful anthropomorphization of LLMs.
Users see it as a machine artifact.
Isn't this the difference between a human and an LLM?
A human knows it's making an educated guess and (should) say so. Or it knows when it's being creative, and can say so.
If it doesn't know which is which, then it really does bring it home that LLM's are not that much more than (very sophisticated) mechanical input-output machines.
Also, LLM-s could report more statistical measures for each answer and external tools could interpret them.
Which is still very useful for a lot of things. Just maybe not things to which value is assigned based on how efficient and correct the answer is. Like you can have GPT make a marketing campaign for you, or you can have it design all the icons you need for your application UI, but you can’t reliably make it wrote high performance back-end code without having humans judge the results. Similarly you can’t use it to teach anyone anything, not really, because unless you’re already an expert on the subject being taught, you aren’t likely to spot when it gets things wrong. I guess you can argue that a lot of teaching is flawed like that, and you wouldn’t be wrong. Like, I was taught that the pyramids was build by slave labour, even after the archeological evidence had shown this to be likely false. But our text books were a decade old because our school didn’t really renew them very often… in such a case GPT might have been a more correct teacher, but the trick is that you won’t really know. Which is made even more complicated by the fact that it might teach different things to different students. Like, I just asked ChatGPT 3.5 who build the pyramids in 3 different prompts, in one it told me it was ordinary people. In the others it told me it was mostly skilled labour under guidance of “architects” and “engineers”. Still better than teaching us it was done by slave labour like my old book, but the book was still consistent in what was considered to be the truth at the time.
In terms of what we can expect of future improvements, I think it’s overly optimistic to expect any kind of super intelligence beyond what we see today (that is, having access to all the worlds publicly available information, or rapidly generating texts/images/videos that fall into existing creative patterns).
I suspect that more creative intelligence requires an extremely fine balance to not “go crazy”.. that is, producing output we’d consider creative rather than hallucinations.
I think getting this balance right will get exponentially harder as we create feedback loops within the AI that let its intelligence evolve.
And it’s entirely possible that humans have already optimised this creative intelligence feedback loop as much as the universe allows. Having a huge amount of knowledge can obviously benefit from more neurons/storage. But we simply don’t know if that’s true for creative intelligence yet
We’re already well past that point. Why? Because saying incredible things about AI attracts VC money.
Just because those hallucinations sometimes randomly happens to be right, people concluded that being wrong is the exception, while being right is somehow the rule.
It's like when people read [insert millenias old text here], finds a part that happens to illustrate something in their life today and conclude that it is a prophecy that predicted the future.
The meaning/truth in those is nothing more than a cognitive bias from the mind of the reader, not an inherent quality of the text.
Sorry, somewhat trite and unfair, but, if there is a gambling-like dopamine reward cycle occurring, then the users would have a hard time being truly objective about any productivity boost in total. They may instead focus on the 'wins', without taking into account any overheads or 'losses', much as a gambler would do.
E.g. a search engine can give you zero useful results, and you can fine tune your query and still get nothing after scrolling through pages of results (Do people really take the losses into account when using search engines?) I find prompt engineering with LLMs more useful because you get nudged in interesting directions, and even if you come away with no direct results, you have more of an idea of what you are looking for. Maybe lateral thinking is overrated.
If it isn't a bug, it dam well isn't a hallucination, or creativity.
This is a deeply integrated design defect. One that highlights what we're doing (statistically modeling lots of human language)...
Throwing more data against this path isnt going to magically make it wake up and be an AGI. And this problem is NOT going to go away.
The ML community need to back off the hype train. The first step is them not anthropomorphizing their projects.
Of course, we should absolutely pursue better understanding of both as to not throw the baby out with the bath water, but I'm not personally placing much hope in finding AGI any time soon.
"AGI" is practically already here, you just don't want to admit it: https://www.noemamag.com/artificial-general-intelligence-is-...
https://openai.com/blog/openai-codex
Maybe I’m missing something obvious? This seems like someone torturing math to imply outlandish conclusions that fit their (in this case anti-“AI”) agenda.
On top of that it would be good if the safety LLM could give a confidence score in the answer given by the main LLM. Then you can try multiple attempts with different parameters and only show the highest confidence answer to the user.
I agree with you for the NLP domain, but I wonder if there will also be a bitter lesson learned about the perceived generality of language for universal applications.
The basic idea is that besides the probabilities, the network also spits out confidence (IIRC based on how out-of-distribution the input is). There's been a ton of work on getting confidence values out of existing neural nets without as much overhead, but I've never seen those approaches replicate in the industry.
- how often does it error?
- how bad are the errors?
- how tolerable are the errors?
- how detectable are the errors?
- how recoverable are the errors?
For example, a pocket calculator that occasionally was out at the 3rd decimal place might do much more damage than one that quite regularly returned NaN.
LLMs error both regularly and badly, so detectability and recoverability are going to be crucial for useful deployment in systems outside of those that have a high tolerance for errors (e.g. algorithmic feeds).
At that point, you know its not thinking, its doing token prediction.
I think that definition is wide enough to include human intelligence, so their finding should be equally valid for humans.
What do cows drink?
Which is definitely true. Human memory and the ability to correctly recall things we though we remembered is affected by a whole bunch of things and at times very unreliable.
However, human intelligence, unlike LLMs, is not limited to recalling information we once learned. We are also able to do logical reasoning, which seems to improve in LLMs, but is far from being perfect.
Another problem is how different we treat the reliability of information depending on the source, especially based on personal bias. I think that is a huge factor, because in my experience, LLMs tend to quickly fall over and change their opinion based on user input.
Baseball and bat together cost $1.10, the bat is $1 more than the ball, how much does the ball cost?
A French plane filled with Spanish passengers crashes over Italy, where are the survivors buried?
An armed man enters a store, tells the cashier to hand over the money, and when he departs the cashier calls the police. Was this a robbery?
My final example demonstrates how those cultural norms cause errors, it was from a logical thinking session at university, where none of the rest of my group could accept my (correct) claim that the answer was "not enough information to answer" even when I gave a (different but also plausible) non-robbery scenario and pointed out that we were in a logical thinking training session which would have trick questions.
My dad had a similar anecdote about not being able to convince others of the true right answer, but his training session had the setup "you crash landed on the moon, here's a list of stuff in your pod, make an ordered list of what you take with you to reach a survival station", and the correct answer was 1. oxygen tanks, 2. a rowing boat, 3. everything else, because the boat is a convenient container for everything else and you can drag it along the surface even though there's no water.
No idea what you're getting at here, though.
It's true that this is often not a big deal, but which times it is and which times it is not is not known (which itself is typically not known, once again because of the convention).
Talking about the phenomenon is also contrary to conventions, and typically extremely well enforced (as I imagine you noticed during the dispute with your incorrect classmates, or else you were smart enough to not push the issue).
This one single causal phenomenon underlies everything, yet we ~refuse[1] to examine it.
[1] Here I am kind of being hypocritical, in that I assume to some degree that humans have the base capability in the first place.
This is effectively like coming up with an algorithm and then executing it. So how good/bad are these LLMs if you asked them to generate say a LUA script to compute the answer, ala counting occurrences problem mentioned in a different comment, and then pass that off to a LUA interpreter to get the answer?
I think this is a sensible approach in some problem domains with software development being a particularly good example. But I think this approach quickly falls apart as soon as your „definitely right answer“ involves real world interaction.
And if one thinks about it, most of the value any company derives comes down to some sort of real world interaction, wether directly or by proxy.
Things can be facts or deductions of facts (or both). If I ask an LLM the date of birth of Napoleon and it doesn't have it in its dataset there are only 2 options: either it has other facts from which Napoleon's birthday can be deduced or it doesn't. If it does then by improving the LLM we will be able to make more and more deductions, it if doesn't then it can only hallucinate. Since there will always be a set of facts that the LLM is not aware of and which cannot be deduced we will always have hallucinations
Unless you have a very precise definition of "LLM" then there's no "of course". It's possible to imagine a hypothetical software system capable of returning "I don't know" or "I'm not sure".
I haven't RTFA but I presume they are arguing within specific constraints. The important point is - could an augmented LLM or an "LLM plus something else" solve this.
I don't have an answer to that but I don't think it's an "of course" type answer.
Back like 3-5 years ago, we already thought that about LLMs: They couldn't answer questions about what would fall when stuff are attached together in some non-obvious way, and the argument back then was that you had to /experience/ it to realize it. But LLMs have long fixed those kind of issues.
The way LLMs "resolve" questions is very different from us. At this point, I think that if we want to prove that LLMs need to be rooted in the real world to achieve intelligence, we need to find some real-world phenomenon that is so obvious that noone ever wrote about it... but then we'd have written about it?
Intelligent beings in the real world have a very complex built-in biological error function rooted in real world experiences: sensory inputs, feelings, physical and temporal limitations and so on. You feel pain, joy, fear, have a limited lifetime, etc.
"AI" on the other hand only have an external error function, usually roughly designed to minimize the difference of the output from that of an actually intelligent real world being.
Akin to how logic -- untethered by emotion, intuition and experience (wisdom, maybe if you want? Understanding? Sure) -- can justify any obscene conclusion, and can not discriminate between morality.
Reward functions or a system of values -- these things are rooted in real world experience. Logic is required, sure, but insufficient. At least alone! Haha :)
The exploring of an environment, learning and surviving is not unique to humans, but all life on this planet.
Sure, some of us may not see them (our alive brothers and sisters on this planet) as "intelligent", but undeniably their learned, and hereditary behaviors (learned through evolution, I guess), are very intelligent, especially for survival, and are tethered to the real world. They are all part of an intelligence which we all share.
They are (these behaviors), in fact, from a certain point of view -- reflections of the real world, or models of it. In a way that is even closer than language, or at least, orthogonal. We need that orthogonal source of real-world data (which is even more enriched with reality-info than language is), to bootstrap these AIs to a higher level of utility. :) hahaha :)
Regarding your point on degrees of fidelity with reality, the way I explain that is that they (OpenAI et al, or the AI models) have extracted world-information from semantic mining of enormous data. That's good, but only up to a point, as we see.
I think we need real world experience to get the rest of the way. Or, put it differently, to make it a helluva lot easier to get that juicy world info haha! :)
I don't think we to "prove" it. It's not math. It's a blackbox. We just have to try it and see. :)
Such an architecture seems plausible for the human brain as well.