Skip to content

I'm looking for an article showing that LLMs don't know how they work internally

Technology
80 32 5
  • Here's a book for a different audience. Explains in layman terms why to be wary about this tech, https://thebullshitmachines.com/

  • I found the aeticle in a post on the fediverse, and I can't find it anymore.

    The reaserchers asked a simple mathematical question to an LLM ( like 7+4) and then could see how internally it worked by finding similar paths, but nothing like performing mathematical reasoning, even if the final answer was correct.

    Then they asked the LLM to explain how it found the result, what was it's internal reasoning. The answer was detailed step by step mathematical logic, like a human explaining how to perform an addition.

    This showed 2 things:

    • LLM don't "know" how they work

    • the second answer was a rephrasing of original text used for training that explain how math works, so LLM just used that as an explanation

    I think it was a very interesting an meaningful analysis

    Can anyone help me find this?

    EDIT: thanks to @theunknownmuncher
    @lemmy.world
    https://www.anthropic.com/research/tracing-thoughts-language-model its this one

    EDIT2: I'm aware LLM dont "know" anything and don't reason, and it's exactly why I wanted to find the article. Some more details here: https://feddit.it/post/18191686/13815095

    It's the anthropic article you are looking for, where they performed open brain surgery equivalent to find out that they do maths in very strange and eerily humanlike operations, like they will estimate, then if it goes over calculate the last digit like I do. It sucks as a counting technique though

  • Improper comparison; an audio file isn’t the basic action on data, it is the data; the audio codec is the basic action on the data

    “An LLM model isn’t really an LLM because it’s just a series of numbers”

    But the action of turning the series of numbers into something of value (audio codec for an audio file, matrix math for an LLM) are actions that can be analyzed

    And clearly matrix multiplication cannot reason any better than an audio codec algorithm. It’s matrix math, it’s cool we love matrix math. Really big matrix math is really cool and makes real sounding stuff. But it’s just matrix math, that’s how we know it can’t think

    LOL you didn't really make the point you thought you did. It isn't an "improper comparison" (it's called a false equivalency FYI), because there isn't a real distinction between information and this thing you just made up called "basic action on data", but anyway have it your way:

    Your comment is still exactly like saying an audio pipeline isn't really playing music because it's actually just doing basic math.

  • I agree. This is the exact problem I think people need to face with nural network AIs. They work the exact same way we do. Even if we analysed the human brain it would look like wires connected to wires with different resistances all over the place with some other chemical influences.

    I think everyone forgets that nural networks were used in AI to replicate how animal brains work, and clearly if it worked for us to get smart then it should work for something synthetic. Well we've certainly answered that now.

    Everyone being like "oh it's just a predictive model and it's all math and math can't be intelligent" are questioning exactly how their own brains work. We are just prediction machines, the brain releases dopamine when it correctly predicts things, it self learns from correctly assuming how things work. We modelled AI off of ourselves. And if we don't understand how we work, of course we're not gonna understand how it works.

    I agree. This is the exact problem I think people need to face with nural network AIs. They work the exact same way we do.

    I don't think this is a fair way of summarizing it. You're making it sound like we have AGI, which we do not have AGI and we may never have AGI.

  • Even if LLM "neurons" and their interconnections are modeled to the biological ones, LLMs aren't modeled on human brain, where a lot is not understood.

    The first thing is that how the neurons are organized is completely different. Think about the cortex and the transformer.

    Second is the learning process. Nowhere close.

    The fact explained in the article about how we do math, through logical steps while LLMs use resemblance is a small but meaningful example. And it also shows that you can see how LLMs work, it's just very difficult

    I agree, but I'm not sure it matters when it comes to the big questions, like "what separates us from the LLMs?" Answering that basically amounts to answering "what does it mean to be human?", which has been stumping philosophers for millennia.

    It's true that artificial neurons are significant different than biological ones, but are biological neurons what make us human? I'd argue no. Animals have neurons, so are they human? Also, if we ever did create a brain simulation that perfectly replicated someone's brain down to the cellular level, and that simulation behaved exactly like the original, I would characterize that as a human.

    It's also true LLMs can't learn, but there are plenty of people with anterograde amnesia that can't either.

    This feels similar to the debates about what separates us from other animal species. It used to be thought that humans were qualitatively different than other species by virtue of our use of tools, language, and culture. Then it was discovered that plenty of other animals use tools, have language, and something resembling a culture. These discoveries were ridiculed by many throughout the 20th century, even by scientists, because they wanted to keep believing humans are special in some qualitative way. I see the same thing happening with LLMs.

  • Improper comparison; an audio file isn’t the basic action on data, it is the data; the audio codec is the basic action on the data

    “An LLM model isn’t really an LLM because it’s just a series of numbers”

    But the action of turning the series of numbers into something of value (audio codec for an audio file, matrix math for an LLM) are actions that can be analyzed

    And clearly matrix multiplication cannot reason any better than an audio codec algorithm. It’s matrix math, it’s cool we love matrix math. Really big matrix math is really cool and makes real sounding stuff. But it’s just matrix math, that’s how we know it can’t think

    Can humans think?

  • More than enough people who claim to know how it works think it might be "evolving" into a sentient being inside it's little black box. Example from a conversation I gave up on...
    https://sh.itjust.works/comment/18759960

    Maybe I should rephrase my question:

    Outside of comment sections on the internet, who has claimed or is claiming that LLMs have the capacity to reason?

  • Can’t help but here’s a rant on people asking LLMs to “explain their reasoning” which is impossible because they can never reason (not meant to be attacking OP, just attacking the “LLMs think and reason” people and companies that spout it):

    LLMs are just matrix math to complete the most likely next word. They don’t know anything and can’t reason.

    Anything you read or hear about LLMs or “AI” getting “asked questions” or “explain its reasoning” or talking about how they’re “thinking” is just AI propaganda to make you think they’re doing something LLMs literally can’t do but people sure wish they could.

    In this case it sounds like people who don’t understand how LLMs work eating that propaganda up and approaching LLMs like there’s something to talk to or discern from.

    If you waste egregiously high amounts of gigawatts to put everything that’s ever been typed into matrices you can operate on, you get a facsimile of the human knowledge that went into typing all of that stuff.

    It’d be impressive if the environmental toll making the matrices and using them wasn’t critically bad.

    TLDR; LLMs can never think or reason, anyone talking about them thinking or reasoning is bullshitting, they utilize almost everything that’s ever been typed to give (occasionally) reasonably useful outputs that are the most basic bitch shit because that’s the most likely next word at the cost of environmental disaster

    The environmental toll doesn’t have to be that bad. You can get decent results from single high-end gaming GPU.

  • I found the aeticle in a post on the fediverse, and I can't find it anymore.

    The reaserchers asked a simple mathematical question to an LLM ( like 7+4) and then could see how internally it worked by finding similar paths, but nothing like performing mathematical reasoning, even if the final answer was correct.

    Then they asked the LLM to explain how it found the result, what was it's internal reasoning. The answer was detailed step by step mathematical logic, like a human explaining how to perform an addition.

    This showed 2 things:

    • LLM don't "know" how they work

    • the second answer was a rephrasing of original text used for training that explain how math works, so LLM just used that as an explanation

    I think it was a very interesting an meaningful analysis

    Can anyone help me find this?

    EDIT: thanks to @theunknownmuncher
    @lemmy.world
    https://www.anthropic.com/research/tracing-thoughts-language-model its this one

    EDIT2: I'm aware LLM dont "know" anything and don't reason, and it's exactly why I wanted to find the article. Some more details here: https://feddit.it/post/18191686/13815095

    Humans don't know how they work internally either. I mean what are you expecting?

  • I agree. This is the exact problem I think people need to face with nural network AIs. They work the exact same way we do.

    I don't think this is a fair way of summarizing it. You're making it sound like we have AGI, which we do not have AGI and we may never have AGI.

    Let's get something straight, no I'm not saying we have our modern definition of AGI but we've practically got the original definition coined before LLMs were a thing. Which was that the proposed AGI agent should maximise "the ability to satisfy goals in a wide range of environments". I personally think we've just moved the goal posts a bit.

    Wether we'll ever have thinking, rationalised and possibly conscious AGI is beyond the question. But I do think current AI is similar to existing brains today.

    Do you not agree that animal brains are just prediction machines?

    That we have our own hallucinations all the time? Think visual tricks, lapses in memory, deja vu, or just the many mental disorders people can have.

    Do you think our brain doesn't follow path of least resistance in processing? Or do you think our thoughts comes from elsewhere?

    I seriously don't think animal brains or human to be specific are that special that nurural networks are beneath. Sure people didn't like being likened to animals but it was the truth, and I as do many AI researches, liken us to AI.

    AI is primitive now, yet it can still pass the bar, doctors exams, compute complex physics problems and write a book (soulless as it may be like some authors) in less than a few seconds.

    Whilst we may not have AGI the question was about math. The paper questioned how it did 36+59 and it did things in an interesting way where it half predicted what the tens column would be and 'knew' what the units column was, then put it together. Although thats not how I or even you may do it there are probably people who do it similar.

    All I argue is that AI is closer to how our brains think, and with our brains being irrational quite often it shouldn't be surprising that AI nural networks are also irrational at times.

  • They work the exact same way we do.

    Two things being difficult to understand does not mean that they are the exact same.

    Maybe work is the wrong word, same output. Just as a belt and chain drive does the same thing, or how fluorescent, incandescent or LED lights produce light even though they're completely different mechanisms.

    What I was saying is that one is based on the other, so similar problems like irrational thought even if the right answer is conjured shouldn't be surprising. Although an animal brain and nural network are not the same, the broad concept of how they work is.

  • Let's get something straight, no I'm not saying we have our modern definition of AGI but we've practically got the original definition coined before LLMs were a thing. Which was that the proposed AGI agent should maximise "the ability to satisfy goals in a wide range of environments". I personally think we've just moved the goal posts a bit.

    Wether we'll ever have thinking, rationalised and possibly conscious AGI is beyond the question. But I do think current AI is similar to existing brains today.

    Do you not agree that animal brains are just prediction machines?

    That we have our own hallucinations all the time? Think visual tricks, lapses in memory, deja vu, or just the many mental disorders people can have.

    Do you think our brain doesn't follow path of least resistance in processing? Or do you think our thoughts comes from elsewhere?

    I seriously don't think animal brains or human to be specific are that special that nurural networks are beneath. Sure people didn't like being likened to animals but it was the truth, and I as do many AI researches, liken us to AI.

    AI is primitive now, yet it can still pass the bar, doctors exams, compute complex physics problems and write a book (soulless as it may be like some authors) in less than a few seconds.

    Whilst we may not have AGI the question was about math. The paper questioned how it did 36+59 and it did things in an interesting way where it half predicted what the tens column would be and 'knew' what the units column was, then put it together. Although thats not how I or even you may do it there are probably people who do it similar.

    All I argue is that AI is closer to how our brains think, and with our brains being irrational quite often it shouldn't be surprising that AI nural networks are also irrational at times.

    “the ability to satisfy goals in a wide range of environments”

    That was not the definition of AGI even back before LLMs were a thing.

    Wether we’ll ever have thinking, rationalised and possibly conscious AGI is beyond the question. But I do think current AI is similar to existing brains today.

    That's doing a disservice to AGI.

    Do you not agree that animal brains are just prediction machines?

    That's doing a disservice to human brains. Humans are sentient, LLMs are not sentient.

    I don't really agree with you.

    LLMs are damn impressive, but they are very clearly not AGI, and I think that's always worth pointing out.

  • “the ability to satisfy goals in a wide range of environments”

    That was not the definition of AGI even back before LLMs were a thing.

    Wether we’ll ever have thinking, rationalised and possibly conscious AGI is beyond the question. But I do think current AI is similar to existing brains today.

    That's doing a disservice to AGI.

    Do you not agree that animal brains are just prediction machines?

    That's doing a disservice to human brains. Humans are sentient, LLMs are not sentient.

    I don't really agree with you.

    LLMs are damn impressive, but they are very clearly not AGI, and I think that's always worth pointing out.

    The first person to be recorded talking about AGI was Mark Gubrud. He made that quote above, here's another:

    The major theme of the book was to develop a mathematical foundation of artificial intelligence. This is not an easy task since intelligence has many (often ill-defined) faces. More specifically, our goal was to develop a theory for rational agents acting optimally in any environment. Thereby we touched various scientific areas, including reinforcement learning, algorithmic information theory, Kolmogorov complexity, computational complexity theory, information theory and statistics, Solomonoff induction, Levin search, sequential decision
    theory, adaptive control theory, and many more.
    Page 232 8.1.1 Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability

    As UGI largely encompasses AGI we could easily argue that if modern LLMs are beginning to fit the description of UGI then it's fullfilling AGI too. Although AGI's definition in more recent times has become more nuanced to replicating a human brain instead, I'd argue that that would degrade the AI trying to replicate biology.

    I don't beleive it's a disservice to AGI because AGI's goal is to create machines with human-level intelligence. But current AI is set to surpase collective human intelligence supposedly by the end of the decade.

    And it's not a disservice to biological brains to summarise them to prediction machines. They work, very clearly. Sentience or not if you simulated every atom in the brain it will likely do the same job, soul or no soul. It just brings the philosophical question of "do we have free will or not?" And "is physics deterministic or not". So much text exists on the brain being prediction machines and the only time it has recently been debated is when someone tries differing us from AI.

    I don't believe LLMs are AGI yet either, I think we're very far away from AGI. In a lot of ways I suspect we'll skip AGI and go for UGI instead. My firm opinion is that biological brains are just not effective enough. Our brains developed to survive the natural world and I don't think AI needs that to surpass us. I think UGI will be the equivalent of our intelligence with the fat cut off. I believe it only resembles our irrational thought patterns now because the fat hasn't been striped yet but if something truely intelligent emerges, we'll probably see these irrational patterns cease to exist.

  • I agree. This is the exact problem I think people need to face with nural network AIs. They work the exact same way we do. Even if we analysed the human brain it would look like wires connected to wires with different resistances all over the place with some other chemical influences.

    I think everyone forgets that nural networks were used in AI to replicate how animal brains work, and clearly if it worked for us to get smart then it should work for something synthetic. Well we've certainly answered that now.

    Everyone being like "oh it's just a predictive model and it's all math and math can't be intelligent" are questioning exactly how their own brains work. We are just prediction machines, the brain releases dopamine when it correctly predicts things, it self learns from correctly assuming how things work. We modelled AI off of ourselves. And if we don't understand how we work, of course we're not gonna understand how it works.

    They work the exact same way we do.

    Citation needed.

  • Maybe work is the wrong word, same output. Just as a belt and chain drive does the same thing, or how fluorescent, incandescent or LED lights produce light even though they're completely different mechanisms.

    What I was saying is that one is based on the other, so similar problems like irrational thought even if the right answer is conjured shouldn't be surprising. Although an animal brain and nural network are not the same, the broad concept of how they work is.

    What I was saying is that one is based on the other

    Not in any direct way, no. At least not in any way more rigorous than handwavey analogies.

  • But why male models?

    Because Blue Steel.

  • People that can not do Matrix multiplication do not possess the basic concepts of intelligence now? Or is software that can do matrix multiplication intelligent?

    People that can not do Matrix multiplication do not possess the basic concepts of intelligence now?

    As a mathematician (at least by education), I think that's a great definition, yes.

  • I agree. This is the exact problem I think people need to face with nural network AIs. They work the exact same way we do. Even if we analysed the human brain it would look like wires connected to wires with different resistances all over the place with some other chemical influences.

    I think everyone forgets that nural networks were used in AI to replicate how animal brains work, and clearly if it worked for us to get smart then it should work for something synthetic. Well we've certainly answered that now.

    Everyone being like "oh it's just a predictive model and it's all math and math can't be intelligent" are questioning exactly how their own brains work. We are just prediction machines, the brain releases dopamine when it correctly predicts things, it self learns from correctly assuming how things work. We modelled AI off of ourselves. And if we don't understand how we work, of course we're not gonna understand how it works.

    You're definitely overselling how AI works and underselling how human brains work here, but there is a kernel of truth to what you're saying.

    Neural networks are a biomimicry technology. They explicitly work by mimicking how our own neurons work, and surprise surprise, they create eerily humanlike responses.

    The thing is, LLMs don't have anything close to reasoning the way human brains reason. We are actually capable of understanding and creating meaning, LLMs are not.

    So how are they human-like? Our brains are made up of many subsystems, each doing extremely focussed, specific tasks.

    We have so many, including sound recognition, speech recognition, language recognition. Then on the flipside we have language planning, then speech planning and motor centres dedicated to creating the speech sounds we've planned to make. The first three get sound into your brain and turn it into ideas, the last three take ideas and turn them into speech.

    We have made neural network versions of each of these systems, and even tied them together. An LLM is analogous to our brain's language planning centre. That's the part that decides how to put words in sequence.

    That's why LLMs sound like us, they sequence words in a very similar way.

    However, each of these subsystems in our brains can loop-back on themselves to check the output. I can get my language planner to say "mary sat on the hill", then loop that through my language recognition centre to see how my conscious brain likes it. My consciousness might notice that "the hill" is wrong, and request new words until it gets "a hill" which it believes is more fitting. It might even notice that "mary" is the wrong name, and look for others, it might cycle through martha, marge, maths, maple, may, yes, that one. Okay, "may sat on a hill", then send that to the speech planning centres to eventually come out of my mouth.

    Your brain does this so much you generally don't notice it happening.

    In the 80s there was a craze around so called "automatic writing", which was essentially zoning out and just writing whatever popped into your head without editing. You'd get fragments of ideas and really strange things, often very emotionally charged, they seemed like they were coming from some mysterious place, maybe ghosts, demons, past lives, who knows? It was just our internal LLM being given free rein, but people got spooked into believing it was a real person, just like people think LLMs are people today.

    In reality we have no idea how to even start constructing a consciousness. It's such a complex task and requires so much more linking and understanding than just a probabilistic connection between words. I wouldn't be surprised if we were more than a century away from AGI.

  • You're definitely overselling how AI works and underselling how human brains work here, but there is a kernel of truth to what you're saying.

    Neural networks are a biomimicry technology. They explicitly work by mimicking how our own neurons work, and surprise surprise, they create eerily humanlike responses.

    The thing is, LLMs don't have anything close to reasoning the way human brains reason. We are actually capable of understanding and creating meaning, LLMs are not.

    So how are they human-like? Our brains are made up of many subsystems, each doing extremely focussed, specific tasks.

    We have so many, including sound recognition, speech recognition, language recognition. Then on the flipside we have language planning, then speech planning and motor centres dedicated to creating the speech sounds we've planned to make. The first three get sound into your brain and turn it into ideas, the last three take ideas and turn them into speech.

    We have made neural network versions of each of these systems, and even tied them together. An LLM is analogous to our brain's language planning centre. That's the part that decides how to put words in sequence.

    That's why LLMs sound like us, they sequence words in a very similar way.

    However, each of these subsystems in our brains can loop-back on themselves to check the output. I can get my language planner to say "mary sat on the hill", then loop that through my language recognition centre to see how my conscious brain likes it. My consciousness might notice that "the hill" is wrong, and request new words until it gets "a hill" which it believes is more fitting. It might even notice that "mary" is the wrong name, and look for others, it might cycle through martha, marge, maths, maple, may, yes, that one. Okay, "may sat on a hill", then send that to the speech planning centres to eventually come out of my mouth.

    Your brain does this so much you generally don't notice it happening.

    In the 80s there was a craze around so called "automatic writing", which was essentially zoning out and just writing whatever popped into your head without editing. You'd get fragments of ideas and really strange things, often very emotionally charged, they seemed like they were coming from some mysterious place, maybe ghosts, demons, past lives, who knows? It was just our internal LLM being given free rein, but people got spooked into believing it was a real person, just like people think LLMs are people today.

    In reality we have no idea how to even start constructing a consciousness. It's such a complex task and requires so much more linking and understanding than just a probabilistic connection between words. I wouldn't be surprised if we were more than a century away from AGI.

    Maybe I am over selling current AI and underselling our brains. But the way I see it is that the exact mechanism that allowed intelligence to flourish within ourselves exists with current nural networks. They are nowhere near being AGI or UGI yet but I think these tools alone are all that are required.

    The way I see it is, if we rewound the clock far enough we would see primitive life with very basic nural networks beginning to develop in existing multicellular life (something like jellyfish possibly). These nural networks made from neurons neurotransmitters and synapses or possibly something more primitive would begin forming the most basic of logic over centuries of evolution. But it wouldn't reassemble anything close to reason or intelligence, it wouldn't have eyes, ears or any need for language. At first it would probably spend its first million years just trying to control movement.

    We know that this process would have started from nothing, nural networks with no training data, just a free world to explore. And yet over 500 million years later here we are.

    My argument is that modern nural networks work the same way that biological brains do, at least the mechanism does. The only technical difference is with neurotransmitters and the various dampening and signal boosting that can happen along with nuromodulation. Given enough time and enough training, I firmly believe nural networks could develop reason. And given external sensors it could develop thought from these input signals.

    I don't think we would need to develop a consciousness for it but that it would develop one itself given enough time to train on its own.

    A large hurdle that might arguably be a good thing, is that we are largely in control of the training. When AI is used it does not learn and alter itself, only memorising things currently. But I do remember a time when various AI researchers allowed earlier models to self learn, however the internet being the internet, it developed some wildly bad habits.

  • They work the exact same way we do.

    Citation needed.

    Formatting might be off on some of these, had to convert some papers to text as some were only scanned and I couldn't be bothered writing it all out by hand:

    "The PDP models are inspired by the structure and function of the brain. In particular, they are based on networks of neuron-like units whose interactions resemble those among neurons in the cerebral cortex."
    Parallel Distributed Processing: Explorations in the Microstructure of Cognition McClelland, Rumelhart, & the PDP Research Group

    "The design of artificial neural networks was inspired by knowledge of the brain, in particular the way biological neurons are interconnected and the way they communicate through synapses."
    Deep Learning LeCun, Bengio, Hinton

    "The design of deep learning architectures owes much to our understanding of the hierarchical structure of the visual cortex, particularly as revealed by Hubel and Wiesel’s work on simple and complex cells."
    Neuroscience-Inspired Artificial Intelligence Hassabis et al.

    "The relationship between biological and artificial neural networks has now become a central issue in both neuroscience and AI. Deep networks trained with backpropagation may offer a plausible model of some aspects of human cognition."
    Cognitive computational neuroscience Kriegeskorte & Douglas (2018)

    "Goal-driven deep learning models, when trained to solve behavioral tasks, can develop internal representations that match those found in the brain."
    Using goal-driven deep learning models to understand sensory cortex Yamins & DiCarlo