4 ms·

There is debate over whether or not every person has an internal monologue, and a nontrivial number of people self-report that they don't experience conscious d

• nomel 3y ago
  > It seems pretty safe to say that those stimuli are not being processed using a language model. Could you expand on that? I don’t see how it follows. If there is unconscious output, why wouldn’t it suggest there’s something analogous to a “unthinking” language model stuck behind our conscious, self reflective, bits? How does unconscious output prove that it’s not something similar to a language model? Or, are you speaking on a technicality, saying it’s not a literal language model? I think the idea is something similar and, obviously, multi modal, not a literal LLM/text predictor implementation like runs on GPUs.
  • danShumway 3y ago
    Short answer, the responses are happening too fast for that stimuli to be going through the language centers of our brain. In order to say that the processing of that stimuli is like a language model, we would need to widen our definition of a language model to encompass basically any communication between neurons happening within the brain. ---- So on one hand, you could call that a technicality, on the other hand I would say that if we reduce "like an LLM" to mean "responds to a stimuli by generating a response", at that point "like an LLM" is so broad as to be meaningless. Sure, we can define "humans use a language model" as "human brains generate signals to other parts of the human brain in response to stimuli" -- but that's not really a useful thing to say when we're comparing human reasoning to GPT-4. If we're talking so broadly about reasoning, then it would be just as accurate to say that human brains are like Markov chains. After all, Markov chains interpret inputs and transform them into other signals that they send to other parts of the chain. But if I told you that GPT-4 was just a Markov chain you would (rightfully) disagree with me on that and you would (rightfully) say that I was oversimplifying what's going on within GPT-4's model. Typically, when people say that human beings are text generators, they're talking in the context of GPT-4; they're trying to claim that GPT-4's behavior is analogous to human behavior. Whether or not GPT-4 can reason under any definition of reasoning (direct analogy to human thought should not be the only way we think about reasoning), the way GPT works and the way that it's trained and taught is extremely alien to how human brains work and how humans learn. Broadening the definition doesn't change the fact that LLMs and human brains are very different in practical ways that influence their observable behaviors. ---- I don't think people are consciously trying to do this, but the debate over whether humans are LLMs often ends up feeling like an attempt to simultaneously broaden and narrow a definition at the same time. What people seem to want to do is to describe LLMs extremely broadly but then assume that conclusions they draw from that broad definition are necessarily applicable to a very narrow definition of an LLM (specifically, GPT and similar models). But once we define an LLM broadly enough to encompass the entirety of human reasoning, we have also broadened the category enough to encompass a lot of processes that no one would call reasoning. What doesn't transform inputs into other forms and pass them along to another object? An AC/DC electrical converter does that, but that doesn't mean it can think. It's like claiming, "computers are made of carbon and so are humans, so computers are basically the same as humans." Well, wait a second -- the category as you've defined it is so broad that objects within that category can no longer be assumed to share other attributes with each other.
    • nomel 3y ago
      > to be going through the language centers of our brain. > "like an LLM" is so broad as to be meaningless. I don't think this is a fair interpretation, in the context of this comment chain. To me, a charitable interpretation is that "like an LLM" means no feedback loop, no refinement, just input to output, like an LLM. I point my brain at problems, using my conscious intent, and I see answers. I don't really consider myself consciously involved in those answers, and I do have to check them, and sometimes iterate, with that iteration almost always involving some solidification of the answers/concepts by externalizing them in writing, drawing, etc. Point at solidification, get response, repeat. The "me" in my head is the one talking, pointing, receiving, checking, and mixing. The "here's your answer" intuition that I experience is the "like an LLM thing". My understanding of my own thought process follows the historic interpretation: the pre-frontal cortex is a new attachment, and probably the one that is the active participant, and planner, that is "me". The fast, but unrefined, input->output system being the rest of it, and is probably closer to the experience of a monkeys day to day activities.
      • danShumway 3y ago
        > "like an LLM" means no feedback loop, no refinement, just input to output, like an LLM. But that's not what an LLM is, an LLM is a large language model. The distinguishing characteristics between an LLM and other AIs are not whether or not an LLM has a feedback loop. Lots of AI categories (arguably the majority of AIs) generate output from a single set of inputs as a single "operation" (ignoring the fact that most neural networks including LLMs internally have multiple layers and do in fact have multiple transformation steps, but whatever, we can treat that as one step for the purposes of conversation). If anything, LLMs are the exception here in that they very often do have a feedback loop during normal usage; they are most commonly used in a conversational context where they generate the next "chunk" of a conversation after being fed back their previous answers alongside the followup responses of the person working with them. Arguably the feedback loop of an LLM is that as a conversation progresses, its future output is based on its previous output, which very literally becomes its new input after being marked up and extended by a human being. That is notably more feedback than many other AIs get. Doubly so if you're trying to argue that an LLM can encompass a multi-modal setup, because suddenly refinement and feedback between multiple models is a core part of the final product. So I'm not sure I agree that an LLM fits into that category in the first place, but even assuming it does, if your definition of an LLM is is just that it takes an input and turns it into an output as a single step without help... that is just such a broad category, I would guess the majority of AIs fall into that category. It's not what makes LLMs special; most predictive neural networks take a single set of inputs and generate a single set of outputs without intermediary human input. What makes LLMs interesting as a category is the training and structure and quirks of how they work, what makes them interesting is the differences between LLMs and other AI techniques. To jump back to the Markov chain again, Markov chains do not have a feedback loop or refinement step: they take an input and map it to an output as a single step. Are Markov chains LLMs? Is any data transformer that operates in a single step an LLM? That is a really broad definition to use. ---- And we run into the same problems, because if you're arguing that a human brain is like an LLM and what you're really saying is, "it's kind of like an AI in general" -- well, that doesn't really say anything about whether a human brain is specifically similar to a system like GPT. There are lots of different ways to build a neural network, LLMs are one strategy. When you say this: > The "here's your answer" intuition that I experience is the "like an LLM thing". What this sentence is actually saying is that you experience conclusions and knowledge where you don't know the source or where the process of your brain generating a decision or information happens unconsciously. The sentence is just saying that there are unconscious parts of your brain where you are not an active participant in the thinking process and it feels like the a spontaneous transformation from input to output. But does that really strike you as a strong indicator that your brain is like an LLM, or does that sound more like a general description of almost any black-box system where the mechanisms are hidden and you can only examine the inputs and outputs? To say that there are parts of our thinking process that we're not able to consciously observe or examine is really just saying that parts of our thinking process resemble a black-box oracle. It's not making a strong claim about GPT.
  • eichin 3y ago
    Isn't this the essence of teddy-bear/rubber-duck debugging? You're forcing yourself to run your abstract concepts through the narrow channel of language and instantiating it in a way you haven't previously done (because if you had, you'd have solved the problem already and wouldn't need to be talking it through.) (I'm very specifically talking about the case where you explain and noone else contributes/responds, and you come up with the answer solely through trying to explain/speak it.)
• the_gipsy 3y ago
  Internal monologue can only be self-reported, and I kind of feel like telling someone that says they are ill that it's just all in their head, but... I don't think that any functioning human could possibly have no internal monologue, just lack of introspection.