What can we take away from the ‘stochastic parrot’ saga?
inferencemagazine.substack.com
inferencemagazine.substack.com
This debate isn’t about the computations underlying cognition, it’s about wanting to feel special.
The contention that “it’s a stochastic parrot” usually implied “it’s merely a stochastic parrot, and we know that we must be so much more than that, so obviously this thing falls short”.
But… there never was any compelling proof that we’re anything more than stochastic parrots. Moreover, those folks would say that any explanation of cognition falls short because they can always move the goal posts to make sure that humans are special.
Gather enough parrots on a stage and at least one can theoretically utter a series of seemingly meaningfuly word-like sounds that is legitimately novel, that has never been uttered before.
But I doubt any randomly picked HN user will actually accomplish that in fact before say age 40. Most people just don’t ever get enough meaningful speaking opportunities to make that statistically likely. There’s just too many tens of billions of people that have already existed and uttered words.
It seems to be plausible, to me, given enough parrots.
Edit: actually that looks like it's just an offhand mention of Google's initial report, but I don't really feel like spending more time tracking down details to rebut so silly a claim.
Do you have an argument that makes sense?
That's an assertion that has not been proven one way or the other. It's certainly true that progress has leveled off after an extraordinary climb, and most people would say it's still not a fully general intelligence yet. But we don't know if the next step requires incremental work or if it requires a radically different approach.
So it's just taking a stance on an unproven assertion rather than defining anything fundamental.
It seems eminently plausible that the way cognition works is to take in current context and select the most appropriate next action/token. In fact, it’s hard to think of a form of cognition that isnt “given past/context, predict next thing”
Because we (humans) weren't born into a world with computers, internet, airplanes, satellites, etc
"Complete next token" means that everything is already in the data set. It can remix things in interesting ways, sure. But that isn't the same as creating something new
Edit: I would love to hear someone's idea about how you could "parrot" your way into landing people on the moon without any novel discovery or invention
Everything that is physical is nothing but remixes and recombinations of a very small set of tokens.
I'm not conflating symbolic systems with the physical substrate: they're obviously different levels of abstraction. What I am saying is that symbolic reasoning, language, creativity, and knowledge all emerge from the same underlying physical processes. They're not magic. They're not floating in some Platonic realm. They’re instantiated in real, measurable patterns, whether in neurons or silicon.
You can't have metaphysics without physics. And we have solid evidence, from neuroscience, from pharmacology, from evolutionary biology, that the brain's symbolic output is fundamentally a physical phenomenon. Injuries, chemicals, electrical stimulation, they all modulate “metaphysical” experience in completely physical ways.
Emergence matters here. Yes, atoms aren’t thoughts, but enough atoms arranged the right way do start behaving like a thinking system. That’s the whole point of complex systems theory, chaos theory, and even early AI work like Hofstadter and Dennett. I recommend "Gödel, Escher, Bach", or Melanie Mitchell's "Complexity: A Guided Tour", if you're curious.
If you're arguing there's something else, some kind of unphysical or non-emergent component to knowledge or creativity, I'd honestly love to hear more, because that's a bold claim. But waving away the physical substrate as irrelevant doesn’t hold up under scrutiny.
We're not talking about "physical" with LLMs, we're talking about knowledge and creativity and reasoning, which are metaphysical.
The sum total of human knowledge cannot possibly be purely composed of remixes and recombinations, there has to be some baseline that humans invented for there to even be something to remix!
Lnowledge and creativity absolutely are physical things. It's clear from brain injury studies that there are very localized and specific functions to this creativity.
Drugs also clearly have a very physical affect on these attributes.
I think it's better to look at large language models in the context of Wittgenstein. Humans are more than next token predictors because we participate in “language games” through which we experimentally build up a mental model for what each word means. LLMs learn to “rule follow” via a huge corpus of human text but there’s no actual intelligence there (in a Wittgensteinian analysis) because there’s no “participation” beyond RLHF (in which humans are playing the language games for the machine). There’s a lot to unpack there but that’s the gist of my opinion.
Until we get some rigorous definitions for intelligence or at least break it up into many different facets, I think pie in the sky philosophy is the best we can work with.
An ensemble of LLMs trained identically would generate the same next token(s) forever. But we don't - we generate different sequences.
We are not LLMs.
Best to not assume humans are LLMs.
Not the best human behavior, but certainly human behavior.
LLMs are built with the intention of mimicry. It's no surprise they look like mimicry. If we show a new trick they can't do, and their makers train it to mimic that, can we be blamed for calling that mimicry... mimicry?
The current generation of LLM falls pretty heavy in the mimicary family of AI.
You introduce two categories, state that LLMs are not part of one category, and then conclude that it must be in the other. In reality, the distinction between the two classes is not so clear.
The transformer architecture is quite something, and the number of layers and nodes involved in a typical LLM is staggering. This goes way beyond linear regression.
I don't know whether to acknowledge that Kurt Goedel was a very special parrot
Because "moving the goalposts" (exceedingly intricately, one must add) was indeed a specialty of his (or Cantor's).
Otherwise, imho, humans are superior to parrots because we can kinda derive "understanding" in the face of paradox.. (also Einstein,Lobachevsky, special parrots?)
YMMV, I personally haven't tried to get fresh in this specific way with a you-get-whadya-pay-for chatbot..
(Though some say the average person is also an user of tools like Occam's razor, so maybe we are all stochastic crows?)
Since humans are just stochastic parrots, we don't need to add features or change anything about LLMs. Innovation is for the weak and stupid. We can just scale LLMs by doing nothing except adding more data, parameters and training time.
The status quo bias is unreal. I don't even know what purpose it serves other than discouraging technological progress. The people claiming to champion LLMs by denying the differences between humans and LLMs are their biggest enemies.
It's less "humans are special" and more "this ain't it, chief".
Because, with a reasonable understanding of how LLMS work, the way that they produce text is nothing like the way that my mind works.
A good amount of people spend most of their lives operating that way.
I don't actually like the stochastic parrot argument either to be fair.
I feel like the author is ignoring the various knobs (randomization factors may be a better term) applied to the models during inference that are tuned specifically to make the output more believable or appealing.
Turn the knobs too far and the output is unintelligible garbage. Don't then them far enough and the output feels very robotic or mathematical, its obvious that the output isn't human. The other risk of not turning the knobs far enough would be copyright infringement, but I don't know if that happens often in practice.
Claiming that LLMs aren't stochastic parrots without dealing with the fact that we forced randomization factors into the mix misses a huge potential argument that they are just cleverly disguised stochastic parrots.
What you have described, a probability distribution with carefully-tuned parameters, is perfectly captured by the word stochastic as it's commonly used by statisticians.
I mentioned I don't like the stochastic parrot argument, and that I find this article's argument lacking. Both are for the same reason, the arguments are making claims that we simply can't make while missing the fundamental understanding of what intelligence really is and how human (and other animals) brains work.
Does that mean most people are merely parrots too?
https://math-perturb.github.io/
https://arxiv.org/abs/2502.06453
Leaderboard: https://math-perturb.github.io/#leaderboard
Anyone who continues to use the parrot metaphor should support it with evidence at least as strong as the “On the Biology of a Large Language Model” research by Anthropic which the article refers to:
https://transformer-circuits.pub/2025/attribution-graphs/bio...
What makes you believe that is the case?
Most definitions and measurements of intelligence by most laypeople and psychologists include the ability to reason, with mathematical reasoning widely accepted as part of or a proxy for it. They are imperfect but “intelligence” does not have a universally accepted definition.
Do you have a better measurement or definition?
Not when you already know all of the answers and just have to draw a line between the questions and the answers!
A relevant bit:
“for MATH-P-Hard, we make hard perturbations, i.e., small but fundamental modifications to the problem so that the modified problem cannot be solved using the same method as the original problem. Instead, it requires deeper math understanding and harder problem-solving skills.”
A child could score the same knowing the outcomes and guessing randomly which ones go to which questions
“Could you explain this sentence concisely?
For the skeptics: Scoring just 10% or so in Math-Perturb-Hard below the original MATH Level 5 (hardest) dataset seems in line with or actually better than most people would do.”
Gemini 2.5 Pro:
“The sentence argues that even if a model's score drops by about 10% on the "Math-Perturb-Hard" dataset compared to the original "MATH Level 5" (hardest) dataset, this is actually a reasonable, perhaps even good, outcome. It suggests this performance decrease is likely similar to or better than how most humans would perform when facing such modified, difficult math problems.”
A human society may develop their own math in a base 13 system, or an entirely different way of representing the same concepts. When they can't solve our base 10 math problems in a way that matches how we expect does that mean they are parrots?
Part of the problem here is that we still have yet to land on a clear, standard definition of intelligence that most people agree with. We could look to IQ, and all of its problems, but then we should be giving LLMs an IQ test to answer rather than a math test.
Check out the problems in the MATH dataset, especially Level 5 problems. They are fairly advanced (by most people’s standards) and most are not dependent on which N in the base-N system used to solve them. The answers would be different of course but the structures of the problems and solutions remain largely intact.
Website for tracking IQ measurements of LLMs:
The best one already scores higher than all but the top 10-20% of most populations.
Except, it doesn't. Maybe some math problems do -- or maybe all of them do, when the text isn't in the training set -- but it turns out that most problems can be solved by a machine that regurgitates text, randomly, from all the math problems ever written down.
One of the ways that this debate ends in a boring cul-de-sac is that people leap to conclusions about the meaning of the challenges that they're using to define intelligence. "The problem has only been solved by humans before", they exclaim, "therefore, the solution of the problem by machine is a demonstration of human intelligence!"
We know from first principles what transformer architectures are doing. If the problem can be solved within the constraints of that simple architecture, then by definition, the problem is insufficient to define the limits of capability of a more complex system. It's very tempting to instead conclude that the system is demonstrating mysterious voodoo emergent behavior, but that's a bit like concluding that the magician really did saw the girl in half.
Even a thermostat can show intelligent behaviour. The issue for the thermostat is that all the intelligence has happened ahead of time.
Similarly for feedback. All the operator needs to do is log each input in a file marked for that user and then when new input arrives the old input is used as context in the lookup table. Ultimately, arbitrarily sophisticated intelligent behaviour can be produced without the operator ever having any understanding of it.
Just think about it. All the person in the room does are mechanical manipulations. The person's understanding or not understanding of Chinese language is causally disconnected from everything including functioning of the room. There's zero reasons to look at their understanding to make conclusions about the room.
The second point is that it's somehow about syntactic manipulation specifically. But why? What would change if the person in the room is solving QM-equations of your brain quantum state? Would it mean that the perfect model of your brain doesn't understand English language?
The Chinese Room is also an argument that mere symbolic manipulation is insufficient to model a human mind.
As for the QM-equations, the many-body problem in QM is your enemy. You would need a computer far larger than the entire universe to simulate the quantum states of a single neuron, never mind a human brain.
Please, read again what I wrote.
Regarding "larger than Universe": "the argument" places no restrictions on runtime or space complexity of the algorithm. It's just another intuitive notion: syntactic processing is manageable by a single person, other kinds of processing aren't.
I'm sorry for the confrontational tone, but I really dismayed that this thing keeps floating and keeps being regarded as a foundational result.
Theoretical computer science established early that input-output behavior captures the essence of computation. The causal mechanism underlying the computation does not matter, because all plausible mechanisms seem to be fundamentally equivalent.
The Chinese room argument showed that this does not extend to intelligence. That intelligence is fundamentally a causal rather than computational concept. That you can't use input-output behavior to tell the difference between an intelligent entity and a hard-coded lookup table.
On one level, LLMs are literally hard-coded lookup tables. But they are also compressed in a way that leads to emergent structures. If you use the LLM through a chat interface, you are interacting with a Chinese room. But if the LLM has other inputs beyond your prompt, or if it has agency to act on its own instead of waiting for input, it's causally a different system. And if the system can update the model on its own instead of using a fixed lookup table to deal with the environment, this is also a meaningful causal difference.
Searle's argument doesn't actually show anything. It just illustrates a complex system that appears intelligent. Searle then asserts, without any particular reasoning, that the system is not intelligent, simply because, well, how could it be, it's just a bunch of books and a mindless automaton following them?
It's a cyclic argument: A non-human system can't be intelligent because, uhh, it's not human.
This is wrong. The room as a whole is intelligent, and knows Chinese.
People have, of course, made this argument, since it is obivous. Searle responds by saying "OK, well now imagine that the man in the room memorizes all the books and does the entire computation in his head. Now where's the intelligence???" Ummm, ok, now the man is emulating a system in his head, and the system is intelligent and knows Chinese, even though the man emulating it does not -- just like how a NES emulator can execute NES CPU instructions even though the PC it runs on doesn't implement them.
Somehow Searle just doesn't comprehend this. I guess he's not a systems engineer.
As to whether a lookup table can be intelligent: I assert that a lookup table that responds intelligently to every possible query is, in fact, intelligent. Of course, such a lookup table would be infinite, and thus physically impossible to construct.
Intelligence, as we commonly understand it, is something humans have but we currently can't define. Turing proposed a definition based on the observable behavior of a system. We take an aspect of human behavior people consider intelligent and test the behavior of other systems against that. If we can't tell the difference between human behavior and the behavior of an artificial system, we consider the artificial system intelligent.
Searle used a thought experiment to argue that Turing's definition was not useful. That it did not capture the concept of intelligence in the way people intuitively understand it. If it turns out there was a person speaking Chinese answering the questions, the behavior is clearly intelligent. But if there was only a simple mechanism and a precomputed lookup table, it doesn't feel intelligent.
Maybe we need a better definition of intelligence. Maybe intelligence in the sense people intuitively understand it is not a useful concept. Or maybe something else. We don't know that, because we don't really understand intelligence.
I think ultimately I decided the Chinese room experiment was actually flawed and didn't reveal what it purported to reveal. From a neurophysiological viewpoint: The chinese room is very much the cartesian theater, and Searle places himself as the little man watching the screen. Since the cartesian theater does not exist, he's never going to see a movie.
I might be missing a more subtle point of Searle's though; maybe the chinese room experiment should be read differently?
I think a flaw of the argument is that the way it is framed makes it sound like the system is simple (like a "lookup table") which tricks people's intuitions into thinking it doesn't sound "intelligent". But the actual algorithm implemented by the "Chinese room" would in fact be insanely complex. In any case, I think Searle's intuition here is simply wrong. The system is in fact intelligent, even if it's just a series of lookup tables.
What this sounds like to me is that you don't place much value on the system actually understanding what it is doing. The system does not understand the input or the output, it is just a series of lookup tables
If you ask it about the input you just gave it, can it remember that input?
If you ask it to explain your previous input, and explain the output, can it do that? Do those have to be made into new entries in the lookup table first? Does it have the ability to create new entries in the lookup table without being told to do so?
It seems to me you consider "intelligence" a very low bar
yes. Searle's fundamental misunderstanding is that "syntax is insufficient for semantics", but this is just nonsense that could be only believed by someone that has never actually tried to derive meaning from syntactic transformation (e.g. coding/writing a proof)
What? Why? Of course it understands.
> If you ask it about the input you just gave it, can it remember that input?
The system Searle describes has memory, yes.
Perhaps you are getting at the fact that LLMs performing inference don't have memory, but actually they can be given memory via context. You might argue that this is not the same as human memory, but you don't know this. Maybe the way the brain works is, we spend each day filling our context, and then the training happens when we sleep. If that is true, are humans not intelligent then?
However, the stochastic parrot arguments (and Gary Marcus in many of his writings) made specific, unambiguous empirical predictions about how LLMs would never be able to do many things, such as 'add numbers'. For example, in the original Bender & Koller 2020 paper "Climbing towards NLU: On Meaning, Form, and Understanding in the Age of Data" (which lays out the core ideas which are the justification for the 2021 paper that actually introduced the rhetoric of 'stochastic parrot') made many clear, falsifiable statements about what LLMs would never be able to do; here's one of them: https://aclanthology.org/2020.acl-main.463.pdf#page=14
> To get a sense of how existing LMs might do at such a task, we let GPT-2 complete the simple arithmetic problem 'Three plus five equals'. The five responses below, created in the same way as above, show that this problem is beyond the current capability of GPT-2, and, we would argue, any pure LM.
One could say many things about this claim, but not that it is a 'strictly philosophical one'.
I disagree right away. There are more sophisticated probability models than lookup tables.
> It'd be running a search for the most similar pattern in its training data and copying this.
Also untrue. Sophisticated probability models combine probabilities based on combining all the bits of context, and by fuzzing similar tokens together via compressing (i.e. you don't care what particular token is used, just that a similar one is used.)
They're parrots, just better parrots than this person can conceive of.
Continuing the metaphor, we never wanted to work in a pet shop in the first place. We wanted to be … lumberjacks! Floating down the mighty rivers of British Columbia! With our best girls by our side!
So what’s “dead?” Overconfidently assuming you can know how an LLM does something without actually investigating it.
Problem is, we don’t even know what makes us think. So you can jump to any conclusion and nobody could really tell if you’re wrong.
We do know how transformers and layers work. They’re algorithms that crunch numbers. A great deal of numbers. And we can use the training set to generate plausible outputs given some input. Yes, stochastic parrot is a reduction of all the technical sophistication in LLMs. But it’s not entirely baseless. At the end of the day it is copying what’s in the training data. In a very clever way.
However, resist the temptation to believe we understand human brains and human thought. And resist the temptation to anthropomorphize algorithms. It’s data and patterns.
…
The problem with this theory, is that, alas, it isn’t true.
If a language model was just a stochastic parrot, when we looked inside to see what was going on, we’d basically find a lookup table. … But it doesn’t look like this.
But does that matter? My understanding is that, if you don’t inject randomness (“heat”) into a model while it’s running, it will always produce the same output for the same input. In effect, a lookup table. The fancy stuff happening inside that the article describes is, in effect, [de]compression of the lookup table.
Of course, maybe that’s all human intelligence is too (the whole ‘free will is an illusion in a deterministic universe’ argument is all about this) - but just because the internals are fancy and complicated doesn’t mean it’s not a lookup table.
I guess it means that we can throw away the notion that "it can be represented as a lookup table" has some profound meaning. Without further clarifications, at least. Finite/infinite lookup table, can/can't be constructed in time polynomial in the number of entries. Things like that.
So far, LLM models have not demonstrated grasp on dual language phonetic jokes and false cognates.
Humans learn a second language very quickly, and false cognates that work on phonetics are the first steps in doing so, doesn't require a genius to understand.
I am yet to see an LLM that can demonstrate that. They can translate it, or repeat known false cognates, but can't come up with new ones on the spot.
If they do acquire that, we will come up with another creative example of what humans can do that machines can't.
I am sure there are false cognate signs among them, and dual users of both sign languages can appreciate them.
She knows ASL as well, but I don't think she knows any other dialect of sign language (is dialect the right term? I'm not actually sure).
The page says "Something has gone terribly wrong :(".
He's not dead, he's resting.
LLMs are stochastic parrots and so are humans - but humans still get to be special. Humans are more stochastic as we act on far more input than a several-thousand token prompt.
It's true, and trivial, that all next word predictors are stochastic and are designed to generate output based on information from their training data.
The claim that this generation merely "parrots" the training data is more significant, but obviously false if you interact with these models at all.
As someone else pointed out, I think there's deep philosophical issues about intelligence and consciousness underlying all this and I'm not sure it can be resolved this way. In some sense, we all might be stochastic parrots — or rather, I don't think the problem can be waved away without deeper and more sophisticated treatments on the topic.
This author has no idea what's going on.
The AI didn't just start trying to predict the next word in other poems, it was explicitly instructed to do so. It then sucked in a bunch of poems and parroted them out.
And... the author drastically over-represents its success with a likely cherry-picked example. When I gave Claude lines to rhyme with, it gave me back "flicker" to rhyme with "killer" and "function" to rhyme with "destruction". Of the 10 rhymes I tried, only two actually matched two syllables ("later/creator" and "working"/"shirking")I'm not sure how many iterations the author had to run to find a truly unusual rhyme like "rabbit/grab it", but it pretty obviously is selection bias.
And...
I actually agree with the other poster who says that part of this stochastic parrot argument is about humans wanting to feel special. Exceptionalism runs deep: we want to believe our group (be it our nation, our species, etc.) are better than other groups. It's often wrong: I don't think we're particularly unique in a lot of aspects--it's sort of a combination of things that makes us special if we are at all.
AI are obviously stochastic parrots if you know how they work. The research is largely public and unless there's something going on in non-public research, they're all just varieties of stochastic parroting.
But, these systems were designed in part off of how the human brain works. I do no think it's in evidence at all that humans aren't stochastic parrots. The problem is that we don't have a clear definition of what it means to understand something that's clearly distinct from being a stochastic parrot. At a certain level of complexity of stochastic parroting, a stochastic parrot is likely indistinguishable from someone who truly understands concepts.
I think ultimately, the big challenge for AI isn't that it is a stochastic parrot (and it is a stochastic parrot)--I think a sufficiently complex and sufficiently trained stochastic parrot can probably be just as intelligent as a human.
I think the bigger challenge is simply that entire classes of data simply have not been made available to AI, and can't be made available with current technology. Sensory data. The kind of data a baby gets from doing something and seeing what happens. Real-time experimentation. I think a big part of why humans are still ahead of AI is that we have a lot of implicit training we haven't been able to articulate, let alone pass on to AI.