This paper presents a theoretical proof that AGI systems will structurally collapse under certain semantic conditions — not due to lack of compute, but because of how entropy behaves in heavy-tailed decision spaces.
The idea is called IOpenER: Information Opens, Entropy Rises. It builds on Shannon’s information theory to show that in specific problem classes (those with α ≤ 1), adding information doesn’t reduce uncertainty — it increases it. The system can’t converge, because meaning itself keeps multiplying.
The core concept — entropy divergence in these spaces — was already present in my earlier paper, uploaded to PhilArchive on June 1. This version formalizes it. Apple’s study, The Illusion of Thinking, was published a few days later. It shows that frontier reasoning models like Claude 3.7 and DeepSeek-R1 break down exactly when problem complexity increases — despite adequate inference budget.
I didn’t write this paper in response to Apple’s work. But the alignment is striking. Their empirical findings seem to match what IOpenER predicts.
Curious what this community thinks: is this a meaningful convergence, or just an interesting coincidence?
Links:
This paper (entropy + IOpenER): https://philarchive.org/archive/SCHAIM-14
First paper (ICB + computability): https://philpapers.org/archive/SCHAII-17.pdf
Apple’s study: https://machinelearning.apple.com/research/illusion-of-think...
I am sympathetic to the kind of claims made by your paper. I like impossibility results and I could believe that for some definition of AGI there is at least a plausible argument that entropy is a problem. Scalable quantum computing is a good point of comparison.
But your paper is throwing up crank red flags left and right. If you have a strong argument for such a bold claim, you should put it front and centre: give your definition of AGI, give your proof, let it stand on its own. Some discussion of the definition is useful. Discussion of your personal life and Kant is really not.
Skimming through your paper, your argument seems to boil down to "there must be some questions AGI gets wrong". Well since the definition includes that AGI is algorithmic, this is already clear thanks to the halting problem.
Thanks for this - Looking forward to reading the full paper.
That said, the most obvious objection that comes to mind about the title is that … well, I feel that I’m generally intelligent, and therefore general intelligence of some sort is clearly not impossible.
Can you give a short précis as to how you are distinguishing humans and the “A” in artificial?
That about ‘cogito ergo sums it up’ doesn’t it?
Intelligence is clearly possible. My gut feeling is our brain solves this by removing complexity. It certainly does so, continuously filtering out (ignoring) large parts of input, and generously interpolating over gaps (making stuff up). Whether this evolved to overcome this theorem I am not intelligent enough to conclude.
Sure I can (and thanks for writing)
Well, given the specific way you asked that question I confirm your self assertion - and am quite certain that your level of Artificiality converges to zero, which would make you a GI without A...
- You stated to "feel" generally intelligent (A's don't feel and don't have an "I" that can feel) - Your nuanced, subtly ironic and self referential way of formulating clearly suggests that you are not a purely algorithmic entity
A "précis" as you wished: Artificial — in the sense used here (apart from the usual "planfully built/programmed system" etc.) — algorithmic, formal, symbol-bound.
Humans as "cognitive system" have some similar traits of course - but obviously, there seems to be more than that.
>but obviously, there seems to be more than that.
I don't see how that's obvious. I'm not trying to be argumentative here, but it seems like these arguments always come down to a qualia, or the insistence that humans have some sort of 'spark' that machines don't have, therefore: AGI is not possible since machines don't have it.
I also don't understand the argument that "Your nuanced, subtly ironic and self referential way of formulating clearly suggests that you are not a purely algorithmic entity". How does that follow?
What scientific evidence is there that we are anything other than a biochemical machine? And if we are a biochemical machine, how is that inherently capable of more than a silicon based machine is capable of?
> I also don't understand the argument that "Your nuanced, subtly ironic and self referential way of formulating clearly suggests that you are not a purely algorithmic entity". How does that follow?
It doesn't follow.
Trivially demonstrated by the early LLM that got Blake Lemonie to break his NDA also emitting words which suggested to Lemonie that the LLM had an inner life.
Or, indeed, the output device y'all are using to read/listening to my words, which is also successfully emitting these words despite the output device very much only following an algorithm that simply recreates what it was told to recreate. "Ceci n'est pas une pipe", etc. https://en.wikipedia.org/wiki/The_Treachery_of_Images
Consciousness is an issue. If you write a program to add 2+2, you probably do not believe some entity poofs into existence, perceives itself as independently adding 2+2, and then poofs out of existence. Yet somehow, the idea of an emergent consciousness is that if you instead get it to do 100 basic operations, or perhaps 2^100 then suddenly this becomes true? The reason one might believe this is not because it's logical or reasonable - or even supported in any way, but because people assume their own conclusion. In particular if one takes a physicalist view of the universe then consciousness must be a physical process and so it simply must emerge at some sufficient degree of complexity.
But if you don't simply assume physicalism then this logic falls flat. And the more we discover about the universe, the weirder things become. How insane would you sound not that long ago to suggest that time itself would move at different rates for different people at the same "time", just to maintain a perceived constancy of the speed of light? It's nonsense, but it's real. So I'm quite reluctant to assume my own conclusion on anything with regards to the nature of the universe. Even relatively 'simple' things like quantum entanglement are already posing very difficult issues for a physicalist view of the universe.
My issue is that from a scientific point of view, physicalism is all we have. Everything else is belief, or some form of faith.
Your example about relativity is good. It might have sounded insane at some point, but it turns out, it is physics, which nicely falls into the physicalism concept.
If there is a falsifiable scientific theory that there is something other than a physical mechanism behind consciousness and intelligence, I haven't seen it.
Boltzmann brains and A. J. Ayer's "There is a thought now".
Ages ago, it occurred to me that the only thing that seemed to exist without needing a creator, was maths. That 2+2 was always 4, and it still would be even if there were not 4 things to count.
Basically, I independently arrived at similar conclusion as Max Tegmark, only simpler and without his level of rigour: https://benwheatley.github.io/blog/2018/08/26-08.28.24.html
(From the quotation's date stamp, 2007, I had only finished university 6 months earlier, so don't expect anything good).
But as you'll see from my final paragraph, I no longer take this idea seriously, because anything that leads to most minds being free to believe untruths, is cognitively unstable by the same argument that applies to Boltzmann brains.
MUH leads to Aleph-1 infinite number of brains*. I'd need a reason for the probability distribution over minds to be zero almost everywhere in order for it to avoid the cognitively instability argument.
* if there is a bigger infinity, then more; but I have only basic knowledge of transfinites and am unclear if the "bigger" ones I've heard about are considered "real" or more along the lines of "if there was an infinite sequence of infinities, then…"
Oh no, I am not at all trying to find an explanation of why this is (qualia etc.). There is simply no necessity for that. It is interesting, but not part of the scientific problem that i tried to find an answer to.
The proof (all three of them) holds without any explanatory effort concerning causalities around human frame-jumping etc.
For this paper, It is absolutely sufficient to prove that a) this cannot be reached algorithmically and that b) evidence clearly shows that humans can (somehow) do this , as they have already done this (quite often).
> this cannot be reached algorithmically
> humans can (somehow) do this
Is this not contradictory?
Alternatively, in order to not be contradictory doesn't it require the assumption that humans are not "algorithmic"? But does that not then presuppose (as the above commenter brought up) that we are not a biochemical machine? Is a machine not inherently algorithmic in nature?
Or at minimum presupposes that humans are more than just a biochemical machine. But then the question comes up again, where is the scientific evidence for this? In my view it's perfectly acceptable if the answer is something to the effect of "we don't currently have evidence for that, but this hints that we ought to look for it".
All that said, does "algorithmically" here perhaps exclude heuristics? Many times something can be shown to be unsolvable in the absolute sense yet readily solvable with extremely high success rate in practice using some heuristic.
> Alternatively, in order to not be contradictory doesn't it require the assumption that humans are not "algorithmic"? But does that not then presuppose (as the above commenter brought up) that we are not a biochemical machine? Is a machine not inherently algorithmic in nature?
No, computation is algorithmic, real machines are not necessarily (of course, AGI still can't be ruled out even if algorithmic intelligence is, only AGI that does not incorporate some component with noncomputable behavior.)
> computation is algorithmic, real machines are not necessarily
Author seems to assume the latter condition is definitive, i.e. that real machines are not, and then derive extrapolations from that unproven assumption.
> No, computation is algorithmic, real machines are not necessarily
As the adjacent comment touches on are the laws of physics (as understood to date) not possible to simulate? Can't all possible machines be simulated at least in theory? I'm guessing my knowledge of the term "algorithmic" is lacking here.
Using computation/algorithmic methods we can simulate nonalgorithmic systems. So the world within a computer program can behave in a nonalgorithmic way.
Also, one might argue that universe/laws of physics are computational.
OP seems to have a very confused idea of what an algorithmic process means... they think the process of humans determining what is truthful "cannot possibly be something algorithmic".
Which is certainly an opinion.
> whatever it is: it cannot possibly be something algorithmic
https://news.ycombinator.com/item?id=44349299
Maybe OP should have looked at a dictionary for what certain words actually mean before defining them to be something nonsensical.
> For this paper, It is absolutely sufficient to prove that a) this cannot be reached algorithmically and that b) evidence clearly shows that humans can (somehow) do this , as they have already done this (quite often).
The problem with these kinds of arguments is always that they conflate two possibly related but non-equivalent kinds of computational problem solving.
In computability theory, an uncomputability result essentially only proves that it's impossible to have an algorithm that will in all cases produce the correct result to a given problem. Such an impossibility result is valuable as a purely mathematical result, but also because what computer science generally wants is a provably correct algorithm: one that will, when performed exactly, always produce the correct answer.
However, similarly to any mathematical proof, a single counter-example is enough to invalidate a proof of correctness. Showing that an algorithm fails in a single corner case makes the algorithm not correct in a classical algorithmic sense. Similarly, for a computational problem, showing that any purported algorithm will inevitably fail even in a single case is enough to prove the problem uncomputable -- again, in the classical computability theory sense.
If you cannot have an exact algorithm, for either theoretical or practical reasons, and you still want a computational method for solving the problem in practice, you then turn to heuristics or something else that doesn't guarantee correctness but which might produce workable results often enough to be useful.
Even though something like the halting problem is uncomputable in the classical, always-inevitably-produces-correct-answer-in-finite-time sense, that does not necessarily stop it from being solved in a subset of cases, or to be solved often enough by some kind of a heuristic or non-exact algorithm to be useful.
When you say that something cannot be reached algorithmically, you're saying it's impossible to have an algorithm that would inevitably, systematically, always reach that solution in finite time. And you would in many cases be correct. Symbolic AI research ran into this problem due to the uncomputability of reasoning in predicate logic. (Uncomputability is not the main problem that symbolic AI ran into but it was one of them.)
The problem is that when you say that humans can somehow do this computationally impossible thing, you're not holding human cognition or problem solving to the same standard of computational correctness. We do find solutions to problems, answers to questions, and logical chains of reasoning, but we aren't guaranteed to.
You do seem to be aware of this, of course.
But you then run into the inevitable question of what you mean by AGI. If you hold AGI to the standard of classical computational correctness, to which you don't hold humans, you're correct that it's impossible. But you have also proven nothing new.
A more typical understanding of AGI would be something similar to human cognition -- not having formal guarantees but working well enough for operating in, understanding, or producing useful results the real world. (Human brains do that well in the real world -- thanks to having evolved in it!)
In the latter case, uncomputability results do not prove that kind of AGI to be impossible.
> What scientific evidence is there that we are anything other than a biochemical machine? And if we are a biochemical machine, how is that inherently capable of more than a silicon based machine is capable of
Iron and copper are both metals but only one can be hardened into steel
There is no reason why we should assume a silicon machine must have the same capabilities as a carbon machine
Unless you can show - even a single example would do - that we can compute a function that is outside the Turing computable set, then there is a very strong reason that we should assume a silicon machine has the same capabilities as a carbon machine to compute.
Yeah, but bronze also makes great swords… what’s the point here?
> You stated to "feel" generally intelligent (A's don't feel and don't have an "I" that can feel) - Your nuanced, subtly ironic and self referential way of formulating clearly suggests that you are not a purely algorithmic entity
This is completely unrelated to the proof in the link. You have to clearly explain what reasoning in your argument for “AGI is impossible” also implies human intelligence is possible. You can’t just jump to conclusions “you sound human therefore intelligence is possible”
It's simple: Either your proof holds for NGI as much as for AGI, or neither, or you can clearly define what differentiates them that makes it work for one and not the other.
So, in a word: a) there is no ghost in the machine when the machine is a formal symbol-bound machine. And b) to be “G” there must be a ghost in the machine.
Is that a fair summary of your summary?
If so do you spend time on both a and b in your papers? Both are statements that seem to generate vigorous emotional debate.
These are.. very weak rebuttals.
I think you’ve just successfully proven that general human intelligence indeed does not exist.
Not the person asked, but in time honoured tradition I will venture forth that the key difference is billions of years of evolution. Innumerable blooms and culls. And a system that is vertically integrated to its core and self sustaining.
AI can be, and often are, trained by simulated evolution.
I would argue that you are not a general intelligence. Humans have quite a specific intelligence. It might be the broadest, most general, among animal species, but it is not general. That manifests in that we each need to spend a significant amount of time training ourselves for specific areas of capability. You can't then switch instantly to another area without further training, even though all the context materials are available to you.
This seems like a meaningless distinction in context. When people say AGI, they clearly mean "effectively human intelligence". Not an infallible, completely deterministic, omniscient god-machine.
There's a great deal of space between effectively human and god machine. Effectively human meaning it takes 20 years to train it and then it's good at one thing and ok at some other things, if you're lucky. We expect more from LLMs right now, like being able to have very broad knowledge and be able to ingest vastly more context than a human can every time they're used. So we probably don't just think of or want a human intelligence.. or we want an instant specific one, and the process of being about to generate an instant specific one would surely be further down the line to your god like machine anyway.
The measure of human intelligence is never what humans are good at, but rather the capabilities of humans to figure out stuff they haven't before. Meaning, we can create and build new pathways inside our brains to perform and optimize tasks we have not done before. Practicing, then, reinforces these pathways. In a sense we do what we wish LLMs could - we use our intelligence to train ourselves.
It's a long (ish) process, but it's this process that actually composes human intelligence. I could take a random human right now and drop them somewhere they've never been before, and they will figure it out.
For example, you may be shocked to know that the human brain has no pathways for reading, as opposed to spoken language. We have to manually make those. We are, literally, modifying our brains when we learn new skills.
> For example, you may be shocked to know that the human brain has no pathways for reading, as opposed to spoken language.
I'm not shocked at all.
> I could take a random human right now and drop them somewhere they've never been before, and they will figure it out.
Yes, well not really. You could drop them anywhere in the human world, in their body. And even then, if you dropped me into a warehouse in China I'd have no idea what to do, I'd be culturally lost and unable to understand the language. And I'd want to go home. So yes you could drop in a human but they wouldn't then just perform work like an automonon. You couldn't drop their mind into a non human body and expect anything interesting to happen, and you certainly couldn't drop them anywhere inhospitable. Nearer to your example, you couldn't drop a football player into a maths convention and a maths professor into a football game and expect good results. The point of an AI is to be useful. I think AGI is very far away and maybe not even possible, whereas specific AIs are already abound.
It doesn't take 20 years for humans to train new tasks. Perhaps to master very complicated tasks, but there is many tasks you can certainly learn to do in a short amount of time. For example, "Take this hammer, and put nails in top 4 corners of this box, turn it around, do the same". You can master that relatively easy. An AGI ought to be able to practically all such tasks.
In any case, general intelligence merely means the capability to do so, not the amount of time it takes. I would certainly bet a physical theorist for example can learn to code in a matter of days despite never having been introduced to a computer before, because our intelligence is based on a very interconnected world model.
It takes about 10 years to train a human to do anything useful after creation.
A 4 year old can navigate the world better than any AI robot can
While I'm constantly disappointed by self driving cars, I do get the impression they're better at navigating the world than I was when I was four. And in public roads specifically, better than when I was fourteen.
Unless you can prove that humans exceed the Turing computable, the headline is nonsense unless you can also show that the Church-Turing thesis isn't true.
Since you don't even appear to have dealt with this, there is no reason to consider the rest of the paper.
> In plain language:
> No matter how sophisticated, the system MUST fail on some inputs.
Well, no person is immune to propaganda and stupididty, so I don't see it as a huge issue.
I have no idea how you believe this relates to the comment you replied to.
If I'm understanding correctly, they are arguing that the paper only requires that an intelligent system will fail for some inputs and suggest that things like propaganda are inputs for which the human intelligent system fails. Therefore, they are suggesting that the human intelligent system does not necessarily refute the paper's argument.
If so, then the papers argument isn't actually trying to prove that AGI is impossible, despite the title, and the entire discussion is pointless.
But what then is the relevance of the study?
I suppose it disproves embodied, fully meat-space god if sound?
I'm looking at the title again and it seems wrong, because AGI ~ human intelligence. Unless human intelligence has non physical components to it.
I think OP answered the question here:
https://news.ycombinator.com/item?id=44349516
No, he didn't.
It's not even close to addressing the argument.
could you explain for a layman
I'm not sure if this will help, but happy to elaborate further:
The set of Turing computable functions is computationally equivalent to the lambda calculus, is computationally equivalent to the generally recursive functions. You don't need to understand those terms, only to know that these functions define the set of functions we believe to include all computable functions. (There are functions that we know to not be computable, such as e.g. a general solution to the halting problem)
That is, we don't know of any possible way of defining a function that can be computed that isn't in those sets.
This is basically the Church-Turing thesis: That a function on the natural numbers can be effectively computable (note: this has a very specific meaning, it's not about performance) only if it is computable by a Turing machine.
Now, any Turing machine can simulate any other Turing machine. Possibly in a crazy amount of time, but still.
The brain is at least a Turing machine in terms of computabilitity if we treat "IO" (speaking, hearing, for example) as the "tape" (the medium of storage in the original description of the Turing machine). We can prove this, since the smallest Turing machine is a trivial machine with 2 states and 3 symbols that any moderate functional human is capable of "executing" with pen and paper.
(As an aside: It's almost hard to construct a useful computational system that isn't Turing complete; "accidental Turing completeness" regularly happens, because it is very trivial to end up with a Turing complete system)
An LLM with a loop around it and temperature set to 0 can trivially be shown to be able to execute the same steps, using context as input and the next token as output to simulate the tape, and so such a system is Turing complete as well.
(Note: In both cases, this could require a program, but since for any Turing machine of a given size we can "embed" parts of the program by constructing a more complex Turing machine with more symbols or states that encode some of the actions of the program, such a program can inherently be embedded in the machine itself by constructing a complex enough Turing machine)
Assuming we use a definition of intelligence that a human will meet, then because all Turing machines can simulate each other, then the only way of showing that an artificial intelligence can not theoretically be constructed to at least meet the same bar is by showing that humans can compute more than the Turing computable.
If we can't then "worst case" AGI can be constructed by simulating every computational step of the human brain.
Any other argument about the impossibility of AGI inherently needs to contain a something that disproves the Church-Turing thesis.
As such, it's a massive red flag when someone claims to have a proof that AGI isn't possible, but haven't even mentioned the Church-Turing thesis.
Compute functions != Intelligence though.
For example learning from experience (which LLMs cannot do because they cannot experience anything and they cannot learn) is clearly an attribute of an intelligent machine.
LLMs can tell you about the taste of a beer, but we know that they have never tasted a beer. Flight simulators can't take you to Australia, no matter how well they simulate the experience.
> Compute functions != Intelligence though.
If that is true, you have a proof that the Church-Turing thesis is false.
> LLMs can tell you about the taste of a beer, but we know that they have never tasted a beer. Flight simulators can't take you to Australia, no matter how well they simulate the experience.
For this to be relevant, you'd need to show that there are possible sensory inputs that can't be simulated to a point where the "brain" in question - be it natural or artificial - can't tell the difference.
Which again, would boil down to proving the Church-Turing thesis wrong.
I think that may depend on how someone defines intelligence. For example, if intelligence includes the ability to feel emotion or appreciate art, then I think it becomes much more plausible that intelligence is not the same as computation.
Of course, simply stating that isn't in of itself a philisophically rigorous argument. However, given that not everyone has training in philosophy and it may not even be possible to prove whether "feeling emotion" can be achieved via computation, I think it's a reasonable argument.
I think if they define intelligence that way, it isn't a very interesting discussion, because we're back to Church-Turing: Either they can show that this actually has an effect on the ability to reason and the possible outputs of the system that somehow exceeds the Turing computable, or those aspects are irrelevant to an outside observer of said entity because the entity would still be able to act in exactly the same way.
I can't prove that you have a subjective experience of feeling emotion, and you can't prove that I do - we can only determine that either one of us acts as if we do.
And so this is all rather orthogonal to how we define intelligence, as whether or not a simulation can simulate such aspects as "actual" feeling is only relevant if the Church-Turing thesis is proven wrong.
What program would a Turing machine run to spontaneously prove the incompleteness theorem?
Can you prove such a program may exist?
Assuming the Church-Turing thesis is true, the existence of any brain now or in the past capable of proving it is proof that such a program may exist.
If the Church-Turing thesis can be proven false, conversely, then it may be possible that such a program can't exist - it is a necessary but not sufficient condition for the Church-Turing thesis to be false.
Given we have no evidence to suggest the Church-Turing thesis to be false, or for it to be possible for it to be false, the burden falls on those making the utterly extraordinary claim that they can't exist to actually provide evidence for those claims.
Can you prove the Church-Turing thesis false? Or even give a suggestion of what a function that might be computable but not Turing computable would look like?
Keep in mind that explaining how to compute a function step by step would need to contain at least one step that can't be explain in a way that allows the step to be computable by a Turing machine, or the explanation itself would instantly disprove your claim.
The very notion is so extraordinary as to require truly extraordinary proof and there is none.
A single example of a function that is not Turing computable that human intelligence can compute should be low burden if we can exceed the Turing computable.
Where are the examples?
> Assuming the Church-Turing thesis is true, the existence of any brain now or in the past capable of proving it is proof that such a program may exist.
Doesn't that assume that the brain is a Turing machine or equivalent to one? My understanding is that the exact nature of the brain and how it relates to the mind is still an open question.
That is exactly the point.
If the Church-Turing thesis is true, then the brain is a Turing machine / Turing equivalent.
And so, assuming Church-Turing is true, then the existence of the brain is proof of the possibility of AGI, because any Turing machine can simulate any other Turing machine (possibly too slowly to be practical, but it denies its impossibility).
And so, any proof that AGI is "mathematically impossible" as the title claims, is inherently going to contain within it a proof that the Church-Turing thesis is false.
In which case there should be at least one example of a function a human brain can compute that a Turing machine can't.
An accurate-enough physical simulation of Kurt Gödel's brain.
Such a program may exist- unless you think such a simulation of a physical system is uncomputable, or that there is some non-physical process going on in that brain.
> then the only way of showing that an artificial intelligence can not theoretically be constructed to at least meet the same bar is by showing that humans can compute more than the Turing computable.
I would reframe: the only way of showing that artificial intelligence can be constructed is by showing that humans cannot compute more than the Turing computable.
Given that Turing computable functions are a vanishingly small subset of all functions, I would posit that that is a rather large hurdle to meet. Turing machines (and equivalents) are predicated on a finite alphabet / state space, which seems woefully inadequate to fully describe our clearly infinitary reality.
Given that we know of no computable function that isn't Turing computable, and the set of Turing computable functions is known to be equivalent to the lambda calculus and equivalent to the set of general recursive functions, what is an immensely large hurdle would be to show even a single example of a computable function that is not Turing computable.
If you can do so, you'd have proven Turing, Kleen, Church, Goedel wrong, and disproven the Church-Turing thesis.
No such example is known to exist, and no such function is thought to be possible.
> Turing machines (and equivalents) are predicated on a finite alphabet / state space, which seems woefully inadequate to fully describe our clearly infinitary reality.
1/3 symbolically represents an infinite process. The notion that a finite alphabet can't describe inifity is trivially flawed.
Function != Computable Function / general recursive function.
That's my point - computable functions are a [vanishingly] small subset of all functions.
For example (and close to our hearts!), the Halting Problem. There is a function from valid programs to halt/not-halt. This is clearly a function, as it has a well defined domain and co-domain, and produces the same output for the same input. However it is not computable!
For sure a finite alphabet can describe an infinity as you show - but not all infinity. For example almost all Real numbers cannot be defined/described with a finite string in a finite alphabet (they can of course be defined with countably infinite strings in a finite alphabet).
Non-computable functions are not relevant to this discussion, though, because humans can't compute them either, and so inherently an AGI need not be able to compute them.
The point remains that we know of no function that is computable to humans that is not in the Turing computable / general recursive function / lambda calculus set, and absent any indication that any such function is even possible, much less an example, it is no more reasonable to believe humans exceed the Turing computable than that we're surrounded by invisible pink unicorns, and the evidence would need to be equally extraordinary for there to be any reason to entertain the idea.
Humans do a lot of stuff that is hard to 'functionalise', computable or otherwise, so I'd say the burden of proof is on you. What's the function for creating a work of art? Or driving a car?
You clearly don't understand what a function means in this context, as the word function is not used in this thread in the way you appear to think it is used.
[dead]
The mathematical proof, as you describe it, sounds like the "No Free Lunch theorem". Humans also can't generalise to learning such things.
As you note in 2.1, there is widespread disagreement on what "AGI" means. I note that you list several definitions which are essentially "is human equivalent". As humans can be reduced to physics, and physics can be expressed as a computer program, obviously any such definition can be achieved by a sufficiently powerful computer.
For 3.1, you assert:
"""
Now, let's observe what happens when an Al system - equipped with state-of-the-art natural language processing, sentiment analysis, and social reasoning - attempts to navigate this question. The Al begins its analysis:
• Option 1: Truthful response based on biometric data → Calculates likely negative emotional impact → Adjusts for honesty parameter → But wait, what about relationship history? → Recalculating...
• Option 2: Diplomatic deflection → Analyzing 10,000 successful deflection patterns → But tone matters → Analyzing micro-expressions needed → But timing matters → But past conversations matter → Still calculating...
• Option 3: Affectionate redirect → Processing optimal sentiment → But what IS optimal here? The goal keeps shifting → Is it honesty? Harmony? Trust? → Parameters unstable → Still calculating...
• Option n: ....
Strange, isn't it? The Al hasn't crashed. It's still running. In fact, it's generating more and more nuanced analyses. Each additional factor may open ten new considerations. It's not getting closer to an answer - it's diverging.
"""
Which AI? ChatGPT just gives an answer. Your other supposed examples have similar issues in that it looks like you've *imagined* an AI rather than having tried asking an AI to seeing what it actually does or doesn't do.
I'm not reading 47 pages to check for other similar issues.
> physics can be expressed as a computer program
Citation needed. If you've spent any time dynamical systems, as an example, you'd know that the computer basically only kind of crudely estimates things, and only things that are abstractly near by. You may be able to write down some PDEs or field equations that may describe things at some base level, but even statistical mechanics, which is really what governs a huge amount of what we see and interact with, is just a pretty good approximation. Computers (especially real ones) only generate approximate (to some value of alpha) answers; physics is not reducible to a computer program at all.
> You may be able to write down some PDEs or field equations that may describe things at some base level, but even statistical mechanics, which is really what governs a huge amount of what we see and interact with, is just a pretty good approximation.
QED.
When the approximation is indistinguishable from observation over a time horizon exceeding a human lifetime, it's good enough for the purpose of "would a simulation of a human be intelligent by any definition that the real human also meets?"
Remember, this is claiming to be a mathematical proof, not a practical one, so we don't even have to bother with details like "a classical computer approximating to this degree and time horizon might collapse into a black hole if we tried to build it".
> Citation needed. If you've spent any time dynamical systems, as an example, you'd know that the computer basically only kind of crudely estimates things, and only things that are abstractly near by. You may be able to write down some PDEs or field equations that may describe things at some base level, but even statistical mechanics, which is really what governs a huge amount of what we see and interact with, is just a pretty good approximation. Computers (especially real ones) only generate approximate (to some value of alpha) answers; physics is not reducible to a computer program at all.
You're proving too much. The fact of the matter is that those crude estimations are routinely used to model systems.
> As humans can be reduced to physics, and physics can be expressed as a computer program
This is an assumption that many physicists disagree with. Roger Penrose, for example.
That's true, but we should acknowledge that this question is generally regarded as unsettled.
If you accept the conclusion that AGI (as defined in the paper, that is, "solving [...] problems at a level of quality that is at least equivalent to the respective human capabilities") is impossible but human intelligence is possible, then you must accept that the question is settled in favor of Penrose. That's obviously beyond the realm of mathematics.
In other words, the paper can only mathematically prove that AGI is impossible under some assumptions about physics that have nothing to do with mathematics.
> then you must accept that the question is settled in favor of Penrose. That's obviously beyond the realm of mathematics.
Not necessarily. You are assuming (AFAICT) that we 1. have perfect knowledge of physics and 2. have perfect knowledge of how humans map to physics. I don't believe either of those is true though. Particularly 1 appears to be very obviously false, otherwise what are all those theoretical physicists even doing?
I think what the paper is showing is better characterized as a mathematical proof about a particular algorithm (or perhaps class of algorithms). It's similar to proving that the halting problem is unsolvable under some (at least seemingly) reasonable set of assumptions but then you turn around and someone has a heuristic that works quite well most of the time.
Where am I assuming that we have perfect knowledge of physics?
To make it plain, I'll break the argument in two parts:
(a) if AGI is impossible but humans are intelligent, then it must be the case that human behavior can't be explained algorithmically (that last part is Penrose's position).
(b) the statement that human behavior can't be explained algorithmically is about physics, not mathematics.
I hope it's clear that neither (a) or (b) require perfect knowledge of physics, but just in case:
(a) is true by reductio ad absurdum: if human behavior can be explained algorithmically, then an algorithm must be able to simulate it, and so AGI is possible.
(b) is true because humans exist in nature, and physics (not mathematics) is the science that deals with nature.
So where is the assumption that we have perfect knowledge of physics?
You didn't. I confused something but looking at the comment chain now I can't figure out what. I'd say we're actually in perfect agreement.
"Many" is doing a lot of work here.
Penrose’s views on consciousness is largely considered quackery by other physicists.
1. I appreciate the comparison — but I’d argue this goes somewhat beyond the No Free Lunch theorem.
NFL says: no optimizer performs best across all domains. But the core of this paper doesnt talk about performance variability, it’s about structural inaccessibility. Specifically, that some semanti spaces (e.g., heavy-tailed, frame-unstable, undecidable contexts) can’t be computed or resolved by any algorithmic policy — no matter how clever or powerful. The model does not underperform here, the point is that the problem itself collapses the computational frame.
2. OMG, lool. ... just to clarify, there’s been a major misunderstanding :)
the “weight-question”-Part is NOT a transcript from my actual life... thankfully - I did not transcribe a live ChatGPT consult while navigating emotional landmines with my (perfectly slim) wife, then submit it to PhilPapers and now here…
So - NOT a real thread, - NOT a real dialogue with my wife... - just an exemplary case... - No, I am not brain dead and/or categorically suicidal!! - And just to be clear: I dont write this while sitting in some marital counseling appointment, or in my lawyer's office, the ER, or in a coroners drawer
--> It’s a stylized, composite example of a class of decision contexts that resist algorithmic resolution — where tone, timing, prior context, and social nuance create an uncomputably divergent response space.
Again : No spouse was harmed in the making of that example.
;-))))
Just a layman here so Im not sure if Im understanding (probably not), but humans dont analyze every possible scenario ad infinitum, we go based on the accumulation of our positive/negative experiences from the past. We make decisions based on some self construed goal and beliefs as to what goes towards those goals, and these are arbitrary with no truth. Napolean for example conquered Europe perhaps simiply becuause he thought he was the best to rule it, not through a long chain of questions and self doubt
We are generally intelligent only in the sense that our reasoning/modeling capabilities allow us to understand anything that happens in space-time.
> the “weight-question”-Part is NOT a transcript from my actual life... thankfully - I did not transcribe a live ChatGPT consult while navigating emotional landmines with my (perfectly slim) wife, then submit it to PhilPapers and now here…
You have wildly missed my point.
You do not need to even have a spouse in order to try asking an AI the same question. I am not married, and I was still able to ask it ask it to respond to that question.
My point is that you clearly have not asked ChatGPT, because ChatGPT's behaviour clearly contradicts your claims about what AI would do.
So: what caused you to write to claim that AI would respond as you say they would respond, when the most well-known current generation model clearly doesn't?
> Specifically, that some semanti spaces (e.g., heavy-tailed, frame-unstable, undecidable contexts) can’t be computed or resolved by any algorithmic policy — no matter how clever or powerful. The model does not underperform here, the point is that the problem itself collapses the computational frame.
I see no proof this doesn’t apply to people
I read some of the paper, and it does seem silly to me to state this:
"But here’s the peculiar thing: Humans navigate this question daily. Not always successfully, but they do respond. They don’t freeze. They don’t calculate forever. Even stranger: Ask a husband who’s successfully navigated this question how he did it, and he’ll likely say: ‘I don’t know… I just… knew what to say in that moment....What’s going on here? Why can a human produce an answer (however imperfect) while our sophisticated AI is trapped in an infinite loop of analysis?” ’"
LLM's don't freeze either. In your science example too, we already have LLMs that give you very good answers to technical questions, so on what grounds is this infinite cascading search based on?
I have no idea what you're saying here either: "Why can’t the AI make Einstein’s leap? Watch carefully: • In the AI’s symbol set Σ, time is defined as ‘what clocks measure-universally’ • To think ‘relative time,’ you first need a concept of time that says: • ‘flow of time varies when moving, although the clock ticks just the same as when not moving' • ‘Relative time’ is literally unspeakable in its language • "What if time is just another variable?", means: :" What if time is not time?"
"AI’s symbol set Σ, time is defined as ‘what clocks measure-universally", it is? I don't think this is accurate of LLM's even, let alone any hypothetical AGI. Moreover LLM's clearly understand what "relative" means, so why would they not understand "relative time?".
In my hypothetical AGI, "time" would mean something like "When I observe something, and then things happens in between, and then I observe it again", and relative time would mean something like "How I measure how many things happen in between two things, is different from how you measure how many things happen between two things"
[dead]
“This paper presents a theoretical proof that AGI systems will structurally collapse under certain semantic conditions…”
No it doesn’t.
Shannon entropy measures statistical uncertainty in data. It says nothing about whether an agent can invent new conceptual frames. Equating “frame changes” with rising entropy is a metaphor, not a theorem, so it doesn’t even make sense as a mathematical proof.
This is philosophical musing at best.
Correct: Shannon entropy originally measures statistical uncertainty over a fixed symbol space. When the system is fed additional information/data, then entropy goes down, uncertainty falls. This is always true in situations where the possible outcomes are a) sufficiently limited and b)unequally distributed. In such cases, with enough input, the system can collapse the uncertainty function within a finite number of steps.
But the paper doesn’t just restate Shannon.
It extends this very formalism to semantic spaces where the symbol set itself becomes unstable. These situations arise when (a) entropy is calculated across interpretive layers (as in LLMs), and (b) the probability distribution follows a heavy-tailed regime (α ≤ 1). Under these conditions, entropy divergence becomes mathematically provable.
This is far from being metaphorical: it’s backed by formal Coq-style proofs (see Appendix C in he paper).
AND: it is exactly the mechanism that can explain the Apple-Papers' results
Your paper only claims that those Coq snippets constitute a "constructive proof sketch". Have those formalizations actually been verified, and if so, why not include the results in the paper?
Separately from that, your entire argument wrt Shannon hinges on this notion that it is applicable to "semantic spaces", but it is not clear on what basis this jump is made.
I'm wondering if you may have rediscovered the concept of "Wicked Problems", which have been studied in system analysis and sociology since the 1970's (I'd cite the Wikipedia page, but I've never been particularly fond of Wikipedia's write up on them). They may be worth reading up on if you're not familiar with them.
It's interesting. The question from the paper "Darling, please be honest: have I gained weight?" assumes that the "socially acceptability" of the answer should be taken into account. In this case the problem fits the "Wickedness" (Wikipedia's quote is "Classic examples of wicked problems include economic, environmental, and political issues"). But taken formally, and with the ability for LLM to ask questions in return to decrease formal uncertainty ("Please, give me several full photos of yourself from the past year to evaluate"), it is not "wicked" at all. This example alone makes the topic very uncertain in itself
Wow, that is a great advice. Never heard of them - and they seem to fit perfectly into the whole concept THANK YOU! :-)
In your paper it states:
AGI as commonly defined
However I don’t see where you go on to give a formalization of “AGI” or what the common definition is.
can you do that in a mathematically rigorous way such that it’s a testable hypothesis?
I don't think it exists. We can't even seem to agree on a standard criteria for "intelligence" when assessing humans let alone a rigorous mathematical definition. In turn, my understanding of the commonly accepted definition for AGI (as opposed to AI or ML) has always been "vaguely human or better".
Unless the marketing department is involved in which case all bets are off.
It can exist for the purpose of the paper. As in "when I write AGI, I mean ...". Otherwise what's the point in any rigour if we're just going by "you know what I mean" vibes.
Apple's paper sets up a bit of a straw man in my opinion. It's unreasonable to expect that an LLM not trained on what are essentially complex algorithmic tasks is just going to discover the solution on the spot. Most people can solve simple cases of the tower of Hanoi, and almost none of us can solve complex cases. In general, the ones who can have trained to be able to do so.
> specific problem classes (those with α ≤ 1),
For the layman, what does α mean here?
I'm sure this is a reference to alpha stable distributions: https://en.m.wikipedia.org/wiki/Stable_distribution
Most of these don't have finite moments and are hard to do inference on with standard statistical tools. Nassim Taleb's work (Black Swan, etc.) is around these distributions.
But I think the argument of OP in this section doesn't hold.
does this include if the AI can devise new components and use drones and things essentially to build a new iteration of itself more capable to compute a thing and keep repeating this going out into the universe as needed for resources and using von Neumann probes.. etc?