ICBTheory

il y a 7 moisdiscuss

I'd say that conclusion is a manifestation of pragmatic wisdom.

Anyway: I agree. The paper certainly doesn't argue that AI is useless, but that autonomy in high-stakes domains is mathematically unsafe.

In the text, I distinguish between operating on an 'Island of Order' (where hallucinations are cheap and correctable, like fixing a syntax error in code) versus navigating the 'Fat-Tailed Ocean' (where a single error is irreversible).

Tying this back to your comment: If an AI hallucinates a variable name — no problem, you just fix it. But I would advise skepticism if an AI suggests telling your boss that 'his professional expertise still has significant room for improvement.'

If hallucinations are structural (as the Coq proof in Part II indicates), then 'living with them' means ensuring the system never has the autonomy to execute that second type of decision.

ICBTheory

il y a 7 moisdiscuss

Author here.

This paper is Part III of a trilogy investigating the limits of algorithmic cognition. Given the recent industry signals regarding "scaling plateaus" (e.g., Sutskever etc.), I attempt to formalize why these limits appear structurally unavoidable.

The Thesis: We model modern AI as a Probabilistic Bounded Semantic System (P-BoSS). The paper demonstrates via the "Inference Trilemma" that hallucinations are not transient bugs to be fixed by more data, but mathematical necessities when a bounded system faces fat-tailed domains (alpha ≤ 1).

The Proof: While this paper focuses on the CS implications, the underlying mathematical theorems (Rice’s Theorem applied to Semantic Frames, Sheaf Theoretic Gluing Failures) are formally verified using Coq.

You can find the formal proofs and the Coq code in the companion paper (Part II) here:

https://philpapers.org/rec/SCHTIC-16

I’m happy to discuss the P-BOSS definition and why probabilistic mitigation fails in divergent entropy regimes.

ICBTheory

il y a 12 moisdiscuss

And finally 7. On “But humans are finite too—so why not replicable?”

Yes. Humans are finite. But we’re not symbol-bound, and we don’t wait for the frame to stabilize before we act.We move while the structure is still breaking, speak while meaning is still assembling, and decide before we understand—then change what we were deciding halfway through.

NOT because we’re magic. Simply because we’re not built like your architecture (and if you think everything outside your architecture is magic, well…)

If your system needs everything cleanly defined, fully mapped, and symbolically closed before it can take a step, and mine doesn’t— then no, they’re not the same kind of thing.

Maybe this isn’t about scaling up? … Well, it isn’t It’s about the fact that you can’t emulate improvisation with a bigger spreadsheet. We don’t generalize because we have all the data. We generalize because we tolerate not knowing—and still move.

But hey, sure, keep training. Maybe frame-jumping will spontaneously emerge around parameter 900 billion.

Let me know how that goes

ICBTheory

il y a 12 moisdiscuss

6. On “This is just a critique of current models—not AGI itself”

No.

This isn’t about GPT-4, or Claude, or whatever model’s in vogue this quarter. Neither is it about architecture. It’s about what no symbolic system can do—ever.

If your system is: a) Finite b)Bounded by symbols C) Built on recursive closure

…it breaks down where things get fuzzy: where context drifts, where the problem keeps changing, where you have to act before you even know what the frame is.

That’s not a tuning issue, that IS the boundary. (And we’re already seeing it.)

In The Illusion of Reasoning (Shojaee et al., 2025, Apple), they found that as task complexity rises: - LLMs try less - Answers get shorter, shallower - Recursive tasks—like the Tower of Hanoi—just fall apart - etc

That’s IOpenER in the wild:Information Opens. Entropy Rises. The theory predicts the divergence, and the models are confirming it—one hallucination at a time.

ICBTheory

il y a 12 moisdiscuss

5. On “Kolmogorov and Chaitin are misused”

It’s a fair concern.Chaitin does get thrown around too easily — usually in discussions that don’t need him.

But that’s not what’s happening here.

– Kolmogorov shows that most strings are incompressible. – Chaitin shows that even if you find the simplest representation, you can’t prove it’s minimal. – So any system that “discovers” a concept has no way of knowing it’s found something reusable.

That’s the issue. Without confirmation, generalization turns into guesswork. And in high-K environments — open-ended, unstable ones — that guesswork becomes noise. No poetic metaphor about the mystery of meaning here. It’s a formal point about the limits of abstraction recognition under complexity.

So no, it’s not a misuse. It’s just the part of the theory that gets quietly ignored because it doesn’t deliver the outcome people are hoping for.

ICBTheory

il y a 12 moisdiscuss

4. On “This is just the No Free Lunch Theorem again”

Well … not quite. The No Free Lunch theorem says no optimizer is universally better across all functions. That’s an averaging result.

But this paper is not at all about average-case optimization. It’s about specific classes of problems—social ambiguity, paradigm shifts, semantic recursion—where: a)The tail exponent alpha is = or < 1 —>no mean exists, b) Kolmogorov complexity is incompressible, and c) the symbol space lacks the needed abstraction

In these spaces, learning collapses not due to lack of training, but due to structural divergence. Entropy grows with depth. More data doesn’t help. It makes it worse.

That is what “IOpenER” means: Information Opens, Entropy Rises.

It is NOT a theorem about COST… rather a structure about meaning. What exactly is so hard to understand about this?

ICBTheory

il y a 12 moisdiscuss

3. On “He redefines AGI to make his result inevitable”

Sure. I redefined AGI. By using… …the definition from OpenAI, DeepMind, Anthropic, IBM, Goertzel, and Hutter.

So unless those are now fringe newsletters, the definition stands:

- A general-purpose system that autonomously solves a wide range of human-level problems, with competence equivalent to or greater than human performance -

If that’s the target, the contradiction is structural: No symbolic system can operate stably in the kinds of semantic drift, ambiguity, or frame collapse that general intelligence actually requires. So if you think I smuggled in a trap, check your own luggage because the industry packed it for me.

ICBTheory

il y a 12 moisdiscuss

2. On “This is just philosophy with no testability”

Yes, the paper is also philosophical. But not in the hand-wavy, incense-burning sense that’s being implied. It makes a formal claim, in the tradition of Gödel, Rice, and Chaitin: Certain classes of problems are structurally undecidable by any algorithmic system.

You don’t need empirical falsification to verify this. You need mathematical framing. Period.

Just as the halting problem isn’t “testable” but still defines what computers can and can’t do, the Infinite Choice Barrier defines what intelligent systems cannot infer within finite symbolic closure.

These are not performance limitations. They are limits of principle.

ICBTheory

il y a 12 moisdiscuss

1. On “The brain obeys physics, physics is computable—so AGI must be possible”

This is the classical foundational syllogism of computationalism. In short:

   1.The brain obeys the laws of physics.
   2.The laws of physics are (in principle) computable.
   3.Therefore, the brain is computable.
   4.Therefore, human-level general intelligence is computable, and AGI is  
     inevitable and a question of time, power and compute.

This seems elegant, tidy, logically sound. And: it is patently false — at step 3… And this common mistake is not technical, but categorical: Simulating a system’s physical behavior is not the same as instantiating its cognitive function.

The flaw is in the logic — it’s nothing less than a category error. The logic breaks exactly where category boundaries are crossed without checking if the concept still applies. That by no means inference, this is mere wishful thinking in formalwear. It happens when you confuse simulating a system with being the system. It’s in the jump from simulation to instantiation.

Yes, we can simulate water. -> No, the simulation isn’t wet.

Yes, I can “simulate” a fridge. ->But if I put a beer in myself, and the beer doesn’t come out cold after some time,then what we’ve built is a metaphor with a user interface, not a cognitive peer.

And yes: we can simulate Einstein discovering special relativity. -> But only after he’s already done it. We can tokenize the insight, replay the math, even predict the citation graph. But that’s not general intelligence, that’s a historical reenactment, starring a transformer with a good memory.

Einstein didn’t run inference over a well-formed symbol set. He changed the set, reframed the problem from within the ambiguity. And that is not algorithmic recursion, is it? Nope… That’s cognition at the edge of structure.

If your model can only simulate the answer after history has solved it, then congratulations: you’ve built a cognitive historian, not a general intelligence.

ICBTheory

il y a 12 moisdiscuss

Hey all, apologies for the delayed response. I was on a flight, then had guests, then had to make some rapid decisions involving actual real-world complexity (the kind that is not easily tokenized).

I’ve now had time to read through the thread properly, and I appreciate the range of engagement—even the sharp-edged stuff. Below, I’ve gathered a set of structured responses to the main critique clusters that came up.

ICBTheory

l’année dernièrediscuss

You’re misreading what I’m doing, and I suspect you’re also misdefining what a “proof” in this space needs to be.

I’m not assuming humans exceed the Turing computable. I’m not using human behavior as a proof of AGI’s impossibility. I’m doing something much more modest - and much more rigorous.

Here’s the actual chain:

1. There’s a formal boundary for algorithmic systems. It’s called symbolic containment. A system defined by a finite symbol set Σ and rule set R cannot generate a successor frame (Σ′, R′) where Σ′ introduces novel symbols not contained in Σ. This is not philosophy — this is structural containment, and it is provable.

2. Then I observe: in human intellectual history, we find recurring examples of frame expansion. Not optimization, not interpolation — expansion. New primitives. New rules. Special relativity didn’t emerge from Newton through deduction. It required symbols and structures that couldn’t be formed inside the original frame.

3. That’s not “proof” that humans exceed the Turing computable. That’s empirical evidence that human cognition appears to do something algorithmic systems, as formally defined, cannot do.

4. This leads to a conclusion: if AGI is an algorithmic system (finite symbols, finite rules, formal inference)then it will not be capable of frame jumps.And it is not incapable of that, because it lacks compute. The system is structurally bounded by what it is.

So your complaint that I “haven’t proven humans exceed Turing” is misplaced. I didn’t claim to. You’re asking me to prove something that I simply don’t need to assert .

I’m saying: algorithmic systems can’t do X (provable), and humans appear to do X (observed). Therefore, if humans are purely algorithmic, something’s missing in our understanding of how those systems operate. And if AGI remains within the current algorithmic paradigm, it will not do X. That’s what I’ve shown.

You can still believe humans are Turing machines, fine for me. But if this belief is to be more than some kind of religious statement, then it is you that would need to explain how a Turing machine bounded to Σ can generate Σ′ with Σ′ \ Σ ≠ ∅. It is you that would need to show how uncomputable concepts emerge from computable substrates without violating containment (->andthat means: witout violating its own logic - as in formal systems, logic and containment end up as the same thing: Your symbol set defines your expressive space, step outside that, and you’re no longer reasoning — you’re redefining the space, the universe you’re reasoning in).

Otherwise, the limitation stands — and the claim that “AGI can do anything humans do” remains an ungrounded leap of faith.

Also: if you believe the only valid proof of AGI impossibility must rest on metaphysical redefinition of humanity as “super-Turing,” then you’ve set an artificial constraint that ensures no such proof could ever exist, no matter the logic.

That’s intellectually trading epistemic rigor for insulation.

As for your claim that I misunderstand Turing machines, please feel free to show precisely which part fails. The statement that a TM cannot emit symbols not present in its alphabet is not a misunderstanding — it’s the foundation of how TMs are defined. If you think otherwise, then I would politely suggest you review the formal modl again.

ICBTheory

l’année dernièrediscuss

You’re flipping the logic.

I’m not assuming humans are beyond Turing-computable and then using that to prove that AGI can’t be. I’m saying: here is a provable formal limit for algorithmic systems ->symbolic containment. That’s theorem-level logic.

Then I look at real-world examples (Einstein is just one) where new symbols, concepts, and transformation rules appear that were not derivable within the predecessor frame. You can claim, philosophically (!), that “well, humans must be computable, so Einstein’s leap must be too.” Fine. But now you’re asserting that the uncomputable must be computable because humans did it. That’s your circularity, not mine. I don’t claim humans are “super-Turing.” I claim that frame-jumping is not computation. You can still be physical, messy, and bounded .. and generate outside your rational model. That’s all the proof needs.

ICBTheory

l’année dernièrediscuss

Yes, of course — if you define Ω² as “English + All of Science,” then congratulations, you have defined an unbounded oracle. But you’re just shifting the burden.

No sysem starting from Ω₁ can generate Ω₂ unless Ω₂ is already implicit. ... If you build a system trained on all of science, then yes, it knows Einstein because you gave it Einstein. But now ask it to generate the successor of Ω² (call it Ω³ ) with symbols that don’t yet exist. Can it derive those? No, because they’re not in Σ². Same limitation, new domain. This isn’t about “a small frame can’t do AGI.” It’s about every frame being finite, and therefore bounded in its generative reach. The question is whether any algorithmic system can exeed its own Σ and R. The answer is no. That’s not content-dependent, that’s structural.

ICBTheory

l’année dernièrediscuss

The claim isn’t that humans maintain a consistent metascience. In fact, quite the opposite. Frame jumps happen precisely because human cognition is not locked into a consistent formal system. That’s the point. It breaks, drifts, mutates. Not elegantly — generatively. You’re pointing to HOL-in-HOL or other meta-theoretical modeling approaches. But these aren’t equivalent. You can model a frame-jump after it has occurred, yes. You can define it retroactively. But that doesn’t make the generative act itself derivable from within the original system. You’re doing what every algorithmic model does: reverse-engineering emergence into a schema that assumes it. This is not sloppiness. It’s making a structural point: a TM with alphabet Σ can’t generate Σ′ where Σ′ \ Σ ≠ ∅. That is a hard constraint. Humans, somehow, do. If you don’t like the label “frame jump,” pick another. But that phenomenon is real, and you can’t dissolve it by saying “well, in HOL I can model this afterward.” If computation is always required to have an external frame to extend itself, then what you’re actually conceding is that self-contained systems can’t self-jump — which is my point exactly...

ICBTheory

l’année dernièrediscuss

No problem here is you proof - although a bit long:

1. THEOREM: Let a semantic frame be defined as Ω = (Σ, R), where

Σ is a finite symbol set and R is a finite set of inference rules.

Let Ω′ = (Σ′, R′) be a candidate successor frame.

Define a frame jump as: Frame Jump Condition: Ω′ extends Ω if Σ′\Σ ≠ ∅ or R′\R ≠ ∅

Let P be a deterministic Turing machine (TM) operating entirely within Ω.

Then: Lemma 1 (Symbol Containment): For any output L(P) ⊆ Σ, P cannot emit any σ ∉ Σ.

(Whereas Σ = the set of all finite symbol strings in the frame; derivable outputs are formed from Σ under the inference rules R.)

Proof Sketch: P’s tape alphabet is fixed to Σ and symbols derived from Σ. By induction, no computation step can introduce a symbol not already in Σ. ∎

2. APPLICATION: Newton → Special Relativity

Let Σᴺ = { t, x, y, z, v, F, m, +, · } (Newtonian Frame) Let Σᴿ = Σᴺ ∪ { c, γ, η(·,·) } (SR Frame)

Let φ = “The speed of light is invariant in all inertial frames.” Let Tᴿ be the theory of special relativity. Let Pᴺ be a TM constrained to Σᴺ.

By Lemma 1, Pᴺ cannot emit any σ ∉ Σᴺ.

But φ ∈ Tᴿ requires σ ∈ Σᴿ \ Σᴺ

→ Therefore Pᴺ ⊬ φ → Tᴿ ⊈ L(Pᴺ)

Thus:

Special Relativity cannot be derived from Newtonian physics within its original formal frame.

3. EMPIRICAL CONFLICT Let: Axiom N₁: Galilean transformation (x′ = x − vt, t′ = t) Axiom N₂: Ether model for light speed Data D: Michelson–Morley ⇒ c = const

In Ωᴺ, combining N₁ and N₂ with D leads to contradiction. Resolving D requires introducing {c, γ, η(·,·)}, i.e., Σᴿ \ Σᴺ But by Lemma 1: impossible within Pᴺ. -> Frame must be exited to resolve data.

4. FRAME JUMP OBSERVATION

Einstein introduced Σᴿ — a new frame with new symbols and transformation rules. He did so without derivation from within Ωᴺ. That constitutes a frame jump.

5. FINALLY

A: Einstein created Tᴿ with Σᴿ, where Σᴿ \ Σᴺ ≠ ∅

B: Einstein was human

C: Therefore, humans can initiate frame jumps (i.e., generate formal systems containing symbols/rules not computable within the original system).

Algorithmic systems (defined by fixed Σ and R) cannot perform frame jumps. But human cognition demonstrably can.

QED.

BUT: Can Humans COMPUTE those functions? (As you asked)

-> Answer: a) No - because frame-jumping is not a computation.

It’s a generative act that lies outside the scope of computational derivation. Any attempt to perform frame-jumping by computation would either a) enter a Goedelian paradox (truth unprovable in frame),b) trigger the halting problem , or c) collapse into semantic overload , where symbols become unstable, and inference breaks down.

In each case, the cognitive system fails not from error, but from structural constraint. AND: The same constraint exists for human rationality.

ICBTheory

l’année dernièrediscuss

Wow, that is a great advice. Never heard of them - and they seem to fit perfectly into the whole concept THANK YOU! :-)

ICBTheory

l’année dernièrediscuss

Very good point.

I in fact had thought of describing the problem from a systems theoretical perspective as this is another way to combine different paths into a common principle

That was a sketch, in case you are into these kind of approaches:

2. Complexity vs. Complication In systems theory, the distinction between 'complex' and 'complicated' is critical. Complicated systems can be decomposed, mapped, and engineered. Complex systems are emergent, self-organizing, and irreducible. Algorithms thrive on complication. But general intelligence—especially artificial general intelligence (AGI)—must operate in complexity. Attempting to match complex environments through increased complication (more layers, more parameters) leads not to adaptation, but to collapse. 3. The Infinite Choice Barrier and Entropy Collapse In high-entropy decision spaces, symbolic systems attempt to compress possibilities into structured outcomes. But there is a threshold—empirically visible around entropy levels of H ≈ 20 (one million outcomes)—beyond which compression fails. Adding more depth does not resolve uncertainty; it amplifies it. This is the entropy collapse point: the algorithm doesn't fail because it cannot compute. It fails because it computes itself into divergence. 4. The Oracle and the Zufallskelerator To escape this paradox, the system would need either an external oracle (non-computable input), or pure chance. But chance is nearly useless in high-dimensional entropy. The probability of a meaningful jump is infinitesimal. The system becomes a closed recursion: it must understand what it cannot represent. This is the existential boundary of algorithmic intelligence: a structural self-block. 5. The Organizational Collapse of Complexity The same pattern is seen in organizations. When faced with increasing complexity, they often respond by becoming more complicated—adding layers, processes, rules. This mirrors the AI problem. At some point, the internal structure collapses under its own weight. Complexity cannot be mirrored. It must either be internalized—by becoming complex—or be resolved through a radically simpler rule, as in fractal systems or chaos theory.

6. Conclusion: You Are an Algorithm An algorithmic system can only understand what it can encode. It can only compress what it can represent. And when faced with complexity that exceeds its representational capacity, it doesn't break. It dissolves. Reasoning regresses to default tokens, heuristics, or stalling. True intelligence—human or otherwise—must either become capable of transforming its own frame (metastructural recursion), or accept the impossibility of generality. You are an algorithm. You compress until you can't. Then you either transform, or collapse

ICBTheory

l’année dernièrediscuss

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.

;-))))

ICBTheory

l’année dernièrediscuss

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).

ICBTheory

l’année dernièrediscuss

Why? 1. Basically because physical laws obviously allow more than algorithmic cognition and problem solving. (And also: I am bound by thermodynamics as my mother in Law is, still i get disarranged by her mere presence while I always have to put laxatives in her wine to counter that)

2. human rationality is equally limited as algorithms. Neither an algorithm nor human logic can find itself a path from Newton to Einsteins SR. Because it doesn't exist.

3. Physical laws - where do they really come from? From nature? From logic? Or from that strange thing we do: experience, generate, pattern, abstract, express — and try to make it communicable? I honestly don’t know.

In a nutshell: there obviously is no law that forbids us to innovate - we do this, quite often. There only is a logical boundary, that says that there is no way to derive something out of a something that is not part of itself - no way for thinking to point beyond what is thinkable.

Imagine little Albert asking his physics teacher in 1880: "Sir - for how long do I have to stay at high speed in order to look as grown up as my elder brother?" ... i guess "interesting thought" would not have been the probable answer... rather something like "have you been drinking? Stop doing that mental crap - go away, you little moron!"

Submissions

Formal Proof: LLM Hallucinations Are Structural, Not Statistical (Coq Verified)

AGI Is Mathematically Impossible (3): Kolmogorov Complexity

AGI is Mathematically Impossible 2: When Entropy Returns

comments