Abhishek S.
Shipping in public. Listening in private.

Abhishek

I lead women’s Indo-Western & Premium at Max Fashion. I also wrote the AI that runs the buying floor.

Rare profile. Category operator who ships production code.

Senior Buying Leader · Max Fashion Women’s Indo-Western & Premium · 530+ India stores NIFT ’12 · Twelve years on the floor

abhishek@bengaluru ~ %
>role: senior buying lead
>dept: women’s indo-western + premium
>floor: 530+ stores india

The Halting Problem — When Computation Hits a Wall

A Turing machine with 748 states exists that halts if and only if the Riemann Hypothesis is false. Nobody will ever know whether it halts. Not because the computation is too long — because the question is structurally beyond ZFC set theory, the mathematics underwriting all of science.

This is what Turing proved in 1936: certain questions about computation have no algorithmic answer. Not slow. Not hard. Impossible by construction. The wall is real, and it shows up in quantum physics, in AI alignment, and in any system rich enough to describe itself.

The proof, in five lines

Given a candidate algorithm H(P, I) that decides whether program P halts on input I:

  1. Build program D(P): runs H(P, P). If H says "halts," D loops. If H says "loops," D halts.
  2. Now ask: what does H(D, D) return?
  3. "Halts" → D loops. Contradiction.
  4. "Loops" → D halts. Contradiction.
  5. H cannot exist.

The argument is diagonal self-reference, the same shape Gödel used in 1931 for concept godel incompleteness. No amount of computational power resolves it. A computer the size of the universe running for the age of the universe is no closer to a general halting decider than a pocket calculator.

A small historical correction: Turing's 1936 paper never uses the word "halt." The phrase "halting problem" was coined by Martin Davis around 1952 and popularized by Hartley Rogers in 1957. The proof as taught today is a cleaned-up cousin of Turing's original diagonal argument about printable real numbers.

Rice's Theorem: the full blast radius

Henry Gordon Rice (1951) generalized the wall: any non-trivial semantic property of programs is undecidable. Semantic means "about what the program computes." Non-trivial means at least one program has the property and at least one does not.

The consequence list is brutal:

The last one is why Rice's theorem walked into machine learning research in 2025.

The Busy Beaver: finite numbers you cannot reach

The Busy Beaver function BB(n) asks: of all Turing machines with n states that eventually halt, what is the maximum number of steps any of them takes?

n BB(n)
1 1
2 6
3 21
4 107
5 47,176,870
6 > 10^36,534

BB(5) was settled in July 2024 by the bbchallenge.org collaboration using machine-checked Coq proofs — fourteen years after Heiner Marxen's last conjectured bound. BB(6) jumped from 47 million to a number with 36,534 digits. BB(7) is suspected to be larger than any specific number humans will ever write down.

The mathematically interesting cliff: there exists a Turing machine with roughly 748 states that halts iff the Riemann Hypothesis is false (Yedidia and Aaronson, 2016; later optimized). Knowing BB(748) would resolve RH. But Aaronson, Yedidia, and others have shown BB(n) becomes independent of ZFC somewhere around n=745 — for sufficiently large n, the value of BB(n) is unprovable in standard mathematics.

This is the sharp tip: questions about specific, concrete, finite integers — not infinities, not paradoxes — are formally beyond the axioms we use for physics.

Undecidability in physics

The halting problem doesn't stay in computer science. It has been formally embedded in physical systems.

Spectral gap (Cubitt, Pérez-García, Wolf, Nature 2015; extended in Forum of Mathematics, Pi, 2022). The spectral gap of a quantum many-body Hamiltonian — whether the ground state is separated from excited states by an energy gap — is undecidable for the general class of 2D translationally-invariant lattice Hamiltonians. The proof embeds a universal Turing machine into the ground state structure. A physicist handed an arbitrary lattice model cannot, in principle, determine whether it is gapped or gapless.

Thermalization (Nature Communications, 2021). Whether a given quantum many-body system thermalizes with respect to a given observable is undecidable in general. The eigenstate thermalization hypothesis — the dominant story about why isolated quantum systems reach equilibrium — cannot be algorithmically verified for arbitrary Hamiltonians.

The pattern is consistent: when a physical system is rich enough to simulate universal computation, questions about its long-run behavior inherit the halting wall. Phase boundaries, equilibrium properties, ground state energies — these become not just unsolved but unsolvable.

AI alignment as a Rice's-theorem problem

A 2025 Scientific Reports paper ("Machines that halt resolve the undecidability of artificial intelligence alignment," doi: 10.1038/s41598-025-99060-2) made the formal case: the inner alignment problem — whether an AI model satisfies a non-trivial alignment criterion of its outputs given inputs — is undecidable by Rice's theorem. So is the harming problem, so is the containment problem.

The claim is narrower than it sounds. It says: post-hoc verification of alignment for arbitrary AI architectures is impossible. It does not say alignment is impossible. The proposed escape is that alignment must be a design property built into a system with provable termination, not a behavioral characteristic audited after training. A halting machine can in principle carry formal guarantees; an arbitrary one cannot.

This recasts mechanistic interpretability research as a specific tactical move: not solving the halting problem, but building enough transparency into particular systems that their relevant properties become locally decidable. See concept transformer architecture.

What's contested

The interpretation, not the proofs. Three live arguments:

The first is whether physical undecidability matters in practice. Critics note that the spectral gap construction uses pathological Hamiltonians not found in nature. Defenders reply that the class being undecidable means any general theory of phase transitions will hit cases it cannot resolve — and we don't get to pre-decide which Hamiltonians nature hands us.

The second is whether Rice's theorem genuinely closes the door on AI safety verification. Several researchers argue that real AI systems are not arbitrary Turing machines; they are bounded, structured, and trained on specific distributions. Approximate verification on the actual distribution may be tractable even if exact verification on all inputs is not. The counter-claim is that adversarial inputs are precisely the distribution shift case where the undecidability bites.

The third is whether self-reference is the deep cause or a surface artifact. Some philosophers of mathematics argue that undecidability is symptomatic of expressive power generally, and any sufficiently rich formalism — self-referential or not — will hit it. This is the open frontier between concept godel incompleteness and the broader question of formal limits.

Why this has to do with other realms

The halting problem is the connective tissue between computer science and the rest of the wiki. If concept simulation hypothesis is taken seriously, the simulator inherits the wall: questions about our universe's long-run behavior may be formally undecidable to whoever is running the simulation. If consciousness is a semantic property of neural computation, Rice's theorem suggests no algorithm reads it off the wiring — the move at the heart of concept hard problem consciousness. And phase transitions, the workhorse of concept emergence, are exactly where physical undecidability enters: emergence may in some cases be not just hard to compute from microstates but formally impossible.

The unifying observation: any system powerful enough to model itself contains questions it cannot decide. Turing machines, axiomatic mathematics, possibly physics, possibly minds.

An open question

If a sufficiently powerful reasoner cannot fully model itself, what does this say about the limits of introspection — for humans, for AIs trained on human introspection, or for any cognitive system trying to verify its own alignment from the inside?

Key sources

Further reading

See Also