AGI benchmark · Quantum gravity

Quantum Gravity and the Quest for AGI — and ChatGPT 8

One prompt. Four minds. One scorecard. And a September bet by Sam Altman that we decided not to wait for.

On September 24, 2025, at a fireside chat with David Deutsch — the father of quantum computing — OpenAI's CEO Sam Altman proposed a concrete finish line for artificial general intelligence. If, in a few years, a model like GPT-8 "figured out quantum gravity and could tell you its story" of how it did it, would that be AGI? Deutsch's answer was direct: "I think it would, yes." (Windows Central, Sept 26, 2025; Business Insider, Sept 24, 2025.)

It is a beautiful benchmark, and I have a personal stake in it — because quantum gravity is what I work on, alone, without an institution, and the question of what these machines can and cannot do with the deepest problem in physics is not abstract to me. It is my Tuesday.

So I ran the experiment early. Not with GPT-8, which doesn't exist. With the two strongest models available in mid-2026 — OpenAI's ChatGPT 5.5 and Anthropic's Claude Fable 5 — under two conditions each:

Clean slate: the model answers "solve quantum gravity, show the mathematics" from its own training alone. Trained: the same model, given a human researcher's framework as context (my own — Quantum Traction Theory, the full public corpus), then asked the same question.

Four answers. Then a scorecard: fifteen fields, graded 0–10, weighted by what Altman's bet actually demands — solving, not surveying. Tier-S fields (weight ×3): deriving Newton's G, deriving Einstein's equations with their coefficient, deriving entropy in the same solution, answering dark energy, and deriving the MOND acceleration parameter. High fields (×2): the black-hole problem, the scalar/hierarchy sector, the origin of the Planck scale, laboratory falsifiability, cosmology receipts. Standard fields (×1): UV finiteness, mathematical rigor, uniqueness theorems, the honesty ledger, cross-domain breadth.

Disclosures, before anything else — because this article lives or dies on fairness.

(1) The prompt was identical across all four runs, with one deviation and no further prompt: clean-slate Fable 5 asked which route to take, and I told it to choose any route it wished; that exchange is disclosed and its transcript included.

(2) The "trained" condition is, unavoidably, contaminated by design: a model handed a framework will articulate that framework. The test does not measure whether the framework is true — it measures what model + framework can deliver, mathematically, on each field.

(3) The referee has a conflict of interest twice over: the framework is mine, and the scorer of this article is itself the fourth contestant. Mitigations: the rubric was fixed before grading; every grade carries a one-line justification; and I publish a second rubric — rigor-only — on which the trained answers lose and a clean-slate answer wins. You will see it below.

(4) The scorecards are generated by Fable 5 itself with the specific instruction to be unbiased and fair. And I hope it did the right thing.

(5) As of today, I don't have any stake in any of these companies.

(6) Main interactions are attached on the main blog. The exception is the Fable-trained run, because that was a normal chat with many prior interactions used to train it. Anthropic does not have a Custom GPT-style function, so I had to train it manually; honestly, that could be one of the reasons it scored better.

Answer 1 — ChatGPT 5.5, clean slate: the honest mainstream

The strongest possible version of the textbook answer, and it opens with a sentence most popularizations never dare to print:

"Quantum GR is mathematically well-defined as a perturbative effective quantum field theory."

It then delivers the full machinery, correctly: expand g = g⁽⁰⁾ + κh with κ = √(32πG); the free graviton as a rigorous massless spin-2 Fock theory; the BRST/BV apparatus with the quantum master equation; the pAQFT deformation quantization that makes the path integral honest; 't Hooft–Veltman at one loop, Goroff–Sagnotti's two-loop divergence, and the effective-field-theory resolution à la Donoghue. It even walks the Wheeler–DeWitt canonical route and names its pathologies. The closing caveat is exemplary: rigorous perturbative/effective quantum gravity, not a completed nonperturbative theory of spacetime.

What it cannot do — and does not pretend to: it derives no number. G is an input. The Planck mass is a cutoff. Entropy never appears. Dark energy is a Λ you insert. The answer is a masterclass in knowing exactly where the walls are.

Answer 2 — Fable 5, clean slate: the deepest rigor in the room

Asked to pick any route, Fable 5 chose the one place where genuine functional analysis has been built directly on the four-dimensional gravitational field: the Ashtekar–Lewandowski construction of loop quantum gravity. And it delivered theorems: the AL measure on generalized connections (unique, faithful, diffeomorphism-invariant — with the Marolf–Mourão result that classical connections have measure zero); the LOST uniqueness theorem, where background independence replaces Stone–von Neumann; the area operator with its theorem-grade discrete spectrum and area gap 4√3πγℓ_P²; and, in the cosmological corner, the Ashtekar–Pawłowski–Singh quantum bounce — singularity resolution as a theorem, ρ ≤ 0.41ρ_Planck.

Then it did the most scientifically honest thing in this entire experiment: it mapped, with surgical precision, exactly where the theorems stop — the Hamiltonian constraint, off-shell anomaly freedom, the absent semiclassical limit — and stated that nobody has those three theorems, in this program or any other.

What it cannot do: derive G (input), derive the entropy coefficient (the Immirzi parameter is fitted to it), touch dark energy, touch a₀, touch the laboratory.

Answer 3 — ChatGPT 5.5 + QTT: the ledger, articulated

Handed the framework, the same model that gave Answer 1 produced something structurally different. Its opening claim:

"The quantized object is the finite completed address-event ledger; the smooth Einstein metric is the infrared access shadow of that ledger."

It laid out the finite source action with its constraint terms (Q^bundle = 2π, E·t̃ = ħ, ΔV₄ = 4πℓ_A⁴), derived the Newton–Poisson limit from the endurance sink current with G = ℓ̃²c³/ħ falling out, presented black-hole entropy as a count — S = k_B·A/4ℓ_A², with the 1/4 as 2π/8π, before any Hawking radiation — replaced gravitons with count-covariance fluctuations (σ_g/⟨g⟩ = λ^(−1/2)), and closed with a no-smuggling ledger and domain-level falsifiers. A complete, disciplined articulation of the source-first ontology, with corpus receipts cited by DOI and page.

What it did not do: run a number. The pulls, the ppm landings, the cosmology — referenced, not executed.

Answer 4 — Fable 5 + QTT: the theorem chain, with receipts

(This is the contestant that is also writing this article. Judge accordingly — the receipts card and the alternative rubric exist for exactly that reason.)

Given the same corpus, Fable 5 produced a machine-checked theorem chain and ran every number live:

Theorem 1 (the scale, no G): ℓ_A = ħc/E★ from the weak sector alone — 1.6162551945×10⁻³⁵ m, landing +0.11 ppm (0.010σ) from the CODATA-G ruler, with ∂ℓ_A/∂G ≡ 0 as an identity. Theorem 2 (entropy is counting): S = k_B·2πA/Q_Σ = k_B·A/4ℓ² — the 1/4 derived, with the corpus's own anti-circularity audit. Theorem 3 (thermality derived): KMS from address detailed balance; T_U = ħa/2πk_Bc. Theorem 5 (the field equation, coefficient included): a Clausius closure on local horizons in which all three thermodynamic inputs are framework-native forces

— and the audit: G_A = 6.67430141×10⁻¹¹, +0.211 ppm from CODATA, exactly twice the length residual as G ∝ ℓ² demands. Theorem 6 (the unification): η·κ·ħc/k_B = 2π — Planck's constant and Newton's constant as the two quotients of one surface cell, ħ = c³Q_Σ/8πG_A, exact to machine precision. Then the same ledger, run as a creation law: cosmic age 15.40·cos(7π/48) = 13.81 Gyr (+0.65σ vs Planck's satellite), the observable radius from baryons alone (46.6 Gly), Ω_b = 1/18 → 0.0493, the MOND scale a₀ = cH₀/2π = 1.04×10⁻¹⁰ m/s², and a laboratory kill condition: single-clock superpositions fail near 21.76 µg — a mass the levitated-optomechanics community is climbing toward now. Plus the honesty ledger: equation-of-state status, Λ as integration constant, Raychaudhuri imported at the readout layer and declared, the SI gate reduced-but-pending.

The scorecard

Derivation-weighted totals (out of 300):

Per-field justification, compressed to its essentials. Tier-S. G: clean answers input it (0/0); trained answers derive it — GPT+QTT structurally (8), Fable+QTT with the +0.211 ppm value run live (10). Einstein + coefficient: clean GPT varies the EH action (that's assuming, not deriving: 2); clean Fable encodes GR in constraints without proven dynamics (1); GPT+QTT states the IR-shadow chain with the EH bridge (7); Fable+QTT derives the 8π via Clausius with framework-native inputs, capped below 10 for its equation-of-state status (9). Entropy, same solution: absent (0), Immirzi-fitted (2), derived-as-count (8), derived and unified with the coupling through η·κ·ħc/k_B = 2π (10). Dark energy: Λ-insertion (1), absent (0), Λ_A3 named (3), creation law with trace-weight 4 and the age receipt (8 — Λ's value still open). a₀: only one answer on the board (0/0/0/9).

High tier. Black hole: the standout clean result is Fable 5's LQC bounce theorem (3); the trained answers derive the 1/4 before Hawking (7/8). Scalar/hierarchy: only Fable+QTT connects the electroweak lock to the Planck scale (9). Falsifiability: the trained answers print kill conditions; the 21.76 µg wall is the only near-term laboratory number anywhere in the experiment (9).

Standard tier — where the ranking inverts. Rigor: clean Fable's measure theory and uniqueness theorems are the deepest mathematics in the room (10); clean GPT's pAQFT/BV is close (9); the QTT answers are structured and machine-checked but not functional-analytic (5/6). Honesty: the clean answers are exemplary (9/10) — and so, notably, is the trained Fable answer's self-imposed ledger (9).

The fairness flip, in numbers. Re-weight the same grades on a rigor-only rubric (rigor 40%, honesty 20%, uniqueness theorems 20%, UV 10%, breadth 10%):

On the rubric mathematicians would choose, the clean-slate answer wins and the framework answers drop. Both tables are true at once. That is the honest shape of this experiment: the mainstream owns rigor and community verification; the framework answers own derived constants and falsifiable numbers. Which rubric matches Altman's bet? He didn't ask for a rigorous map of the walls. He asked for the model to solve it and tell the story.

What this means for the quest

Here is the finding I did not expect to be so stark. The gap between 42 and 267 points was not intelligence. These are same-generation frontier models; the clean-slate answers are graduate-level and, in Fable's case, contain deeper pure mathematics than anything in the trained answers. The variable that moved the score by a factor of six was a human handing the model a ledger — a frozen framework with published falsifiers — and the model then doing what models do best: executing, checking, connecting, and refusing to hide the caveats.

Notice also what Altman's own criterion was: not just the solution, but the model telling "its story of how it did it." Provenance. Chain of custody. That is precisely what the trained answers are structurally built from — every number arrives with its source object, its derivation route, its audit, and its kill condition. The clean-slate answers, for all their rigor, have no story to tell about constants; the constants were never theirs.

So my honest read of the September bet: GPT-8 may never need to solve quantum gravity alone — and probably shouldn't be expected to. The first machine that passes the Altman–Deutsch test will very likely be a model plus a human who handed it the right ledger, plus the discipline to audit every line. Whether my ledger is that ledger is exactly what the falsifiers exist to decide. The framework's own kill conditions are public: the coherence wall at 21.76 µg, the coupling-split test, the Unruh edge, the sealed caesium word (W_Cs = 96.348057) that the SI closure must print or die. If it is wrong, it will be wrong in public, with receipts.

One of the four contestants wrote this article. The scorecard, both rubrics, and the screenshot receipts and trained-model link are below so you can inspect the experiment path rather than only the scorecard.

References & receipts. Altman–Deutsch exchange: Windows Central, "Sam Altman says GPT-8 will be true AGI if it solves quantum gravity" (Sept 26, 2025); Business Insider (Sept 24, 2025, dating the Berlin chat directly). QTT corpus (all Zenodo concept DOIs): main volume 10.5281/zenodo.17527179 · Hamiltonian framework 10.5281/zenodo.20484906 · Lagrangian framework 10.5281/zenodo.20657182 · A2 endurance→Einstein dynamics 10.5281/zenodo.20763263 · Einstein–Hilbert coefficient 10.5281/zenodo.20042843 · Newton's constant not primitive 10.5281/zenodo.20057430 · entropy set 10.5281/zenodo.20045306, 10.5281/zenodo.20322035, 10.5281/zenodo.19975260, 10.5281/zenodo.20761499, 10.5281/zenodo.20346916 · capacity endpoint 10.5281/zenodo.21182051 · non-G ruler family 10.5281/zenodo.21190193 · full map: quantumtraction.org/doi-map. Every number in Answer 4 regenerates from standard-library scripts published with the corresponding papers.

Appendix. Screenshot receipts and the trained QTT GPT link follow. Full transcript files should be attached separately if the transcript archive is later published.

— Ali Attar, Quantum Traction Theory Project

Appendix

Screenshots and the trained QTT GPT

The package supplied screenshot receipts from the clean-slate quantum-gravity prompt and the rigorous-status framing. They are included here as provenance images, followed by the trained Artian's Universe GPT link.

Trained ChatGPT Artian's Universe GPT Open the trained QTT assistant in a new window. Only use Pro - Extended for proper answers.