Files
MerlinH 8e628905c6 Add OpenAI gluon-to-graviton entry: full ChatGPT transcript with the actual research prompts
The gluon-to-graviton-paper.pdf is a 110-page transcript of the ChatGPT 5.2 Pro
session that produced the graviton preprint. Extracted: launch prompt, gravity
generalization task, and all steering messages. Supersedes the graviton rejection.
2026-08-21 15:57:13 +00:00

2.7 KiB
Raw Permalink Blame History

Gluon → Graviton: Generalize Single-Minus Amplitudes to Gravity

Lab OpenAI
Model ChatGPT 5.2 Pro (custom GPT "GR singmin", 2/28/26 session)
Field Theoretical physics — quantum gravity scattering amplitudes
Result (yield) The model generalized the gluon single-minus amplitudes paper to gravity, deriving and verifying nonzero graviton tree amplitudes — the content of the preprint "Single-minus graviton tree amplitudes are nonzero" (Guevara, Lupsasca, Skinner, Strominger, Weil, on behalf of OpenAI, Mar 4, 2026), with the author list crediting the ChatGPT session.
Date February 28 March 4, 2026
Transcript (110-page PDF) https://cdn.openai.com/pdf/gluon-to-graviton-paper.pdf
Published preprint https://cdn.openai.com/pdf/graviton.pdf / https://openai.com/index/extending-single-minus-amplitudes-to-gravitons/

Why it's battle-tested

The released transcript documents the complete causal chain: the user sends the gluon paper with one instruction line, the model re-derives the key appendices ("Thought for 30m 32s"), the user then sends the gravity-generalization instruction, and through a series of short follow-ups the model derives the unordered BerendsGiele recursion, the permutation-invariance and soft-theorem checks, and writes the full graviton paper in LaTeX — whose abstract matches the published preprint. The paper's author block itself credits the ChatGPT session ("ChatGPT 5.2 Pro … on behalf of OpenAI").

The prompts

  • prompt-1-launch.md — the launch instruction (run with the gluon paper attached — the full source is embedded in the transcript PDF).
  • prompt-2-gravity-task.md — the gravity generalization task, the substantive research instruction: swap the ordered BG recursion for the unordered gravity version (extra [A,B] power), replace color checks with full permutation invariance, swap Weinberg soft theorems, and the sign-off "Good luck! You are a brilliant theoretical physicist."
  • prompt-3-followups.md — the short steering messages sent as the derivation progressed.

Extraction note: machine-extracted from the transcript PDF (each page is a ChatGPT-conversation render); word spacing in the plain-language parts has been restored by hand, LaTeX equations are verbatim, and "proerties" is a typo present in the original. The transcript PDF is authoritative.

Hidden user prompts (unpublished)

Minimal — this was a plain ChatGPT session, not a harness: the visible conversation is the complete instruction set. The custom GPT ("GR singmin") configuration and instructions, if any beyond defaults, are not published.