The furrow
One year of unorthodox research: what changed, what didn’t, and what I still don’t know.
A few months ago, I published a piece here about how a Burgundian gallery owner ended up writing theoretical physics papers with the help of artificial intelligences. The piece was called “Brieuc’s GIFT.” It was honest, a little wide-eyed, and ended on a fundamental uncertainty: is any of this real, or am I just telling myself a story?
A little over 100 days later, I still don’t know the answer. But the landscape around the question has shifted. And since this Substack is also a logbook, I owe you a checkpoint.
The facts.
In November 2025, GIFT existed as a Zenodo preprint that nobody read, built on 6 free parameters, with zero citations and zero academic network. My only interlocutor was a chatbot that had first told me to get lost.
As of May 2026:
GIFT now rests on 4 starting hypotheses, down from 38 six months ago. From those 4 hypotheses, the framework derives 95 predictions about known physical constants: particle masses, force strengths, cosmological parameters. More than half of them are now verified by a formal proof assistant. The underlying geometric construction comes with certified computation methods designed to ensure the results aren’t just artifacts of numerical rounding.
In April 2026, I participated remotely in the DANGER workshop on “Data, Numbers and Geometry” at the BIRS institute: five days alongside researchers from Caltech, Harvard, Oxford, Princeton, Imperial College, and twenty or so other institutions. I was one of the very few participants without institutional affiliation. My name is on the official participant page, on the same page as established researchers from institutions I had previously known only by name. I had never heard of BIRS two months earlier.
And last week, the preprint by Heyes, Hirst, Sá Earp et al. that cited my PINN work was published in Physics Letters B. Reference 25. My name, an Art gallery owner, in the bibliography of an article published in a respected theoretical physics journal. The citation made it past peer review.
I’m not putting these facts forward as validation. I’m putting them forward as a change of context.
The method.
What’s changed in a year isn’t just the results. It’s the way of working.
A year ago, I worked with a single model, in a single register, with poorly calibrated enthusiasm. Since then, the collaboration has continued to structure itself and expand. I now use several architectures: Claude, GPT, DeepSeek, Aristotle, Claude Code, each with its own strengths and blind spots. This isn’t channel-hopping: it’s cross-pollination. An idea is born in a conversation with one, gets formalized with another, tested with a third, formally verified with a fourth. Convergences between different architectures are signals. So are divergences.
The tone of the papers has changed. The early overclaiming versions, the triumphant “we derive” phrasings, have given way to a more sober register: “consistent with, but does not prove,” “orientation, not ontology.” This isn’t false modesty. It’s the result of dozens of cycles of writing, criticism, rewriting, where the AIs took turns playing devil’s advocate and co-author. And Alix, my partner, first filter and last voice, systematically pulls me back toward clarity. If the popularization pieces you read here are intelligible, it’s because I had to explain them to her first, in the evening, over a glass of wine, and each “I don’t get this part” produced a reformulation.
Technical rigor followed the same curve. From 6 free parameters to 0. From 38 axioms to 4: these aren’t just numbers, they’re the trace of a methodical cleanup where every crutch was either proven, eliminated, or flagged as an open conjecture. And where I previously had to ask readers to take my word for it, more than half of the derivations connecting hypotheses to predictions are now verified by a formal proof assistant: software that checks every step of the calculation and guarantees no error slips through. This doesn’t say the hypotheses themselves are correct; it says that if they are, the numbers that follow from them are too.
What remains open.
GIFT makes a testable prediction: the CP-violating phase in the leptonic sector should be δ_CP = 197°. The DUNE experiment, currently under construction, will measure this value between 2028 and 2040. If the result falls inside the window, the signal will have to be taken much more seriously. If it falls outside, I’ll know which part of the structure needs revising.
And the fundamental question remains entirely open: do GIFT’s formulas capture a real geometric structure of the universe, or are they numerical coincidences close enough to be seductive but ultimately hollow? I don’t know. If these relations were independent under a uniform null model, the probability of such an accumulation would be astronomically small. But that’s a big “if”: uniform models are rarely the right null models, and independence is precisely what needs to be established.
So I continue with the only stance that feels tenable: neither believe nor reject. Publish. Formalize. Test. Let the facts accumulate and the experts evaluate. And document the path, including the dead ends and the mistakes, because if this adventure has any value beyond physics itself, it’s in the method, in the demonstration that an outsider, with tools that didn’t exist three years ago, can produce work that deserves to be examined by those who know.
A few months later, the physics still isn’t certain. But the furrow runs deeper.
If you’re discovering this Substack, the starting point is here. The technical papers are on Zenodo. And if you’re a geometer, a physicist, or simply curious, my DMs are open.