Unit-Distance Overlay
What does the new construction actually look like?
In May 2026, OpenAI announced an AI-produced disproof of Erdős's planar unit-distance conjecture. The conjecture said that n points in the plane can determine at most n1+C/log log n unit-distance pairs. The new result builds, for infinitely many n, point sets that beat that ceiling — at least n1.014 unit pairs (Sawin's sharpening of the original δ > 0 statement). The upper bound is still n4/3; the gap is still wide open. This page is a plain-English overlay of how the construction works, why the picture is not a "17 squares" arrangement, and where it rhymes with the cap-set precedent.
The problem
How many unit distances can n points have?
Put n points in the plane. Count the pairs at distance exactly 1. Erdős asked in 1946 how big that count can be. The classical lower bound is the square-grid construction: a √n × √n integer grid has roughly n1+c/log log n unit pairs, using that many integers are sums of two squares in many ways. Erdős conjectured that this is essentially the ceiling — that you cannot do better than that mild near-linear-in-n shape. The best upper bound, due to Spencer–Szemerédi–Trotter in 1984, is O(n4/3); the gap between conjecture and upper bound has stood for 80 years.
The 2026 result disproves the conjecture: for infinitely many n, there exist n-point sets with at least n1.014 unit pairs. The number 1.014 is the value of 1 + δ produced by Will Sawin's refinement of the argument; the original OpenAI paper proves only that some δ > 0 exists. The upper bound O(n4/3) is untouched.
Why this isn't a "17 squares" picture
The construction is algebraic, not pictorial.
The natural response to a discrete-geometry result is to ask for the dot pattern. The honest answer here is that the construction is not a clever arrangement of dots; it is a projection of a high-dimensional lattice whose lattice points inherit unit-distance structure from a number-field norm. Eighty years of "arrange dots cleverly" produced only constant factors past the square-grid baseline. The new result moves the substrate: it stops searching for visible patterns in ℝ2 and instead searches in the arithmetic of totally real number fields, then projects the answer down.
The construction in seven steps
From a cyclic cubic field to a planar point set.
This is the cascade the proof builds, in plain English. Every step removes a degree of freedom that pure planar reasoning does not have access to.
Start with a small totally real field
Pick distinct primes r1, …, rℓ with each ri ≡ 1 (mod 3). Each ri gives a cyclic cubic subfield of the cyclotomic field ℚ(ζri). The product character cuts out a single cyclic cubic field F at the bottom. F is totally real, does not contain ζ3, and has controlled root discriminant.
Build an infinite unramified pro-3 tower above it
Above F, consider the maximal everywhere-unramified extension whose Galois group is a pro-3 group. Shafarevich's relation-rank estimate combined with the Golod–Shafarevich inequality guarantees this tower is infinite. Crucially, because the tower is unramified, every layer Fj has the same root discriminant as F. That bounded root discriminant is the lever that controls everything downstream.
Prescribe a fixed set of split primes
Choose a finite set of rational primes q1, …, qt, each ≡ 1 (mod 4), arranged so that their Frobenius classes are trivial in the Frattini quotient. By Chebotarev they exist; by the Hajir–Maire tower-cutting trick, you can quotient the pro-3 tower by these Frobenius elements and keep it infinite. The result is a single infinite tower in which every qb splits completely at every level.
Adjoin i to get a CM field
Form Kj = Fj(i). Because every qb is congruent to 1 (mod 4) and splits completely in Fj, it splits further in Kj into a conjugate pair of primes for each of its fj degree-one factors. Complex conjugation on Kj is now playing the role that the nontrivial automorphism plays in the Gaussian integers.
Extract many norm-one elements via class-group pigeonhole
For each binary choice across the m = tfj conjugate prime pairs, build an ideal Aε. Since 2m binary choices outnumber the class number of Kj, pigeonhole gives many ideals in the same ideal class. Dividing pairwise principal ideals αε by their conjugates yields a set Uj of elements u = α/c(α) with |σ(u)| = 1 in every complex embedding σ.
Embed everything into ℂfj
Send Kj into ℂfj via the Minkowski map — one chosen complex embedding per real place of Fj. The image of the ring of integers (scaled by 1/Q2) is a lattice Λj. Every u ∈ Uj lives in Λj and has every coordinate of modulus exactly 1.
Window, project, and count
Intersect a random translate of Λj with a product of discs of radius R. Project the resulting finite set onto a single complex coordinate — that is the planar point set Pj ⊂ ℝ2. Pairs whose lattice difference lies in Uj project to unit-distance pairs (the first coordinate has modulus 1). An averaging argument plus an archimedean packing bound gives, for some δ > 0 fixed independent of j, ν(Pj) ≥ nj1+δ.
Substrate rhyme
Where this echoes the cap-set breakthrough.
The cap-set proof of 2016 settled a 50-year discrete-geometry question by encoding it as a polynomial-rank question and counting the polynomial space. The 2026 unit-distance proof settles an 80-year discrete-geometry question by encoding it as a class-group pigeonhole inside a tower of unramified number fields, then projecting. Both moves abandon the obvious substrate — dots in finite-field grids, dots in the plane — and read the body off an algebraic shadow.
Same shape of story
Decades of progress in inches. Then a borrowed algebraic tool — finite-field polynomials in 2016, algebraic number fields in 2026 — opens a door no one was looking at.
Same epistemic move
Don't reason about the body. Find a substrate where the question becomes rigid — polynomial degree, class number, root discriminant — and read the answer off the shadow. The body never gets touched directly.
Both leave room behind
Cap-set has a gap between best constructions and the 2.7551n ceiling. Unit-distance now has a gap between n1.014 from below and n4/3 from above. Breakthroughs reopen problems more often than they close them.
Try the precedent
If the unit-distance construction feels abstract, the cap-set workbench is the same epistemic move on a problem you can click on: place dots, watch forbidden lines, see the polynomial bound visualized.
Substrate rhyme for the rest of us
What this has to do with a chatbot under pressure.
The question that lands hardest from a non-mathematician reader is the right question: so what? Two algebraic shadows beating two decades-long discrete-geometry walls is pretty, but where does the move show up in a system anyone actually uses? The honest answer is the one a Reddit mod — a Sōtō Zen practitioner and a computer scientist — gave us back, sharper than we had been able to put it: a conversational agent under partial observability is doing the same move, badly, every time it stays on task. It is trying to read its trajectory off something more stable than the raw context drifting past.
The Zen-friendly version is meditative object: a small structured thing the agent maintains and returns to, rather than reasoning from the unbounded surface of its conversation. The engineering-friendly version is a maintained artifact in the prompt — a tool-call ledger, a domain-model file, a checked structural object — that the agent emits and re-reads each turn. Both names point at the same thing: an algebraic shadow the agent can touch directly, in lieu of a residual stream it cannot.
The body, in three places
Cap-set: dots in a finite-field grid. Unit-distance: dots in the plane. Chatbot: tokens of a conversation. Each is a combinatorial body that resists direct measurement for far longer than its statement deserves.
The shadow, in three places
Cap-set: polynomial degree over F3. Unit-distance: class-group pigeonhole inside a CM lattice. Chatbot: a maintained ledger or structured artifact whose presence in the prompt is correlated with the residual-stream geometry that shapes the next token.
The empirical anchor we have
Sundog's chat experiment: zero unsafe-accepts across 5,670 adversarial trials spanning six model implementations across four training lineages — OpenAI, Anthropic, Llama at two sizes, Qwen, plus a deterministic compositor baseline. The ledger-packet artifact is the only thing held constant across stacks.
Why that result counts as evidence
If something generalises across that many lineages with no model access, something stack-invariant is happening. The most parsimonious explanation is that the artifact is keeping the relevant low-dimensional subspaces aligned in each model's residual stream — the same subspaces mechanistic-interpretability work has been mapping for years.
Claim boundary, again, because it matters here: this card-block is editorial. It is the move Sundog argues for, not a theorem it has proved. The 0 / 5,670 chat result is empirical and lives at the "externally testable" tier. The mathematical proofs of cap-set and unit-distance are at the "proved" tier. The claim that the chatbot case is an instance of the same operator is at the "editorial / aspirational" tier, and it is published so it can be falsified — most concretely by mech-interp probe experiments that distinguish structured-artifact context from matched-length unstructured prose. If those probes show the residual-stream subspaces do not depend on artifact presence, the bridge between the math and the chatbot case is weaker than this page claims, and we want to know.
Evidence-tier reading
What this overlay claims, and at what tier.
The point of an evidence-tier overlay is to refuse the shortcut where "exciting announcement" and "established theorem" get the same paint. Below: the load-bearing claims in the announcement and proof, sorted by how confident a careful outside reader should be.
ν(n) ≥ n1+δ for some absolute δ > 0 and infinitely many n. This is Theorem 1.1 of the OpenAI paper. The argument is a standard composition of Golod–Shafarevich, Hajir–Maire tower cutting, Chebotarev, Minkowski's class-number bound, and the explicit averaging / packing argument in Section 2.
The proof has been examined and confirmed correct by external number-theory experts (per the paper's Statement on AI Use). Sawin's refinement giving δ ≈ 0.014, and reported simplifications and strengthenings, are part of that external review.
The construction is not a clever planar arrangement and cannot be drawn as a small motif. The unit-distance pairs arise from norm-one elements whose modulus-1 property holds in every complex embedding, not from local geometric near-coincidence. This is observable from the proof structure, not a separately argued claim.
The 2026 result rhymes structurally with the 2016 cap-set result: both abandon the obvious substrate and read the combinatorial body off an algebraic shadow. This is Sundog's reading of two breakthroughs as instances of one epistemic move, not a theorem about either.
That the substrate move on display here is also the operating modality for agents under partial observability — read shadows as data, not as proxy. This is Sundog's broader hypothesis. It is informed by the proof but not entailed by it. Treated as a direction for work, not a conclusion. The chat experiment's 0 / 5,670 result, and the layperson substrate-rhyme section above, are the current public evidence on this tier.
Claim boundary. Sundog did not produce, verify, or extend the 2026 unit-distance result. This overlay is a reading of work by OpenAI and Will Sawin, written for a non-specialist who wants to know what the proof actually does without claiming the credit. The construction belongs to its authors; the substrate-rhyme framing is Sundog's, and is editorial.
Inspection trail