The obvious connection: 2D crystals

In 2D materials, atoms sit on a plane. Nearest-neighbor bond lengths are unit distances. The classic 2D lattices rank-ordered by coordination number:

• Square lattice — coordination 4
• Triangular lattice — coordination 6, densest planar packing
• Honeycomb (graphene) — coordination 3, a sublattice of the triangular

The unit-distance problem asks how many “nearest-neighbor bonds” a planar point set can have. The 80-year consensus was: periodic lattices are optimal. That consensus just broke.

The hard boundary

Before drawing the wrong conclusion: real material geometry is not a pure optimization of planar point configurations. Carbon sp² hybridization is three-coordinate because of electron structure, not because a better geometric arrangement has not been found. Thermodynamic stability, electronic constraints, and symmetry rules govern what configurations physically exist.

OpenAI’s new configuration family does not predict new materials. Worth being explicit about that.

But the quasicrystal parallel is real

In 1982, Dan Shechtman found diffraction patterns in a rapidly-cooled Al–Mn alloy that showed fivefold rotational symmetry — impossible under the crystallographic dogma that all long-range ordered solids must be periodic. The field’s response was initially rejection. Shechtman eventually won the 2011 Nobel Prize in Chemistry.

Quasicrystals exist because the assumption “ordered = periodic” turned out to be wrong. The unit-distance result is the same category of lesson: the intuition about what optimal planar arrangements look like had a systematic blind spot for eighty years.

Here’s where it gets personal

My undergraduate research was on gas-phase metal clusters: magnetron sputtering to generate clusters in the gas phase, photoelectron spectroscopy for experimental signals, and genetic algorithms to search potential energy surfaces (PES) for lowest-energy configurations. Theoretical spectra are generated from candidate geometries and compared against experiment; when they match, the geometry is confirmed as the structure.

In this workflow, the geometry is what is being solved for, not what is assumed at the start.

The case I keep coming back to is Ag₁₀ versus Au₁₀. Same group, same valence electron configuration, same nominal coordination chemistry. If structure were purely geometric, they should converge to the same lowest-energy configuration. They do not.

Gold’s relativistic effects — 6s orbital contraction, 5d orbital expansion, strong s–d hybridization — mean that Au clusters stay planar to significantly larger sizes than Ag clusters. Au₁₀ is essentially flat. Ag₁₀ has already transitioned to a 3D configuration. The genetic algorithm exploring the same geometric search space falls into different basins because the shape of the potential energy surface is different.

Geometry is downstream of electronic structure, not upstream.

What the unit-distance result actually means for structure search

The direct implication is limited. The unit-distance problem is a hard-constraint, potential-free, combinatorial geometry problem — categorically different from DFT global optimization on a potential energy surface.

The indirect implication, for me at least, is methodological. Genetic algorithms for structure search depend on priors about what “reasonable configurations” look like — how the initial population is seeded, how crossover and mutation operators are designed. Many implementations carry implicit preferences for high-symmetry offspring, which makes sense as a heuristic but is not obviously grounded in anything more principled than convention.

The cluster search community has been moving away from this for years: random seeding, basin-hopping, particle swarm optimization, machine-learning potentials for accelerated PES sampling. The unit-distance result is probably more supporting argument for an existing trend than paradigm shift. But it is a clarifying argument, and those are not always easy to find.

One sentence

OpenAI showed that the eighty-year intuition about optimal planar arrangements had a blind spot. For me, it was a useful occasion to think about how much geometric intuition is quietly embedded in the tools one reaches for most automatically — and how rarely that gets examined.

Note: OpenAI’s proof has not yet undergone formal peer review. Princeton’s Will Sawin has independently refined parts of the result, which is a good sign. Status should be monitored.

The structural parallel

Nature op-eds, Cell Press perspective pieces, and the Viewpoints section of Accounts of Materials Research all share an underlying architecture:

1. One counter-intuitive proposition stated in the first paragraph
2. A defense built on anomalous data, not consensus data
3. An exit that reframes the whole field

This is exactly the structure of a good Subway Take. The constraint — one stop, one transit card, one claim — is not a limitation. It is the thing that forces clarity. A research proposal that cannot survive the Subway Take test probably has not found its core claim yet.

Counter-intuitive entry points

In energy materials, the default mode of literature-reading is consensus extraction: what do most papers agree on? This is useful for situational awareness. It is not useful for generating paradigm-shifting ideas.

The Subway Take requires the opposite: where does the literature contradict itself? Where do two papers with almost identical experimental setups report results that cannot both be true? Where is the anomaly that everyone has been footnoting and nobody has been following?

The same logic applies across the field: localized defects in SRR catalysts, semi-soluble transition layers in SEI formation, polysulfide corrosion dynamics in Li-S systems. The literature treats these as problems. The Subway Take asks: what if they are the mechanism?

The one-sentence test

There is a version of this for research: whether the core claim can be stated in one sentence that a peer in the field can immediately engage with — agree, disagree, push back. Something like:

The protective SEI layer is not passive — it is an active material that can be programmed from the electrolyte side.

That sentence can be argued. It makes a specific prediction. It implies an experiment. Whether the analogy to a subway take holds up under scrutiny is another question, but I find the exercise useful for checking whether I have actually landed on a claim yet, or whether I am still in the part of the project where I am collecting evidence for a question I have not clearly posed.

The visual anchor

In the interview program, the transit card is the unmistakable prop — the visual signal that we are in argument mode. In high-impact scientific communication, this function belongs to the TOC figure.

The TOC figure is not a results summary. It is a physical model that makes the counter-intuitive claim visible before the reader encounters the evidence. It should be minimalist — one clear spatial or mechanistic relationship, rendered in a style that communicates “conceptual” rather than “data.” The figure serves as a Schelling point: when people disagree about what the paper claims, there is something concrete to point to.

Systematic anomaly hunting

Generating these entry points requires a different relationship to the literature than consensus-mapping does. Automated literature filtering — tracking conflicting datasets, unexplained fluctuations, mechanisms that the authors flag but do not resolve — is not a shortcut. It is a way of running the anomaly detection at a scale that individual reading cannot sustain.

When a Subway Take-ready anomaly surfaces in the automated stream, the question is always the same: is this a measurement artifact, or is this the experiment that is quietly right about something the field has been quietly ignoring?

The transit card is in someone’s hand. The train is moving.

The vanishing moat

For a generation, the moat of a working scientist was a stack of small, learned advantages — a feel for which solvent to try first, a familiarity with the quirks of a given instrument, a cultivated network of collaborators who knew which dataset to trust. None of these were glamorous. All of them were, until very recently, genuinely scarce. Scarcity is what makes a moat a moat.

What I have observed over the last two years — first as an annoyance, then with mounting alarm, then with a strange sort of relief — is that almost every one of these advantages is being commoditized in real time. Using Claude to assist with batch DFT calculations has been a game-changer: from iterating Linux scripts to generating input files, the efficiency is staggering. The work I used to charge a graduate student with for three weeks now happens between morning coffees.

The moat is not gone. It is evaporating — slowly enough that we can still pretend, fast enough that the pretending is starting to feel theatrical.

From scripts to ecosystems

The shift that matters is not “AI helps with parts of research.” That phase ended around 2024. What is happening now is the assembly of autonomous research ecosystems: pipelines where a hypothesis enters at one end, and a benchmark-anchored, peer-verified manuscript exits at the other. The orchestrator is no longer a postdoc; it is an agent graph. AtomPub is one node in such a graph, and it is far from the only one.

What remains, then?

The honest answer is: a small set of things, but a very important set. Choice of question. Choice of dataset. Choice of what counts as a good answer. These are the upstream decisions that the ecosystem cannot make for you, because the ecosystem has no preferences of its own. It will optimize whatever objective you give it.

Eq. 1 is an admittedly cute way of saying: as compute becomes free, the researcher’s contribution is increasingly the taste embedded in the gradient itself — the angle of attack, not the magnitude.

Taste as the last moat

I do not mean taste as aesthetic preference. I mean taste as the accumulated, mostly-unverbalized intuition for what is worth doing. It includes knowing which experiments are real, which results would actually change your behavior, and which questions are merely fashionable. This is exactly the part of our work that ecosystems cannot synthesize from scratch — not because they are weak, but because they have no skin in the game.

A note on the social-media tax

A small adjacent observation, because it bears on how this essay itself will be discovered. Since 2022 I have run a paper-bot on X for the battery community, and four years of reading Open Graph crops has convinced me that publishers vary wildly in how seriously they take the front-end of dissemination.

The technical fix is not exotic — proper og:image at 1200×630, a summary_large_image Twitter card, and a TOC figure designed with safe-area center in mind. AtomPub treats this as table stakes. When a high-impact paper is throttled by an algorithm responding to a 480×270 thumbnail, the throttling is not ideological. It is just bad packaging. We can fix the packaging.