Polyhedra.net has been quietly serving printable nets for a few hundred geometric solids, free, with no login, no paywall, and no JavaScript framework in sight. The site’s front page lists its categories in a flat, unglamorous way: Platonic Solids, Archimedean Solids, Kepler-Poinsot Polyhedra, Other Uniform Polyhedra, Compounds, Pyramids, Concave Pyramids, Truncated Pyramids, Dipyramids, Prisms, Antiprisms, Concave Prisms, Twisted Prisms, Kaleidocycles. Click a picture, get a PDF net, print it, fold it. That is the entire product.
This is not a news event. Nothing shipped this week. But the site is a useful specimen for anyone thinking seriously about what frontier AI models are actually trained on, and what they are not. It is the kind of page that shows up in a web crawl, gets tokenized, and disappears into a 15-trillion-token pretraining corpus without anyone noticing. The question worth asking is whether anything useful survives that trip.
What the site actually contains
The archive is small by web standards. A few hundred models, organized by mathematical family. The taxonomy is precise and mostly correct: Platonic Solids are the five regular convex polyhedra; Archimedean Solids are the thirteen convex uniform polyhedra with regular faces but not all identical; Kepler-Poinsot Polyhedra are the four regular star polyhedra; Kaleidocycles are the ring-shaped flexahedra that rotate through themselves. Each entry links to a net, a flat 2D unfolding that folds into the 3D solid.
That last part is the interesting bit. A net is not a picture of a polyhedron. It is a construction plan. To go from net to solid, a reader has to mentally fold faces along edges, track which vertices meet, and verify that the dihedral angles close. This is spatial reasoning, and it is exactly the kind of task that large language models have historically been bad at, despite scoring well on text benchmarks.
Why this is an AI story
Frontier labs have spent 2024 and 2025 pouring compute into reasoning. OpenAI’s o-series, DeepMind’s Gemini 2.5 thinking modes, Anthropic’s extended thinking in Claude, and DeepSeek’s R1 all bet that test-time compute can buy better multi-step inference. The benchmark suite has followed: AIME, GPQA, FrontierMath, SWE-bench Verified. These are hard problems, and the scores have climbed fast.
Geometry is a different beast. The standard measure is Geometry3K and its harder cousin PGPS9K, both built from textbook problems. Models do well on those. Give a model an unfolded net and ask it to identify the solid, or ask it to produce a valid net for a truncated icosahedron, and the failure rate jumps. The reason is not vocabulary. It is that the model has to hold a 3D object in working memory and manipulate it, and token prediction is a poor substrate for that.
Polyhedra.net is a reminder that the web contains a lot of this kind of structured, spatial, non-verbal knowledge. It is sitting there, freely licensed in practice if not explicitly, in a format that a crawler can ingest. Whether that ingestion translates into capability is an open empirical question, and the labs are not publishing the answer.
The data problem underneath
There is a harder point. Most of the high-quality spatial and geometric content on the open web is thin. Wikipedia’s polyhedron articles are dense and well-sourced. Wolfram MathWorld is authoritative but commercially licensed. Polyhedra.net is one of a small number of sites that offers the actual construction artifacts, the nets, at scale and for free. If a lab wanted to build a spatial-reasoning training set from scratch, it would find itself scraping sites like this one, plus a long tail of university course pages, hobbyist CAD forums, and origami archives.
That tail is exactly what the current crawl-and-filter pipeline tends to underweight. Deduplication heuristics downrank small sites. Quality filters trained on perplexity against a reference model penalize unusual formatting. A page that is mostly images with short captions, like most of polyhedra.net, gets flattened or dropped. The result is that frontier models see a lot of Reddit and a lot of Stack Overflow and comparatively little of the hand-built, specialist web.
This is not a new complaint. It is the same one researchers have made about low-resource languages, about niche scientific subfields, and about non-English content. The polyhedra case is just unusually clean, because the knowledge is visual and the site is small enough to count.
What it means for builders
For anyone building AI tools in 2026, the practical takeaway is narrower than “the web is dying.” It is that spatial and geometric reasoning remain a genuine gap, and that gap is not closing as fast as the language benchmarks suggest. If you are building a CAD assistant, a robotics planner, a materials-design copilot, or an education product that touches geometry, do not assume a frontier model has absorbed the relevant structure just because it can name the Platonic solids. It can name them. It cannot reliably fold them.
The other takeaway is about data provenance. Labs that want to close this gap will need to license or carefully curate specialist corpora, not just scale the crawl. Polyhedra.net has no robots.txt problem and no paywall, but it also has no API, no structured metadata, and no explicit license. That ambiguity is the norm for the specialist web, and it is going to become a bigger deal as the easy tokens run out.
The site’s own framing is modest. “Polyhedra are beautiful 3-D geometrical figures that have fascinated philosophers, mathematicians and artists for millennia,” it says. “On this site are a few hundred paper models available for free. Make the models yourself.” That last instruction is the one no current model can follow end to end. Print the net, fold the faces, tape the edges, check the solid. Watch a model try it and you will see where the reasoning actually stops.