Vectron: I stopped asking them to imagine it

I teach an artificial intelligence course almost every month: four sessions, an hour and a half each. At some point in the first session I always end up in the same place: I ask the audience to imagine a language model. Not as a traditional database of tables and rows, but as something closer to a space. The problem is I don't control that imagination. Everyone in the room builds a different metaphor, and several of those metaphors are wrong.

I noticed it fast: when I showed it to them instead of asking them to imagine it — even if it was just a rough sketch on the whiteboard — the whole session changed. From there, over time, a fully didactic system grew. That system is Vectron.

What it is, without metaphors

Vectron is a public 3D map of meaning, live at vectron.kilowatto.com. Every particle in the cube is a real word from a corpus of 20,473 concepts, and its position comes from a real 1,024-dimensional embedding generated by bge-m3 (Workers AI), compressed down to 3 coordinates via PCA.

It's not a mockup or a prerecorded animation. You type a sentence and it actually gets tokenized, with the same BPE (cl100k_base) a real model uses. You pin a particle and its neighbors come from an actual vector search in Cloudflare Vectorize, across the full 1,024 dimensions, not the 3 you see on screen. You compare two concepts and the cosine similarity it shows you is the real cosine similarity. The whole stack — Workers, Vectorize, D1, R2, and Workers AI, with the frontend in Three.js on top of WebGPU (20,473 particles in a single draw call) — and a corpus that keeps growing on its own, every 30 minutes, via cron.

Why I didn't leave it as a weekend curiosity

I've spent more than 25 years in technology, most of it building companies — Ignia Cloud, my investments, Yucatech Festival. Vectron wasn't born from that same ambition. It came from exactly the same place as my classroom frustration: if something can be shown instead of described, it should be shown properly, even if that means going deep into shaders, PCA, and GPU rendering instead of leaving it as a weekend experiment.

Teaching where it lies, too

The part of the project I'm proudest of isn't the part that looks pretty. It's the part that measures, with real numbers over the corpus, how much you should trust what you're looking at.

What Vectron measures about itself: the 3 visible axes retain just 11.69% of the real variance (the full 16 components don't even reach 26%). Around 31% of the neighbors the screen suggests are invented by the projection (trustworthiness of 0.694 at k=10). The cosine similarity between two completely unrelated words is 0.412 — zero doesn't exist. And there's real clustering: one concept shows up in 77 neighbor lists against an average of 10.

That's where the product's most important design decision comes from: the promise is tied to the query you make, not to where the points happen to land. The lines connecting concepts are the honest channel — real cosine similarity in 1,024 dimensions; position is the dishonest channel — 11.69% of the variance. That's why we kept the lines, with corrections: a cap on how many get drawn, thickness proportional to the real cosine similarity, a label declaring what you asked, and lines cut short when the relationship isn't reciprocal.

The moment I knew it was going to work

It wasn't when I finished the rendering, or when the corpus hit its target of 20,473 concepts. It was in a course, showing it to my students, when they pinned a word and watched its real neighbors light up in space. That "ahh, now I get it" look I'd been chasing for years with whiteboard drawings — I finally saw it happen in real time, without having to explain anything further.

What's next

Vectron doesn't stop at a single level. There's a version for someone who has never touched a language model — no jargon, with 300 curated concepts instead of the full 20,473, because fewer elements produce more real learning. There's an intermediate version, with tokens, embeddings, and a failure lab. And there's an advanced version, for whoever wants the full instrument: the unedited math and control over the corpus. Splitting by level of expertise, rather than by difficulty, is what separates a novice looking at surface features from an expert looking at deep principles.

And the project keeps correcting itself. When one of the audits I ran on it — six in total, with more than 190 verified references — found that the corpus proved nothing about real bilingualism, I ran the proper protocol and it turned out Spanish scored correctly 50.7% of the time against 100% for English: a gap entirely between languages. It got fixed by embedding both forms, and today it's 99.3%.

That, I think, is the real idea behind Vectron: an educational tool about artificial intelligence that applies to itself the same scrutiny it teaches people to demand from any AI. Come in, play with it, and tell me how well — or how badly — it explains something you used to be asked only to imagine: vectron.kilowatto.com.