Lab · Interactive

The YC Atlas

Every Y Combinator company in the public directory — 6,215 of them — turned into a 384-dimension vector by a model running on my own CPU, then flattened onto one screen. Companies sit near each other when they describe themselves alike. Nothing here is ranked, scored, or predicted.

Snapshot 14 Sep 2026 · source: YC-OSS community mirror · 32 clusters · download the dataset (850 KB)

2005 2026

How it was built

The directory snapshot comes from the YC-OSS community mirror of YC's public company index. It is not an official YC API. Each company's description, tags and sector are concatenated and embedded with paraphrase-multilingual-MiniLM-L12-v2 (384 dimensions) through fastembed/ONNX, on CPU, locally. No embedding API was called and nothing left the machine.

Names, dates and outcome-flavoured marketing phrases are stripped before embedding, so the vector describes the activity rather than the result. That scrubbing is imperfect: a confident press-release tone still leaks through.

Clusters come from KMeans on the full vectors; the map is t-SNE over the first 50 principal components. The nearest-neighbour list in the panel is computed by cosine similarity in all 384 dimensions — not from the two-dimensional picture. That distinction is the whole point: the map is a convenience, the vectors are the data.

What this cannot tell you

The analysis built on this: AGI is almost here. What does YC actually fund in 2026?