The Gap Map

Landscape slides of the AI industry are a genre unto themselves: logo grids ranked by clout, obsolete the week they ship. Current AI — the $400M-plus public-interest nonprofit born at last year's Paris summit — has done something duller and far more useful. Its Gap Map grades open-source AI projects on three plain axes (how open, how adopted, how capable), and ships the whole thing as reproducible, MIT-licensed data where every score cites a primary source.

The point isn't to celebrate the ecosystem. It's to find the holes. And the headline hole is a bus factor: the open inference layer — the software that actually runs open models — rests on just three mature, widely-used projects (vLLM, llama.cpp, and SGLang). Nearly everything else routes through those three. If any one falters, the layer beneath much of open AI is suddenly exposed. The map inverts the usual free-rider story too: categories like agent orchestration were pioneered in open source, yet remain structurally under-maintained.

Treat it as a v0.1 census, not the last word. Of 24,626 projects surveyed, only 421 are actually scored — the rest sit as an un-triaged long tail awaiting citations, and headline counts still differ between the repo and the blog because the map is being edited live. What makes it worth opening anyway is the discipline: the data is genuinely inspectable. Simon Willison loaded its 16,000-plus tracked repos into a browser database and poked at them himself — which is the whole idea. A map you can audit is a map you can trust to point at the thin spots.