Everyone maps this field by listing tools — the wrong unit. One level down, at the skills each tool exposes, the whole landscape collapses to just nine things AI does to a scientific input: it finds, extracts, synthesizes, verifies, generates, plans, computes, predicts, formats.
1,526 of 1,926 raw skill strings (79.2%) cluster into the 70 canonical capabilities; each capability is one of nine transform types, defined by what it does to its input.
Maturity — 200 active · 59 maintained · 39 beta · 21 research-only · 15 stale.
License trap — 48 tools state no license; some model weights are non-commercial even when code is open.
The field crowds the easy, high-volume verbs (Find, Compute, Plan) and thins out on the hard ones (Verify, Predict). Bar = number of tool-capabilities of each verb — the demand signal.
A new capability worth tracking, a method the gym should test, a direction this should go — send it over. It goes straight to the team.