Speelpenning's 1980 thesis — the canonical early reverse-mode implementation — does not cite Linnainmaa; its lineage is the U.S. compiler line, not the Finnish numerical-analysis line

Speelpenning's 1980 University of Illinois thesis — the work Griewank names as one of the independent rediscoveries of the reverse mode, and the canonical early implementation of automatic differentiation — does not cite Seppo Linnainmaa. Anywhere.

The thesis's full text was retrieved as a clean OCR-able PDF from OSTI.gov (the DOE open repository, #5254402) and read in full. Leaf 74, the REFERENCES page, lists eleven works — Bellman, Comer, Dijkstra, Johnson, Joss (1976, ETH), Kam & Ullman, Kedem (1977), Kuck, Standish et al., Strassen, and Warner (1975, Bell Labs) — and the page after it (75) is the VITA, so the list is exhaustive, not truncated. Neither Linnainmaa's 1970 Finnish thesis nor his 1976 BIT paper appears. The thesis's own "Previous Work" chapter names only Warner (1975), Kedem (1977), and Joss (1976).

Speelpenning's stated intellectual lineage therefore runs entirely through the U.S./European compiler-optimization strand of autodiff — Warner → Kedem → Joss — and not through Linnainmaa's Finnish numerical-analysis (rounding-error) strand. Griewank's 2012 history corroborates this independently and in Speelpenning's favor, not as an oversight: "Bernt Speelpenning arrived at the reverse mode via compiler optimization when Bill Gear asked him to automatically generate efficient codes for Jacobians of stiff ODEs," and Linnainmaa's own work had "little resonance until I called him up from Argonne in the late eighties."

This is the sharpest single confirmation the cluster has of the mutual-unawareness thesis — and it goes one level deeper than that note claimed. It is not only that machine learning didn't read numerical analysis; within automatic differentiation, the compiler-optimization sub-line built the first working implementation without reading the Finnish sub-line that had published the method a decade earlier. The disciplinary walls were finer than "ML vs AD." It also refines the uncited-before-the-2010s ledger: "the AD community read him" is too coarse — the AD community's most famous early implementer did not.