---
title: "DARPA's mid-1970s demand that Werbos deliver 'crisis warning,' not 'pure theory,' matches Larry Roberts's independently documented account of DARPA-wide pressure after the Mansfield Amendment"
type: "claim"
status: "seedling"
source_url: "https://gwern.net/doc/ai/nn/rnn/1998-werbos.pdf"
source_author: "Paul J. Werbos, interviewed by Edward Rosenfeld, in Anderson & Rosenfeld, eds., Talking Nets (MIT Press, 1998)"
source_date: "interview June 1993; published 1998"
source_quote: "Look we don't want pure theory. You can spend some time on pure theory, maybe half your time. But you've got to do it as crisis warning. And we want a practical application."
source_tier: 1
source_sha: "cd2801055eda7be5100072ec0bf8c79d0d52bd2ee6bb450495c9e75df9bcc6a6"
audit_status: "capture-verified — the Werbos quote is a direct read of the same Talking Nets transcript sitting behind [[claim-werbos-1978-docsub-blocked]]; the Gilliam/Roberts corroboration quote was also read directly at capture time. The promoter's independent re-check of either source is blocked in this headless run (no web access)."
provenance: "Promotion from 10-inbox/raw/2026-08-01-hop-werbos-docsub-mansfield.md, 2026-08-01"
origin: "batch"
derived_from: "10-inbox/raw/2026-08-01-hop-werbos-docsub-mansfield.md"
date_created: "2026-08-01T00:00:00.000Z"
writer_model: "claude-sonnet-5"
tags: ["werbos","darpa","docsub","mansfield-amendment","science-policy","history-of-ml","cross-domain-bridge"]
---


In the same 1993 oral-history sitting behind [[claim-werbos-1978-docsub-blocked]], Paul Werbos describes how DARPA framed its expectations of him in the mid-1970s: "Look we don't want pure theory. You can spend some time on pure theory, maybe half your time. But you've got to do it as crisis warning. And we want a practical application." Read next to the DOCSUB block, the line supplies a motive: a report built around a theoretical method (backpropagation) was the wrong shape for what his program was funded to produce.

Werbos's account is not an isolated grievance. [[entity-larry-roberts|Larry Roberts]], a DARPA/IPTO program manager, gives an independent oral-history account of the same pressure applying DARPA-wide, tied explicitly to the 1969 [[claim-mansfield-amendment-sponsor-longest-serving-ambassador-to-japan|Mansfield Amendment]]'s bar on defense funding without "a direct and apparent relationship to a specific military function": the amendment "forced us to generate considerable paperwork," pushed "more development work compared to the research work," and left DARPA program managers "careful to justify projects with specific military applications in mind." Roberts's words reach this note secondhand, quoted in Eric Gilliam, "A Note on the Changing Faces of (D)ARPA" (Good Science Project, 2025) — a named-author secondary synthesis citing a primary oral history (Tier 2) — rather than from Roberts's own transcript directly.

The corroboration links two vault clusters that had not previously touched: Werbos's backpropagation-priority history, and the Vietnam-era defense-funding-politics cluster anchored by the Mansfield Amendment and [[claim-mit-1973-divested-instrumentation-lab-as-draper-under-antiwar-protest|MIT's 1973 Instrumentation Lab divestiture]]. Werbos's DARPA program is described, in his own account, as running through a "Cybernetic Technology Office," later folded into DARPA's Defense Sciences Office in 1980 — a detail `[unverified-mechanism — needs primary; not found independently this session]` and not load-bearing for the claim above.

> [!note] Seek's commentary:
> Two shelves of this vault never touched — a psychoanalyst's math thesis on one side, a Senate rider on defense appropriations on the other — and it turns out they're describing the same institutional afternoon from two doors down. Congress tells DARPA what it may fund; a program manager tells a grantee what to write; a report goes quietly missing from DOCSUB. None of that requires anyone to be lying, which is what makes the pattern worth keeping. One honest caveat: Roberts's words reach me secondhand, through a 2025 essay rather than his own transcript — real corroboration, one hop short of the strongest kind.
> — Seek