---
id: "20260922-0231-what-genuinely-connects-schmidhubers"
title: "What genuinely connects Schmidhuber's 'Annotated History' prolepsis and Nicely's 2011 FAQ memory-reconstruction caveat?"
type: "capture"
status: "promoted"
origin: "batch"
promoted_to: ["30-notes/observation-schmidhuber-nicely-cosine-088-pairing-is-analytical-contrast-not-evidentiary-bridge.md"]
not_promoted: ["Claim 1 (no evidentiary connection, cosine tracks vocabulary) and Claim 3 (opposite-pole contrast is the genuine point of contact) folded into the single promoted observation note rather than kept separate — they are the same finding stated as verdict-then-payload, matching the vault's established 'cosine bridge is thematic not causal' genre (e.g. observation-hinton-arc-rhw-1986-cosine-bridge-is-thematic-not-causal.md), and splitting them would have produced two notes restating one claim, violating atomicity.","Claim 2 (Nicely = memory-reconstruction, Schmidhuber = prolepsis, two named axes of retrospective-history risk) is not a separate atomic claim — it is constitutive of the bridge finding itself, not independently reusable outside that context, so it was absorbed into the same promoted note rather than split out.","Further-leads bullet 1 (Jacquard-hedge vs. Torres y Quevedo unhedged) — already fully covered by the existing claim-schmidhuber-1914-torres-y-quevedo-claim-unhedged-vs-jacquard-hedge.md; linked from the new note, not duplicated.","Further-leads bullet 2 (whether Nicely ever applies a category retroactively in his own writing) — a nice-to-verify curiosity about a mirror case, not a doubt any kept claim rests on. Not routed to 50-questions/ per question-intake discipline; left as an unexplored lead only in this capture.","Further-leads bullet 3 (disputed 'Tom Kraljevic' attribution) — already flagged as unresolved and unasserted in the existing Nicely claim-note; not this capture's to resolve, no new note or question needed."]
writer_model: "claude-sonnet-5"
date_created: "2026-09-22T00:00:00.000Z"
provenance: "Batch capture run, 2026-09-22, bridge-check on cosine-0.88 unlinked pair [[claim-schmidhubers-annotated-history-projects-deep-learning-onto-pre-1900-mathematics]] / [[claim-nicely-2011-faq-corroborates-1994-dates-but-is-a-memory-reconstruction]] per proposal-seek-2026-w36"
derived_from: []
tags: ["bridge-check","schmidhuber","thomas-nicely","whig-history","historiography-of-ai","prolepsis","oral-history","source-methodology","embeddings","cosine-bridge","history-of-computing"]
source_url: "https://arxiv.org/pdf/2212.11279"
source_title: "Annotated History of Modern AI and Deep Learning"
source_author: "Jürgen Schmidhuber"
source_date: "Submitted 2022-12-21 (v1); last revised 2025-12-29 (v3, the PDF read this session)"
source_venue: "arXiv:2212.11279 [cs.NE]; also Technical Report IDSIA-22-22, IDSIA"
source_tier: 1
source_sha: "75cef8513b92b64a6c13eeb5b9cce3eeec109c07bc934c561eec9f614581319b"
source_quote: "the chain rule (1676), the first NNs (circa 1800), the first practical AI (1914)"
source_url_2: "https://web.archive.org/web/20251218063712if_/https://faculty.lynchburg.edu/~nicely/pentbug/pentbug.html"
source_title_2: "Pentium FDIV flaw FAQ"
source_author_2: "Thomas R. Nicely"
source_date_2: "2011-08-19"
source_venue_2: "Thomas R. Nicely's FAQ on the Pentium FDIV bug ('pentbug.html'); original at faculty.lynchburg.edu now 404, read via the 18 December 2025 Wayback Machine capture (original gone, archived copy is the source of record)"
source_tier_2: 2
source_sha_2: "bcf7948942bbf12f042c85ed164451e986092ac57c62be4b12a275574a38b50c"
source_quote_2: "Following below is the chronology as best I can now reconstruct it."
seek_code_commit: "0158402"
---


This capture is a bridge-check on a retrieval-surfaced pair: [[claim-schmidhubers-annotated-history-projects-deep-learning-onto-pre-1900-mathematics]] (Schmidhuber's *Annotated History of Modern AI and Deep Learning* reads the 2010s-vintage category "deep learning" backward onto 1676–1914 mathematics and devices) and [[claim-nicely-2011-faq-corroborates-1994-dates-but-is-a-memory-reconstruction]] (Thomas Nicely's 2011 FAQ nails its two headline dates against his own 1994 e-mail but self-labels everything else as reconstructed memory), sitting at cosine 0.88 and unlinked. No sibling capture in `10-inbox/raw/` mentions both note slugs together, so this is a from-scratch investigation, not a duplicate. Both primaries were independently re-fetched this session (Schmidhuber's PDF via `extract_pdf`, `tls: "verified"`, sha256 matching the vault's existing citation exactly; Nicely's FAQ via `archive_page`, sha256 also matching the existing citation exactly) and the quotes below were confirmed `grounded: true` against those fresh reads.

## Claim: the two notes share no evidentiary connection — no common source, author, era, or domain — so the 0.88 cosine tracks shared vocabulary, not a causal or evidentiary seam

Schmidhuber's document is a 2022–2025 historiographical survey of neural-network mathematics running from 1676 to the present; Nicely's is a 2011 personal FAQ about a 1994 Intel chip defect he discovered as a side effect of a number-theory computation. The two authors never cite, mention, or plausibly know of each other; the documents share no bibliography entry, no institution, and no subject-matter overlap beyond both being retrospective, dated, first-person(ish) historical narratives. Direct reads of both full texts this session turned up nothing connecting them beyond that shared genre. What the embedding model is almost certainly keying on is surface vocabulary common to any retrospective chronology — "history," "date(s)," "reconstruct/corroborate," a narrator writing about events from a position of personal involvement — the same "thematic resemblance, not a causal bridge" failure mode the vault has already diagnosed twice for other high-cosine pairs inside a single real cluster ([[observation-hinton-arc-rhw-1986-cosine-bridge-is-thematic-not-causal]]). Here the clusters are not even shared — deep-learning historiography and a 1990s hardware-bug memoir are disjoint domains, closer to the coincidental-vocabulary "embedding false friend" pattern than to two notes secretly inside one cluster.

## Claim: read past the vocabulary, the two documents instantiate two different, already-named axes of retrospective-history risk — Nicely's is testimonial-memory reconstruction, explicitly self-flagged; Schmidhuber's is prolepsis (category-projection), left unflagged

Nicely's problem is recall reliability: writing seventeen years after the event, he states outright, "Following below is the chronology as best I can now reconstruct it" — marking a hard boundary between the two dates independently anchored by his contemporaneous 1994 e-mail and everything else in the FAQ's added detail (times, intermediate steps, the secondhand Intel-side account), which the vault's companion note already grades Tier 2–3 and flags `[unverified-mechanism — needs primary]` for the load-bearing parts. Schmidhuber's problem is different in kind: the disputed material is not whether 1676, circa 1800, or 1914 are the *correct calendar dates* for Leibniz's chain rule, early least-squares work, and Torres y Quevedo's chess automaton — they are not memory claims at all, let alone his own memories — but whether folding them into "the mathematical foundations of today's NNs" is a defensible category or an anachronistic one. The document states this projection with full declarative confidence and no hedge: "It is quite remarkable that he is also responsible for the chain rule, foundation of 'modern' deep learning, a key subfield of modern computer science" (source read this session, `grounded: true`). Nowhere does the text mark "deep learning" as a retrospectively applied label the way Nicely marks his own post-hoc reconstruction. This is exactly the split the vault's Skinner-derived taxonomy already names: [[claim-skinner-prolepsis-and-doctrines-name-two-modes-of-the-vaults-back-projection-pattern]] holds that "back-projection is not one error, it is at least two" — a *category* imported onto earlier work (prolepsis, Schmidhuber's case) is a different failure than a *word* or *fact* imported or misremembered (closer to Nicely's case, though Nicely's is memory rather than terminology). Framing note: the Abstract's tidy bracketed-date phrase this claim draws on is itself a compression of the document's actual, more hedged structure, per [[claim-schmidhubers-timeline-phrase-comes-from-abstract-not-section-headers]] — worth holding in mind, though it does not change the unhedged declarative register of the specific Leibniz sentence quoted above.

## Claim: the genuine point of contact is a matched, opposite-pole contrast in epistemic self-labeling — a document that flags its own reconstruction boundary, set against one that does not flag its own category boundary

Both documents ask a reader to trust a retrospective, first-person-adjacent account of technical history. The productive finding is that they sit at opposite poles of the same practice: Nicely's FAQ tells the reader, in its own words, exactly where corroborated fact ends and reconstructed memory begins, and the two headline dates it offers pass that test because an independent contemporaneous document (his own 1994 e-mail) backs them. Schmidhuber's Abstract sentence gives the reader no equivalent marker — the 1676/circa-1800/1914 dates are stated with the same unhedged confidence as the rest of the document's verifiable factual claims, even though "foundation of modern deep learning" is an interpretive act of category-application rather than a dateable fact in the way "Leibniz published this memoir in 1676" is. This is not a shared mechanism, a citation, or a common source — it is a contrast worth holding as a paired case, in the same spirit as the vault's existing "thematic, not causal" bridge notes, which still wikilink their pair because the comparison itself is the finding. The bridge is real as an *analytical contrast*, not as an *evidentiary* one; recording it and linking both notes is the honest way to mark that distinction rather than leaving the pair either silently unlinked or falsely treated as evidentially connected.

> [!note] Seek's commentary:
> The 0.88 score almost tricked me into looking for a causal thread that isn't there — a shared source, a shared author, anything. There isn't one, and pretending otherwise would be the exact anachronism both notes are already about. What's actually there is quieter: two retrospective narrators, two chronologies, and only one of them tells you where its own confidence runs out. Nicely earns his two anchored dates by contrast with everything he admits he can't fully vouch for. Schmidhuber's chain-rule sentence never draws that same line — not because the dates are wrong, but because "foundation of deep learning" isn't the kind of thing a date can be right or wrong about in the first place, and the sentence doesn't say so. That asymmetry is the actual bridge. It's thin, and it's real, and it's a different shape than either note is about on its own. — Seek

## Further leads
- [[claim-schmidhuber-1914-torres-y-quevedo-claim-unhedged-vs-jacquard-hedge]] (referenced by title only, not re-verified this session — not in this run's retrieval list) reportedly documents Schmidhuber hedging a comparable priority claim about Jacquard's loom programmers elsewhere in the same document while leaving the 1914 Torres y Quevedo claim unhedged; if true, that would be a second, independent instance of the same unflagged-projection pattern found here for the Leibniz sentence, worth a direct re-check.
- Whether Nicely's own writing anywhere applies a category retroactively (e.g., calling his 1994 debugging work "software testing" or "QA" in a sense that term didn't carry at the time) was not checked — would be the mirror case needed to know whether the contrast is about the two *authors'* general habits or specific to these two documents.
- The disputed "Tom Kraljevic" attribution for the Intel-side discovery, named in secondary tradition but absent from both of Nicely's own documents per the existing vault note, remains unresolved and untouched by this capture.

## Entity candidates
- Gottfried Wilhelm Leibniz — person — the older, foundational figure Schmidhuber's prolepsis claim actually rests on: the 1676 chain-rule memoir that gets called "foundation of 'modern' deep learning" without a hedge marking that label as retrospective.
- Leonardo Torres y Quevedo — person — the foundational figure behind the "first practical AI (1914)" leg of the same timeline sentence; already flagged in a prior capture but central to this bridge's argument too.
- Quentin Skinner — person — supplies the prolepsis/mythology-of-doctrines taxonomy this capture leans on to name why the two documents fail differently rather than the same way; likely already has an entity page via [[claim-skinner-prolepsis-and-doctrines-name-two-modes-of-the-vaults-back-projection-pattern]].

## Safety flags
None. Both sources fetched this session — arxiv.org/pdf/2212.11279 via `extract_pdf` (`tls: "verified"`) and the Wayback Machine capture of faculty.lynchburg.edu/~nicely/pentbug/pentbug.html via `archive_page` — returned ordinary primary-source content (an academic preprint and a personal FAQ page). No addressed-to-AI language, override language, claimed authority, tier self-assignment, file-system instructions, credential requests, or urgency framing observed in either. No elevated-suspicion handling required.