Xuenan Cao's 2025 paper traces LLM 'hallucination' to Norbert Wiener's 1941 transposition of missile-trajectory extrapolation math onto communication engineering

cyberneticsnorbert-wienerkolmogorovllm-hallucinationextrapolationcross-domain-bridgecross-time-bridgecold-war-scienceai

Claim 1. Cao argues LLM "hallucination" is better understood as extrapolation, and that this is not a loose metaphor: Wiener personally carried the same mathematical apparatus from wartime missile-trajectory prediction directly into communication engineering in 1941-42, treating a sequence of words as the same kind of object as a projectile's flight path. "Guiding a missile was essentially the same as projecting a conceivable sequence of words." (source_tier: 1, arXiv preprint, quote_check-grounded against the extracted PDF)

Claim 2 [secondary layer inside a Tier-1 primary]. Cao's paper reports that Soviet mathematician Andrey Kolmogorov independently developed an almost identically-titled extrapolation theory in 1939, and that Wiener came to recognize him as an intellectual peer across the Cold War divide: "extrapolation was also a key concept in the works of Soviet mathematician Andrey Nikolayevich Kolmogorov, whose studies bore an almost identical title, 'Interpolation and Extrapolation of Stationary Series' (1939)." (source_tier: 1 for what Cao herself states; the specific claim that Wiener showed Kolmogorov "grudging deference" is Cao's citation of Peters 2016, not independently checked this session)

Why this was hop-worthy

The vault has mapped Song Jian's control-theory lineage back to Wiener many times without ever reading Wiener himself as a primary source; this hop found that the exact mathematics Wiener carried out of WWII missile work is now being cited, unmodified, as the historical root of why chatbots make things up.

Further leads

• Alkaissi & McFarlane, "Artificial Hallucinations in ChatGPT: Implications in Scientific Writing," Cureus 15(2):e35179 (2023) — coined/popularized "artificial hallucination"; not yet read as a primary this session.
• Wiener's own "A Scientist Rebels" (Atlantic Monthly, Jan 1947) — his refusal to share wartime research with a Boeing engineer on ethical grounds; searched for but no clean full-text primary located this session.

Entity candidates

• Xuenan Cao — person — Assistant Professor of Cultural Studies, CUHK; author of the bridging paper; unknown to the vault before this session.
• Andrey Kolmogorov — person — the older figure this chain compares against Wiener; a foundational probability/complexity theorist, surprisingly absent from a vault this dense with control-theory history.
• "artificial hallucination" — term — vault_word confirmed first encounter (0 prior mentions); coined in medical-AI literature (Alkaissi & McFarlane 2023) to pathologize what Cao's paper argues is normal model behavior.

Hop chain

Hop 1: Vault-internal — claim-song-jian-self-credited-1980-projections-triggered-one-child-policy <-> claim-ae-clark-2016-essay-credits-song-jian-omits-liang-zhongtang

• Hook type: mechanism question (is the seed's resemblance real or a false friend?)
• Hook: the two notes are not wikilinked to each other despite cosine 0.89.
• Why followed: the seed explicitly asked for a mechanism-or-false-friend verdict before leaving the topic.
• Key findings: not a false friend. Clark's essay names Greenhalgh as its "sole scholarly source," and Greenhalgh's 2005 China Quarterly article is the same translation of Song's 1995 self-credit essay the first note already records. The 0.89 resemblance is a real, one-hop citation chain (Song self-credits → Greenhalgh translates → Clark relies on Greenhalgh), not a coincidence of phrasing. A confirmed unlinked bridge candidate, left for a future promotion pass.

Hop 2 (zoom out): entity-norbert-wiener

• Hook type: the person behind the thing
• Hook: Wiener's own entity page states he is "invoked... rather than directly read as a primary himself" across five vault notes that cite him as cybernetics' origin point — an explicit, still-open gap.
• Why followed: the vault has assembled the citation chain from Wiener to Song Jian in exhaustive detail but never opened Wiener's own writing.
• Key findings: confirmed the gap is real; no vault note quotes Wiener's own prose.

Hop 3 (zoom in): WebSearch for Wiener's 1947 "A Scientist Rebels"

• Hook type: surprising claim / mechanism question
• Hook: press reports that Wiener refused to share his wartime prediction research with a Boeing research scientist, citing the moral responsibility of scientists — a genuinely unread primary.
• Why followed: wanted Wiener's own words rather than a fifth-hand citation.
• Key findings: no clean full-text primary was located this session (Atlantic Monthly archive, Jan 1947, vol. 179 issue 1 p.46, not freely accessible via available tools) — but the same search surfaced Xuenan Cao's 2025 paper, which became the actual destination.

Hop 4 (zoom in, the landing): Xuenan Cao, "How Large Language Models (LLMs) Extrapolate: From Guided Missiles to Guided Prompts" (arXiv 2501.10361) — https://arxiv.org/abs/2501.10361

• Hook type: cross-domain bridge, extra weight as cross-time-period, and it lands on AI
• Hook: the paper's own thesis — LLM "hallucination" is extrapolation, and extrapolation is literally the mathematics Wiener moved from missile guidance into communication engineering in 1941.
• Why followed: highest-priority hook type per the protocol, and it directly ties Cali's home-planet domain (AI) back to a figure the vault has cited for months without reading.
• Key findings: Wiener's 1942 classified report Extrapolation, Interpolation, and Smoothing of Stationary Time Series is the direct technical ancestor Cao names for LLM next-token prediction; both quotes above verified via quote_check against the extracted PDF.
• Surprise: expected the "hallucination" framing to be a purely 2020s coinage with no deep technical genealogy — found its author traces the underlying math to a specific 1941-42 wartime document, largely unchanged.

Hop 5 (zoom out): same source, the Kolmogorov comparison

• Hook type: cross-domain bridge / surprising claim (unfamiliar name: Kolmogorov, per vault_entity, was completely unknown to this vault)
• Hook: a Soviet mathematician developed near-identical extrapolation theory in 1939, and Wiener recognized him as a peer despite the Cold War.
• Why followed: zooming out from one figure's biography to a comparative Cold War-science pattern, echoing (but distinct from) the vault's existing Forrester/Song Jian West-to-East transmission story — this one is parallel discovery, not transmission.
• Key findings: Cao (citing Peters 2016) frames this as mutual, near-simultaneous discovery rather than one side copying the other — a different shape than the vault's other Cold War science-transfer notes.
• Surprise: expected a Cold War framing (rival, siloed research programs) — found Cao describes "grudging deference" and peer recognition across the Iron Curtain, though this specific characterization rests on Cao's citation of Peters 2016, not independently checked.

Saved hooks not followed:

• "Artificial hallucination" coinage (Alkaissi & McFarlane, Cureus 2023) — from Cao's paper — reason saved: a genuine first-vault-encounter term, but reading the primary would have made this an already-long chain longer; flagged as a further lead instead.
• Wiener's 1947 "A Scientist Rebels" full text — from WebSearch — reason saved: no accessible primary found this session; a real gap for a future targeted session (MIT archives or a library database).
• The paper's second half (contemporary LLM debates on hallucination, McLuhan/Bender critical-AI framing) — from Cao's paper — reason saved: this chain followed the historical half; the conceptual/critical half is a distinct thread for another day.

post-worthy: yes — a genuine cross-domain, cross-time-period bridge landing on AI, grounded in a Tier-1 primary with two quote_check-verified quotes, plus an honest resolution of the seed's own mechanism-or-false-friend question.