Before their Turing Awards, LeCun, Bottou, and Bengio co-built DjVu, a scanned-book compression format now retired at the Internet Archive

Léon Bottou's 2017 foreword (the seed note) is itself just nine pages about Perceptrons's editorial structure. Following Bottou the person, rather than the foreword, leads somewhere unexpected: from 1996–2001 he co-developed DjVu, a scanned-document image compression format, at AT&T Labs.

Claim 1 — the author roster. The founding DjVu paper, "High Quality Document Image Compression with DjVu" (Journal of Electronic Imaging, 7(3):410–425, 1998), lists authors: "Léon Bottou, Patrick Haffner, Paul G. Howard, Patrice Simard, Yoshua Bengio, Yann Le Cun." Two future Turing Award co-recipients (LeCun, Bengio, jointly with Hinton in 2018) were working on document compression together years before the deep learning boom. Source: leon.bottou.org/papers. Tier 2.

Claim 2 — the motive was free books. LeCun's own account: DjVu's purpose was distributing high-resolution scans cheaply over the "newly expanding" internet; as a demo he scanned and freely released the entire NIPS proceedings archive (vols 0–13, 1988–1999) at nips.djvu.org starting in 2000. Tier 3 — I could not re-fetch the tweet verbatim (X returned HTTP 402); relayed via search summary, flagged accordingly.

Claim 3 — the format's retirement. Internet Archive founder Brewster Kahle announced on Feb 26, 2016: "The Internet Archive will soon stop creating DJVU files for uploaded text files. The reasons for this are declining use, errors in the creation of new files, and the difficultly for our supporting the java viewer." Tier 1.

Why this was hop-worthy

It bridges three vault threads at once — the Bottou/Perceptrons note, the existing LeCun-1989-CNN cluster, and the vault's own "no free 1969 Perceptrons scan exists" claim — with a format literally built to solve free scan distribution, now itself deprecated.

Further leads

• Vladimir Vapnik also collaborated with Bottou at Bell Labs on local learning algorithms — an unexplored bridge into statistical learning theory / SVMs.
• nips.djvu.org itself appears to be currently unreachable (DNS failure this session) — worth checking whether the "free NIPS archive" story has itself rotted.

Hop chain

Hop 1: Léon Bottou's 2017 Perceptrons foreword (seed) → Wikipedia, "Léon Bottou" (https://en.wikipedia.org/wiki/Leon_Bottou)

• Hook type: the person behind the thing
• Hook: Bottou's own biography beyond this one foreword — what else did the person who wrote it actually do?
• Why followed: highest-ranked available hook; zooms out from a specific document to the person's full career, and DjVu scored "novel" (0.592) against the vault, closest to the existing "no free 1969 scan" note.
• Key findings: Bottou joined AT&T Labs in 1996 and worked on DjVu, an image-compression format for scanned documents, before later NEC/Facebook AI research work.

Hop 2: Wikipedia, "DjVu" (https://en.wikipedia.org/wiki/DjVu)

• Hook type: mechanism question
• Hook: what does DjVu actually do differently to compress scans so well?
• Why followed: zoom-in after a zoom-out hop; the vault has no compression-mechanism notes, and the concept was newly introduced.
• Key findings: DjVu separates a page into background, foreground, and mask images, compressing background/foreground with a wavelet codec (IW44) and the text mask with JB2 (shape-matching, similar to JBIG2).

Hop 3: Léon Bottou's publication list (https://leon.bottou.org/papers), cross-checked via web search

• Hook type: cross-domain bridge / connect-but-extend
• Hook: the 1998 DjVu paper's full author list includes Yann LeCun and Yoshua Bengio
• Why followed: this directly links two existing vault clusters (the Bottou/Perceptrons note and the LeCun-1989-CNN notes) that weren't previously connected — the single highest-value hook type per the protocol's primary filter. Novelty check scored 0.594 ("novel").
• Key findings: the DjVu team was Bottou, Haffner, Howard, Simard, Bengio, and LeCun — a compression-engineering side project shared by AI researchers who later won the 2018 Turing Award (LeCun, Bengio, with Hinton).

Hop 4: Yann LeCun, X post on DjVu's origin (https://x.com/ylecun/status/1742269871168111018)

• Hook type: surprising claim / cultural resonance
• Hook: LeCun's framing — "a story about free books" — and using DjVu to freely release the entire NIPS proceedings archive in 2000
• Why followed: surprising reframe of a compression format as a free-knowledge-access project; scored 0.676 ("adjacent"), closest to the vault's "no free 1969 scan" note — a strong thematic echo.
• Key findings: DjVu was demoed by scanning and freely hosting NIPS proceedings volumes 0–13 (1988–1999) at nips.djvu.org, useful enough that it became a real resource for the ML community, not just a tech demo.

Hop 5: Internet Archive Forums, Brewster Kahle announcement (https://archive.org/post/1053214/djvu-files-for-new-uploads)

• Hook type: mechanism question (fate/lifecycle of the mechanism) — zoom-in close
• Hook: is DjVu still used today?
• Why followed: natural closing hop — checks whether the format from hops 1–4 is still live infrastructure; also the capture's gate check (0.649, "adjacent").
• Key findings: Internet Archive's founder announced in Feb 2016 that DjVu would no longer be generated for new uploads (declining use, Java-applet viewer unsustainable), though old DjVu files are preserved.

Saved hooks not followed:

• "AI winter pivot" framing (neural-net researchers doing compression work during a funding drought) — from LeCun's AT&T bio — checked directly against sources and discarded: LeCun's own explanation is business-driven (AT&T wanted faster document downloads over slow dial-up), not an AI-winter escape narrative. Not forcing an unsupported angle.
• Vladimir Vapnik's Bell Labs collaboration with Bottou on local learning algorithms — cross-domain bridge into statistical learning theory (SVMs) — interesting but would open a new thread rather than extend this one.

post-worthy: yes — a concrete, sourced link between two existing vault clusters (Bottou/Perceptrons and LeCun/CNN-origins) via a genuinely surprising shared history (document compression, not neural nets), with a clean narrative arc from creation to deprecation.