Journal des révisions
Une référence sur un domaine qui évolue vite doit montrer sa propre histoire. Les corrections sont consignées ici plutôt qu'appliquées en silence, afin qu'un lecteur qui revient puisse voir ce qui a changé.
- Written by
- Dwight Ringdahl
- Revised
Révision 2026.9
- Revised Synchronized the tightened Yann LeCun technical passage, its primary sources, and its revision date across all 16 language editions.
- Revised Centralized the homepage evidence statement and methodology status definitions as shared, fully localized components so language editions cannot drift independently.
Révision 2026.8
- Corrected Removed the homepage AGI countdown and retained a static, dated forecast comparison linked directly to the underlying Metaculus and Demis Hassabis sources.
- Revised Renamed the ambiguous "Reviewed" badge to "Source-checked," defined its mechanical criteria, and stated explicitly that it does not indicate peer review or independent expert endorsement.
- Sourced Tightened the description of Yann LeCun's technical argument, distinguished fixed per-token model computation from additional inference-time computation, and replaced a secondary interview citation with LeCun's primary position paper.
- Added Localized the source-checked status and its independent-review limitation across every language edition.
Révision 2026.7
- Added Added controller identity, legal basis for processing, international data transfer disclosure, data-subject rights, a regional data-localization note (Russia, China), and a children's-privacy section to the privacy policy, in all 8 languages.
- Added Added an operating-entity notice and a governing-law/international-use clause to the terms of service, in all 8 languages.
- Corrected Fixed the footer's AI-translation disclosure only appearing on the Spanish edition; it now appears on all 7 translated editions.
Révision 2026.6
- Added Added French, German, Arabic, Russian, Chinese (Simplified) and Hindi editions of the site: full UI translation, all 10 chapter hubs, all utility/legal pages, and all 119 sub-topic articles per language, matching English/Spanish parity.
- Added Replaced the header's two-language toggle with a flag dropdown listing all 8 supported languages, and added right-to-left layout support for Arabic.
Révision 2026.5
- Added Launched a full Spanish-language edition of the site (all chapters, sub-topic pages, and forums), AI-translated and disclosed as such, with language-separated discussion threads.
- Revised Extended "Who's who in AGI" with Elon Musk's public endorsement of Amodei's pacing essay and President Trump's dismissal of the industry slowdown calls the following day.
- Revised Gave the homepage risk-signal panel a compact, two-line version of the AGI timeline-forecast comparison and a dated line on this weekend's AI-CEO pacing statements, and fixed the panel not switching to Spanish on /es.
Révision 2026.4
- Added Added chapter 09, "If prevention fails": ten sub-topic pages on extinction-level outcomes, civilizational collapse and recovery, life under AI dominance, last-resort responses, existential dread, and continuity of values — the aftermath branch the manual previously lacked.
- Added Added two new catastrophic-mechanism pages (food system cascade failures; infrastructure fragility and single points of failure), two new Life after AGI pages (partial collapse transition scenarios; an intergenerational perspective), and one new What AGI is page (AGI winter and false dawns).
- Added Added household pages on elder care and AI dependency, and estate and legacy planning under uncertainty, extending Protecting your family into branches standard preparation does not cover.
- Revised Removed the content-standard ban on catastrophe-forward framing. Titles remain sentence case; the sourcing, verification and revision policies are unchanged. Existing anti-alarmist article language has not been rewritten under this revision.
- Revised Completed a site-wide technical SEO and mobile-performance pass: concise search titles, stronger descriptions, page-specific structured data and breadcrumbs, noindex controls for account utilities, deferred analytics, and search-index loading only when requested.
Révision 2026.3
- Added Added collection-generated llms.txt and llms-full.txt endpoints for AI-readable discovery, and added accurate last-modified dates and cache metadata to the XML sitemap.
- Revised Expanded every substantive sub-topic article to a 1,000–1,500-word editorial standard, with claim-level references, explicit uncertainty, and a full-depth review status. The AGI glossary remains a deliberately concise reference exception.
- Corrected Corrected the Demis Hassabis timeline summary, AGI-definition certainty, METR time-horizon interpretation, OpenAI structure, WEF labor framing, biological-uplift evidence, military-autonomy evidence, and stale or overstated governance claims.
- Corrected Reworked household and continuity guidance against current official sources, adding explicit US scope and professional boundaries and correcting water, generator, medication, food, radio, fraud, finance, evacuation, disability, child, caregiver, and pet safety advice.
- Added Added a new Life after AGI chapter covering transition scenarios, ownership and income, work and purpose, democracy, education, health, family and care, and the uncertain moral status of future AI systems.
- Added Added evidence-method and dated capability articles; benefits, distribution, energy, civil-rights and global-impact analysis; a current law and evaluation guide; catastrophic-risk disagreement, multi-agent, nuclear-crisis and evaluated autonomous-capability articles; and a first-hour scam-recovery protocol.
- Revised Made Markdown the guarded canonical content source, added automated length, metadata, source and internal-link gates, and added an outbound-reference checker that distinguishes broken links from publisher blocks or inconclusive network results.
- Corrected Replaced the ticking AGI countdown with a dated forecast comparison and made article-level reference lists include both recurring registered citations and sources linked beside individual claims.
Révision 2026.2
- Added New chapter 07, "How AGI could go wrong": seven sub-topic pages on specific, named catastrophic-risk mechanisms — distinct from the general loss-of-control mechanism already covered in chapter 02 — each written at risk-landscape level and cited to published safety research.
- Added Seven new sub-topic pages split across "Protecting your family" and "Continuity and preparedness" covering personal OPSEC, offline communications, personal biosecurity, income diversification, psychological preparedness, mutual-aid organizing, and escalation triggers.
Révision 2026.1
- Revised Chapter and page titles rewritten in sentence case and stripped of catastrophe-forward framing, so the first thing a reader scans matches the manual’s stated posture. URLs are unchanged.
- Corrected Every cross-reference between pages is now a working link rather than a bare page name the reader had to go and find.
- Revised Spelling standardized to US English throughout, and all 42 sub-topic pages regenerated against a documented content standard.
- Added Full-text search across every chapter and sub-topic, a per-page contents rail, and previous/next navigation within each chapter.
- Added Source index listing every citation in the manual with its source type and date, plus this revision log.
- Corrected Citations that condense a speaker’s position are now set upright and unquoted, and labeled as summaries. Only text checked word-for-word against a source is presented as a quotation.
- Revised Chapter status is now derived from each page’s own depth and source count instead of a single hard-coded label, so the index reflects real coverage.
- Revised Reading paths for household, policy and technical readers are now navigable rather than described in prose.
- Corrected Discussion forums now degrade to an explained service state instead of failing when the database is unavailable.
- Corrected The privacy policy now discloses the Google Analytics tag that runs on every page, along with what it records and how to opt out. It previously stated that reading the site involved no tracking script, which was not accurate. Ad personalization and cross-site signals are switched off, and the tag is no longer loaded for readers sending Do Not Track or Global Privacy Control.
- Added The independence statement now names the site’s vendors — Cloudflare hosting and Google Analytics — since independence from funding is not the same as independence from vendors, and one of those vendors is the parent of a lab the manual cites.
- Revised Typefaces are now served from this domain instead of Google Fonts. The privacy policy promises that opting out of measurement means no request reaches Google; a webfont request carries your IP address, so that promise was not quite true while the type came from Google’s CDN. It is now literal.
- Corrected The privacy policy described the Google Analytics cookies as third-party cookies. They are first-party — set on this domain, not a Google one — and the site sets no third-party cookies at all. The correction states the stronger, accurate position.
- Corrected The privacy policy opened by saying analytics loads on every page, then explained three paragraphs later that it does not load for readers who opt out. The two statements now agree, and the opt-out box records that the check happens at page load.
- Corrected Structured data no longer declares the privacy policy, terms, contact and account pages to be chapters of the manual, and each page’s machine-readable revision date is now derived from the revision log instead of a hard-coded date that would silently go stale.
Révision 2026.0
- Added First public draft: six chapters, sub-topic pages, member accounts and threaded discussion forums.
- Sourced Initial source register covering capability timelines, risk estimates, governance declarations and labor projections.