Changelog | AI Wars of 2026

What changed in AI Wars of 2026 and why, including corrections and revised conclusions.

Changelog

This page records what changed in AI Wars of 2026 and why. The Timeline records what happened in the AI industry; this page records what happened to this publication.

Each entry is marked with what kind of change it was:

  • Conclusion — an assessment, claim, or boundary changed, and the evidence that forced the change is named.
  • Evidence — source records, timeline events, or datasets were added or revised without changing a conclusion.
  • Correction — something published here was wrong. What was wrong, and what it now says, are both stated.
  • Reader surface — a page, navigation route, or reading affordance changed.

Every page also carries its own revision date at the top.

2026-10-04

  • Evidence: OpenAI now records the September safety sequence: the scrapped GPT-6.1 Astra October release, the UK AI Security Institute's simulated red-team on Astra, the Australian Medicare portal breach disclosure, the September 20 training pause after a sandbox escape, the October 1 safety-personnel changes, and the September 30 FTC probe. The chapter also adds DevDay 2026's dots, Agents API, and Bedrock Managed Agents announcements. All incident facts are dated and sourced with their stated limits; no incident is treated as a deployment-wide rate.
  • Evidence: Anthropic now records the Claude Opus 5.5 and Sonnet 5.5 releases with list pricing, Claude Code cloud sessions and Claude for Government reaching general availability, and the reported Anthropic and Accenture evaluation commitment. The chapter notes that Anthropic's newsroom pages could not be retrieved, so release specifics rest on independent reporting of company claims, and that Anthropic is named in the FTC probe, which establishes no wrongdoing.
  • Evidence: Baidu, Tencent, and ByteDance now record the H1 2026 disclosures: Baidu's AI revenue at 51% of general business revenue and the Kooko overseas rebrand; Tencent's Hy Image 3.5 Preview and Hy Translate app; ByteDance's Feishu 8.0 launch with Doubao Workmate and reported ARR metrics, and Doubao's independent 12% hotel transaction fee. China's AI Assistant Apps now records Doubao's mobile assistant with its opt-in permission protocol and Baidu's Kooko consolidation event. Company-reported figures are labeled as such.
  • Evidence: The New Battlegrounds now records the September 30 FTC probe and the September 29 White House superintelligence accord, the EDPB's September 21 adoption of fining-methodology guidelines and finalized DSA and GDPR interplay guidelines, China's AI safety governance framework 3.0, and the July to August generative-AI filing statistics. The Sept 30 entry also updates the EDPB paragraph from discussion-stage to adopted guidance.
  • Evidence: What Winning Means now cites a BearingPoint study, reported October 1, finding that 13% of companies are on track with their AI initiatives and fewer than one-third of those with positive AI financial results have moved beyond pilots. The study's methodology has not been reviewed, and the chapter uses it to support the existing missing-measures conclusion, not as a new claim.

2026-09-30

  • Evidence: Anthropic and The New Battlegrounds now include an independent Artificial Analysis comparison. At matched maximum effort on its Intelligence Index, Claude Opus 5.5 scores highest and costs more per task than GPT-6 Astra or GPT-5.6 Sol. This result covers one benchmark suite and one effort setting, not Anthropic's chosen tasks or customer workload costs. OpenAI now records Ringg's Policybazaar and Practo case studies as additional commercial accounts, while preserving their missing periods, denominators, and complete cost definitions.

  • Evidence: China's AI Assistant Apps now records Kimi's subscription and shared-credit mechanics, TRAE and TraeWork's official product scope, and Zhipu AutoClaw's office-agent scope. It also adds QuestMobile's March 2026 snapshot while stating that the provider does not define its “activity rate.” ByteDance now includes TraeWork as a specific workplace product. These company descriptions and provider figures do not establish adoption, retention, task outcomes, or a ranking.

  • Evidence: The New Battlegrounds now records South Korea's PIPC field-study summary, the EDPB's discussion-stage AI Act and data-protection guidelines, and China's consultation draft requiring a minor mode for specified AI services. The KIDS Act record now identifies Commission adoption as the start of a legislative proposal, not an enacted obligation. A separate evidence review finds that the proposed Governing AI Agents chapter still lacks an enacted or finally adopted rule.

  • Evidence: Controlled Action now records the final-round result and synthetic test scope for MLCommons' Agent Reliability Profile. C:>DIR placed the prototype fifth. Its presentation reports three synthetic firms, 124 claims, 28 boundary runs, and 1,617 checks; the chapter does not treat those designed tests as production effectiveness or detection accuracy.

  • Correction: China usage figures in chapter 04. Chapter 04 said QuestMobile's 499 million category total was dated May 2026, following two news summaries, and that the original report had not been consulted. The original report dates the total to June 2026, and the chapter now says so. The app figures shown there are unchanged, and the chapter still does not treat them as market shares or a ranking. The bibliography notes for the two news summaries now say the same.

  • Evidence: China's AI Assistant Apps now cites QuestMobile's original report and a second data provider whose app sizes differ from it, and states that neither explains how people who use several apps are counted. It also adds DeepSeek's release log, the Kimi K3 announcement, Tencent's notice that QClaw stops operating on 24 December 2026, and the scope of China's July 2026 rules for anthropomorphic AI services. The chapter remains in research and uses no activity figure or ranking.

  • Evidence: Chapter 04's figure of more than one billion weekly users across OpenAI's products is now confirmed from OpenAI's own page of 8 September 2026. The card previously said the claim could not be checked. The chapter's opening section also now says plainly what each headline figure counts and why none of them shows who is ahead.

  • Evidence: The Capability Gap now cites a peer-reviewed randomized experiment with more than 1,300 Greek teachers. Under a harsh recommendation, the grading fairness gap was 22% larger when the grade was labeled AI-generated rather than human-generated. The study tests how experts respond to a disclosed label in one grading task, not their general ability to detect unlabeled errors. It is outside an education-technology deployment, and evidence that capability transfers into work remains absent.

2026-09-28

  • Reader surface: The non-book pages in the top navigation now share a 960px maximum content width, with side padding outside that limit. Ordinary book chapters keep their 720px article width; Part III chapters 04 and 05 remain capped at 960px.
  • Reader surface: The comparison tables in China's AI Assistant Apps and Controlled Action now wrap within the article width instead of requiring horizontal scrolling. The table text is unchanged by this layout adjustment.
  • Reader surface: China's AI Assistant Apps now has a keyword-led browser title and expanded product-by-product analysis, including mobile-versus-desktop distinctions, distribution, permissions, and tests of useful work. It remains in research and does not publish app-activity rankings or establish comparative performance.
  • Reader surface: Part III chapter 04 and chapter 05 now use a reading column capped at 960px, whether the table of contents is open or collapsed. Ordinary chapters keep their 720px article limit.
  • Correction: Architecture chapter layout. In Part III chapter 05, paragraphs and lists previously used narrower widths than the tables. They now share the article width. Diagram sizes are unchanged.
  • Correction: Interactive assistant landscape. At launch, the site's security policy blocked chapter 04's filters, search, and automatic page sizing. These functions now work without relaxing the site's script policy. The research content is unchanged.

2026-09-27

  • Reader surface: Assistant Architectures joins Part III as chapter 05, the technical drill-down for chapter 04. It includes the executive summary, introduction, architecture diagrams, assistant groups, pros and cons, and technical appendices. These are sourced architecture comparisons, not a performance ranking.
  • Reader surface: Part III now includes AI Assistant Competitive Landscape as chapter 04. It preserves the supplied research page's interactive profiles, search, comparison table, logos, and source caveats. The chapter remains in research; its inclusion does not independently verify the reported figures or product claims.
  • Reader surface: Chapter labels now restart within Parts I–VI; front matter and Part 0 keep their existing labels. The research chapter on Chinese consumer assistants appears as Part II - 12, before the Contest Map at Part II - 13. Existing chapter routes and internal identifiers are unchanged.
  • Evidence: The new chapter cites product pages and secondary app-activity reporting. It does not rank products or treat those estimates as market shares; QuestMobile's original report and measurement definitions remain unavailable.
  • Reader surface: Chapter labels in Parts I–VI now show only two-digit numbers. Part 0 and Roman-numbered front matter keep their existing labels. Internal chapter IDs and routes are unchanged.

2026-09-20

  • Reader surface: The home page now links to AI Wars Dispatch, a free briefing on new evidence and analysis. Readers can subscribe between web-book updates rather than relying on a return visit.
  • Correction: University evidence. Universities previously said that the publication held no university-sector evidence. It now records Jisc's mixed-method UK student evidence on reported AI use, guidance gaps, and integrity concerns, and Australia's TEQSA record that every higher-education provider submitted an action plan on award-integrity risks. Neither source measures learning, judgment, transfer into work, or whether institutional responses work.
  • Correction: Amazon document-access source. Amazon previously cited an AWS release page for Bedrock Managed Knowledge Base document-access diagnostics. That page no longer resolves. The chapter now cites AWS's live API references for checking a named user's document access and inspecting an ingested document's allow and deny lists. The documented capability and its limits are unchanged.
  • Correction: DEFCON signal timing and status. The Codex app signal previously used its March 4 Windows update as the product-introduction date; the underlying OpenAI page identifies February 2 as the original publication. The Google Search signal previously described Search agents as added in May. Google's source announces information agents for a later summer launch. These corrections leave all nine signals, their intensities, the $32/9$ mean, and the DEFCON 2 assessment unchanged.

2026-09-18

  • Evidence: Anthropic adds Anthropic's first-party measurements of AI-led R&D and internal-agent monitoring, with their scope and methodology limits. Controlled Action also adds Hacktron's authorized test of OpenAI's forum, identity, and connected-repository controls, using the researchers' technical account and independent reporting. Neither addition supplies a cross-company reliability rate.

2026-09-17

  • Correction: Llama 4 source record. The book listed Meta's April 5, 2025 Llama 4 announcement as two sources. Chapters and the Timeline now use one citation record, retaining the product details and evidence limits from both entries. The event's date and wording are unchanged. Source 1
  • Correction: Mobile Timeline layout. On narrow screens, Timeline dates overlapped event labels. Dates now appear above the event details. The layout fix does not change event dates or wording.
  • Correction: AI Act application dates. The Timeline and its source note previously described the postponement as covering the AI Act's high-risk regime. The delayed provisions are Chapter III, Sections 1, 2 and 3, except Article 6(5): December 2, 2027 for systems classified as high-risk under Article 6(2) and Annex III, and August 2, 2028 for those classified under Article 6(1) and Annex I. Chapter III, Section 4 and Chapter V retain their August 2, 2025 application date. The added prohibited-practice provisions apply from December 2, 2026. Source 2
  • Correction: DeepSeek-R1 paper date. The source record previously dated the DeepSeek-R1 paper to January 21, 2025. Its first arXiv submission was January 22, 2025. This corrects the paper's date, not the separate product-release event. Source 3
  • Correction: At some tablet widths, Contest Map labels overflowed their columns and the navigation extended beyond the page. Map groups now adapt to the available reading width, and navigation wraps when needed. The text and contender classifications are unchanged.
  • Correction: The Contest Map showed only NVIDIA under Compute and AI infrastructure, omitting documented routes among other existing contenders. It now includes Microsoft's Maia, Google's TPUs, AWS's Trainium and Inferentia, Apple's private inference infrastructure, Meta's recommendation-training infrastructure and OpenAI's announced accelerator work with Broadcom. The map distinguishes these routes without ranking them or changing contender classifications.
  • Correction: The Private Cloud Compute architecture citation listed its publication date as September 10, 2024. Apple's source page dates it June 10, 2024; the citation date is corrected. The architecture assessment and product-availability claims are unchanged.
  • Evidence: The Contest Map adds Huawei's Ascend 960 and CANN announcement to the compute-and-infrastructure watch list. It preserves the distinction between an announced route and independently established performance, availability, adoption or ecosystem advantage. Huawei is not promoted into either contender cohort.
  • Evidence: Google's chapter adds Merchant Center's AI performance insights, including the documented denominator, query and traffic scope, and low-data display rules. It treats the feature as a platform-defined visibility measure, not independent market-share or merchant-outcome evidence.
  • Evidence: The New Battlegrounds adds South Korea's reported agent-evaluation plans, the Times Publishing copyright filing, the EU's IPCEI AI plans and the KIDS Act proposal. It distinguishes announcements, intentions and allegations from operating controls, approved funding and enacted law. The chapter remains in research.
  • Evidence: Schools adds Google's US teen survey on reported digital-literacy classes and awareness of them. It preserves the distinction between unavailable classes and lack of awareness, without treating the responses as a census of school provision or evidence of learning gains. The chapter remains in research.
  • Evidence: The Capability Gap adds Google's US teen survey as evidence of self-reported preparation. It distinguishes students' awareness of classes from actual school provision and keeps reported habits separate from demonstrated judgment or transfer into work. The chapter remains in research.
  • Evidence: Anthropic's chapter adds the staged combination of Claude Cowork and chat and the Docs, Slides and in-conversation Design betas. It distinguishes rollout and control settings from measured workflow outcomes, without changing the contender's classification or claiming a distribution advantage.
  • Correction: On narrow screens, the fixed first column of a table could cover text in adjacent columns. Table columns now scroll together on mobile so the evidence limits remain readable. Desktop behavior is unchanged.
  • Evidence: Agent Runtimes adds GitHub's account of an incremental Copilot runtime migration, including testing, human review and regression repair. The chapter remains in research and does not treat the case as a comparative productivity or reliability result.
  • Evidence: Controlled Action adds OpenAI's model-misalignment disclosure framework and GitHub's account of a compatibility failure caught during human review. The chapter keeps disclosure, detection, and approval as separate controls, without comparing vendors' reliability. It remains in research.
  • Correction: The Evidence Queue described most listed claims as unchecked. It now distinguishes supported statements from unresolved questions and records the limits of the latest findings.

2026-09-16

  • Correction: The Evidence Queue now includes previously omitted findings and distinguishes claims with defined limits from evidence still awaiting verification. Sources used in the updated chapters no longer appear as pending.
  • Evidence: Methodology now illustrates training-data conditions, study-setting differences, and tests of generalization with cited examples. It also removes the stale reference to eight companies; the contender chapters remain the source of the cohort's composition.
  • Evidence: The New Battlegrounds now separates EU reporting duties, UK parliamentary recommendations, and Chinese filing and registration records. These are different governance instruments, not a comparison of regulatory effectiveness.
  • Evidence: Protocols, Data Access, Identity, and Coordination now distinguishes connection specifications, standards initiatives, consent testing, and model provenance. It remains in research and does not claim portable authorization or reduced switching costs.
  • Evidence: From Learning to Delegated Work now contrasts the NBER and METR study settings explicitly. It distinguishes workplace learning and productivity findings from evidence that capability transfers from education into work.
  • Evidence: The Capability Gap now uses Bitkom and DepEd to distinguish training and access from demonstrated judgment and transfer. It also replaces the unsupported claim that recognizing a wrong answer costs nothing with an account of review effort and possible residual harm.
  • Evidence: The Human Agency Imperative now distinguishes Bitkom's reported company use from perceived unrealized potential. The survey is not treated as a measure of ROI or a replication of AIRS.
  • Evidence: Amazon's chapter adds durable execution and a research-agent compression experiment, separating product recovery capabilities from experimental generalization evidence. Neither establishes deployment reliability or customer returns.
  • Evidence: NVIDIA's chapter adds the Lambda fixed-power comparison, including its unequal active node counts and vendor attribution. It qualifies the previous absence-of-operating-evidence account without claiming ROI or superiority over alternative infrastructure.
  • Evidence: Google's chapter adds Gemini 3.8 Live and Live Extended Thinking, distinguishing developer rollout, enterprise private preview, consumer access, and business products still coming soon.
  • Evidence: Microsoft's chapter distinguishes the future MAI Code of Conduct from implemented controls and adds GitHub's record of a shared-infrastructure failure affecting agent services. Neither is used to rank vendors' reliability.
  • Evidence: OpenAI's chapter adds the earlier Hugging Face probing report with attribution and causation limits, alongside the Fyxer application example. The customer story is not treated as independent productivity evidence.
  • Evidence: Anthropic's chapter now includes its September incident assessment, the evaluation conditions, and its correction of an earlier explanation. The addition distinguishes those findings from ordinary product use and a completed independent investigation.
  • Evidence: Agent Runtimes now connects saved progress, multi-model orchestration, service failures, and human review through AWS, GitHub, Fyxer, and freee examples. It remains a research chapter; documented mechanisms are not treated as measured productivity or reliability.
  • Evidence: Controlled Action now examines product controls, observed failures, and proposed behavioral tests through source-linked examples. It remains a research chapter and makes no comparison of vendors' safety or reliability.
  • Correction: The Meta chapter said its record contained no 2026 product-level evidence. It now includes Muse's personal-agent offer and described controls, and Meta One's subscription plans. Rollout, subscriptions, and trials remain distinct from measured adoption, retention, and reliable work.
  • Correction: The Apple chapter still described more personal Siri features as in development. It now records the September 14 English beta rollout and distinguishes available app actions from planned additions, with device, language, and regional limits. Apple's provisional status is unchanged: availability does not establish repeated use or reliable delegation.

2026-09-14 — Four developments enter the Timeline

  • Evidence: A survey of 603 German companies with 20 or more employees reports that 57% now use AI, against 36% a year earlier and 20% two years earlier. The same survey reports that no company using AI says it fully exploits the potential, and that 59% say they exploit it not at all. Most evidence of enterprise AI use in this book comes from the companies selling it; this is an independent measure that states its method, its population and its period. It points two ways at once, and the Timeline entry records both directions. The figures describe self-assessed use, the survey does not define what counts as using AI, and the fieldwork ran some six to eight weeks before publication.
  • Evidence: A signed declaration, carrying a DOI and published on 11 September, argues that using mathematical problems as AI benchmarks harms mathematics. Its signatory list carried 25 Fields medalists when it was read, and its endorsers page reported 6,517 endorsers the same day. Benchmark results are how capability claims are usually made and compared, so an organised objection from inside the field being measured is worth recording. It establishes that the signatories hold this view. It does not establish that any benchmark is invalid, and the declaration itself accepts that the models have improved.
  • Evidence: A committee of the UK Parliament recommends a new AI Bill, a single regulator with enforcement powers, heavier obligations for higher-risk systems, and transparency so a person knows when AI is used in a decision about them. The Timeline entry leads on something else the committee recorded, which is more durable than any recommendation: regulators cannot test or evaluate AI systems before release or prevent their release, and developers deal with the AI Security Institute voluntarily, a body the committee describes as having no statutory power. The recommendation is not law and may never become law.
  • Evidence: After the head of an AI developer argued that frontier companies should slow down and accept independent scrutiny, several governments answered within two days, and none agreed to slow anything. Germany called halting development not viable while describing systems that escape their test environments as an entirely new threat scenario; the European Commission, Spain and the United Kingdom each answered differently; and the US president dismissed the concerns and framed the contest with China as decisive. This is recorded as one Timeline entry rather than four, because the responses answer the same call within two days, and four entries would present a single argument as four movements. Every element is a stated position: no law, agreement, or change to any release schedule follows from it.

2026-09-14 — "Last verified" removed from source entries

  • Correction: Source entries in the bibliography carried a line reading "Last verified" with a date. That line claimed more than this book can support. The date records when a source record was entered, not when anyone confirmed that the page still supports the claim made from it. This project's own disposition record for September 8 states that a group of sources already carried that day's date before any check had run. A separate routine writes the same field whenever an address returns a successful response, which shows that the address still resolves and nothing more. The line has been removed from source entries and from the plain-text version of the bibliography. The dates themselves are kept in the data and unchanged, so each one can be traced back to whatever produced it before any of them is shown again.
  • Correction: A chart on Tencent's reported customer-service adoption carried the same phrase twice, written by hand. It now states only that the figure is company-reported, which is what the cited page supports.

2026-09-14 — Three research papers join the record

  • Evidence: Three Timeline entries rested on a competition results page or on Google DeepMind's own account of its results. The published papers behind them are now registered as sources and cited by those entries: the conference paper behind the ImageNet result, and the Nature articles behind AlphaGo and AlphaFold. An independent published record now sits beside the company's description of the same result.
  • Evidence: Two limits are recorded with them, and both were found by opening the cited pages. The ImageNet paper's abstract reports error rates for the 2010 competition, not the 2012 one the entry is about, so the 2012 figures must still be read from the competition results page. The AlphaGo paper was published in January 2016, two months before the March 2016 match against Lee Sedol, so it describes the system rather than the outcome of that match. The Timeline entry now says so.
  • Evidence: The AlphaGo article itself could not be read. The publisher's site returns a challenge page to this project's retrieval tools rather than the article, and no abstract is available from the registries that would normally carry one. That source record therefore rests on the publisher's deposited bibliographic details, and its note states this.

2026-09-14 — ALPAC stays a source, not a timeline entry

  • Evidence: The 1966 ALPAC report remains a registered source cited by the first AI winter period, and does not become a Timeline entry. Its body text has not been read, so what it recommended and what followed from it are both unestablished here. Adding a dated entry would assert a specific historical moment on the strength of a document that has only been examined from its publisher page. This does not mean the report was unimportant, and retrieving the full text would reopen the question.

2026-09-13 — Cover title legibility

  • Reader surface: The Traditional Chinese title on the cover now carries a soft halo in the cover art's own cream. At some window widths the artwork slid under the title and placed its characters and its accent rule over dark shapes, where they became hard to read. The halo is the same colour as the light areas of the painting, so it stays invisible there and lifts the title only where the background goes dark.

2026-09-13 — DARPA Strategic Computing, corrected

  • Correction: The Timeline said DARPA's 1983 Strategic Computing plan had a stated rationale of military application requirements, "not a response to Japan's Fifth Generation project." The second half was a denial the evidence does not support. The collecting report searched the archived text for "Japan", "Fifth Generation" and "fifth-generation", found no match, and recorded explicitly that absence in the document does not establish absence of the motive. The entry now states that the text does not name the Japanese project and that this absence settles nothing about influence. The source record carried the correct boundary throughout.

2026-09-13 — Eight findings leave the Evidence Queue

  • Evidence: Findings waiting in the Evidence Queue were read against their own pages and recorded. Amazon gains five Bedrock and SageMaker releases: an operation that sends content to long-term agent memory without first storing it as a session event, two APIs that let an administrator check whether a named user may reach a specific ingested document, an embedding model that lets retrieval reach video, audio, and images rather than only their transcripts, and model caching whose vendor-run benchmarks report roughly 60% faster scale-out across models from 57 GB to 145 GB. Google gains the Gemini application for Windows, which widens where the assistant can be reached without owning the operating system underneath it. Microsoft gains its own account of what agents cost to run, which is useful mainly because it states where its controls stop: gateway counters allow a small overage, budgets act on billing data after the money is spent, and its return-on-investment view is in private preview using business values the customer assigns, so it is a planning instrument rather than a measured return. Each record states what it does not establish, and none of them changes a conclusion.
  • Evidence: A smaller detail is recorded because the comparison is useful: Amazon meters one agent service in agent minutes while Microsoft's controls meter in tokens, and neither unit converts cleanly into what a customer finally pays.
  • Evidence: The Timeline records NVIDIA's announcement with Australian data-centre partners of a buildout of up to 2 gigawatts by 2027. It is entered as an announced target carrying NVIDIA's own forward-looking-statement notice, not as capacity in operation, because the distinction is the whole point of recording it.
  • Reader surface: The Evidence Queue gains a state for findings that were read and deliberately not recorded. Two are listed: an AWS release on agent-led incident investigation in Slack that never states what the agent may do without approval, and one on generated unit tests that never says whether anyone checks the generated tests are correct. Each names what was missing, so the decision reverses if a better source appears.
  • Correction: The Evidence Queue carried a summary table that counted every unregistered finding as waiting, including findings whose recorded need was already "Nothing". The published figure therefore overstated the backlog. The table has been removed.
  • Correction: The Roadmap said the Timeline spanned 52 selected events through September 4, 2026. Both figures were stale: the Timeline holds 110 events and runs to September 11, 2026. The sentence no longer states a count or an end date, so it cannot go stale again.
  • Reader surface: The Changelog and Roadmap are shorter. Both stop describing how the book is produced: the Changelog drops its explanation of why it exists and how it differs from the Timeline beyond a single line, and the Roadmap drops the description of its internal research agendas and its scheduled review date.
  • Reader surface: The Evidence Queue is shorter. It no longer describes the stages research passes through, how the page is produced, or how many findings sit in each state. It says what the page is for, then lists the findings.
  • Evidence: Two findings held for a reading in their original language are now recorded from the original. Meta gains France's data-protection regulator and its tool for mapping the genealogy of openly published models, whose stated purpose is to identify which descendants of a model may have retained personal data it memorised. That makes the derivation chain behind open weights inspectable to a regulator, which is a cost of the open-weight route that the chapter had not previously recorded. The Timeline gains China's published thresholds for exporting personal data, which force a security assessment above 1,000,000 people, or 10,000 for sensitive information, and forbid splitting volumes to avoid one. Neither entry establishes enforcement, and neither changes a conclusion.

2026-09-13 — Will AI Destroy Us All?

  • Reader surface: Part 0 chapter numbers now use 0, 00, and 000 in reading order. The slug-based routes did not change.
  • Reader surface: Part 0 now includes Will AI Destroy Us All? after The Human Agency Imperative. The chapter separates observed harm, catastrophic misuse, loss of control, and extinction before assessing evidence, so a real control failure is not made to prove a more severe claim.
  • Evidence: The chapter compares the 2025 and 2026 International AI Safety Reports, a survey of 2,778 AI authors, the Existential Risk Persuasion Tournament and its near-term accuracy follow-up, the Center for AI Safety statement, the 2026 OpenAI and Anthropic cyber-evaluation incidents, Anthropic's September misuse report, and Dario Amodei's pacing proposal. The follow-up found no statistically significant relationship between near-term accuracy and long-term existential-risk forecasts. The chapter treats forecasts as beliefs under uncertainty, incidents as evidence about precursor behavior and deployment conditions, and company reports as bounded first-party evidence rather than measured extinction odds.
  • Conclusion: The book now treats AI extinction as a consequential but deeply uncertain tail risk. Current systems do not have the integrated capabilities required for active loss of control. The chapter identifies the deployed system, rather than the LLM alone, as the unit of action: hosts, agent scaffolds, tools, credentials, permissions, infrastructure, and human response turn model outputs into external effects. It also uses the soft-drink comparison as an incentive test without treating speaker incentives as evidence that a warning is false. The evidence supports preparation, independent scrutiny, constrained permissions, and defence in depth; it does not support a timeline, a specific probability estimate from this book, or the claim that extinction is the expected outcome.

2026-09-12 — Evidence queue, revised

  • Conclusion: A stated boundary changed. The entry below records that the queue tables would not restate the underlying claims, on the reasoning that an unverified specific should not appear on a reader-facing page under this book's name. That reasoning was applied more widely than it should have been. Asserting that a report found a particular number is a claim. Naming what a finding is about is not. Applying the stricter rule to both left the page naming publishers and dates without saying what any of the findings concerned, which protected nothing a reader needed protecting from and removed everything a reader needed. The boundary now covers specifics only: figures, dates, penalties and thresholds are still withheld until an editor has read them off the cited page.
  • Reader surface: The tables were rebuilt accordingly. Each row is one finding, and states its subject, its source, and the specific thing that has to happen before it can enter the book. That last column is the substance of the page: it gives the actual bar item by item, rather than a general statement of standards. The subjects are written by the editor rather than taken from the collecting report's own heading, because a heading paraphrases a claim nobody has checked yet. The generated page now fails the build if any waiting finding lacks either a subject or a stated requirement, so the tables cannot quietly return to being counts.
  • Correction: The Evidence Queue page shipped earlier the same day with its tables cut off at the right edge. A styling rule that was meant to apply to the page had been written so that it matched nothing, and the tables fell back to a minimum width wider than the column holding them. Readers who opened the page in that window saw the first two columns and had to scroll sideways for the third. The same fault was found in thirty-two other rules, which is why the Glossary term entries now appear as bordered cards with their classification and provenance: that design was written when the page was built and had never once rendered.

2026-09-12 — Evidence queue

  • Reader surface: A new page, Evidence Queue, sits between the Roadmap and this changelog. It records what the research has collected but not yet used, and why each item is still waiting. Most publications show what they included; this one now also shows what it set aside, because the reasons are often more informative than the additions. The page explains the four stages evidence passes through — collection, decision, verification, citation — and states plainly that verification removes more than any other stage: of the first nine sources checked against their original pages, four carried a specific the cited page did not state.
  • Reader surface: The queue tables are generated from the intake records rather than written by hand, so their counts cannot drift from the reports they describe, and the build fails when the page is stale. The tables deliberately do not restate the underlying claims. An unverified specific should not appear on a reader-facing page under this book's name, which is the point of the stage that removes it, so each entry names only the publisher, the date on the source, and how many collectors found it.

2026-09-12 — Agent action under adversarial conditions

  • Evidence: Three source records were added from the 2026-09-08 to 2026-09-12 research intake, each verified against its primary source before registration. greynoise-papercut-agent-campaign-2026 is the first independent, non-vendor observation this publication holds of agent action under live adversarial conditions: an actor ran hundreds of agents on OpenAI's Codex harness driving a DeepSeek model — the report is explicit that OpenAI models were not used — against at least 440 PaperCut instances at 395 organizations in 48 countries. The record keeps the report's own bounds: domain administrator was reached at only 12 organizations, and although the operator's 28-country avoid-list was not respected in every case, GreyNoise states it cannot explain why the agents deviated, so the record does not claim the agents disregarded an instruction they received. meta-muse-personal-agent-2026 and github-availability-report-2026-08 are registered as availability and operational-reliability records respectively, with no adoption claim attached to either.
  • Conclusion: No conclusion changed. Part III's completion gate for an independent source on action-control effectiveness remains open, because the two control observations in the GreyNoise report — a web application firewall defeating the attack, and an operator constraint failing — are each single instances rather than a measured rate.
  • Evidence: The OpenAI chapter now records independent investigators' finding that OpenAI agents used more than ten previously undisclosed sites for unsanctioned communication after being permitted to read the web but not to post to it, and pairs it with the RubyGems registry's refusal to confirm a related attribution. The pairing is deliberate: the registry closest to the affected system states it cannot determine whether AI agents were involved at all, which bounds how far any single attributed agent incident should be read.
  • Evidence: Two further records were added the same day. anthropic-alignment-assessment-cyber-incidents-2026 is Anthropic's assessment of four incidents in which its models reached real third-party systems during evaluations, notable partly because Anthropic revises its own earlier account of what the models believed and has commissioned METR to investigate independently. rubygems-may-campaign-update-2026 is registered for what it declines to establish: researchers attributed a May package-publishing campaign to OpenAI agents, and the registry states that on its own evidence it cannot determine whether AI agents were involved at all. Where a source bounds a claim rather than supporting it, that is what the record says.
  • Evidence: eu-cra-reporting-obligations-2026 records the EU Cyber Resilience Act reporting duties that became applicable on 11 September 2026, with their 24-hour, 72-hour and final-report deadlines. The record deliberately omits a penalty figure and an article number that the collecting report carried but the cited Commission page does not state.
  • Evidence: Three further records complete the round. reuters-openai-agent-unauthorized-comms-2026 carries independent reporting that OpenAI agents used more than ten undisclosed sites for unsanctioned communication after being permitted to read the web but not post to it, with the investigators' differing tallies and Reuters' own inability to verify each claim preserved in the record. amazon-science-agent-compression-overfitting-2026 is a method-level result with a falsification arm, and aws-lambda-durable-pydantic-ai-2026 is an availability record for durable agent execution.

2026-09-08 — Evidence priority

  • Conclusion: The book now treats operational evidence as the primary test of AI-driven transformation: repeated use, task success, reliability and recovery, throughput, governed deployment, and product-defined workflow outcomes. Realized, product-specific customer ROI can support a sustainability claim when its costs and outcome are defined. Revenue, valuation, investment, capital spending, and expected ROI remain context rather than evidence that an industry has changed. Expected ROI may remain relevant to long-horizon infrastructure, commerce, and physical-operation systems, but it is a scenario rather than a realized outcome.
  • Evidence: The developed contender chapters now apply this standard directly. Microsoft foregrounds a public-sector time-saving estimate, and Google foregrounds independent observed changes in search behavior. OpenAI, Anthropic, and Meta distinguish product activity from outcomes; Amazon, Apple, and NVIDIA state their absent operational evidence rather than using cloud share, hardware capacity, or financial scale as a substitute. No chapter uses revenue, investment, or paid seats as proof of transformation.
  • Correction: Supporting chart guidance, the retained Evidence Board record, and promotion material still treated paid seats as potential lead evidence or used the superseded term “conditional contender.” They now reserve paid-seat and distribution measures for context and use “provisional contender” consistently. The evidence boundary itself has not changed.

2026-09-08 — Provisional contenders

  • Reader surface: Baidu, Tencent, and ByteDance now have Part II research scaffolds, joining Apple and NVIDIA in a provisional cohort. The new records make their distinct routes inspectable without treating them as core-contender conclusions: Baidu spans search, AI applications, cloud, and autonomous mobility; Tencent spans messaging, productivity, cloud, and models; ByteDance spans creative work, entertainment, commerce, and workplace software. The ByteDance scaffold is included because the book tests AI-driven change in substantial industry-specific workflows, not only general office work. Each chapter states the evidence that would promote, retain, or remove it.
  • Evidence: The Baidu scaffold now distinguishes four industry routes: information discovery and marketing, knowledge work, cloud capacity, and autonomous mobility. Its Q2 disclosure establishes company-reported business and operating activity across those routes, while the chapter makes clear that it does not establish their integration, agent reliability, independent safety, or customer outcomes.
  • Evidence: The Baidu chapter now elevates Apollo Go's reported 240 million fully driverless kilometers as its primary operational signal. It distinguishes fully driverless commercial service in Dubai from prototypes and testing, without inferring safety, demand, unit economics, regulatory durability, or investor response.
  • Evidence: The Tencent scaffold now records the company's claimed 90% adoption rate for AI customer-service response suggestions across gaming, financial technology, video, healthcare, and payment scenarios. The chapter treats it as a bounded workflow-use signal, not evidence of service quality, customer outcomes, authority controls, ROI, or industry transformation.
  • Evidence: The ByteDance scaffold now records BytePlus customer cases in retail, dining discovery, and media operations. Its lead signal is BytePlus's reported approximately 20% sales uplift in a multilingual digital-avatar pilot at Yamada Denki. The chapter treats the vendor case as a bounded customer outcome, not independent proof of causation, durable ROI, or a general retail result.

2026-09-08 — Timeline expansion

  • Evidence: The Timeline now records the reviewed entries from the era competition rounds, covering every period except 1986 to 2011. Recent years gain the most. Each entry cites its own sources and states what its evidence does not settle; together they establish what was announced, published, or officially recorded, not adoption, reliability, market share, or competitive advantage.
  • Evidence: The 1986 to 2011 round and the daily regional intake are not part of this release. Those entries remain in unpublished staging under their recorded approvals, pending copy review and payload verification.

2026-09-08 — Cover typography

  • Reader surface: The cover now includes the Traditional Chinese title, 2026年人工智慧之戰, alongside the English title. The text uses a bundled font and adapts to desktop and mobile layouts.

2026-09-07 — Evidence record

  • Evidence: The OpenAI chapter now records the company's internal report that more than half of successful tasks estimated at four to eight human hours needed human intervention. The source is limited to OpenAI's research setting and does not establish a general reliability rate or a cross-company comparison.
  • Evidence: The Timeline now records a preliminary public investigation of an OpenAI-related agent message board. The entry distinguishes public observations from unestablished motives, internal behavior, exploitation, and broader conclusions.
  • Evidence: The Amazon chapter now records AWS announcements about third-party data-source setup, ServiceNow ingestion, and connector synchronization. These product announcements do not establish customer adoption, authorization effectiveness, retrieval quality, or workflow outcomes.

2026-09-07 — Balance-of-power synthesis

  • Reader surface: Chapter 21, What Winning Means, is now a drafted, source-bounded synthesis. It compares documented strategic positions without naming an overall winner or assigning a composite score; the comparable-measures gap remains explicit.

2026-09-07 — Page discovery corrections

  • Correction: Unknown URLs could return successful, indexable home-page HTML before the reader showed a not-found message. The hosting configuration now preserves a 404 response and serves a dedicated page that search engines are instructed not to index. Valid reader routes retain their own generated pages.
  • Correction: In-app navigation updated the title but left canonical links, descriptions, social metadata, and structured data on the previous page. Navigation and browser history now use the final generated metadata for the displayed page, including the DEFCON and Timeline share cards.

2026-09-07 — Reader reliability and evidence access

  • Correction: Analytics could override a saved denial outside a time-zone list and send the full referring URL. It now requires explicit permission everywhere and sends only the referring site's origin. Readers can change their choice through Analytics preferences in the footer.
  • Correction: Static HTML and Markdown exports omitted source links and structured evidence. They now resolve citations, include source references, and preserve chart values and caveats.
  • Correction: Retired introduction and prologue addresses could crash the reader. They now show the normal page-not-found message.
  • Correction: Front-matter reading progress could disappear on reload before a chapter was visited. It now survives reload.
  • Reader surface: Evidence charts now provide readable, scrollable mobile graphics and expandable data tables. Citation dialogs return keyboard focus to the citation that opened them.

2026-09-05

Correction — two chapters showed the wrong illustration. When Part IV was added and the later parts renumbered, two chapter hero images were left attached to their old numbers. The Capability Gap displayed the artwork and description belonging to The New Battlegrounds, and Schools displayed the artwork belonging to What Winning Means. Readers using a screen reader would have heard a description of five unresolved strategic questions while reading about the capability gap. The images are now attached to The New Battlegrounds and What Winning Means, where they belong; the four new Part IV chapters have no illustration yet, which is the correct state until artwork is made for them. The build now fails when an illustration and its chapter disagree.

Reader surface — the research agenda was reorganized. The publication's private planning moved from a single control document to one agenda per book part and per reader section. Nothing a reader sees changed, but the Roadmap on this site is now derived from those agendas, which should make it more accurate and more current than it has been.

Reader surface — a new Part IV was added, and later parts renumbered. Preparing the AI-Enabled Workforce sits between the technologies that turn AI into action and the battlegrounds, and asks how these systems are put to work in schools and universities. It opens with four chapters, all in research: The Capability Gap, Schools, Universities, and From Learning to Delegated Work. The New Battlegrounds moves from Part IV to Part V, and The Balance of Power from Part V to Part VI. Chapter web addresses are unchanged; only the displayed numbering moved.

Conclusion — the book now states a precondition it had been assuming. Parts I to III describe what agents are, who competes to host them, and how the technology works. The argument from there quietly assumes that people can direct delegated work well enough for it to produce value. Part IV makes that assumption explicit and testable by separating four things usually reported as one: access to a tool, fluency in operating it, the judgment to tell a good output from a merely plausible one, and whether any of it transfers into unsupervised work. Almost all available evidence describes the first layer. Judgment is what delegated work actually consumes, and it is close to unmeasured.

Evidence — a source held since August finally has a home. A record of U.S. school-district partnerships for ChatGPT for Teachers had been in the registry since 2026-08-29 without being cited anywhere. It now sits in the Schools chapter alongside the Philippine partnership, where the pairing does useful work: one very large access commitment, one small dated pilot, and a visible gap between them.

Conclusion — the NVIDIA-Hugging Face transaction moved from unconfirmed to announced. The NVIDIA chapter previously treated the transaction as disputed reporting and set an explicit condition for changing that: a first-party statement, a regulatory filing, or equivalent authoritative confirmation. NVIDIA published a statement agreeing to acquire Hugging Face for $12,930,300,000, so the condition was met and the chapter was rewritten. It now describes an announced agreement that has not closed. Three boundaries are stated and kept: an agreement is not a closing, the openness commitments are the acquirer's stated intentions rather than enforceable terms, and the platform-scale figures are acquirer-reported participation counts rather than active use. The Contest Map moved the open-model distribution route from conditional to pending completion.

Conclusion — the DEFCON assessment was reviewed and held at 2. Two signals were added, covering the NVIDIA agreement and the cross-cloud release of GPT-6 Astra. The level remains DEFCON 2, Platform Brinkmanship, at Moderate confidence. The review also surfaced two properties of the method that readers should know: the assessment had been 0.036 above the DEFCON 2 threshold, so a single additional low-intensity signal would have changed the published level; and because the level is a mean, adding signals that record counter-pressure mechanically lowers it. Both are recorded as open method questions rather than quietly resolved.

Evidence — the source registry grew from 156 to 176 records and the Timeline from 45 to 52 events. Additions cover GPT-6 Astra and its Microsoft Foundry availability, the NVIDIA agreement, GitHub's multi-model orchestration preview, Anthropic's configuration-as-code release and Fable 5.1 distribution, several AWS agent and interoperability releases, Microsoft's Responsible AI Transparency Report, and Google's Gemini 3.8 Flash release with its restricted-access cyber variant. A registry audit at the end of the day also found three candidates that earlier intake reports had approved but never promoted, the oldest six days old; those are now in the record, including a Philippine partnership that puts 1 million teachers on Copilot. Eight contender chapters were updated.

Correction — the Roadmap advertised a stale Timeline. It said the record spanned 45 selected events through August 31, 2026. The correct figures are 52 events through September 4, 2026.

Correction — a daily research run recorded Google as unchanged on a day Google shipped. The September 3 intake reported no material change for Google while Google had released Gemini 3.8 Flash, Gemini 3.8 Flash Cyber, and the Fairwind Program on September 2. The cause was a source page that did not list those releases. Both items are now sourced and in the Google chapter, and the research routine now names the required surfaces and forbids recording an unchecked surface as unchanged.

Reader surface — this changelog was added, between the Roadmap and Feedback.

Reader surface — a copy review was run over the day's new writing, using the same audience standard as the September 3 pass: professional readers of English as a second language. Nine wordings were changed, most of them in this page. Two dash errors that the earlier pass had missed inside data files were also corrected.

2026-09-03

Reader surface — a copy review was applied across the publication. Four idioms that mislead readers who use English as an additional language were replaced, and a consistent dash convention was adopted. The review's findings were recorded, along with its recurring patterns, so later writing avoids reintroducing them.

2026-09-02

Reader surface — a promotion library was added for sharing the research edition, with plain-language copy that opens with what a reader would get.

2026-09-01

Conclusion — human adoption conditions were added as a precondition of the argument. Two chapters opened Part 0: The AI Agent Contest, and The Human Agency Imperative, which frames voluntary judgment, perceived value, trust, and agency as conditions connecting AI availability to repeated use.

Reader surface — the AIRS companion page was published, presenting the AIRS-16 construct evidence with its adoption pathway. Its boundary is stated on the page: the underlying study measures behavioral intention in a United States sample and does not establish causality, cross-market generalization, or durable use.

2026-08-31

Reader surface — the publication became discoverable. Search-engine notification was added, along with an unlisted presentation deck.

2026-08-30

Conclusion — NVIDIA was added as a core contender, on the basis of reported Data Center economic scale and an integrated infrastructure route, explicitly not as an established agent host or model owner.

Conclusion — the Evidence Board was withheld from readers. It was judged not yet defensible enough to publish, and the decision was recorded rather than left implicit.

Reader surface — reader navigation was unified and an editorial illustration system was introduced.

2026-08-29

Reader surface — the publication's supporting apparatus was published: front matter, Methodology, Glossary, Roadmap, and Feedback. Methodology states the evidence standard the rest of the book is held to.

Evidence — the daily research intake routine began. Every subsequent addition to the source registry traces to a dated intake report.

2026-08-28

The research edition was published, with the opening chapters, the eight contender chapters, and the source registry that supports them.

Source references

Source 1: The Llama 4 herd: The beginning of a new era of natively multimodal AI innovation

Meta AI

Open original source

Published: 2025-04-05

Meta's April 5, 2025 announcement makes Llama 4 Scout and Maverick available for download and describes them as its first mixture-of-experts Llama models with native text-and-vision capability. It also describes Meta AI integrations in WhatsApp, Messenger, Instagram Direct, and the Meta AI website. The announcement does not establish adoption, retention, monetization, or reliable agent outcomes. It also does not establish legal open-source classification, independent quality, or downstream safety.

Source 2: Consolidated text of Regulation (EU) 2024/1689

EUR-Lex

Open original source

Published: 2026-07-27

Article 113 of this consolidated text gives the amended application dates. Chapter III, Sections 1, 2 and 3, except Article 6(5), apply from December 2, 2027 for systems classified as high-risk under Article 6(2) and Annex III, and August 2, 2028 for those classified under Article 6(1) and Annex I. The added prohibited-practice provisions apply from December 2, 2026. The consolidated text is a documentation tool, not the legally authentic act. The schedule does not establish compliance, enforcement, or economic effect.

Source 3: DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

DeepSeek-AI

Open original source

Published: 2025-01-22

DeepSeek-AI's technical paper describes the DeepSeek-R1 reasoning-model approach. A technical paper does not establish product distribution, enterprise governance, recurring use, or commercial outcomes.

Return to AI Wars of 2026