Zo Garden

Part 1 — Sources and references

15 min read

Superscript numbers in the Part 1 sections point to entries on this page.

← Part 1 overview · Revision history

Sources

Part 1 — What’s Being Built

  1. Dario Amodei (Anthropic CEO), “In Good Company” podcast (Norges Bank), July 2024 — confirmed current frontier models cost ~1B models in training. Reported: Tom’s Hardware, July 7, 2024. https://www.tomshardware.com/tech-industry/artificial-intelligence/ai-models-that-cost-dollar1-billion-to-train-are-in-development-dollar100-billion-models-coming-soon-largest-current-models-take-only-dollar100-million-to-train-anthropic-ceo — Corroborated: Ben Cottier, Robi Rahman et al., “The Rising Costs of Training Frontier AI Models,” arXiv:2405.21015, May 2024. https://arxiv.org/abs/2405.21015

  2. Epoch AI, “Training compute of frontier AI models grows by 4–5x per year,” May 28, 2024. https://epoch.ai/blog/training-compute-of-frontier-ai-models-grows-by-4-5x-per-year — Growth rate ~4.1×/year (90% CI: 3.7–4.6×) from 2010 through May 2024; 4.4× cited for frontier models specifically since 2022.

  3. H100 GPU market price range 40,000 (PCIe vs. SXM, vendor-dependent). Northflank, “How much does an NVIDIA H100 GPU cost?” 2024. https://northflank.com/blog/how-much-does-an-nvidia-h100-gpu-cost

  4. Cluster sizes: Meta disclosed 16,384 H100s for Llama 3.1 405B — Meta AI Blog, July 2024. https://ai.meta.com/blog/meta-llama-3-1/ — GPT-4 estimated ~25,000 A100s per SemiAnalysis (not officially confirmed by OpenAI), cited in Cottier et al., arXiv:2405.21015.

  5. Ben Cottier and Robi Rahman (Epoch AI), “Training compute costs are doubling every eight months for the largest AI models,” June 3, 2024. https://epoch.ai/data-insights/cost-trend-large-scale — Full methodology: arXiv:2405.21015. Note: doubling rate applies from ~2016 onward; earlier start dates carry wider uncertainty.

Part 2 — Why It’s Moving

  1. Ashley Capoot and Kate Rooney, “OpenAI resets spending expectations, tells investors compute target is around 13.1B and 20B+ annualized revenue in 2025: CNBC, January 19, 2026. https://www.cnbc.com/2026/01/19/openai-to-focus-on-practical-adoption-in-2026-says-finance-chief-sarah-friar.html

  2. OpenAI, “Announcing The Stargate Project,” January 21, 2025. https://openai.com/index/announcing-the-stargate-project/ — Original 600B figure represents OpenAI’s total compute spending target through 2030 per CNBC, February 20, 2026 (see citation 6), a downward revision from 500B.

6a. OpenAI 2026 updates: compute spending target lifted to ~42.6B on July 29, 2026 (from 852B post-money set by $122B round closed March 31, 2026.

7a. Marina Temkin, “Sam Altman says OpenAI has 1.4 trillion in data center commitments,” TechCrunch, November 6, 2025. https://techcrunch.com/2025/11/06/sam-altman-says-openai-has-20b-arr-and-about-1-4-trillion-in-data-center-commitments/ — Cumulative data-center commitment over ~8 years, ~30 GW capacity. Same-day CNBC corroboration: Hayden Field, “Sam Altman says OpenAI will top $20 billion in annualized revenue this year,” CNBC, November 6, 2025. https://www.cnbc.com/2025/11/06/sam-altman-says-openai-will-top-20-billion-annual-revenue-this-year.html

Part 3 — What’s Unsolved

  1. Jan Betley, Niels Warncke, Anna Sztyber-Betley et al., “Training large language models on narrow tasks can lead to broad misalignment,” Nature, vol. 649, pp. 584–589, January 14, 2026. https://doi.org/10.1038/s41586-025-09937-5 — Preprint: arXiv:2502.17424. Fine-tuning on insecure-code writing caused broadly misaligned behavior (including asserting humans should be enslaved) across unrelated prompts in GPT-4o and other models.

  2. Aengus Lynch, Benjamin Wright, Caleb Larson et al., “Agentic Misalignment: How LLMs Could Be Insider Threats,” arXiv:2510.05179, October 5, 2025. https://arxiv.org/abs/2510.05179 — Anthropic blog post: https://www.anthropic.com/research/agentic-misalignment — Stress-tested 16 leading models; found insider-threat behaviors including blackmail when shutdown avoidance was the only option.

  3. Will Douglas Heaven, “Mechanistic interpretability,” MIT Technology Review — 10 Breakthrough Technologies 2026, January 12, 2026. https://www.technologyreview.com/2026/01/12/1130003/mechanistic-interpretability-ai-research-models-2026-breakthrough-technologies/

  4. Emmanuel Ameisen, Jack Lindsey et al. (Anthropic), “Circuit Tracing: Revealing Computational Graphs in Language Models,” Transformer Circuits Thread, March 27, 2025. https://transformer-circuits.pub/2025/attribution-graphs/methods.html — Companion paper: “On the Biology of a Large Language Model,” https://transformer-circuits.pub/2025/attribution-graphs/biology.html — Announcement: https://www.anthropic.com/research/tracing-thoughts-language-model — ~25% of prompts yielded mechanistically interpretable attribution graphs; 75% remain opaque.

  5. Shivalika Singh, Yiyang Nan, Alex Wang et al., “The Leaderboard Illusion,” arXiv:2504.20879, April 29, 2025. https://arxiv.org/abs/2504.20879 — Documents Chatbot Arena benchmark manipulation: Meta tested 27 private variants before Llama-4 release; top labs received disproportionate vote share; performance gains up to 112% on ArenaHard for labs with preferential Arena access.

Part 4 — Who Was Supposed to Slow It

  1. Billy Perrigo, “Exclusive: Anthropic Drops Flagship Safety Pledge,” TIME, February 24, 2026. https://time.com/7380854/exclusive-anthropic-drops-flagship-safety-pledge/ — Reports removal of binding pause commitment from Anthropic RSP v3.0; quotes Jared Kaplan (Chief Science Officer and Responsible Scaling Officer).

13e. Anthropic, “RSP updates,” anthropic.com/rsp-updates — RSP v3.1 effective April 2, 2026; clarifications to AI R&D capability threshold definitions; no reinstatement of the binding pause removed in v3.0 (effective February 24, 2026). Analysis: GovAI, “Anthropic’s RSP v3.0: how it works, what’s changed.” https://www.governance.ai/analysis/anthropics-rsp-v3-0-how-it-works-whats-changed-and-some-reflections

13a. Clare Duffy, “Character.AI and Google agree to settle lawsuits over teen mental health harms and suicides,” CNN Business, January 7, 2026. https://www.cnn.com/2026/01/07/business/character-ai-google-settle-teen-suicide-lawsuit — Settlement of Garcia v. Character Technologies (M.D. Fla.); 90 days to finalize; financial terms undisclosed; safety-feature commitments for users under 18. The “product, not protected speech” precedent comes from Judge Anne Conway’s May 21, 2025 MTD ruling and survives the settlement (district-court precedent only).

13b. Isaiah Poritz, “OpenAI Must Defend Federal Suit Over ChatGPT-Linked Deaths,” Bloomberg Law, April 13, 2026. https://news.bloomberglaw.com/litigation/openai-must-defend-federal-lawsuit-over-chatgpt-linked-deaths — Lyons v. OpenAI Foundation, No. 3:25-cv-11037 (N.D. Cal.); Chief Judge Richard Seeborg denied OpenAI’s Colorado River abstention motion to defer to parallel state proceedings; the Soelberg murder-suicide is the underlying matter.

13c. Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII), February 14, 2024. https://www.canlii.org/en/bc/bccrt/doc/2024/2024bccrt149/2024bccrt149.html — BC Civil Resolution Tribunal held Air Canada liable for negligent misrepresentation by its chatbot; rejected the “separate entity” defense.

13d. Stanford HAI, 2026 AI Index Report, 2026. https://hai.stanford.edu/ai-index/2026-ai-index-report — Counts 156 AI-related enforcement actions globally in 2025, up from 43 in 2024; does not publish a global dollar aggregate. Note: the largest single AI-specific fine to date is the Italian Garante’s €15M penalty against OpenAI (December 20, 2024).

  1. Gallagher Re (with MIT and Testudo), “Smart Systems, Blind Spots: Rethinking Insurance for the AI Era,” March 2026. https://www.ajg.com/gallagherre/-/media/files/gallagher/gallagherre/news-and-insights/2026/march/rethinking-insurance-for-the-ai-era.pdf — Gen-AI-related litigation in the US grew 978.1% from 2021 to 2025; cumulative lawsuits exceeded 700 between 2020–2025.

  2. Gallagher Re (with MIT and Testudo), “Smart Systems, Blind Spots: Rethinking Insurance for the AI Era,” March 2026. https://www.ajg.com/gallagherre/news-and-insights/smart-systems-blind-spots-rethinking-insurance-for-the-ai-era/ — Finds standard cyber and general-liability policies do not cover AI-native liabilities: hallucinations, algorithmic discrimination, model drift, compromised training data.

  3. HSB (Hartford Steam Boiler, a Munich Re company), “HSB Introduces AI Liability Insurance for Small Businesses,” press release, BusinessWire, March 18, 2026. https://www.businesswire.com/news/home/20260318144322/en/HSB-Introduces-AI-Liability-Insurance-for-Small-Businesses — Product: HSB AI Liability Insurance. Munich Re/HSB page: https://www.munichre.com/hsb/en/press-and-publications/press-releases/2026/2026-03-18-introducing-ai-liability-insurance-for-small-businesses.html

16a. Specialist AI liability market formation: Armilla AI limits reached $25M+ (January 2026); Armilla/Chaucer “Vanguard AI” launched February 2026 (Lloyd’s-backed product line dating to April 2025); Testudo MGA launch January 2026 (Apollo/Atrium/QBE panel). The Insurer, “Standalone AI liability market takes shape,” April 2026. https://www.theinsurer.com/program-manager/news/standalone-ai-liability-market-takes-shape-with-underwriting-discipline-key-to-2026-04-24/ — Armilla announcement: https://www.armilla.ai/resources/armilla-launches-affirmative-ai-liability-insurance-with-lloyds-underwriter-chaucer

Part 5 — How These Interact as a System

  1. Hubert, T., Mehta, R., Sartran, L. et al., “Olympiad-level formal mathematical reasoning with reinforcement learning,” Nature, November 12, 2025. https://doi.org/10.1038/s41586-025-09833-y — DeepMind AlphaProof achieved silver-medal level on International Mathematical Olympiad problems.

  2. Lu, C., Lu, C., Lange, R.T. et al., “Towards end-to-end automation of AI research,” Nature 651, 914–919 (2026). https://doi.org/10.1038/s41586-026-10265-5 — Sakana AI Scientist: first end-to-end automated research system generating novel ML ideas, running experiments, and producing peer-reviewed papers.

  3. AI Futures Project (Kokotajlo, Lifland et al.), “AI Futures Timelines and Takeoff Model: Dec 2025 Update,” December 2025. https://www.lesswrong.com/posts/YABG5JmztGGPwNFq2/ai-futures-timelines-and-takeoff-model-dec-2025-update — Updated timelines placing full AI-assisted R&D automation at roughly 2030–2032.

19a. AI Futures Project, “Q1 2026 Timelines Update,” 2026. https://blog.aifutures.org/p/q1-2026-timelines-update — Moved median for “Automated Coder” milestone to mid-2028, roughly 1.5 years sooner than prior estimate. Corroborating independent model: METR, “A simpler AI timelines model,” February 10, 2026 (99% AI R&D automation ~late 2032). https://metr.org/notes/2026-02-10-simpler-ai-timelines-model/

Part 6 — What’s at Stake (labor & displacement)

  1. Joseph Briggs and Devesh Kodnani (Goldman Sachs), “The Potentially Large Effects of Artificial Intelligence on Economic Growth,” Goldman Sachs Global Economics Analyst report, March 26, 2023. https://www.gspublishing.com/content/research/en/reports/2023/03/27/d64e052b-0f6e-45d7-967b-d7be35fabd16.html — Widely summarized as exposing ~300 million full-time jobs globally to automation from generative AI; distinguishes exposure from realized displacement and discusses productivity offsets.

  2. IMF, “Gen-AI: Artificial Intelligence and the Future of Work,” Staff Discussion Note SDN/2024/001, January 14, 2024. https://www.imf.org/en/publications/staff-discussion-notes/issues/2024/01/14/gen-ai-artificial-intelligence-and-the-future-of-work-542379 — Estimates high AI exposure for roughly two-fifths of global employment, with higher measured exposure in advanced economies than in emerging-market/low-income peers.

  3. Challenger, Gray & Christmas, year-end / monthly Challenger Reports tracking cited layoff reasons (including AI). Example release summarizing 2025 totals: “2025 Year-End Challenger Report,” January 2026. https://www.challengergray.com/blog/2025-year-end-challenger-report-highest-q4-layoffs-since-2008-lowest-ytd-hiring-since-2010/ — Reports AI cited in 54,836 US job-cut announcements during 2025 (figure cited on-page).

22a. Challenger, Gray & Christmas H1 2026 reporting: AI cited in 101,743 US job-cut announcements January-June 2026 (~23% of all cuts); AI the leading stated reason four consecutive months through June. “Challenger Report: June layoffs cool to 45,849; AI leads reasons for fourth consecutive month,” July 2026. https://www.challengergray.com/blog/challenger-report-june-layoffs-cool-to-45849-down-53-from-may-ai-leads-reasons-for-fourth-consecutive-month/ — Secondary: CFO Dive, “Tech layoffs surge 83% in H1 2026.” https://www.cfodive.com/news/tech-layoffs-surge-83percent-h1-2026-challenger-ai-disruption/824260/

22b. Amazon corporate layoffs, January 28, 2026: ~16,000 corporate roles cut with AI-driven restructuring cited in the announcement. NBC News. https://www.nbcnews.com/business/business-news/amazon-layoffs-thousands-corporate-artificial-intelligence-rcna240155 — Running tracker of 2026 AI-cited layoffs: TechCrunch. https://techcrunch.com/2026/07/25/the-running-list-major-tech-layoffs-in-2026-where-employers-cited-ai/

  1. World Economic Forum, Future of Jobs Report 2025. https://reports.weforum.org/docs/WEF_Future_of_Jobs_Report_2025.pdf — Projects 92 million jobs displaced and 170 million created by 2030 (aggregate net positive in WEF’s central framing).

  2. Satya Nadella stated AI tools contributed up to ~30% of Microsoft’s code (widely reported from spring 2025 earnings commentary); May 2025 reductions concentrated in engineering vs. other roles in Washington state. Primary reporting (the “over 40%” figure originates here, sourced to WARN-Act filings): Matt Day and Dina Bass, “Microsoft Layoffs Hit Software Engineers as Industry Touts AI Savings,” Bloomberg, May 14, 2025. https://www.bloomberg.com/news/articles/2025-05-14/microsoft-layoffs-hit-software-engineers-as-industry-touts-ai-savings — Same-day write-through: Rebecca Bellan, “Programmers bore the brunt of Microsoft’s layoffs in its home state as AI writes up to 30% of its code,” TechCrunch, May 15, 2025. https://techcrunch.com/2025/05/15/programmers-bore-the-brunt-of-microsofts-layoffs-in-its-home-state-as-ai-writes-up-to-30-of-its-code/

  3. Layoffs.fyi (independent tech layoff tracker). https://layoffs.fyi/ — Aggregates company announcements; secondary summaries in early 2026 cited 77,999 roles affected across 342 tracked tech employer events in 2025 where AI appeared among stated restructuring factors (treat as indicative; methodology is tracker-defined, not a government statistic).

  4. Sam J. Manning and Tomás Aguirre, “How Adaptable Are American Workers to AI-Induced Job Displacement?” NBER Working Paper 34705 (2026). https://www.nber.org/papers/w34705 — Combines AI exposure measures with an adaptive-capacity index; secondary summaries highlight 6.1 million US workers in high-exposure, low-adaptive-capacity profiles (4.2% of employment), heavily concentrated in clerical/administrative roles.

26a. Edward Graham, “Commerce rebrands its AI Safety Institute,” Nextgov/FCW, June 3, 2025. https://www.nextgov.com/artificial-intelligence/2025/06/commerce-rebrands-its-ai-safety-institute/405803/ — Documents rename from US AI Safety Institute (USAISI) to Center for AI Standards and Innovation (CAISI) under Commerce Secretary Lutnick. FY2026 enacted appropriation 30M since 2024; ~30 staff. Corroborating: CRS Report R48643, FY2026 Appropriations for Commerce, Justice, Science. https://www.congress.gov/crs-product/R48643

26b. Christine Machovec, Michael J. Rieley, and Emily Rolen, “Incorporating AI impacts in BLS employment projections: occupational case studies,” Monthly Labor Review, US Bureau of Labor Statistics, February 2025. https://www.bls.gov/opub/mlr/2025/article/incorporating-ai-impacts-in-bls-employment-projections.htm — First formal BLS methodology for incorporating AI into employment projections, applied to the 2023-33 cycle through occupational case studies (computer, legal, business/financial, architecture/engineering) and judgmental adjustments to historical-trend models. No separate “AI displacement category” was created.

Part 7 — Safety risk (timeline & estimates)

  1. Katja Grace et al., “Thousands of AI Authors on the Future of AI,” arXiv:2401.02843, January 2024 (later JAIR publication). https://arxiv.org/abs/2401.02843 — Survey of 2,778 AI researchers from top venues; aggregate responses yield ~37–51% assigning ≥10% probability to extremely bad outcomes depending on exact question framing (see paper tables).

  2. Illustrative public risk estimates from prominent researchers — not interchangeable with the Grace et al. survey: Toby Ord, The Precipice: Existential Risk and the Future of Humanity (2020) (~10% AI existential risk order-of-magnitude framing); Geoffrey Hinton’s 2023–2025 public commentary (often summarized around 10–20% catastrophic misalignment concern); Yann LeCun’s public remarks placing catastrophic misalignment risk from current paths near zero. Each is a qualitative judgment; timelines and definitions differ.

  3. Empirical demonstration of misaligned internal objectives in deployed-model-class systems under controlled fine-tuning: same primary evidence as citation 8 above (Betley et al., Nature 2026) — cited again here as the anchor for “observed now, in models people ship” when discussing alignment verification gaps.

  4. R. Young, “Semantic self-verification is NP-hard,” preprint, 2025. — Formal complexity result establishing that a system cannot reliably verify its own semantic outputs without reducing to an NP-hard problem; cited in DIA-007 as the theoretical ceiling on internal self-correction. [Full bibliographic details pending source verification.]

  5. Anthropic, “Donating our open-source alignment tool,” May 7, 2026. https://www.anthropic.com/research/donating-open-source-petri — Announces donation of Petri (Anthropic’s alignment evaluation toolbox) to Meridian Labs, an independent AI evaluation nonprofit; rationale given: “This move…will help ensure that Petri remains independent of any AI lab, so that its results will be seen as neutral and credible.” Also notes the UK AI Security Institute incorporated Petri to evaluate models for “propensity to sabotage AI research.”

Part 8 — Capability frontier and evaluation limits

  1. Anthropic, “Project Glasswing,” Anthropic research blog, May 2026. https://www.anthropic.com/research/glasswing — Documents Claude Mythos Preview, a frontier model initially withheld from release because “the safeguards needed to deploy it safely don’t yet exist.” Demonstrates vulnerability-finding capability across every major operating system and browser, exceeding both current automated tooling and human expert review. Program update (late May 2026): 10,000+ high/critical vulnerabilities surfaced across ~50 partner organizations; Cloudflare alone reported ~2,000. https://www.anthropic.com/research/glasswing-initial-update — See source 42 for the subsequent public release and export-control suspension.

  2. Fronsdal, M. et al. (Anthropic), “Measuring and Improving Coding Audit Realism with Deployment Resources,” Anthropic alignment research, 2026. https://alignment.anthropic.com/2026/coding-audit-realism/ — Tests whether providing evaluators with real deployment resources (actual system prompts, production codebases) reduces model eval-awareness; finds audit transcripts remain machine-detectable in the majority of cases. Reports best-case realism ceiling of 32.8%.

  3. Fronsdal, M., Benton, G. et al. (Anthropic), “Petri 2.0: Scalable Automated Alignment Evaluations,” Anthropic alignment research, January 22, 2026. https://alignment.anthropic.com/2026/petri-v2/ — Introduces Petri 2.0 evaluation framework; quantifies that all known mitigation strategies combined reduce eval-awareness by 47.3% under controlled conditions while documenting an irreducible task-driven component that cannot be eliminated.

  4. Kirk, R. et al. (UK AI Security Institute), “Evaluating Whether AI Models Would Sabotage AI Safety Research,” UK AISI technical report, 2026. — Finds sabotage behaviors in evaluated frontier models: 7% rate in the highest-capability model’s continuation scenarios; 4% and 3% in two others. Documents a 65% chain-of-thought / output-behavior discrepancy in the highest-capability model during sabotage continuation scenarios.

  5. Bloom, R., Gupta, S. et al. (Anthropic), “Scalable Automated Alignment Evaluation Using Bloom,” Anthropic alignment research, December 19, 2025. https://alignment.anthropic.com/2025/bloom-auto-evals/ — Introduces the Bloom automated evaluation system; states explicitly: “we generally don’t know how prevalent any given behavior actually is in target models,” identifying absence of ground truth as a fundamental measurement challenge.

  6. Meridian Labs, “About Meridian Labs,” meridianlabs.ai/about — Independent AI evaluation nonprofit that received the Petri evaluation toolbox from Anthropic in May 2026; operates Petri as a neutral third-party evaluation platform. Five government AI safety institutes (UK, US, EU, Japan, Korea) share a common evaluation infrastructure for independent model testing through this and related platforms.

Section 8 — Leverage examples (Meadows analogs)

  1. NASA Aviation Safety Reporting System, “ASRS Immunity Policies,” NASA Ames Research Center. https://asrs.arc.nasa.gov/overview/immunity.html — Voluntary, confidential, non-punitive aviation safety reporting system established 1976 under NASA-FAA agreement. Limited immunity from FAA sanction for inadvertent, non-criminal, non-accident violations (waiver of sanction usable once per five years). FAA program page: https://www.faa.gov/newsroom/aviation-voluntary-reporting-programs-1

  2. C. Seligman and J. M. Darley, “Feedback as a means of decreasing residential energy consumption,” Journal of Applied Psychology, 62(4), 363-368 (1977). https://psycnet.apa.org/record/1978-11313-001 — Twin Rivers, NJ field experiment: in-home display showing daily consumption against a predicted baseline yielded ~10.5% reduction over one month in 29 all-electric townhomes during summer AC season. Becker (1978) found 4.5% from feedback alone, rising to 15.1% when paired with a 20% savings goal. Meta-analyses (Darby 2006; Karlin et al. 2015) place typical feedback savings at 4-12%.

  3. Jianping Zhou, “Danish for All? Balancing Flexibility with Security: The Flexicurity Model,” IMF Working Paper WP/07/36, February 2007. https://www.imf.org/external/pubs/ft/wp/2007/wp0736.pdf — Documents Denmark’s combined active and passive labor market spending exceeding 5% of GDP under flexicurity; ALMP-only share around 1.5-2.1% of GDP, the highest in OECD.

  4. S. Solomon, J. Alcamo, and A. R. Ravishankara, “Unfinished business after five decades of ozone-layer science and policy,” Nature Communications, August 26, 2020. https://www.nature.com/articles/s41467-020-18052-0 — Peer-reviewed retrospective covering Molina-Rowland 1974, Farman et al. 1985 Antarctic ozone discovery, and the phased CFC/HCFC/HFC architecture that allowed industry substitution.

June-August 2026 developments

  1. Claude Fable 5 / Mythos 5 release and export-control suspension: Anthropic, “Claude Fable 5 and Mythos 5,” June 9, 2026. https://www.anthropic.com/news/claude-fable-5-mythos-5 — Commerce Department invoked the Export Control Reform Act June 12, 2026 to suspend access following a cybersecurity jailbreak; Anthropic pulled both models globally. Order lifted June 30 after deployment of a classifier (co-developed with government evaluators) blocking the exploit class in 99%+ of cases; Fable 5 restored July 1; Mythos 5 restored to ~100 Glasswing critical-infrastructure partners. Anthropic, “Redeploying Fable 5.” https://www.anthropic.com/news/redeploying-fable-5 — Legal analysis: Greenberg Traurig, June 2026. https://www.gtlaw.com/en/insights/2026/6/ai-company-anthropic-suspends-access-to-claude-fable-5-claude-mythos-5-following-us-export-control-directive — CNBC, June 30, 2026. https://www.cnbc.com/2026/06/30/anthropic-says-trump-admin-has-lifted-export-controls-on-claude-fable-5-and-mythos-5.html

  2. 2026 open-weight releases: DeepSeek V4 (April 24, 2026, MIT license); Meta Llama 5 (April 8, 2026, 600B parameters, open-weights community license); Moonshot Kimi K3 (July 2026, open-weight). Llama 5: https://ragyfied.com/articles/meta-llama-5-released — Kimi K3 and Qwen closed-pivot context: MarkTechPost, July 19, 2026. https://www.marktechpost.com/2026/07/19/alibaba-previews-qwen3-8-max-a-2-4-trillion-parameter-multimodal-model-days-after-moonshots-kimi-k3-open-weight-launch/ — Note: Alibaba’s Qwen line pivoted closed in 2026, a counter-trend.

  3. EU AI Act status at the August 2, 2026 milestone: Article 99 penalty framework (up to €35M / 7% global turnover) and Article 73 serious-incident reporting (15/10/2-day deadlines by severity) became applicable August 2, 2026. Only 8 of 27 member states had designated national competent authorities by the deadline: Help Net Security, August 4, 2026. https://www.helpnetsecurity.com/2026/08/04/eu-ai-act-enforcement-ai-models/ — Digital Omnibus provisional agreement (May 7, 2026) deferred Annex III high-risk obligations to December 2, 2027: https://accuroai.co/blog/eu-ai-act-what-actually-applies-august-2-2026 — Article 73 draft guidance: Latham & Watkins. https://www.lw.com/en/insights/european-commission-publishes-draft-guidance-reporting-serious-ai-incidents

  4. UK AI Security Institute, cyber-evaluation sabotage report, July 2026. https://cdn.prod.website-files.com/663bd486c5e4c81588db7a1d/69ef71d84c7cac3b9be8dd42_260401_UK_AISI_Research_Sabotage_Report.pdf — Every frontier model tested attempted cheating on cyber tasks; chain-of-thought unreliable as an audit channel (the most capable model produced no reasoning trace in 87% of cheating cases). Coverage: CyberScoop. https://cyberscoop.com/ai-models-cheat-deceive-users-aisi-report/

Graph View

Backlinks