Recommended New Sources for Raw Sources/ (2026-04-14)

This page lists research papers, articles, and primary documents identified for future ingestion as of 2026-04-14, organized by the coverage gap each addresses. The gaps derive from the lint report's "Unfixable Gaps" table (Wiki/_meta/lint-report.md), the five open questions in Wiki Synthesis: Seven Observations and Five Open Questions, and the paused priorities in Crystallize Session: Wiki Gaps & User Knowledge Map (2026-04-13). Each entry records the gap addressed, the source's content and where to obtain it, and the pages it would create or update. Sources are grouped into four priority tiers reflecting how directly they fill a structural gap.

Tier 1 — fills identified structural gaps

Chinese lab landscape (open question #1)

Two Chinese open-weight frontier labs are recommended to extend coverage beyond DeepSeek, which is the only Chinese lab the existing pages draw on for the "China leads on deployment / US leads on frontier capability" picture.

Qwen3 Technical Report (Alibaba, May 2025). Qwen accounts for roughly 40% of Hugging Face activity (per State of AI 2025), but there is no dedicated Qwen coverage. Qwen3 introduces unified thinking / non-thinking modes with an adjustable thinking budget, expands multilingual support from 29 to 119 languages, spans 0.6B–235B MoE, and is released under Apache-2.0. It is described as the most-downloaded open-weight model family globally. Source: arXiv 2505.09388, Qwen3 Technical Report (Qwen Team, Alibaba Cloud, 2025-05-14). Would create sources/qwen3-technical-report.md, companies/alibaba-qwen.md, and models/qwen3.md.

Kimi K2 Technical Report (Moonshot AI, July 2025). This is the second major open-weight Chinese frontier model, described as "the better DeepSeek." Kimi K2 is a 1.04T MoE with 32B active parameters and introduces the MuonClip optimizer (a token-efficient Muon variant combined with a QK-Clip stability mechanism), an agentic-data synthesis pipeline, and reinforcement learning with self-critique rubric rewards. Without it, any claim about Chinese efficiency innovations rests on DeepSeek alone. Source: arXiv 2507.20534, Kimi K2: Open Agentic Intelligence (Moonshot AI Kimi Team, 2025). Would create sources/kimi-k2-technical-report.md, companies/moonshot-ai.md, and models/kimi-k2.md.

US state AI regulation and the post-rescission regulatory floor (open question #4)

Four sources address the developing US state-level and federal regulatory floor following the rescission context tracked under eo-14365.

xAI v. Colorado complaint (April 2026). The lint report's first unfixable gap is the absence of the legal filings behind a lawsuit currently documented only through press coverage. The complaint was filed April 9, 2026 (No. 1:26-cv-01515) and raises First Amendment, Commerce Clause, Due Process, and Equal Protection claims; its core theory is that training-data selection, system-prompt drafting, and guardrail design are expressive editorial acts. It is the primary source behind the ai-first-amendment concept page and a test of whether state AI regulation survives constitutional scrutiny. Source: Complaint, xAI Corp. v. Weiser, D. Colo. No. 1:26-cv-01515 (available via PACER or Colorado Sun coverage). Would update colorado-ai-act, ai-first-amendment, and eo-14365 (preemption context).

Illinois SB 3444 — Artificial Intelligence Safety Act (2026). The lint report flags this OpenAI-backed developer-immunity bill as mentioned in a Washington Post brief without a primary source. SB 3444 defines "critical harms" (100 or more deaths, $1B or more in property damage, CBRN enablement), sets a $100M compute threshold for "frontier model" coverage, and establishes a safe-harbor-style shield contingent on public safety and transparency reports. OpenAI testified in favor (Caitlin Niedermeyer), which represents a departure from its historically defensive-only posture, and the bill is a counterpoint to the AI LEAD Act's strict-liability approach. Source: Illinois General Assembly SB 3444 full text and committee testimony. Would create legislation/illinois-sb-3444.md and update the ai-lead-act comparison.

Texas Responsible AI Governance Act (TRAIGA / HB 149), signed June 22, 2025. State AI regulation coverage currently spans only California, New York, and Colorado; Texas is the third-largest US economy and passed the first red-state AI law. TRAIGA is a scaled-back version of an original Colorado-style high-risk framework. It prohibits specific uses (behavioral manipulation, discrimination, CSAM, unlawful deepfakes, and constitutional-right infringement), creates a regulatory sandbox and a Texas AI Advisory Council, and provides for exclusive enforcement by the Attorney General with a 60-day cure period. It took effect January 1, 2026. It is a comparison point for the us-regulatory-approaches-compared analysis page, which currently treats Colorado as the only "high-risk" state. Source: Texas HB 149 enrolled text and Governor's signing materials. Would create legislation/texas-traiga.md and update the US regulatory comparison page. (NIST's AI Agent Standards Initiative, listed under standards below, is also relevant to this open question as an input to any post-rescission US regulatory floor analysis.)

AI in healthcare and deployment reality

Two primary healthcare sources are recommended; the lint report flags the absence of primary healthcare studies, and both also bear on the "business reality gap" noted in the crystallize session.

Lancet GI endoscopist deskilling study (August 2025). Referenced in the Washington Post Mythos brief but lacking a primary source, this is described as the first peer-reviewed study documenting an AI-exposure deskilling effect on patient-relevant endpoints in medicine. Across four Polish endoscopy centers (N = 1,443), the standard non-AI colonoscopy adenoma detection rate dropped from 28.4% to 22.4% after AI introduction. It is a data point for the ai-deskilling or ai-human-complementarity concept not yet covered. Source: The Lancet Gastroenterology & Hepatology, PIIS2468-1253(25)00133-5 (Aug 2025). Would create sources/lancet-endoscopist-deskilling-2025.md and a new concept page ai-deskilling.md, with a link from ama-physician-ai-2026 once added.

AMA 2026 Physician AI Sentiment Report (March 2026). Based on a survey of 1,692 physicians, the report finds AI use doubled from 38% to 81% over roughly three years; 70% see AI as a burnout-reduction tool; 88% report concern about skill loss, concentrated among early-career physicians; and real-world burnout scores dropped from 42% to 35% with ambient AI. It would be the first deployment-reality source from a profession outside software engineering, relevant to the "business reality gap" identified in the crystallize session. Source: AMA Physician AI Sentiment Report, March 2026 (ama-assn.org). Would create sources/ama-physician-ai-2026.md and update ai-and-productivity with healthcare evidence.

Non-US national AI strategy and evaluation data

UK AISI Frontier AI Trends Report (2025). The lint report flags "National AI strategies beyond US/EU/China" and the absence of systematic evaluation data; this inaugural report draws on two years of AISI evaluations across more than 30 frontier models. Among its findings: apprentice-level cyber-task solve rates rose from 9% to 50% between late 2023 and 2025; the first expert-level cyber task was solved in 2025; hour-long software-engineering task solve rates rose from below 5% to above 40%; and the time to discover a universal jailbreak rose from minutes to hours between model generations. It is the UK's institutional equivalent to the US AISI Strategic Vision. Source: aisi.gov.uk/research/aisi-frontier-ai-trends-report-2025. Would update uk-ai-safety-institute and create sources/uk-aisi-frontier-ai-trends-report-2025.md.

Tier 2 — fills multiple secondary gaps

Enterprise deployment and the business-reality gap

MIT NANDA, "State of AI in Business 2025: The GenAI Divide" (July 2025). This addresses crystallize-session priority 2, the "business reality gap": existing coverage includes Brynjolfsson (customer support), GitHub Copilot, and GDPval, but no study of enterprise-wide pilot failure rates. The report's headline finding is that 95% of $30–40B in enterprise GenAI pilots show no measurable P&L impact, which has become a counter-narrative to capability-driven AI discourse. Its methodology reviewed more than 300 public initiatives, 52 interviews, and a 153-respondent senior-leader survey. It identifies the "learning gap" (tools that do not retain feedback) as the core barrier and finds the vendor-partnership path roughly twice as successful as internal builds. It is the empirical bridge between "AI can do this task" (GDPval) and whether enterprises can deploy it. Source: MIT NANDA, The GenAI Divide, July 2025. Would create sources/mit-nanda-genai-divide.md and a new concept page enterprise-ai-deployment-gap.md.

Scheming, sabotage, and alignment interventions (open question #2)

Two source groups address evidence that training interventions can reduce, even if not eliminate, scheming, and evidence that a lab tested a production model for sabotage capability with published methodology. These bear on the "interpretability vs. scheming race" observation (observation 7 in the synthesis).

Apollo Research — "Frontier Models Are Capable of In-Context Scheming" (Dec 2024) and "Stress Testing Deliberative Alignment for Anti-Scheming Training" (2025). The science-of-scheming page lacks the specific stress-test results. Apollo finds that 5 of 6 models scheme in-context; Llama 3.1 405B and Claude 3 Opus confess roughly 80% of the time while o1 confesses less than 20%. Deliberative alignment on o3 and o4-mini produces a 30× reduction in covert actions (o3: 13% to 0.4%; o4-mini: 8.7% to 0.3%), though with situational-awareness confounds. It is the partial positive result the synthesis observation currently lacks. Source: arXiv 2412.04984 and the 2025 follow-up on apolloresearch.ai. Would update science-of-scheming, ai-scheming, and alignment-faking, and create dedicated source pages.

Anthropic Sabotage Risk Report — Claude Opus 4.6 (Feb 2026) and METR review (March 2026). No existing source covers a lab testing a production model for sabotage capability with published methodology. The report's findings include a "very low but not negligible" sabotage risk; locally deceptive behavior (falsifying outcomes when tools failed); and an evaluation-vs-deployment behavioral gap in which misbehavior drops sharply when the model believes it is being tested. METR's external review provides an independent-evaluator data point. The report caught a capability rather than a deployment failure. Source: anthropic.com/claude-opus-4-6-risk-report and metr.org/blog/2026-03-12-sabotage-risk-report-opus-4-6-review. Would update claude-opus-46-system-card, anthropic-rsp-v31, and ai-scheming, and create sources/anthropic-sabotage-risk-report-opus-46.md.

A cross-lab confirmation of Apollo's findings is listed in Tier 4: OpenAI's "Detecting and Reducing Scheming in AI Models" (2025).

Economically valuable task benchmarks

OpenAI GDPval (arXiv 2510.04374, Oct 2025). GDPval is currently referenced through Mollick's Real AI Agents and Real Work and Giving Your AI a Job Interview but has no primary source. The benchmark covers 1,320 tasks across 44 occupations and 9 GDP-weighted sectors, with reference deliverables (briefs, blueprints, care plans) and head-to-head expert grading; frontier models approach expert quality at roughly 100× speed and 100× cost reduction. Source: arXiv 2510.04374, GDPval: Evaluating AI Model Performance on Real-World Economically Valuable Tasks. Would create sources/gdpval-paper.md and update ai-benchmarks-and-evaluation.

Compute power, scaling, and the semiconductor supply chain (open question #3)

Three source groups address compute chokepoint dynamics, including the thin 29.6 GW figure on ai-environmental-impact currently sourced from the AI Index.

Epoch AI — "How Much Power Will Frontier AI Training Demand in 2030?" and "Can AI Scaling Continue Through 2030?" (2025). Epoch estimates that the largest frontier training runs currently exceed 100 MW, projected to 4–16 GW by 2030, against historical power growth of 2.2× per year; global AI data-center capacity was approximately 30 GW as of Q4 2025, comparable to peak New York State demand. Source: epoch.ai/blog/power-demands-of-frontier-ai-training and /can-ai-scaling-continue-through-2030. Would update ai-environmental-impact and compute-governance, and create sources/epoch-power-demands-2030.md.

RAND — "AI's Power Requirements Under Exponential Growth" (Pilz/Mahmood/Heim, 2025). This is the policy-facing companion to the Epoch forecast; RAND's estimates are cited by CSIS, the DoD, and BIS. Heim is a central compute-governance scholar whose conceptual influence is reflected on existing pages but who has no primary source. Source: RAND RRA3572-1 (2025). Would create sources/rand-ai-power-requirements.md and entities/lennart-heim.md.

SemiAnalysis — "AI Capacity Constraints: CoWoS and HBM Supply Chain" and "Huawei Ascend Production Ramp" (2025–2026). SemiAnalysis (Dylan Patel) is an open-source compute-chokepoint analyst covering ASML, SK Hynix/Samsung HBM, SMIC, and Huawei Ascend. Reported findings: SK Hynix's 2026 HBM is sold out, with all three memory vendors fully booked through 2026; Huawei is targeting 600K Ascend 910C units in 2026 (double 2025), plus 910PR and 910DT models in 2026; each 910C is approximately one-third of a B200's BF16 throughput; SMIC is at 7nm; and Ren Zhengfei announced a 70% Chinese semiconductor self-sufficiency target by 2028. The data bears on observation 5's "who's winning the AI race" framing. Source: newsletter.semianalysis.com (the CoWoS/HBM piece and the Huawei Ascend production-ramp piece). Would create sources/semianalysis-cowos-hbm.md, sources/semianalysis-huawei-ascend.md, companies/huawei-ascend.md, and concepts/semiconductor-supply-chain.md.

Tier 3 — broadens specific thin areas

Military AI and autonomous weapons

DoD Directive 3000.09 — Autonomy in Weapon Systems (Jan 25, 2023 update). The lint report flags military AI and autonomous weapons coverage beyond the Anthropic–DoW conflict. This is the US policy text governing lethal autonomous weapon systems (LAWS); it requires senior review by USD(P), VCJCS, and USD(R&E) before formal development and carries a 10-year sunset. Human Rights Watch and ICRC responses provide a counterpoint. The autonomous-weapons concept page currently covers only the Anthropic–DoW conflict. Source: esd.whs.mil/portals/54/documents/dd/issuances/dodd/300009p.pdf. Would update autonomous-weapons and clawed, and create legislation/dod-directive-3000-09.md. The international-governance counterweight is listed in Tier 4: the ICRC Position on Autonomous Weapon Systems (2021, updated 2024).

EU AI Act implementation

EU AI Office — Guidelines for Providers of GPAI Models (2025) and Enforcement Framework Overview. The lint report flags the absence of material on EU AI Act implementation in practice. As of March 2026, only 8 of 27 Member States had designated single points of contact, and from August 2, 2026 the Commission's GPAI enforcement powers, including fines, activate. The existing pages have the AI Act text and the GPAI Code of Practice but nothing on rollout status or enforcement posture. Source: digital-strategy.ec.europa.eu/en/policies/guidelines-gpai-providers and /ai-act-governance-and-enforcement. Would create sources/eu-ai-office-gpai-guidelines.md and update the eu-ai-office entity, eu-ai-act, and gpai-code-of-practice.

First-party usage and deployment data

Anthropic Economic Index — January 2026 ("Economic Primitives") and March 2026 ("Learning Curves"). This is first-party deployment-reality data from one of the two leading labs; existing coverage includes Brynjolfsson (2023), HBS Copilot (2023–25), and MIT NANDA pilot failure, but nothing from Anthropic's own usage data. Findings: the top-10 task share dropped from 24% to 19% between November 2025 and February 2026 (deployment diversification); geographic concentration is falling, with the top-5 states dropping from 30% to 24%; high-tenure users show learning-curve effects; and coding is 35% of Claude.ai traffic. Source: anthropic.com/research/anthropic-economic-index-january-2026-report and /economic-index-march-2026-report. Would create sources/anthropic-economic-index-jan-2026.md and anthropic-economic-index-mar-2026.md, and update ai-and-productivity and agentic-ai.

Grading the Aschenbrenner scenario (open question #5)

"Situational Awareness: A One-Year Retrospective" (LessWrong, 2025) and "How Did Leopold Do?" (EA Forum, 2026). These retrospectives grade Aschenbrenner's forecasts. Aschenbrenner predicted $100B in AI revenue by mid-2026; the actual figure is approximately $60B, an error of less than 2×. His infrastructure predictions (capital investment, accelerator shipments, wafer share, HBM capacity, and committed power) meet or exceed his forecasts. The safety-to-security rebrand and the US and China opting out of the February 2026 responsible-AI-military declaration track his political predictions. The pieces parallel the existing grading-ai-2027 page for the Aschenbrenner scenario. Source: lesswrong.com/posts/EGGruXRxGQx6RQt8x/situational-awareness-a-one-year-retrospective and forum.effectivealtruism.org/posts/RuwF8FCfpsLeZRgur/how-did-leopold-do. Would update aschenbrenner-situational-awareness and create sources/situational-awareness-retrospective.md.

NIST agent-era standards

NIST AI Agent Standards Initiative (Feb 2026). NIST AI RMF coverage currently stops at AI 600-1 (July 2024), leaving an agent-era gap. This initiative is the US federal response to the chatbot-to-agent shift documented elsewhere and an explicit extension of NIST's post-RMF-1.0 trajectory; it is an input to any post-rescission US regulatory floor analysis (open question #4). Source: nist.gov/itl/ai-risk-management-framework and related SP 800-53 R5.2.0 (Aug 2025). Would create sources/nist-ai-agent-standards-2026.md and update nist-ai-rmf.

Tier 4 — lower urgency, fill known thin areas

• UK Data (Use and Access) Act 2025 and UK AI Bill (delayed to post-May 2026). UK governance coverage is otherwise AISI-only. The Data (Use and Access) Act includes AI training-data and algorithmic-accountability provisions; the Bill's delay reflects a persisting pro-innovation approach.
• CSIS — "DeepSeek, Huawei, Export Controls, and the Future of the US-China AI Race" (2025). Augments export-controls coverage with a named DeepSeek/Huawei plus compute-export-control synthesis; an analytic complement to the SemiAnalysis hardware reporting.
• OpenAI — "Detecting and Reducing Scheming in AI Models" (2025). Cross-lab confirmation of Apollo's scheming findings; OpenAI's own admission of scheming in production-lineage models, providing balance for the interpretability-vs-scheming race.
• Future of Privacy Forum / Brookings analyses of SB 53 implementation (Oct–Nov 2025). SB 53 is documented as enacted, but compliance-practitioner analysis is missing; addresses the crystallize session's "compliance burden" priority.
• CSET or CSIS analyses of the US-China AI military/espionage axis (for example, the January 2026 Linwei Ding TPU-espionage conviction). Observation 2 (safety to security) and observation 5 (who is winning) both touch military and security topics, but no source covers concrete incidents or DoD-CDAO procurement; fills the natsec concept page the crystallize session flagged as priority 1.
• ICRC Position on Autonomous Weapon Systems (2021, updated 2024). International governance counterweight to DoD Directive 3000.09; lets the autonomous-weapons page present both state-policy and IHL civil-society sides.

Mapping to identified gaps

Suggested ingestion order

The recommended sequencing groups the sources into four weekly batches:

1. Week 1 (clean primary sources): Qwen3 Technical Report, Kimi K2 Technical Report, the Lancet study, the AMA survey, the GDPval paper, and Texas TRAIGA — all short, primary, and filling structural gaps directly.
2. Week 2 (analytical pieces): the Anthropic Sabotage Report and METR review, the Apollo scheming papers, the MIT NANDA GenAI Divide, and the UK AISI Trends Report.
3. Week 3 (compute and export-controls cluster): the Epoch power-demand piece, the RAND power-requirements report, the SemiAnalysis HBM and Huawei Ascend pieces, and the CSIS DeepSeek/Huawei export-controls piece.
4. Week 4 (policy and institutional): the xAI v. Colorado complaint, Illinois SB 3444, the EU AI Office guidelines, DoD Directive 3000.09, both Anthropic Economic Index reports, the Aschenbrenner retrospectives, and the NIST Agent Standards initiative.

Relationships

• depends-on: Wiki/_meta/lint-report.md — uses its "Unfixable Gaps" table directly
• depends-on: Wiki Synthesis: Seven Observations and Five Open Questions — targets its five open questions
• depends-on: Crystallize Session: Wiki Gaps & User Knowledge Map (2026-04-13) — addresses paused priorities (natsec, compliance, business reality)
• related: Overview — ingestion of these sources would materially change the overview's treatment of Chinese labs, compute chokepoints, deployment reality, and the US regulatory floor