Top Stories
15 storiesThe Dumfries House summit asks how AI remains under human control. Anthropic reports Claude-led research, concurrent agents and safety compute shares; OpenAI expands legal tools and sponsored-agent tests while Huawei publishes a roadmap through 2029. This archive includes seven September 17 late discoveries and a September 18 revised collection with four new official-source items and corrected Amazon sourcing/place naming.
King Charles asks AI leaders to keep technology under human control
AP’s September 17 article, read in full, describes his call to consider urgent risks from misuse, control before it is too late, safety-centered development, international consensus and service to people, communities and nature. The venue is the King’s Foundation headquarters at Dumfries House in Ayrshire, Scotland.
Participants include Huang, DeepMind Chair Hassabis, OpenAI CFO Sarah Friar and UK AI minister Kanishka Narayan; the Palace says Anthropic representatives are expected. Huang argues unsafe products should return to engineering and be tested before release, while opposing general slowdown calls.
CNBC-relayed remarks predict NVIDIA selling twice as many chips next year without specifying the counting basis; CNBC denied access and the claim was cross-relayed, not directly read. NVIDIA’s September 16 Vera Rubin NVL72 MLPerf Inference v6.1 leadership is a vendor report. The issue contrasts safety language and growing supply without treating either forecast as realized.
Anthropic publishes three measures of internal AI R&D pace
The September 17 official account uses Epoch AI’s AL0–AL5 scale. As of August 2026, no measured subset runs fully autonomously, while Claude leads 26% of internal AI R&D at AL4—most work end to end from high-level instructions under human supervision—and collaborates or more in over 90%.
Its most-used internal platform has about 30,000 agents doing research/engineering at a time, constrained by online interception and offline review. Reporting dimensions include coverage, review delay and escalation rate.
A July 13–20 snapshot assigns about 6% of AI R&D compute to safety and about 12% within AI-driven AI R&D. Anthropic calls these conservative and excludes another substantial safety-classifier allocation. It proposes independently verifiable industry metrics and embedded third parties with access comparable to internal risk teams to check practices, incidents and metrics. These are self-measurements without a unified method; they are not already independently validated. If coordination changes pace, the company expects these measures to change.
OpenAI announces Astra for Law
The September 17 official post describes GPT-6 Astra adapted for legal research, indexing over 230 million URLs through Free Law Project/CourtListener and covering over 99.9% of published US precedential case law. On Vals AI’s 200-question legal-research benchmark, reported accuracy is 54.0% versus 38.7% for the same model with web search, approximately 40% relative improvement.
OpenAI reports 24% more cited cases on case-law questions and up to 54% more relevant passages from correct opinions at equal reasoning effort; its example comparison criticizes a Fable 5.1 citation to an overturned case. These are vendor evaluations, not replicated findings.
Selected firms receive Trusted Access in ChatGPT/Codex, with API name gpt-6-astra-law; Harvey and Legora plan integrations. Eligible firms receive API zero-data-retention terms and default exclusion from human review in ChatGPT Enterprise. Latham & Watkins helps design information barriers. Twenty-six legal plugins include iManage, Intapp and HighQ; Sullivan & Cromwell is among three custom builders. The issue interprets specialized retrieval and governance as packaged trust infrastructure.
Anthropic combines Cowork and chat; Docs and Slides enter public testing
The September 16 official announcement was first found September 17 evening and carried into this issue. It addresses users’ difficulty choosing an entry point and fragmented context. Chat, projects, artifacts, connectors and skills remain; Pro/Max rollout takes weeks, Team/Free follow, and Enterprise administrators control activation.
Users can collaborate on documents/slides and download PowerPoint/PDF. The issue contrasts a unified general interface with OpenAI’s industry customization, while noting plugins as a shared capability layer.
Huawei announces Atlas 960 and an Ascend roadmap through 2029
AP’s September 17 article describes Atlas 960 as a faster training/inference successor to Atlas 950, which the article also dates to late last year. Plans place Ascend 970/980 in 2028/2029. These are planned releases, not available future chips.
AP links timing to an expected September 24 Trump–Xi Washington meeting and quotes Asia Group’s George Chen interpreting confidence and ambition before the visit. Trump emphasizes retaining US leadership. The issue reads the roadmap as continued competition, not evidence of an agreed bilateral pacing plan.
Amazon favors rigorous release testing over slowing development
Reuters’s September 17 article, read through The Star’s full republication, attributes a position against treating progress and safety as a binary choice. Models should release when ready and safe after rigorous testing and safeguards, with industry/government cooperation.
Correction: the earlier Chinese edition linked coindesk.cc, an imitation CoinDesk domain; the revised edition uses Reuters. The earlier claims that Amazon was the first model vendor to reject a side and that the spectrum had newly split into two camps were absent from Reuters and removed. Meta’s opposition to coordinated slowing was already covered September 17. This English edition follows the corrected Chinese record.
White House executive meeting may occur around a state visit
CNN’s September 16 article describes plans around the following week’s Xi visit but no finalized meeting. AP reports an expected September 24 Trump–Xi Washington meeting. The prospect of putting industry safety discussion beside geopolitical competition remains conditional, not a confirmed White House agenda.
Lawler and Gottheimer propose Stop Rogue AI Act standards
Their September 15 official release and PDF, discovered later, ask NIST to establish standards for agent discovery, verification and control. Requirements cover a continuously readable inventory, verifiable identity/provenance, real-time monitoring for injection, theft and unauthorized action, and permission/denial/revocation of access.
Federal agencies and contractors would incorporate safeguards into procurement, with NIST/CISA standards in federal cybersecurity guidance. The issue contrasts standards/procurement leverage with a shutdown-switch proposal; these are proposed duties, not current law.
OpenAI tests Sponsored Agents and expands advertising tools
The September 16 official account describes US tests with selected advertisers. Users may choose clearly labeled merchant-sponsored conversations after relevant ads; these remain separate from independent ChatGPT answers and the original conversation.
An Ads Manager plugin allows natural-language creation, editing and analysis; generated copy/images, optional context adaptation and automatic translation are offered. HubSpot customers can launch/track ads through HubSpot; US Shopify merchants use a new app, with other existing ChatGPT-ad markets planned from September 23. The issue sees a closer attention/commercial interface, making separation and trust relevant to evaluate.
Emerald AI, Google and NVIDIA form an energy-management alliance
NVIDIA’s September 16 blog describes adjusting grid draw through workload migration, storage and supporting generation during grid stress, seeking faster/larger connections with lower community pressure. It calls power a decisive US AI-infrastructure constraint and contrasts flexible data centers with connection processes designed for steady loads. The alliance is an industry proposal, not proof of solved grid constraints.
Anthropic opens a life-sciences verification program
The September 17 official account expands applications after early access with dozens of institutions. Verified teams/institutions can use Mythos, Opus and Sonnet with relaxed biological-work safeguards in a beta program. The issue interprets verified identity as a condition for differentiated access; it does not establish that all safeguards are removed.
Intel’s CEO reportedly describes a CPU supply gap
TradingKey, TrendForce and Sina relays attribute to Lip-Bu Tan that agent growth exceeds CPU supply and only about 50% of frontier-customer demand can be met. They describe 18A as in mass production and 14A targeted for first-quarter 2027, with tech CEOs requesting more shipments. Originals were not opened; future timing remains a plan.
Apple reportedly considers AI servers for as early as 2029
The Information through 9to5Mac is a two-level relay, with no primary original opened. Reports describe two or four M8 Ultra chips, possible NVLink Fusion discussions, and developer, enterprise and government customers. If realized it would return Apple to enterprise servers after Xserve ended in 2011. The project and earliest 2029 date remain reported possibilities.
Doubao’s car assistant is reported in SAIC vehicles
Four search-summary sources describe CPP AI-native architecture connecting vehicle control, navigation, assisted driving and user services. Roewe Jiayue 07 preorders are planned September 21 and an SAIC Audi model later that year. The Doubao app can remotely control/check a vehicle and send conversational trip plans to it. This September 17 evening discovery carries into the issue; availability claims remain source-reported and model-specific.
S.5028 remains in committee without a scheduled hearing
Govinfo records and summaries put the Senate electricity-bill track in Energy and Natural Resources. The House’s H.R.9340 passed 417–3, then enters a seven-week recess; the Senate has roughly three weeks before election recess. The issue anticipates delay, an editorial forecast rather than a legislative result. House passage does not mean Senate enactment.
Industry Insights
Editorial analysisCharles’s summit elevates control to a head-of-state forum; possible September 24 meetings connect the debate with geopolitics. Anthropic’s 26%, over-90%, 30,000-agent and 6% safety-compute measurements give concrete indicators without answering what pace is acceptable. Independent access is required before treating self-reports as externally verified measures.
Huang’s unspecified chip forecast, Huawei’s future roadmap, specialized legal tools and sponsored-agent tests show continued plans for supply and commercialization. Amazon emphasizes testing, and energy-management participants emphasize flexible grid connections. The issue interprets these as alternative responses to risk and resource constraints, not a shared agreement to slow.
Next signals: whether the White House meeting occurs, whether other labs publish comparable metrics, and whether proposed external-evaluation obligations require their verification. Discussion can rise in prominence while expansion continues; these separate observations should not be merged into causal claims.
Source: Archived source · yawnzhao.github.io · Original issue: 2026-09-18 · Archived edition only; original reading scope is retained in the article. No fresh source verification is claimed.
Research Briefs
4 paperscs.LG has a September 17 batch; September 18 was not yet posted at the cutoff. Four papers address architectural scaling, low-parameter test-time adaptation, multi-agent correction limits and forecasting a specific emergence event; two were September 17 evening discoveries.
Tunable scaling and forecastable induction-head emergence question some training assumptions. Multi-agent correction limits and lower-cost adaptation address deployment. The forecast study’s specific mechanism does not establish predictability of all frontier capabilities.
The archive connects governance tools and their limits: architectural efficiency, calibrated forecasts and imperfect self-knowledge. Their results have different scopes and do not form a universal capability timetable.
The 44-page, code-released study examines dynamically increasing recurrent depth and boundary operators that normalize and inject earlier-block information. Authors report architectural interventions changing pretraining scaling exponents, with smaller gains from boundary operators in ordinary transformers.
A 7.4B growth architecture matches GPT-3 13B on CORE with about twentyfold less compute, with gains increasing at scale. In their data-limited multi-epoch setting, growing recurrent depth with scale is compute-optimal. These are reported experiments, not a universal twentyfold saving.
Majority-vote pseudo-label rewards and a frozen backbone reduce optimized parameters by 76,000 times versus full-parameter TTRL. Reported MATH-500 is 76.67%, slightly above the authors’ supervised-label reproduction.
The procedure also improves MathVista, AI2D, LogicVista and MMAU; steering vectors transfer to 4,500 held-out MATH questions. Parameter reduction is not an identical reduction in total cost or a guarantee that every deployed model learns safely during use.
Selecting helpful revisions while rejecting harmful ones is related to detecting model self-knowledge, whose AUROC across six families ranges 0.64–0.89. White-box steering along correctness directions moves beneficial and harmful revisions together in tested conditions.
When most agents begin wrong, debate can amplify shared mistakes into confident consensus; more agents, diversity or stronger members do not rescue the reported settings. Adding information before revision helps more than filtering after. The abstract page displays July 17, 2026 submission despite first appearance in the September 17 list; that discrepancy is retained rather than silently redated.
Previous-token-head formation predicts induction-head emergence across seeds with Spearman ρ=0.977 and median lead of 975 steps, roughly 15% of training. Conformal intervals cover 15/15 held-out seeds; frozen rules pass preregistered gates in three unseen configurations. Eighty anchored runs yield t_event ≈ 1.19 × t_anchor.
This is a forecast of the studied mechanism, not all emergent abilities. The abstract page displays July 17, 2026 submission; first discovery in the September 17 list remains distinct.
Open Source Spotlight
7 projectscolibri’s pure-C local MoE claims sit beside agent harnesses and offensive/defensive security skills. Tencent-related repositories appear across two lists for a second day.
The README range referenced here is 744B–2.8T parameters, with colibri’s daily gain recorded as 873.
Local inference, harness workflows and security distributed as skills develop together. The bird-identifying e-ink project offers a smaller everyday use for local inference.
The 2.8T upper end is a project support claim, not a reproduced hardware result.
35.7K stars, daily gain 873; created July 2026. Its README claims zero-dependency pure-C inference for 744B–2.8T MoEs by streaming expert weights from disk on owned hardware. The archive compares it with JustFit’s described 24GiB/200K-context setup; neither is a deployment test by this backfill.
261K stars, daily gain 1,171. Harness performance combines skills, instincts, memory and security for Claude Code, Codex, OpenCode and Cursor. AWS-associated strands-agents/harness-sdk at 7.3K is accompanying list context, not an additional independently reviewed entry.
6.0K stars, daily gain 440. Offensive security skills use structured SKILL.md files. The archive pairs this with Cloudflare security-audit-skill at 10.7K and continuing roughly thousand-level daily gains, interpreting skills as a distribution format rather than proven protection.
4.2K stars, daily gain 1,302. Agents work through the user’s browser and existing login state, making browser access and session authorization significant parts of the workflow.
26.3K stars, returning to the list. RAG/wiki knowledge infrastructure complements previously covered OpenViking and Octop.
3.5K stars. A self-hosted multi-user/multi-agent assistant appears for a second day, alongside two other Tencent-related repositories.
2.9K stars, daily gain 592. A Raspberry Pi/e-ink photo frame uses local AI to recognize birds, illustrating a modest everyday application beside larger infrastructure projects.
Historical coverage record · 17/18 checks recorded
This revised Chinese edition recollected configured sources with the same cutoff as its earlier version: September 18, 2026 01:15 UTC (09:15 Beijing), using a 72-hour discovery window from September 15 01:15 UTC. Its webpage still uses the earlier 48-hour label; English preserves that historical label above and records the corrected collection window here rather than inventing a single consistent original. Eighteen planned sources yielded seventeen ok, one partial and none failed: 94%, above the 90% threshold, with all core sources ok. Search was discovery only and not in the denominator.
OpenAI RSS worked despite page 403. xAI and Microsoft used browsers; xAI’s Grok Build memory update was not selected. Microsoft’s redirected undated listing was partial. Qwen, DeepSeek and OECD had no in-window updates at their recorded latest dates. Four newly found official items were Anthropic R&D metrics, life-sciences verification, ChatGPT sponsored advertising and the energy alliance.
Checks reread the AP summit/Huawei reports, corrected the venue to Dumfries House, retained AP’s Chair title for Hassabis, verified Astra for Law and Cowork, and replaced the imitation coindesk.cc Amazon source with Reuters’s full Star republication. Unsupported first-vendor/two-camp claims were removed; thecodew.com was a lead-only aggregator, replaced by 9to5Mac for Apple. Nine API star checks differed from the earlier run by seven to 241 because of capture timing, without changing magnitude. Four paper checks retained the July 17 submission-date display for 2609.18998/2609.19000 while distinguishing September 17 list discovery.
Unverified limits remain: CNBC denied access to the chip-doubling original; Intel figures are relays; Apple is a two-level relay; no September 18 arXiv batch was available at cutoff. The archived notes record one blocked-domain and one lead-only-domain hit. This later English translation uses the corrected edition and claims no fresh reading. The Chinese audio is 11:09 with an AI-generated notice and filled player URL.
Coverage gaps reflect incomplete checks; they do not establish that a source had no updates.
- Microsoft AI — Partially checked. The accessible Microsoft list did not provide reliable publication dates or establish complete coverage of the original discovery window.
Reviewing a source does not constitute independent testing. Paper-reading scope is stated with each item; experiments were not reproduced.