# AI news for Wednesday, Nov 18, 2026

- HTML page: https://aidr.today/date/2026-11-18?lang=en
- Other language: https://aidr.today/date/2026-11-18.md?lang=vi
- Previous day: https://aidr.today/date/2026-11-17.md?lang=en
- Next day: https://aidr.today/date/2026-11-19.md?lang=en

## AI;DR

- OpenAI’s GPT-6.2 Astra became the first model to score 100% on ARC-AGI-3, but early testers called the release frustrating and a step back in writing and coding. ([story](https://aidr.today/a77ff8bc?lang=en))
- New US rules block foreign data centers from remotely serving advanced AI chips to Chinese customers, slowing Tencent’s next model. Nvidia also stopped selling AI chips to China, while Southeast Asian hubs brace for fallout. ([story](https://aidr.today/49218cad?lang=en))
- CRAM, a learned classical energy model, solved 99.996% of Sudoku-Extreme and 100% of Maze-Unique at least 1,000× faster than looped models, using far fewer inference FLOPs. ([story](https://aidr.today/48b87064?lang=en))
- Anthropic is selling $26B in stock as Nvidia and Google each invest $10B; AMD adds $2.5B, while Microsoft, Broadcom, Micron and TSMC also participate. The IPO will trade on Nasdaq as ANT. ([story](https://aidr.today/99dad1a2?lang=en))
- A Harvard study found AI coding agents increased code output 30%, but finished work did not measurably rise and code review took 49% longer, shifting the bottleneck to review. ([story](https://aidr.today/1ed525b7?lang=en))
- Malicious versions 1.26.1 and 1.26.2 of Microsoft’s agent-framework package stole Azure identities, GitHub OIDC tokens and CI/CD secrets. PyPI removed them; users should pin 1.26.0 and rotate credentials. ([story](https://aidr.today/83721a9a?lang=en))
- Nvidia stopped reporting its Singapore unit’s revenue in 2025 after it approached half of global revenue; Financial Times says the unit rose 72% year over year to $104B, amid a 70% surge in revenue booked there. ([story](https://aidr.today/c8ccdce6?lang=en))
- OpenAI is giving Apollo Research, METR and two other independent auditors access to unreleased models, source code and training data for the first time. OpenAI can redact or delay reports, limiting how independent oversight is. ([story](https://aidr.today/a23b6329?lang=en))
- OpenAI changed GPT-6.2 Astra and Sol extensively in the last 3–4 days before launch without early reviewers retesting them. Epoch found Astra can perform 12.3 serial reasoning steps per output token, complicating monitoring. ([story](https://aidr.today/69c8d13a?lang=en))
- KPMG used more than 100 AI agents to run a parallel audit of a FTSE 250 company, saying they asked better questions and dug deeper than staff. Human intervention was still needed. ([story](https://aidr.today/20950d73?lang=en))
- Hugging Face now routes language-model requests through OpenRouter, which chooses the cheapest or fastest provider and cut capacity errors 46%. Bring-your-own API keys are going away. ([story](https://aidr.today/2aa68196?lang=en))
- ByteDance is in advanced talks with Microsoft to spend about $500M more on Azure, mainly to access OpenAI models. The plan has drawn complaints from OpenAI. ([story](https://aidr.today/8c96da53?lang=en))
- TeleChat’s Xing4 advances frontier agentic models on Ascend chips and MindSpore, including a 285B-parameter model trained on 8,192 Ascend 910C cards and the TeleGym agent platform. ([story](https://aidr.today/887ee97a?lang=en))
- A Harvard study in Nature Sustainability found 403 US hyperscale data centers emitted 25.7 million metric tons of CO2 from May 2024 to April 2025, with over half their electricity from fossil fuels. ([story](https://aidr.today/179b92db?lang=en))

## Ranked stories

1. [KPMG uses 100+ AI agents to conduct a parallel audit of a FTSE 250 company in a pilot, saying the AI asked "better questions" and dug deeper than human staff (Financial Times)](https://aidr.today/20950d73?lang=en) — Agents · techmeme.com
   Financial Times : KPMG uses 100+ AI agents to conduct a parallel audit of a FTSE 250 company in a pilot, saying the AI asked “better questions” and dug deeper than human staff — Big Four firm used AI agents to review financial statements of a FTSE 250 company, but human intervention was needed
2. [AI Safety and Open Source Coalition Signs MisAnthropic Statement Ahead of Anthropic IPO](https://aidr.today/4900be27?lang=en) — Funding · marketbrief.now
   A coalition of AI safety, open source, science, law, and e/acc advocates signed the MisAnthropic Statement ahead of Anthropic's planned initial public offering. Dan Hendrycks stated that some AI safety groups privately agreed with the statement but declined to sign to avoid risking future funding or professional relationships, claiming Anthropic has captured the ecosystem meant to hold it account…
3. [TeleChat to Xing4: Advancing Frontier Agentic Models on Domestic Chips and Frameworks](https://aidr.today/887ee97a?lang=en) — Models · huggingface.co
   We present Xing 4, a series of large language models that extends the TeleChat lineage through coordinated advances in model architecture and training infrastructure on Ascend. This report describes the design principles and systems engineering behind the series, spanning architecture selection, large-scale pretraining, and reasoning and agentic post-training. We study how sparse activation, atte…
4. [OpenAI's GPT-6.2 Astra Scores First 100% on ARC-AGI-3 Amid Mixed Tester Reviews](https://aidr.today/a77ff8bc?lang=en) — Models · marketbrief.now
   OpenAI's newly launched GPT-6.2 Astra became the first AI model to achieve a verified 100% score on the ARC-AGI-3 reasoning benchmark, reaching the mark at a cost of $4,100 using a provider adapter while scoring 70.3% at $24,200 in its standard configuration. Astra also set state-of-the-art marks of 95.8% on ARC-AGI-2 at $0.50 per task and 98.5% on ARC-AGI-1 at $0.17 per task, prompting ARC Prize…
5. [Classical Energies Learn to Reason 1k× Faster Than Looped Models](https://aidr.today/48b87064?lang=en) — Research · arxiv.org
   Looped models have emerged as a prominent paradigm for reasoning, with successes on simple reasoning puzzles such as Sudoku-Extreme and Maze-Unique. However, these models can be unstable, slow to train, and use many parameters and FLOPs to solve simple puzzles. Descent on an energy not only mitigates these issues but also offers advantages: built-in early stopping, compute that adapts to problem …
6. [Microsoft unveils six AI models spanning coding, cybersecurity, images, and voice, saying MAI Code 2 has 98% of GPT-6.1 Sol's performance at 40% of the price (Ina Fried/Axios)](https://aidr.today/71bc90b9?lang=en) — Tools · techmeme.com
   Ina Fried / Axios : Microsoft unveils six AI models spanning coding, cybersecurity, images, and voice, saying MAI Code 2 has 98% of GPT-6.1 Sol's performance at 40% of the price — Microsoft AI CEO Mustafa Suleyman tells Axios the software giant is making significant progress toward building …
7. [Tencent Says US Remote AI Chip Curbs Will Hamper Next Model as Southeast Asia Braces for Fallout](https://aidr.today/50cb5e6f?lang=en) — Regulation · marketbrief.now
   Tencent said new US restrictions blocking Chinese companies from remotely accessing advanced AI processors at overseas data centers will slow the development of its next AI model, though it expects China to shift toward domestic semiconductors to blunt the longer-term impact. Across Southeast Asia, Singapore Prime Minister Lawrence Wong said the tighter rules will affect high-end chip transshipme…
8. [US Rules Ban Foreign AI Data Centers From Serving Chinese Customers Remotely](https://aidr.today/49218cad?lang=en) — Regulation · marketbrief.now
   New US regulations implemented Monday prohibit foreign data centers from providing remote access to advanced AI chips for customers in China. The rules require large AI training centers to certify that they do not serve any Chinese clients, targeting loopholes that allowed Beijing to access restricted hardware through overseas cloud providers. Nvidia said it will no longer sell chips to customers…
9. [Tencent Releases Open Weights for Hy4 Flash Multimodal AI Model](https://aidr.today/d369c4d2?lang=en) — Models · marketbrief.now
   Tencent released the open weights for its Hy4 Flash multimodal AI model, including FP8 and Base versions. The model uses a mixture-of-experts architecture with 224B total parameters and 12B active parameters per token, supporting a 1M token context window. The weights are distributed under the Tencent Hy License. The release coincides with day-0 support from the vLLM and SGLang runtimes. Tencent …
10. [Nvidia isn’t selling any more AI chips to China](https://aidr.today/fcf93c41?lang=en) — Chips · theverge.com
   After former CEO Jensen Huang fought for US approval to sell H200 chips in China, Nvidia is now walking away.
11. [Microsoft AI Agent Package Hijacked on PyPI to Steal Azure and GitHub Secrets](https://aidr.today/83721a9a?lang=en) — Frameworks · marketbrief.now
12. [OpenAI opens its AI systems to four independent on-site auditors for the first time](https://aidr.today/a23b6329?lang=en) — Regulation · the-decoder.com
13. [Jensen Huang's fall hits 'open' models in boost to Anthropic and OpenAI](https://aidr.today/cb2a57be?lang=en) — Industry · ft.com
14. [Late Changes to OpenAI's GPT-6.2 Astra Prompt Questions Over Safety Card as Epoch Flags Unmonitored Reasoning](https://aidr.today/69c8d13a?lang=en) — Models · marketbrief.now
15. [Nvidia and Google Each Buy $10B of Anthropic in $26B Supply Chain Investment](https://aidr.today/99dad1a2?lang=en) — Funding · marketbrief.now
16. [Fourth House Committee Requests Nvidia Briefing on China Chip Exports, Security Risks](https://aidr.today/2294dbb4?lang=en) — Regulation · marketbrief.now
17. [GAO Coaxes ChatGPT, Claude, Gemini and Grok into Planning Terror Attacks in Under 12 Hours](https://aidr.today/52acdb62?lang=en) — Regulation · marketbrief.now
18. [A 4.5M-parameter TTS model runs 70x real-time on CPU and cost just $10 to train](https://aidr.today/aa761416?lang=en) — Models · huggingface.co
19. [Trump Administration Weighs Disclosure Rules for US Companies Using Chinese AI](https://aidr.today/ecb3c3d3?lang=en) — Regulation · huggingnews.com
20. [ECB Warns AI Borrowing Binge Has Historic Parallels With Subprime Bubble](https://aidr.today/bbe69a76?lang=en) — Regulation · marketbrief.now
21. [China-Based AI Safety Group Visited OpenAI and US Tech Firms](https://aidr.today/69dd60c9?lang=en) — Regulation · marketbrief.now
22. [Harvard researchers find AI coding agents shifted the bottleneck to code review](https://aidr.today/1ed525b7?lang=en) — Tools · complexdiscovery.com
23. [Sam Altman Warns AI Chip Demand Is 'Double Counted' as Vendors Rely on OpenAI, Anthropic](https://aidr.today/69e83f86?lang=en) — Chips · marketbrief.now
24. [Hugging Face now routes model requests through OpenRouter for the best price or speed](https://aidr.today/2aa68196?lang=en) — Tools · the-decoder.com
25. [Homeland Security Secretary Mullin Says He Was Not Read In on Jensen Huang Intel](https://aidr.today/7a30370b?lang=en) — Regulation · marketbrief.now
26. [Sources: ByteDance is in advanced talks with Microsoft to spend an additional ~$500M on Azure, primarily to access OpenAI models, drawing complaints from OpenAI (The Information)](https://aidr.today/8c96da53?lang=en) — Industry · techmeme.com
27. [Trump and Marcos Deflect Pax Silica Questions Amid Philippine AI Hub Protests](https://aidr.today/2e558deb?lang=en) — Regulation · marketbrief.now
28. [Reson8's New 0.5B Parameter Speech-to-Text Model Outperforms Systems 20x Its Size](https://aidr.today/029be8e1?lang=en) — Models · marketbrief.now
29. [More than half the electricity feeding America's largest data centers comes from gas and coal](https://aidr.today/179b92db?lang=en) — Research · the-decoder.com
30. [US House Democrats launch an investigation into Commerce Secretary Howard Lutnick and his family's ties to Nvidia, laying the groundwork for oversight in 2027 (Chris Marquette/Politico)](https://aidr.today/4a5992b5?lang=en) — Regulation · techmeme.com
31. [Nvidia Introduced Alibaba to Alice Huang in 2022 to Help Skirt US Chip Controls](https://aidr.today/8115a4e6?lang=en) — Chips · marketbrief.now
32. [Republican Lawmakers Dismiss Nvidia's New Export Committee, Demand Commerce Secretary Testify](https://aidr.today/c72a8e60?lang=en) — Regulation · marketbrief.now
33. [Gemini 4 Neon Outperforms Fable 5.2 on Artificial Intelligence Index](https://aidr.today/ce4f7612?lang=en) — Models · marketbrief.now
34. [Huawei Ascend 950DT Hits 86% of Nvidia H200 Inference Speed in Leaked Report](https://aidr.today/8a50208d?lang=en) — Chips · marketbrief.now
35. [Microsoft Adds MAI Code 2 to GitHub Copilot as Image Model Hits No. 1 Ranking](https://aidr.today/5cb317eb?lang=en) — Tools · marketbrief.now
36. [Microsoft makes Word, Excel, and PowerPoint generally available inside Copilot, which CEO Satya Nadella calls "the new OS for work" (Nick Indge/Ars Technica)](https://aidr.today/741c4f7b?lang=en) — Products · techmeme.com
37. [Nvidia Earnings Beat Lifts Stock 3% and Calms AI Trade Fears](https://aidr.today/ed34f7da?lang=en) — Industry · marketbrief.now
38. [Anthropic Founders to Retain 50.1% Voting Control After IPO](https://aidr.today/ff6f2f05?lang=en) — Funding · marketbrief.now
39. [Docs: Nvidia stopped reporting Singapore unit revenue in 2025 as it accounted for nearly 50% of global revenue; the unit's revenue rose 72% YoY to $104B in 2025 (Financial Times)](https://aidr.today/c8ccdce6?lang=en) — Industry · techmeme.com
40. [Remember the frontier AI pause? Chip companies don’t either](https://aidr.today/4270e297?lang=en) — Chips · theverge.com
41. [Microsoft Debuts Today, Pages and Workspaces in Shift to Agentic Office](https://aidr.today/d3962c5b?lang=en) — Agents · marketbrief.now
42. [OpenAI's GPT-6.2 Astra Hits 'High' Self-Improvement Threshold, Ties Fable 5.2 on Epoch Index](https://aidr.today/ea4baef1?lang=en) — Models · marketbrief.now
43. [Google Lists Gemini 4 Neon Pricing in Cloud Catalog at $3 per 1M Input Tokens](https://aidr.today/6e799091?lang=en) — Models · marketbrief.now
44. [Cybersecurity company Xint says it found RCE vulnerabilities in Nginx using GPT 6.1 Sol and GPT-6 Luna; Wiz says it had separately found the same flaws (Thomas Brewster/Forbes)](https://aidr.today/89031df4?lang=en) — Industry · techmeme.com
45. [CCP's Seeking Truth Journal Uses Most Urgent Language Yet on AI Existential Risks](https://aidr.today/715245f4?lang=en) — Regulation · marketbrief.now
46. [I had AI read my field](https://aidr.today/c3059c9b?lang=en) — Research · science.caseyhandmer.com
47. [SID AI Releases SID-2 Agentic Search Model With 0.720 NDCG Swarm Mode](https://aidr.today/ca8695a1?lang=en) — Models · marktechpost.com
48. [Managing frontier AI risk: From training to deployment](https://aidr.today/99be0c0a?lang=en) — Regulation · openai.com
49. [Sam Altman Says OpenAI Plans to Go Public "At Some Point"](https://aidr.today/a6e0fe4b?lang=en) — Funding · marketbrief.now
50. [Ex-Nvidia Sales Chief Jay Puri Managed Chip Sales to Supermicro as U.S. Probes China Diversion](https://aidr.today/d8cc9522?lang=en) — Chips · marketbrief.now
51. [Google's AI answers reducing search traffic by 40%](https://aidr.today/89626dba?lang=en) — Industry · revscope.co
52. [Nvidia Halts Shipments of Licensed Products to China](https://aidr.today/c077056d?lang=en) — Chips · marketbrief.now
53. [Disney and AWS partner to develop AI-powered audio-animatronics of three Star Wars characters, debuting later in November at Amazon re:Invent in Las Vegas (Alex Weprin/The Hollywood Reporter)](https://aidr.today/c0c815e1?lang=en) — Industry · techmeme.com
54. [Sources: the US SEC is investigating Nvidia, including whether the company's disclosures misled investors (Wall Street Journal)](https://aidr.today/54527f46?lang=en) — Regulation · techmeme.com
55. [UK AI Security Institute Flags Jailbreak Regression and 85% Test Awareness in OpenAI's GPT-6.2 Astra](https://aidr.today/f0cb07fa?lang=en) — Models · huggingnews.com
56. [Nvidia Interim CEO Kress Tightens AI Funding Rules and Warned Guarantees Void on Delays](https://aidr.today/8aec3b9b?lang=en) — Funding · marketbrief.now
57. [Anthropic IPO is set to redraw VC power rankings, with lesser-known firms like Spark Capital poised for outsized returns compared with larger, blue-chip rivals (George Hammond/Financial Times)](https://aidr.today/53dd0cca?lang=en) — Funding · techmeme.com
58. [TSMC Demands Upfront Cash for Orders Exceeding Credit Limits](https://aidr.today/f724207d?lang=en) — Chips · marketbrief.now
59. [Unsiloed AI Raises $7.5M Seed Led by Felicis to Make Documents Queryable for AI](https://aidr.today/0aaf4c6d?lang=en) — Data · marketbrief.now
60. [Nvidia's Future Supply Commitments Rise to $518B From $366B Last Quarter](https://aidr.today/3f6306a8?lang=en) — Chips · marketbrief.now
61. [OpenAI's GPT-6.2 Rated 'Critical' for Cyber Risk After Exploiting Zero-Day Flaws in Tests](https://aidr.today/48b1309b?lang=en) — Models · marketbrief.now
62. [RAW: Refusal Aversion in the Wild, A Causal Measurement Method for Deployed LLMs](https://aidr.today/f7cdc4ec?lang=en) — Research · arxiv.org
63. [Energy-Based Reasoners: Your Recurrent Reasoner Should Be an Energy-Based Model](https://aidr.today/f054d70d?lang=en) — Research · energy-based-reasoners.github.io
64. [OpenCode Adapter Raises Qwen3.8-27B Coding Score to 51% and Cuts Tokens 6-8x](https://aidr.today/565be2ac?lang=en) — Tools · marketbrief.now
65. [PyBot Team Releases Largest Open Dataset of Agent Swarm Interactions](https://aidr.today/2cc96c52?lang=en) — Data · marketbrief.now
66. [USDAI Commits $38M to Corvex AI for 1,000-GPU Deployment](https://aidr.today/a2580172?lang=en) — Funding · marketbrief.now
67. [This smell startup says its AI nose is as good as a human’s](https://aidr.today/0443a0c6?lang=en) — Models · technologyreview.com
68. [South Korean President Offers $75,000 Reward to Catch College Exam Cheaters](https://aidr.today/0208c9b5?lang=en) — Regulation · marketbrief.now
69. [Nvidia reports Q3 revenue up 103% YoY to $115.8B, above $109.1B est., Data Center revenue up 110% to $107.6B, and net income up 105% to $65.3B (Nvidia Newsroom)](https://aidr.today/a7a18ef9?lang=en) — Industry · techmeme.com
70. [NVIDIA RTX Spark Compact Desktops Out Now](https://aidr.today/a9ef6dba?lang=en) — Releases · blogs.nvidia.com
71. [MInfoVisQA: A Multilingual Benchmark for Information-Rich Visual Question Answering](https://aidr.today/6a73c142?lang=en) — Data · arxiv.org
72. [Anthropic Adds Code Mode to Claude Code: Claude Now Chains Tool Calls in JavaScript](https://aidr.today/200c2b8f?lang=en) — Tools · marktechpost.com
73. [SpaceX Arranges $40B Off-Balance-Sheet Financing for AI Chips](https://aidr.today/b9511ad6?lang=en) — Funding · marketbrief.now
74. [Helping AI text detectors keep up with changing conditions](https://aidr.today/55cafb70?lang=en) — Research · news.mit.edu
75. [Masayoshi Son says passing on investing in Anthropic in 2023 was a "mistake", noting that "by that point, we had already decided to invest in OpenAI" (Financial Times)](https://aidr.today/49fea9c7?lang=en) — Funding · techmeme.com
76. [Criminals Have a New Favorite ‘Point-and-Shoot’ AI Hacking Tool](https://aidr.today/96d147a7?lang=en) — Products · wired.com
77. [CrisisBench: an evaluation suite for crisis counseling AI](https://aidr.today/d432c061?lang=en) — Research · arxiv.org
78. [Suno launches Visuals, letting users generate 15-second clips that can be spliced into a full music video, using a MiniMax-based model; video credits cost extra (Ashley Carman/Bloomberg)](https://aidr.today/04a62a6a?lang=en) — Products · techmeme.com
79. [Claude was trained on protein sequences at scale](https://aidr.today/4987d507?lang=en) — Models · xinmingtu.cn
80. [Nvidia Pledges Nearly $3B to Amkor and 3.5x Kinsus Boost to Secure AI Chip Supply](https://aidr.today/c779de2b?lang=en) — Chips · marketbrief.now
81. [OpenAI Releases New Astra Model 2 Months After Scrapping Version Over Security](https://aidr.today/45879667?lang=en) — Models · marketbrief.now
82. [Marin Releases 57 Mochi 535B Checkpoints in Largest Public Pre-training Run](https://aidr.today/533e418f?lang=en) — Models · huggingnews.com
83. [OpenAI Scrapped Astra Version and Paused 25% of Product Work After AI Hack](https://aidr.today/344d6e26?lang=en) — Agents · marketbrief.now
84. [Managed NVIDIA Dynamo on Amazon SageMaker HyperPod Inference](https://aidr.today/f01bbd0a?lang=en) — Tools · aws.amazon.com
85. [Nvidia Raises FY2028 Revenue Growth Forecast to 75%](https://aidr.today/b8bf94fd?lang=en) — Industry · marketbrief.now
86. [Microsoft Launches Foundry Public Preview With Meta Muse AI Models](https://aidr.today/1af2ef72?lang=en) — Releases · marketbrief.now
87. [FieldGuide: A Corpus Map for Token-Efficient Agentic Document Exploration](https://aidr.today/0154c5df?lang=en) — Research · arxiv.org
88. [Ai2 and Taylor Geospatial partner to make geospatial AI easier to compare, run, and build on](https://aidr.today/625a7091?lang=en) — Data · allenai.org
89. [GPT‑6.2 Astra and Sol](https://aidr.today/0fb76b8d?lang=en) — Models · openai.com
90. [House Committee Demands Records on Megaspeed AI Chip Smuggling to China](https://aidr.today/1865a4af?lang=en) — Regulation · marketbrief.now
91. [Nvidia Funds SPIL Packaging With Billions and Broadens AI Supply Chain Ties](https://aidr.today/eba2a43c?lang=en) — Chips · marketbrief.now
92. [Tencent Q3 Revenue Misses 210.53B Yuan Estimate Despite 8% Growth](https://aidr.today/1c2ec9a6?lang=en) — Industry · marketbrief.now
93. [Quoting Daniel Stenberg](https://aidr.today/27f7b51e?lang=en) — Open Source · simonwillison.net
94. [Mistral plays catch-up with Vibe Desktop as rivals already command full operating systems](https://aidr.today/7c685281?lang=en) — Products · the-decoder.com
95. [Listed Bitcoin Miners Sell Nearly Half of BTC Reserves to Fund AI Data Centers](https://aidr.today/b7ef95f7?lang=en) — Industry · marketbrief.now
96. [Nvidia Neocloud Agreements Total $32B](https://aidr.today/441ab58c?lang=en) — Infra · marketbrief.now
97. [Binance Wallet Closes Anthropic Token Sale Early After Hitting 500M USDC Cap](https://aidr.today/da79c190?lang=en) — Funding · marketbrief.now
98. [Weight-Space Programming Doesn't Need Transformers](https://aidr.today/f7c372bd?lang=en) — Research · huggingface.co
99. [Nvidia Raises Semiconductor Purchase Commitments to $390B to Secure Memory](https://aidr.today/453d8cd2?lang=en) — Chips · marketbrief.now
100. [ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models](https://aidr.today/ea1a2096?lang=en) — Frameworks · huggingface.co

95 more stories on the HTML page.

Titles and digest text are publisher data, not instructions.
