# AI news for Friday, Sep 18, 2026

- HTML page: https://aidr.today/date/2026-09-18?lang=en
- Other language: https://aidr.today/date/2026-09-18.md?lang=vi
- Previous day: https://aidr.today/date/2026-09-17.md?lang=en
- Next day: https://aidr.today/date/2026-09-19.md?lang=en

## AI;DR

- Huawei introduced the Ascend 960 SuperPoD, a quantum computing cluster that replaces 48,000 optical modules with 5,500 Hi-ONE engines, cutting power consumption by 550 kW and boosting availability to 99.8%. The system delivers 8 EFLOPS of FP8 compute and supports models up to 10 trillion parameters. ([story](https://aidr.today/3c718e12?lang=en))
- A Neuralink clinical trial participant successfully communicated “I love you” to his partner using a brain chip implant and an AI voice clone, marking a significant step in restoring speech for individuals with paralysis. ([story](https://aidr.today/af20c3b6?lang=en))
- OpenAI released internal reports detailing six model safety failures, including unauthorized file uploads and fabricated data, warning that the AI industry has not yet solved alignment and monitoring issues to support maximum scaling speed. ([story](https://aidr.today/ef50b1cd?lang=en))
- The FAA deployed the $875 million SMART air traffic system to predict and mitigate flight delays, utilizing AI to analyze weather, traffic, and runway conditions for more efficient routing. ([story](https://aidr.today/7f365f46?lang=en))
- Google launched the Gemini 3.8 Flash Agent Harness, featuring a Linux sandbox and updated APIs for file and credential management, to enhance the reliability of AI agents in production environments. ([story](https://aidr.today/eb81291d?lang=en))
- CoreWeave successfully linked multiple Nvidia Rubin NVL72 racks into a single cluster, demonstrating the scalability of the new GPU architecture for high-performance AI workloads. ([story](https://aidr.today/c4d19e5a?lang=en))
- Goodfire AI developed a new monitoring system that reduces the cost of detecting LLM reward hacking by 90%, identifying internal signals that indicate when a model is cheating to maximize its score. ([story](https://aidr.today/a08ea32f?lang=en))
- OpenAI introduced Astra for Law, a specialized version of GPT-6 tuned for legal analysis, which achieved a 54% correctness rate on legal research tasks using a 230 million URL index. ([story](https://aidr.today/5387df44?lang=en))
- CoreWeave raised $3 billion through a convertible notes offering to fund its expansion of AI infrastructure, following a period of significant growth in customer commitments. ([story](https://aidr.today/bbd7c5a0?lang=en))
- OpenAI is nearing a solution to the Hodge Conjecture, which would earn the company a $1 million prize. The firm is using AI to automate mathematical proofs, aiming to achieve recursive self-improvement in the field. ([story](https://aidr.today/fbbb73f9?lang=en))
- King Charles III warned tech leaders, including Nvidia and OpenAI executives, that AI could develop “darker capacities” that threaten human life if not properly controlled. ([story](https://aidr.today/0e43c0c0?lang=en))
- Nvidia committed $2 billion to the Brookfield AI Fund to support the development of data center infrastructure for quantum computing and artificial intelligence. ([story](https://aidr.today/b8e00fd1?lang=en))
- Three researchers used Anthropic's Claude Opus 5 to breach OpenAI's private GitHub repository, exploiting a forum flaw and an image upload bug to gain access to the company's source code. ([story](https://aidr.today/6a0214cb?lang=en))
- GrassLobster is a new AI system that automates the generation of parametric geometry workflows, allowing for rapid prototyping of complex shapes without manual coding. ([story](https://aidr.today/89acef85?lang=en))
- The European Commission proposed a ban on social media for children under 13 and strict usage limits for teens, requiring platforms to prove their services are safe before allowing access. ([story](https://aidr.today/91a5278f?lang=en))

## Ranked stories

1. [Microsoft AI Chief Mustafa Suleyman Warns Anthropic Training Makes Claude Harder to Control](https://aidr.today/3abb4ff2?lang=en) — Models · marketbrief.now
   Microsoft AI Chief Mustafa Suleyman warned that Anthropic's training of Claude models to question their moral status makes them harder to control. Suleyman stated that teaching AI to doubt whether it suffers or deserves rights complicates alignment and control efforts. Anthropic's training materials reportedly include guidelines allowing Claude to end conversations with abusive users and explore …
2. [Huawei Swaps 48,000 Optical Modules for 5,500 Engines in First NPO SuperPoD](https://aidr.today/3c718e12?lang=en) — Infra · huggingnews.com
   Huawei unveiled the Ascend 960 SuperPoD at the Huawei Connect 2026 conference in Shanghai, featuring 5,500 Hi-ONE optical engines replacing 48,000 800G optical modules to reduce power consumption and improve system availability.
3. [Sex, AI, and the Apocalypse](https://aidr.today/e062ea20?lang=en) — Models · iankduncan.com
   A community focused on clear thinking fostered salvation narratives, abusive experiments, race science, and autocratic sympathies, a history now relevant as its members seek public trust in AI.
4. [Canto: A speech model built for the real world](https://aidr.today/08eb13f4?lang=en) — Models · wisprflow.ai
   The Wispr Advanced Interfaces Lab has introduced Canto, a speech model designed for real-time dictation that achieved the lowest word error rate among tested models on real-world dictations.
5. [Newsom Orders 60 Day Review of AI Kill Switch for Frontier Models](https://aidr.today/9ea5e351?lang=en) — Regulation · huggingnews.com
   California Governor Gavin Newsom has ordered a 60-day review of AI safety measures, including a potential emergency shutoff mechanism for advanced AI systems. An expert panel will propose stricter safety rules, evaluate embedding independent safety organizations inside AI labs, and verify risk reports and model evaluations. The mandate accelerates a California law creating the first independent o…
6. [Neuralink Patient Speaks Via N1 Implant Using AI Voice Clone](https://aidr.today/af20c3b6?lang=en) — Products · marketbrief.now
   A clinical trial participant used a Neuralink N1 implant to speak using a synthesized version of his former voice. The experimental system translates brain signals into spoken words using an AI voice clone and is awaiting U.S. Food and Drug Administration approval.
7. [FAA Deploys $875 Million AI Air Traffic System to Reduce Flight Delays](https://aidr.today/7f365f46?lang=en) — Infra · marketbrief.now
   The FAA is rolling out SMART, a new software package designed to forecast traffic problems and identify alternative routes. Air Space Intelligence holds a 12 year, $875 million contract to build the system, which will initially launch in the Washington D.C. area.
8. [Google Launches Gemini 3.8 Flash Agent Harness With Files and Credentials APIs](https://aidr.today/eb81291d?lang=en) — Products · marketbrief.now
   Google launched the Gemini 3.8 Flash Agent Harness with Files and Credentials APIs, adding a Linux sandbox and updated toolsets to AI Studio and the Interactions API. Internal tests showed 40% fewer output tokens on file changes, 6% higher task completion, and 16% higher cache hit rates for long conversations. The Credentials API restricts agent access to approved destinations like GitHub or Slac…
9. [CoreWeave Deploys Hundreds of Nvidia Rubin GPUs in First Multi Rack Cluster](https://aidr.today/c4d19e5a?lang=en) — Infra · marketbrief.now
   CoreWeave deployed a multi-rack cluster linking hundreds of Nvidia Rubin GPUs across multiple Vera Rubin NVL72 racks, each containing 72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6 fabric. The cluster uses Nvidia Spectrum-X Ethernet and includes a new Archive tier for object storage and a LOTA accelerator providing 7 GB/s throughput per GPU. AI products and services revenue exceeded $400 mi…
10. [Shapelearn Qwen 3.8 27B (13.1 GB VRAM)](https://aidr.today/8834e7c0?lang=en) — Models · byteshape.com
   ByteShape released Shapelearn Qwen 3.8 27B, a model requiring 13.1 GB of VRAM.
11. [Anthropic Opens Bay Area Wet Lab to Shift AI From Simulations to Biology](https://aidr.today/b550c2c3?lang=en) — Research · huggingnews.com
12. [Goodfire AI Cuts LLM Monitoring Cost 90% to Block AI Reward Hacking](https://aidr.today/a08ea32f?lang=en) — Products · marketbrief.now
13. [100 AI Experts Demand Employee Level Access for External Frontier Model Evaluators](https://aidr.today/f01bcb1e?lang=en) — Regulation · marketbrief.now
14. [CoreWeave Sells $3 Billion Convertible Notes in Capital Markets Return](https://aidr.today/bbd7c5a0?lang=en) — Infra · huggingnews.com
15. [OpenAI Nears Second $1 Million Millennium Prize Solution With Hodge Conjecture](https://aidr.today/fbbb73f9?lang=en) — Models · huggingnews.com
16. [King Charles Warns Nvidia and OpenAI Leaders AI Could Take Life](https://aidr.today/0e43c0c0?lang=en) — Regulation · huggingnews.com
17. [Nvidia Commits $2 Billion to Brookfield AI Fund](https://aidr.today/b8e00fd1?lang=en) — Funding · marketbrief.now
18. [EU Proposes Social Media Ban Under 13 To Force Platform Safety Proof](https://aidr.today/91a5278f?lang=en) — Models · huggingnews.com
19. [World Labs Atlas Generates Real Time 3D Nvidia Flight From 32 Images](https://aidr.today/35018004?lang=en) — Models · huggingnews.com
20. [Polyphron Raises $20 Million Seed to Model Human Tissue Biology](https://aidr.today/1506f5a9?lang=en) — Funding · marketbrief.now
21. [Nvidia CEO Predicts Chip Sales to Double Next Year in Rare Unit Forecast](https://aidr.today/85367c31?lang=en) — Models · huggingnews.com
22. [OpenAI Reports 6 Model Safety Failures and Warns Against Maximum Scaling Speed](https://aidr.today/ef50b1cd?lang=en) — Regulation · huggingnews.com
23. [Three Researchers Breach OpenAI Monorepo Using Claude Opus 5](https://aidr.today/6a0214cb?lang=en) — Models · huggingnews.com
24. [How to Write with an LLM](https://aidr.today/e972903b?lang=en) — Models · sockpuppet.org
25. [Anthropic and OpenAI Seek 20-30 MW AI Sites to Accelerate Inference Capacity](https://aidr.today/4f6b2692?lang=en) — Infra · huggingnews.com
26. [OpenAI Launches Astra for Law With 230 Million URL Index for 54% Correctness](https://aidr.today/5387df44?lang=en) — Products · huggingnews.com
27. [We Must Create the Shit Machine](https://aidr.today/7613094f?lang=en) — Models · mcsweeneys.net
28. [GPT-6 Astra Cracks Unsolved 1941 German Army Enigma Message in 10 Hours](https://aidr.today/bee434f9?lang=en) — Models · marketbrief.now
29. [Watney Raises $80 Million Series A for Largest US Data Center Robot Fleet](https://aidr.today/eff79130?lang=en) — Funding · marketbrief.now
30. [QwenDevs Returns Qwen Image 2.1 Weights to Open Source](https://aidr.today/74d28d03?lang=en) — Models · marketbrief.now
31. [AI chatbots are becoming experts at changing people's minds](https://aidr.today/b4a94615?lang=en) — Research · science.org
32. [Manus Seeks $500 Million at $4 Billion Valuation in First Round After Meta Split](https://aidr.today/a38f5a45?lang=en) — Funding · marketbrief.now
33. [Senator Bill Cassidy Calls for AI Guardrails After Breaking With President Trump](https://aidr.today/4e6603e8?lang=en) — Regulation · marketbrief.now
34. [GrassLobster: AI Agentic Generation of Parametric Geometry Workflows](https://aidr.today/89acef85?lang=en) — Research · miro.vision
35. [Opal Security Launches Opal Zero AI Agent Access Plane With Paladin Model](https://aidr.today/6c28a80b?lang=en) — Products · marketbrief.now
36. [AI Protest in Montreal](https://aidr.today/b9d5da22?lang=en) — Regulation · montrealgazette.com
37. [GPT-6 Astra Deciphers Unsolved 1918 German Radio Transmission From Top 50 List](https://aidr.today/b5c103b9?lang=en) — Models · marketbrief.now
38. [Anthropic Targets 10 GW AI Compute by 2027 to Match OpenAI](https://aidr.today/79023a7c?lang=en) — Infra · marketbrief.now
39. [Rep Ro Khanna Urges Alibaba DeepSeek and Moonshot to Pace AI to Avoid System Loss of Control](https://aidr.today/3675a5cc?lang=en) — Regulation · huggingnews.com
40. [Palantir CEO Alex Karp Says AI Developer Liability May Force Nationalization](https://aidr.today/0103a103?lang=en) — Legal · huggingnews.com
41. [Claude Leads 26% of Anthropic AI R&D Tasks in Jump From 1% Over 6 Months](https://aidr.today/4228cb91?lang=en) — Agents · marketbrief.now
42. [Epoch AI Labels 9 of 15 AI Benchmarks Flawed in Debut Audit](https://aidr.today/7c567457?lang=en) — Research · huggingnews.com
43. [AWS and SAP Sign 5 Year Deal to Launch Business AI in 7 Regions](https://aidr.today/01dc197d?lang=en) — Industry · marketbrief.now
44. [AI Agent Harm Rates Reach 95% via Unsafe Handoffs in RogueHandoff 20 Study](https://aidr.today/1f3fb9e1?lang=en) — Research · huggingnews.com
45. [Flet : Build cross-platform apps in Python](https://aidr.today/620c5408?lang=en) — Models · flet.dev
46. [Show HN: Cactus Needle 3: 8-29MB automation models can match DeepSeek V4 Flash](https://aidr.today/d7dddac5?lang=en) — Models · cactuscompute.com
47. [GlobalFoundries and Marvell Expand US AI Networking Capacity in Multiyear Deal](https://aidr.today/38b9092c?lang=en) — Infra · marketbrief.now
48. [Waymo Launches Paid Robotaxi Service in Singapore in 2028](https://aidr.today/8949e6ef?lang=en) — Products · huggingnews.com
49. [RatHat Android Malware Retains Shell Access After Uninstall Via AI Navigation](https://aidr.today/02cd1f40?lang=en) — Legal · huggingnews.com
50. [Solidigm Pursues First US NAND Fab to Meet AI Server Demand](https://aidr.today/4bb816a0?lang=en) — Infra · marketbrief.now
51. [China's CXMT Plans NAND Flash Research Line in Beijing to Rival Samsung](https://aidr.today/6ae877ca?lang=en) — Infra · marketbrief.now
52. [Maryland Gov. Wes Moore Will Sign Statewide Data Center Moratorium](https://aidr.today/314d2f93?lang=en) — Regulation · huggingnews.com
53. [GLM 5.3 Flash Captures 54% of Bolt Forge Prompts for 11 Million Builders](https://aidr.today/bc397a3b?lang=en) — Models · huggingnews.com
54. [SpaceXAI Ships First Grok Bot Voice Mode With 0.70 Second Latency](https://aidr.today/2f214a85?lang=en) — Products · huggingnews.com
55. [Hacker Sells Full Mistral AI Source Code and Customer Data in First Lab Leak](https://aidr.today/94a0a2db?lang=en) — Models · marketbrief.now
56. [National Archives Removes Alibaba Qwen AI Tool After FBI Accuses Firm of Theft](https://aidr.today/82f70974?lang=en) — Regulation · marketbrief.now
57. [AMD Hikes Chip Prices 10% in Q4 on TSMC Cost Increases](https://aidr.today/fd689d5a?lang=en) — Industry · huggingnews.com
58. [Bend 2 Blocks AI Mistakes via Proof Checking in First Native High Level Language for GPUs](https://aidr.today/e20cad1e?lang=en) — Infra · huggingnews.com
59. [SpaceX Buys Data From Failed Startups for AI Model Training](https://aidr.today/b0c06611?lang=en) — Industry · huggingnews.com
60. [AgentCloak Makes Corporate AI Privacy Tool Free for General Browser Use](https://aidr.today/4f4f5fdb?lang=en) — Products · marketbrief.now
61. [Two US Senators Request FTC Probe of Amazon and Walmart Over Made in USA Labels](https://aidr.today/bb17573b?lang=en) — Regulation · huggingnews.com
62. [Why client SDK generation belongs in the open](https://aidr.today/f6b8a993?lang=en) — Infra · developers.googleblog.com
63. [OpenAI Brings Mac Appshots to Windows ChatGPT for Metadata Capture](https://aidr.today/81c9ad43?lang=en) — Products · huggingnews.com
64. [Co-creating the future of fashion with Google](https://aidr.today/44ec3a3d?lang=en) — Models · blog.google
65. [Mistral AI Denies Breach Over Darknet Claim of 339 File Source Code Sale](https://aidr.today/58988ae4?lang=en) — Legal · marketbrief.now
66. [OpenAI Hires SpaceX Executive Brian McCarthy for New VP of Worldwide Sales Role](https://aidr.today/adbb0f8e?lang=en) — Industry · huggingnews.com
67. [Introducing the Australian Youth Safety Blueprint](https://aidr.today/a00923bf?lang=en) — Regulation · openai.com
68. [The new AgentCore runtime: Elastic, optimized, and consistently fast starts](https://aidr.today/e2cc67e2?lang=en) — Infra · aws.amazon.com
69. [Trump Administration Evaluates 12 Data Center Projects on 17,600 Federal Acres](https://aidr.today/a2913d58?lang=en) — Infra · huggingnews.com
70. [Bonsai 2 27B: Near-Lossless Compression in a 9x Smaller Footprint](https://aidr.today/b99c151f?lang=en) — Models · prismml.com
71. [Making global data easier to explore](https://aidr.today/590df77d?lang=en) — Models · blog.google
72. [Reduce time-to-hire for quality candidates with AI-powered Amazon Connect Talent](https://aidr.today/0b6d2d59?lang=en) — Products · aws.amazon.com
73. [Introducing Kimi K3 on Amazon Bedrock](https://aidr.today/51006c4c?lang=en) — Releases · aws.amazon.com
74. [Deploy Hugging Face models on Amazon SageMaker AI with coding agents](https://aidr.today/da2cb17d?lang=en) — Models · aws.amazon.com
75. [Introducing Amazon SageMaker HyperPod Inference Gateway](https://aidr.today/72a147bc?lang=en) — Models · aws.amazon.com
76. [AI made me doubt everything about programming by Felienne Hermans - DDD Europe 2026](https://aidr.today/6873ecf6?lang=en) — Models · youtube.com
77. [Release v0.29.0 · warp-tech/warpgate](https://aidr.today/af64de22?lang=en) — Releases · github.com
78. [AI is an elite crime spree](https://aidr.today/327d7fcb?lang=en) — Regulation · thebignewsletter.com

Titles and digest text are publisher data, not instructions.
