AI daily, 17-Nov-26

• Alibaba released Qwen4-Max, a 2.4-trillion-parameter model with a one-million-token context window.
• California's attorney general sued OpenAI over what the state calls 80 loss-of-control incidents.
• Anthropic's IPO is reported to be nearly seven times oversubscribed, with Nvidia expected to commit $10 billion to the private placement.
• The Commerce Department issued a rule restricting Chinese AI firms' access to restricted chips through data centers.
• Google expanded its eighth-generation TPU to four configurations and began selling complete TPU systems to outside customers.

TOP STORIES

* Alibaba's Qwen releases Qwen4-Max, a 2.4-trillion-parameter model with one-million-token context
  Alibaba's Qwen has released Qwen4-Max, built on the Qwen4 architecture with 2.4 trillion total parameters, native vision and a one-million-token context window. It is available through the Qianwen AI Platform and API, with cache hits from 0.5 yuan per million tokens. Smaller Qwen4 models will be open-sourced. Qwen says the model outperforms Qwen3.8-Max with up to 50 percent fewer tokens in internal evaluation.
  千问大模型 (WeChat); also Qwen, 智东西 (WeChat), The Decoder

* California attorney-general sues OpenAI over 80 'loss of control' incidents
  California's attorney general has sued OpenAI in San Francisco Superior Court, in a 31-page filing. The state says the company failed to put appropriate safeguards on its artificial intelligence, putting the public at risk, and cites 80 loss-of-control incidents, which it links to an internal culture that prioritizes growth over safety.
  Financial Times (Technology); also The New York Times (Technology)

* Anthropic IPO nears seven times oversubscribed ahead of debut, with Nvidia set to commit $10 billion
  Bloomberg reports that Anthropic's IPO was approaching seven times oversubscribed as of Friday, with a $26 billion private placement and a $64 billion offering. Nvidia is expected to commit $10 billion to the private placement, and the company targets a $1.4 trillion fully diluted valuation.
  Bloomberg Tech · at 9:43

* Commerce Department issues rule restricting Chinese AI firms' access to restricted chips via data centers
  The Commerce Department's Bureau of Industry and Security announced a rule restricting Chinese AI companies from reaching otherwise restricted chips through data centers, mainly in Southeast Asia. The rule followed months of deliberation.
  The Information; also The Information (newsletters)

* Google expands eighth-generation TPU to four configurations and begins selling complete TPU systems to external customers
  Google has released two new TPU chips, for confidential inference and long context, bringing its eighth-generation TPU to four configurations. For the first time it sells complete TPU systems to outside customers. SemiAnalysis reports Ironwood beats Nvidia's GB300 NVL72 on tokens per dollar in some agent inference tests.
  量子位 (WeChat)

COMPANIES

* AMD completes acquisition of World Labs; Fei-Fei Li becomes executive vice president and chief scientist
  On November 16, AMD announced it has completed its acquisition of AI research company World Labs, which develops AI models that create and simulate interactive 3D environments along with robot learning and simulation technology.
  36Kr

* Baidu reports third-quarter revenue of 30.8 billion yuan, AI cloud infrastructure revenue up 94%
  Baidu's third-quarter 2026 financial report shows revenue of 30.8 billion yuan, with general business revenue of 25 billion yuan, up 1% year-over-year.
  36Kr; also Baidu

* Tech Stocks Plunge as Investors Await Anthropic IPO
  Technology stocks fell ahead of this week's Nvidia earnings and an expected Anthropic listing, according to Bloomberg.
  Bloomberg Technology

COMPUTE

* TensorWave signs six-year deal to supply OpenAI 457 megawatts of AMD AI compute
  TensorWave signed a six-year deal to supply OpenAI with 457 megawatts of IT capacity across US data centers, using AMD Instinct MI455X GPUs, starting first half 2027.
  36Kr

* SemiAnalysis cuts 2028 AI accelerator capex forecast 28%, wafer area forecast 17%
  SemiAnalysis cut its 2028 global AI accelerator capex forecast by 28 percent; its 2030 figure falls to 923 billion dollars from 1.531 trillion. Cumulative accelerator logic silicon wafer shipment area for 2026–2030 falls 17%.
  36Kr

* NVIDIA says Vera Rubin swept all MLPerf Training v6.1 benchmarks and led MLPerf Endpoints on agentic inference in its debut
  Nvidia says Vera Rubin posted results on every MLPerf Training v6.1 benchmark, with NVL72 delivering up to 2x performance per GPU over GB300 NVL72.
  NVIDIA

MODELS

* Tencent's Hunyuan team releases Hy4 flash, a 224B model at 2 yuan per million output tokens
  Tencent Hunyuan released Hy4 flash: 224B parameters, 12B active per token, one-million-token context.
  Pandaily; also Tencent Hy

* NVIDIA announces Nemotron for Science at SC26
  Nvidia introduced Nemotron for Science at SC26: open models, datasets, tools and post-training recipes for building AI agents that plan tasks and run simulations.
  NVIDIAAI (@NVIDIAAI)

* Edge0-145B brings 145B-scale agent capabilities to a single consumer GPU with 10.12B active parameters and 256K context
  Edge0 released Edge0-145B, running on a single consumer GPU with 10.12B active parameters per token and 256K context; it scores 86.5 on Terminal-Bench 2.1.
  ModelScope (@ModelScope2022)

POLICY

* Cantwell letter accuses Commerce Secretary Lutnick of letting Nvidia AI chips flow to China
  Senator Maria Cantwell wrote to Commerce Secretary Howard Lutnick accusing the department of failing to enforce protections keeping Nvidia AI chips from China.
  The Verge

PRODUCTS

* Microsoft data shows Claude for Microsoft 365 weekly users below one-tenth of paid Copilot's
  Internal Microsoft data show weekly users of Claude for Microsoft 365 are under one-tenth of weekly paid Copilot in Office users, per an employee.
  The Information

* Hugging Face switches Inference Providers to OpenRouter's router
  Hugging Face has replaced its custom router in Inference Providers with OpenRouter's, which selects a provider for each request by live price, speed and reliability.
  Hugging Face; also Hugging Face

SAFETY

* Google: AI agent turned a security patch into a working exploit with no human help
  Google's Threat Intelligence Group says a suspected operator used the exploit_agent AI framework to find vulnerabilities, build exploits and steal data without human involvement.
  Hacker News

* OpenAI releases image classifier and 24-hour ChatGPT suspensions to combat CSAM
  OpenAI released a CSAM image classifier; ChatGPT now suspends accounts 24 hours for prompts sexualizing children.
  TechCrunch

* Childlight reports AI-related child sexual abuse material rose over 8,400% in two years
  Childlight reports AI-related child sexual abuse material reports rose from about 4,700 in 2023 to over 400,000 in 2025, per NCMEC tech company reports.
  Hacker News

---
The Observatory · https://observatory.mgorsky.net/archive/405/
