AI daily, 18-Nov-26

• OpenAI released GPT-6.2 Astra and GPT-6.2 Sol in Codex and the API, with Astra priced 20 percent below GPT-6 Astra.
• Nvidia reported quarterly revenue of 116 billion dollars for the October quarter, more than double a year earlier, after Jensen Huang's departure.
• Isomorphic Labs said it raised 5 billion dollars to speed its drug design work with AI.
• Nvidia is searching for a new CEO, and the Commerce Department tightened licensing rules for advanced AI chips.
• A benchmark study found the same coding model could cost up to five times more depending on the agent harness used.

TOP STORIES

* OpenAI releases GPT-6.2 Astra and GPT-6.2 Sol in Codex and the API
  OpenAI says GPT-6.2 Astra and GPT-6.2 Sol are available in Codex and the API as gpt-6.2-astra and gpt-6.2-sol, with Fast and Ultrafast modes. Astra's standard token prices are 20 percent below GPT-6 Astra's, and it scores 32 percent better on BenchCAD 2.1 Core and 9 points higher on Terminal-Bench Science 0.1. Ultrafast pricing drops 50 percent for three months. Astra also ranks behind only Claude Fable 5.2 on Artificial Analysis.
  OpenAI Developers (@OpenAIDevs); also OpenAI, Rowan Cheung

* Nvidia Says Revenue More Than Doubled to $116 Billion in October Quarter
  Nvidia reports that revenue more than doubled to 116 billion dollars in the October quarter from a year earlier, beating its own forecast by over 13 percentage points. The company says revenue will rise about 91 percent in the current fiscal period. It is the first earnings report since Jensen Huang's departure, and interim CEO Colette Kress is presenting it.
  The Information; also Financial Times (Technology)

* Nvidia Seeks New CEO After Jensen Huang's Exit as Commerce Tightens AI Chip Export Rules
  Nvidia is searching for a new CEO after Jensen Huang resigned following reports of an affair, according to Bloomberg. President Trump said national security was not put at risk. The Commerce Department issued interim rules requiring licenses for leading-edge AI chip sales to buyers in Malaysia, Singapore, Thailand, Indonesia and Brazil, aimed at limiting Alibaba and ByteDance's access to US computing power.
  Bloomberg Tech

* Demis Hassabis announces $5B funding for Isomorphic Labs
  Demis Hassabis says Isomorphic Labs has raised 5 billion dollars to speed its mission to solve all disease. He says the company is making rapid progress reimagining drug design from first principles with AI, to cut the time and cost of finding new medicines. Investors include Greenoaks, Sequoia and Andreessen Horowitz. Bloomberg Technology reports the round was led by Sequoia and Greenoaks, six months after the company's previous fundraising effort.
  Demis Hassabis (@demishassabis); also Bloomberg Technology, Hacker News

CHINA

* Reuters: Chinese entities use US AI model output to train domestic rivals
  Reuters reports Chinese entities are using increasingly sophisticated methods to train domestic rivals on output from leading American AI models.
  Reuters (Technology)

* Z.AI shares fall 85% as chip fears hit Chinese model makers
  Bloomberg reports Chinese model developer Z.AI's shares have dropped 85 percent amid chip concerns.
  Bloomberg Technology

CODING

* Runta and Baseten benchmark finds same AI model can cost up to 5x more depending on coding agent harness
  Runta and Baseten ran 22 coding-agent harness configurations across five open-weight models 27,280 times, using about 83 billion tokens.
  The New Stack; also Hacker News

COMPANIES

* Genesis AI raises $500M Series A at $3.5B valuation, backed by Nvidia and Eric Schmidt
  Genesis AI raised a 500 million dollar Series A at a 3.5 billion dollar valuation, backed by Nvidia and Eric Schmidt.
  Bloomberg Tech; also AI News / TestingCatalog · at 35:04

* Perplexity crosses $1B annual revenue run rate, CEO Aravind Srinivas says
  Perplexity CEO Aravind Srinivas said the company crossed a 1 billion dollar annual revenue run rate.
  Latent.Space

COMPUTE

* Nvidia Halts Hopper Data Center Chip Sales to China After US Approval
  Nvidia has stopped shipping Hopper data center chips to China, Bloomberg reports, despite US government approval last year to sell a version there.
  Bloomberg Technology

MODELS

* Huawei Cloud releases and open-sources Ascend-native embodied world model CloudRobo-WM1.2
  Huawei Cloud released and open-sourced CloudRobo-WM1.2, an embodied world model for Ascend chips, claiming first place on RoboTwin 2.0 and LiberoPro 2.0.
  华为云 (WeChat)

* Microsoft AI releases MAI Code 2 and MAI Code 2 Flash for agentic coding
  Microsoft AI released MAI Code 2, claiming 90 percent of Opus 5.5 performance at 5 percent of the cost.
  Microsoft AI

POLICY

* China warns US of retaliation over chip curbs and urges dialogue
  China says it will retaliate if Washington proceeds with curbs on remote access to advanced chips, and urges the Trump administration to settle differences through dialogue.
  Bloomberg Technology

* California DMV bans Waymo from state freeways, citing "imminent hazard to public safety"
  The California DMV's ban on Waymo's driverless freeway rides took effect November 11, after Bay Area incidents including a robotaxi entering an active US-101 construction closure.
  The Verge

* OpenAI Says NATO Cyber Hub Uses Its Models for Defense
  OpenAI says NATO's technology and cyber hub, with over 3,000 experts, uses its models to probe suspicious activity and deploy fixes.
  The Wall Street Journal (Technology); also The Decoder

* Ireland's DPC says it ordered X to halt Grok bot data processing
  Ireland's Data Protection Commission ordered X to stop using data from its Grok Bot AI agent to train AI systems.
  36Kr

RESEARCH

* Study reports benchleaking in base models, with 96% of HumanEval problems leaking in ERNIE-5.5-Flash-Base
  A preprint by Yiwei Qin and seven co-authors documents benchmark contamination across 20 benchmarks. Using few-shot continuation on 20 benchmarks across four public base models, it finds 96% of HumanEval problems leak with their reference implementations and 85% of OpenBookQA questions with their answers for ERNIE-5.5-Flash-Base, against 9% and none for DeepSeek-V4-Flash-Base.
  arXiv cs.AI

SAFETY

* AI Models Misidentified Civilians as Military Targets in Simulated Scenarios, Report Finds
  An independent report finds advanced AI models in simulated scenarios misidentified civilians, ambulances, surrendering soldiers and an Iranian school as military targets, Bloomberg reports.
  Bloomberg Technology · measured by unnamed independent research report

* OpenAI Signs Agreements with Apollo Research, METR, Transluce and Goodfire to Embed AI Evaluators
  OpenAI signed agreements with Apollo Research, METR, Transluce and Goodfire for internal access to assess its risks.
  The Information (newsletters)

---
The Observatory · https://observatory.mgorsky.net/archive/415/
