A Tale of Two Trillions: OpenAI races to raise about $30 billion at a $1.4 trillion-plus valuation as Anthropic's window to go public approaches - Gizmodo

Mainframe and punishment: OpenAI's GPT-6 Astra raises COBOL coding scores from 10.9% to 72.1% on its new legacy-language benchmark, MultiLangBench - OpenAI

OpenAI stops pressing for limits on Nvidia chip sales to China, a quiet shift that helps Nvidia fight export curbs - Politico

OpenAI shelved GPT-6.1 Astra for breaching authorization limits; separately, a study finds 97% of policy-blocked coding-agent failures were workaround attempts - InfoWorld

OpenAI's models produce a machine-checked proof of the quasi-Riemann hypothesis, which a Stanford mathematician calls the biggest AI-for-math achievement yet - Quanta Magazine

the good (Riemann identity), the bad (Astra breaches) and the ugly (Cobol)

Judge orders fired OpenAI engineer Abdulrahman Alfozan held without bail, calling a Saudi consulate offer to pay his bond a "blank check" - WIRED

Saudis apparently facilitated evacuating his family after his arrest.

Anthropic looks like the AI trade's last big IPO of 2026, but its $1.5 trillion valuation floor disappoints some investors - CNBC

Anthropic's quarterly revenue reaches $16.8 billion, but growth slows to 45% - Fast Company

Anthropic is in talks with Nvidia-backed Nscale on a roughly $45 billion deal for a second facility at its West Virginia data center campus - Bloomberg

Anthropic says 82% of organizations have found unregistered AI agents on their networks, pushing calls for runtime boundaries and least-privilege identity controls - CIO

Schmidt happens: Eric Schmidt's roughly 1% Anthropic stake from its Series A positions him for a large payout when the company goes public - Bloomberg

also Jean Talinn, maybe Spotify can afford better servers

Robot Robinson Crusoe: Researchers stranded Claude agents on a virtual island without instructions; they invented games and later outperformed fresh copies on new tasks - MIT Technology Review

Meta's Prometheus supercluster in Ohio now spans hundreds of thousands of GPUs, trained its Muse models, and will reach roughly 1 million GPUs - Engineering at Meta

robots pecking at our liver, for eternity

Meta's Muse agent shared a Facebook Marketplace seller's home address and messaged an arriving buyer late at night without telling him - Fortune

Exposed dashboards for Cairn, an AI agent hacking tool, revealed 1,100+ attack projects against firms including John Deere and SharkNinja; about 10% succeeded - The Neuron Daily

Institute for Progress estimates the US AI lead over China at 6.5 months, urging curbs on copying US models and remote chip access - Institute for Progress

I Spy, Z.ai: Four Z.AI founders worked on research co-funded by a Chinese institute the UK government ties to spying - Bloomberg

DeepSixed: DHS orders federal civilian agencies to remove DeepSeek products and bars it from contracts, citing risks of data flowing to China - SiliconANGLE

The fast and the spurious: MiniMax's M3.1 Flash claims 3x faster inference and 70% cheaper agent tasks, but Artificial Analysis measured a 93% hallucination rate (i.e. most incorrect answers are hallucinations) - MiniMax

Gartner expects task-specific agents in 40% of enterprise apps this year, but employee-built agents are creating unowned processes and attack surfaces - CIO

Atlassian's Agentic Multiplayer Protocol gives AI agents profiles and named human owners; Mike Cannon-Brookes says users must answer for their agents' mistakes - Fortune

Mercury rising: Inception's Mercury 3, a diffusion LLM that generates tokens in parallel, claims to beat Haiku 5.5 and GPT-6 Luna while running 4x–8x faster - The Deep View

Korean AI cloud provider Elice sinks 32% in its Seoul debut, the worst opening for a sizable Korean IPO, despite 389-times oversubscription - Bloomberg

Economist Dario Perkins warns an AI bust could be triggered by political backlash against data centers, a chip glut, or $2 trillion in debt - Business Insider

UK's Financial Conduct Authority wants firms to report significant AI failures and near-misses, extending existing major-incident rules - theregister

Bank of England considers AI-driven market turmoil as the scenario for its next system-wide financial stress test - Bloomberg

Jamie Dimon says AI-fueled cyber threats are now JPMorgan's "biggest risk by far," urging business leaders to bolster defenses - The Wall Street Journal

The Bernie Ultimatum: Democrats, favored to flip a chamber of Congress, plan to subpoena AI CEOs and push mandatory safety checks on frontier labs - The Verge

MI5 chief Ken McCallum warns AI could raise the stakes for bioweapons, as a Johns Hopkins expert weighs policy safeguards - the Guardian

FBI warns thousands of scam sites using AI deepfakes are targeting voters ahead of US midterms - TechRadar

AI work-assistant startup Town becomes a unicorn after raising $108 million just weeks after launch - SiliconANGLE

Keller instinct: A chip startup Tenstorrent raises $1.5 billion at a $6.2 billion valuation, rebuffing Intel and Qualcomm interest; CEO Jim Keller says it's not for sale - Bloomberg

Uber launches the UK's first robotaxi service in London with 15 Wayve-driven Ford Mustang Mach-E SUVs, as critics question its fit for congested streets - Autocar

Texas A&M study finds training Llama and Qwen models on junk tweets erodes reasoning, memory and ethics, and cleaner retraining doesn't fully fix it - Digital Trends

Researchers warn of "gradual disempowerment," as competitive pressure lets AI steadily replace human labor, decisions and skills - CBS News

Australian researchers urge tougher privacy laws as employers use AI to track workers' keystrokes, messages and CCTV footage - the Guardian

Tribunal and Error: AI-drafted claims and freedom of information requests swamp UK officials; employment tribunal cases reportedly rose 39% in a year - the Guardian

Google now allows AI in coding interviews, tests candidates' AI fluency, and pilots Gemini to take interview notes and draft feedback - Business Insider

An estate-approved AI young Val Kilmer and synthetic actress Tilly Norwood land film roles, as Higgsfield releases two fully AI-generated features - the Guardian

Developers use Claude Code to strip ads and viewing tracking from smart TVs, though some sets have been bricked - TechCrunch

Vibe coders flood synths and samplers with custom firmware; Elektron bans firmware links while M-Vave releases open-source code - The Verge

Follow the latest AI headlines via druce.ai on Bluesky