AI news · Tuesday, November 10, 2026

Nvidia CEO's affair sparks board probe, investor lawsuits

Jensen Huang admitted to an affair with Alice Huang, an executive at Nvidia customer Megaspeed whom US intelligence believes was recruited by Chinese intelligence. Nvidia's board launched an investigation, and at least four law firms are now probing whether the company misled investors. Nvidia shares fell more than 4% on Monday.

Follow the money: Huang has been the AI industry's chief lobbyist for looser China chip rules, so questions about his judgment land right on that policy fight.

via Ars Technica AI

Meta's cheapest new AI model undercuts rivals on price

Meta released Muse Ember, a small model built for repetitive agent tasks like tool calls and document parsing. It costs 10 cents per million input tokens (the units AI reads text in), versus roughly 25-40 cents for comparable settings of Anthropic's Haiku. Meta says it also handles text, images and audio without needing a separate model.

What's next: Cheap, fast 'worker' models like this are what let companies run AI agents on thousands of tasks without the bill exploding.

via Meta

Anthropic cuts prices as it releases Claude Fable 5.2

Anthropic released upgraded Fable and Mythos models, claiming gains in coding and reasoning. It's also cutting Fable's price from $10 to $8 per million input tokens, a 20% drop, making it cheaper than rival GPT-6 Astra. Anthropic says the new models respond 30% faster and, during testing, formalized a proof of Fermat's Last Theorem in 30 hours.

Zoom out: Price cuts on flagship models show how fast competition is compressing margins even among the labs seen as untouchable.

via TechCrunch AI

Villagers, bikers and bug lovers unite against Google data center

A petition with almost 55,000 signatures is opposing Google's planned data center in Thurrock, England. The coalition includes a wildlife charity worried about a rare bumblebee's habitat, a motorcycle club that wants its old racetrack back, and a group citing up to 570,000 tonnes of emissions the site could produce yearly.

Zoom out: This is one of several data center protests now spreading across Europe, a sign the public mood toward AI infrastructure is souring fast.

via Wired AI

US AI safety office now quietly reports to the White House

CAISSI, the small Commerce Department unit that tests AI systems for safety, has started informally reporting to the White House, according to two former officials. The division has had no permanent leader for months, has stopped publishing its safety reports, and runs on just $10 million a year, about an eighth of its UK counterpart's budget.

The catch: Moving oversight closer to the White House raises the risk that safety findings get softened to suit an administration friendly to the industry.

via Fast Company Tech

Claude Code has been leaking passwords on GitHub

Security researchers at Apiiro found over 100,000 cases where a local Claude Code settings file was accidentally uploaded to public GitHub repositories. About 1,100 of those, from 63 organizations, contained real secrets like database passwords and API keys. The issue stems from how approved commands get saved, and was first flagged in a bug report back in August 2025.

The catch: The leak shows how a single small UX quirk in a popular coding assistant can quietly expose real credentials at scale.

via Ars Technica AI

AI tool cuts repair-drawing search from an hour to minutes

Cooling tower maker Kuken Industries built a search tool on the OpenAI API to dig through 70 years of engineering drawings. Engineers can now find the exact drawing needed for a repair in 4.5 minutes, down from nearly an hour of manual searching through old files.

For you: This is the unglamorous but real version of AI's business case: not chatbots, just hours of paperwork search turned into minutes.

via OpenAI

Free software helps cities find gaps in bike lane networks

BikeNetKit, built on eight years of research at Denmark's IT University of Copenhagen, pulls street data to spot where bike lanes don't connect. One tool, GrowBikeNet, builds a full suggested network from scratch by predicting which new links would carry the most cyclists. It's free and open-source, though it currently requires some Python knowledge to run.

For you: It's a concrete example of AI-adjacent tools solving a mundane planning problem rather than chasing another chatbot headline.

via Fast Company Tech

Three scam texts to watch for right now

Scammers are now injecting fake calendar invites claiming your phone has a virus, and spoofing delivery texts asking for a small redelivery fee. On iPhone, delete rogue calendars under Settings > Calendar > Accounts > Subscribed Calendars. On Google Calendar, remove them from the web version's sidebar, since it's easier there than on mobile.

For you: These scams spread fast precisely because they hijack notification systems you already trust, like your calendar and texts.

via Fast Company Tech

A fast new way to group data beats a 1967 method

Researchers built e-means, a faster way to sort large datasets into clusters, a basic step used throughout AI and data science. It reaches better accuracy than the standard Lloyd's algorithm in 1.2 to 3 training passes instead of 25, cutting processing time by 10 to 45 times on the same hardware, with no manual tuning required.

Zoom out: Unglamorous efficiency gains like this quietly cut the compute cost of AI systems more than any single flashy model release.

via Hacker News (front page)

What are tokens, and why does every AI bill list them?

A token is a chunk of text, often a word or part of one, that an AI model reads and writes one piece at a time. The word "unbelievable" might become three tokens: "un," "believ," "able." Companies charge per token because that's the actual unit of computing work the model performs, like a taxi meter charging per mile instead of per trip.

For you: Understanding tokens explains why AI subscriptions have usage limits and why long documents or chats quietly cost more than quick questions.

Original explainer

Google adds TPU chips built for AI 'agents' and security

Google introduced two new chip systems, TPU 8tx for encrypted, confidential AI processing and TPU 8i Max for handling very large amounts of text at once. They join two systems launched in April that already offered up to 2.7 times faster training and 80% better inference efficiency than the prior generation. Google now has nearly 100 TPU customers.

Follow the money: Custom chips built specifically for AI agents and regulated industries show Google racing to avoid depending entirely on Nvidia.

via Hacker News (front page)