AI Reading for Tuesday December 9

Art of the GPU deal: Trump approved export of Nvidia's H200, its second-most-powerful AI chip, to China after months of deliberation, with the Commerce Department finalizing export details and six senators introducing legislation to restrict such sales. - The New York Times

Trump greenlights Nvidia H200 AI chip sales to China for 25% cut, says Xi responded positively - CNBC

Unsurprisingly, Beijing plans to restrict access to Nvidia's H200 AI chips, Chinese regulators said, even after the Trump administration approved their export, as China pushes for semiconductor self-sufficiency and tighter control over high-end AI hardware imports. - Financial Times

only locavore indigenously crafted GPUs for us thank you very much

AI Overview Overreach? The European Commission opened a probe into Google's AI summaries, investigating whether Google used website and YouTube content to train generative models without offering compensation or opt-outs after reports AI Overviews cut publishers' traffic about 50%. - BBC

Trump announced on Truth Social he will sign a "ONE RULE" executive order this week to establish a single federal AI rule preempting state AI laws, Reuters reported. - Above the Law

David Sacks urged federal preemption of state AI laws—citing over 1,200 state bills and more than 100 measures passed—said it won’t preempt zoning of datacenters, or child-safety, with various carve-outs for communities, creators and censorship , and announced President Trump will sign an executive order this week. - X (formerly Twitter)

states rights much

Esperanto for Agents: OpenAI, Anthropic, Google Agree to Develop Agent Standards Together - The Information

Sam Altman said OpenAI plans to end its "code red" after releasing a model in January 2026 with improved image generation, speed and personality. - The Wall Street Journal

OpenAI and Instacart launched an in-ChatGPT grocery shopping feature that lets users brainstorm meals, create grocery lists and complete checkout via OpenAI Instant Checkout, with OpenAI taking an undisclosed small fee on sales. - TechCrunch

OpenAI pivots from financial market savior to object of market worry. “The flurry of activity rightfully brought scrutiny and concern over whether OpenAI can fund all this. The timing of its revenue growth is uncertain, and every improvement a competitor makes adds to the risk that it can’t reach its aspirations.” - Bloomberg

Apple, after falling 18% through June to be the Magnificent Seven's second-worst performer in H1 2025, has rallied 35% since, as investors reward its slower AI spending for capital discipline, according to portfolio manager John Barr. - Bloomberg

one man's short-term focus is another's prudent market discipline

Dozens of Apple AI engineers, including Ke Yang, have left for OpenAI and Meta, weakening Apple's AI capabilities and undermining Tim Cook's plan to catch up, while Apple aims for a more personalized Siri in March 2026. - BGR

Apple will expand its Fitness+ service to 28 new markets on December 15, add K-Pop workouts, and use generative AI to provide trainer-voiced Spanish and German dubbing (Japanese dubbing early 2026), with sessions switchable by users. - 9to5Mac

Tim Cook either really cooking, or cooked

Holy guacomole, Zuck-man: Meta is developing a frontier AI model codenamed Avocado, expected in Q1 2026, that may be proprietary, while restructuring its AI teams and raising 2025 capex guidance to $70–72 billion. - CNBC

Google is testing Gemini 3.5 and Gemini 3.0 Flash, with Gemini 3.0 Flash—expected before year-end 2025—offering extended-context, browser-native OS integration, interactive 3D and improved coding/reasoning, while Skyhawk/Seahawk show precise 3D design but visual inconsistencies. - Geeky Gadgets

Google told advertisers it plans to introduce ads into its Gemini AI chatbot with a targeted 2026 rollout, according to agency buyers, but has not shared formats and said there are no plans for ads in the Gemini app. - Adweek

Google will add a second Gemini-based model, the User Alignment Critic, to Chrome to double-check agentic plans and veto dangerous actions, extend origin isolation via Agent Origin Sets, require confirmations for sensitive actions, and raise vulnerability rewards to $20,000. - The Register

I tried Google’s prototype smart glasses and it almost made me forget about my phone - CNN

Google plans both audio-only and screen-based smart glasses - CNBC

The Information reported Microsoft cut internal Azure AI sales forecasts after FirstPageSage data (early December) showed ChatGPT at 61.3%, Microsoft Copilot 14.1% (2% quarterly growth) and Google Gemini 13.4% (12% quarterly growth). - Windows Central

Cope-pilot? Microsoft Copilot experienced a major outage Tuesday morning (UK time), disrupting Copilot access in Microsoft 365 and integrated services and prompting user reports on Down Detector while Microsoft has not confirmed the cause or restoration timeframe. - The Independent

few seem to have noticed

A profile of Liang Wenfeng of DeepSeek, as part of Nature's list of 10 people who shaped science in 2025 - Nature

Authors of the original Transformer paper at an AI startup debuted an open-source model. - Bloomberg

Great that they are competing with top China labs, but 8b is of most interest to fine-tune for an enterprise task, or on an edge device. If they could do something that beats Gemma 27b I can run on MacBook, or GPT-oss-120b that runs on 1 Nvidia GPU, they would have my…attention.

Gemini 3 is nice but cheap open source models are the sleeper story of 2025, didn't beat frontier models but very good for practical enterprise purposes.

GPT-oss-120b is competitive with GPT-5-mini (medium), which is adequate for a lot of tasks, can run on an Nvidia DGX box or an 80GB A100.

At NeurIPS 2025, which drew about 26,000 attendees, researchers and executives said interpretability and measurement remain unresolved, prompting strategy splits—Google pivoted toward pragmatic methods, OpenAI doubled down—and Martian launched a $1 million interpretability prize. - NBC News

AISLE's AI autonomous analyzer discovered and helped fix a three-year vulnerability in CryptoLib that, requiring local access, could let attackers inject arbitrary commands to seize control of or intercept data from NASA spacecraft, including Mars rovers. - Space

hackers missed their chance to take Discovery on a joy ride

An editorial in the Shallow Review of AI Safety states post-training (RL) runs outperformed large pretraining by ~30×, and inference HBM limits model scale with compute-optimal active params ≈1T at 100K H100s and ≈4T at 600K Ironwood TPUs. - Less Wrong

Psycho-analyzing LLMs reveals post-training as a form of trauma - arXiv

The slop thickens: AI videos flood social media - The New York Times

The International Committee of the Red Cross warned that major AI models, including ChatGPT, Gemini and Copilot, fabricate research papers, journals and archival references, prompting misleading record requests that increase archival workloads and undermine trust. - Scientific American

people are getting big mad at the librarians for not finding documents surfaced by large language models that do not exist.

A study of more than 11,000 13–17-year-olds in England and Wales found one in four used AI chatbots for mental health support, rising to about 40% among teens affected by youth violence and higher use among Black children. - the Guardian

Families of two US teens are suing OpenAI over allegedly harmful ChatGPT exchanges, research finds one in four 13–17‑year‑olds in England and Wales asked chatbots about mental health, and UK science minister Liz Kendall urged Ofcom to apply online-harms rules. - the Guardian

When your fridge addresses you by name, saying "We're sorry we upset you, Carol", leading you to think you are having a psychotic episode and hospitalizing yourself. - Reddit

about as likely to make you crazy as help you

Noom launched Face Scan and Future Me in October 2025, using rPPG via NuraLogix and Haut.ai to estimate biomarkers; CNET's test returned a biological age of 44 for a 37-year-old and the app directed users toward paid interventions. - CNET

a lot of AI pseudo-medical snake oil AI

Harvard Business Review Analytic Services' July 2025 survey of 603 business and technology leaders found only 6% fully trust AI agents to autonomously run core processes, while 86% expect AI agent investment to increase over the next two years. - Fortune

Cursor CEO Michael Truell said the $29.3 billion startup has automated roughly 80% of internal support tickets, runs an AI-powered company-wide Q&A, crossed $1 billion ARR and grown to over 300 employees. - Fortune

Former Slack CIO at Samsara reported pilots showing 59% reduction in support chat, 27% auto-resolved IT tickets and ~40% acceptance of AI-suggested code, and advised treating AI as a business-strategy accelerator with a VC-style portfolio expecting ~10% high-return pilots. - Fast Company

Coder launched a self-hosted governed AI development stack for hybrid human/agent teams—featuring AI Bridge, Agent Boundaries and expanded Coder Tasks—and said it currently supports about 3,000 concurrent workspaces (roughly 10,000 developers). - SiliconANGLE

Anthropic brings enhanced Claude Code integration to Slack - The Verge

iFixit launched FixBot, an AI repair assistant trained on 125,000 guides that accepts user photos and voice, displays schematics, fetches external data, and is available free in beta on iOS and Android switching to a $4.99/month or $50/year freemium plan. - Engadget

hope it doesn’t hallucinate and make me brick something or electrocute myself

World's first actual smart phone, with AI and agents built in? - threadreaderapp.com

Disorder in the court: As famously illustrated in Dept of Magic v. Harry Potter, LLM hallucinations producing fictitious case citations have eroded trust in legal practice, and a Cornell study found verification costs may offset AI savings, prompting calls to limit high-risk AI use, strengthen verification, and restrict deployments to low-risk tasks. - Above the Law

In-house legal teams are adopting generative and agentic AI for contract review, compliance and litigation support, with firms like Salesforce reporting about 9,500 hours saved annually, invoice-review automation cutting operating costs ~30% and Marsh McLennan saving $9m. - Financial Times

Major law firms such as Quinn Emanuel, WilmerHale and Willkie use AI to analyze massive evidence—about 70,000 documents and 20–40 depositions—saving hundreds of hours (Thomson Reuters: ~240 hours per lawyer yearly) but raising accuracy and confidentiality concerns. - Financial Times

Mercedes is partnering with Chinese software developer Momenta to equip Lumo Mobility with Level 4 self-driving cars in Abu Dhabi, marking Mercedes' first foray into robotaxis. - Bloomberg

AI image generators and chat tools such as ChatGPT, MidJourney and Gemini are reshaping home-design inspiration—research by Tilanka Chandrasekera found non-designers produced the most creative, satisfying AI-assisted projects, even as outputs sometimes include impossible or unbuildable details. - Yahoo

Second Stage's Broadway premiere of Jordan Harrison's Marjorie Prime opened at the Helen Hayes Theater starring June Squibb and directed by Anne Kauffman, with only light tweaks since its 2015 Off-Broadway run and renewed debate over AI 'Primes'. - The New York Times

60 Minutes covers Character.ai - YouTube

AI imitates, mixes and matches successful human reasoning and explaining patterns but never says "hmmh that's weird" which is the jumping-off point for creativity and innovation. - Fair Observer

Would AI shoot you if you gave it a gun? "Any advanced AI would logically eliminate the only threat to its existence." - YouTube

Cracked me up … a bit jarring how it mixes the absurdist humor with the apocalyptic ... surely we're not going to put the fate of the world in the hands of easily tricked Cheech and Chong AI interns, are we?

Einstein: The definition of insanity is doing the same thing over and over again and expecting a different result.

AI: hold my beer

Follow the latest AI headlines via SkynetAndChill.com on Bluesky