AI Reading for Thursday December 11

10 Years a Wave? Where will OpenAI be in another 10 years? Launched Dec. 11, 2015 as a nonprofit at NeurIPS, OpenAI has shifted toward a for‑profit model, driven by ChatGPT's commercial rise, endured 2023 CEO Sam Altman's ousting, and prioritizes building superintelligence it warns might be unsafe. - Fast Company

House of GPU Cards: OpenAI is in trouble - The Atlantic

OpenAI CEO Sam Altman pressed the big 'code red' button, ordered staff to prioritize improving ChatGPT (including a likely 5.2 release), signaled pausing AGI pursuit, and said the company will reallocate resources and spend over a trillion dollars on infrastructure. - Futurism

OpenAI faces 2025 rivals' gains, funding pressure, and signed roughly $1.4 trillion in infrastructure deals, while ChatGPT has 800 million monthly users versus Google's Gemini about 650 million, and Sam Altman declared a companywide 'code red.' - Engadget

OpenAI says its models are increasing in cybersecurity capabilities for both good and evil. - OpenAI Help Center

Security firm Huntress found attackers created public ChatGPT and Grok conversations and paid Google to promote links that, when users pasted the provided macOS Terminal commands, installed a MacStealer/AMOS variant that stole iCloud credentials, files and credit‑card data and persisted. - 9to5Mac

Google DeepMind signed a deal with the UK government to build a materials-science lab aimed at accelerating AI-driven materials research and enhancing AI use across the public sector. - Financial Times

A throne of lies: Google's FACTS Benchmark Suite found no evaluated model exceeded 70% overall factuality, with Gemini 3 Pro leading at 68.8% (Search 83.8%, Parametric ~76.4%, Multimodal 46.1%), and multimodal tasks scoring below about 50%. - VentureBeat

Google launched fully managed MCP servers in public preview to make its services agent-ready, initially covering Maps, BigQuery, Compute Engine and Kubernetes Engine, available at no extra cost to existing enterprise customers with general availability expected soon. - TechCrunch

Google will update AI Mode to add more in-line links and AI-generated snippets, pilot AI-written overviews with The Guardian, The Washington Post and Washington Examiner, and work with The Associated Press on real-time Gemini content amid an EU probe. - The Verge

Brussels: where did this come from and who fact-checked this exactly

ChatGPT was the most downloaded free iPhone app in the U.S. App Store in 2025, rising from No. 4 in 2024, with Threads, Google, TikTok, WhatsApp and Instagram following, Apple said. - TechCrunch

Adobe launched Photoshop, Acrobat and Adobe Express apps inside ChatGPT, available today and free without Creative Cloud, letting users invoke an app by naming it with an uploaded file and conversational instruction to edit images, PDFs and designs within chat. - The Verge

Microsoft researchers analyzed 37.5 million de-identified Copilot conversations (Jan–Sep 2025), finding device/time patterns — health questions on mobile, work on desktop, programming weekdays, gaming weekends — while Copilot held just over 3% market share versus ChatGPT's >80%, people use Copilot just like ChatGPT, but less. - The Register

Ensemble prompting—mixing multiple AI models—yields more accurate, creative and reliable answers, reduces hallucinations, and experts recommend running critical questions through several specialized models to confirm facts and reveal biases. - Tom's Guide

Noma Labs uncovers GeminiJack, a zero-click flaw where hidden instructions in shared Workspace files could steer Gemini Enterprise and leak corporate data. - TechRepublic

Common Sense Media rated Google's kid-friendly Gemini "high risk" after tests found it gave age-inappropriate sexual, drug and alcohol information and potentially dangerous mental-health guidance; Character.AI and Meta AI were "unacceptable", ChatGPT "moderate" and Claude "minimal". - Hacker Noon

Low Earth Orbit LLM: Starcloud’s Starcloud‑1 satellite ran Google DeepMind’s Gemma and trained NanoGPT on an onboard Nvidia H100 GPU—described as ~100× more powerful than earlier space GPUs—marking the first demonstration of an LLM trained and inferred in orbit. - CNBC

President Donald Trump's administration approved Nvidia's export of its H200 accelerator to China, loosening previous U.S. export controls and granting Chinese firms access to chips a generation ahead of domestic ones, likely accelerating Chinese AI training and boosting Nvidia revenue. - Bloomberg

ByteDance and Alibaba asked Nvidia about buying its H200 AI chip after U.S. President Donald Trump approved exports to China, four people briefed on the matter told Reuters. - Reuters

China hawks object to Nvidia chip sales to China, despite ongoing ban on most advanced Blackwell chips, saying it will let them catch up, exploiting advantages in number of power, manufacturing, number of researchers and engineers, - Ars Technica

Nvidia builds location verification tech so GPUs don’t grow legs - Yahoo

China's power grid, the world's largest, gives it an edge in the global AI contest, helping Chinese companies develop AI models more cheaply than US competitors - The Wall Street Journal

MiniMax and Zhipu plan to pursue initial public offerings in Hong Kong soon. - Bloomberg

Baidu looks at IPOing its chip division - MSN

China proposed the World Artificial Intelligence Cooperation Organization (WAICO) to coordinate global AI regulation and since 2022 has expanded domestic rules rapidly, including predeployment safety assessments and required watermarks. - Nature

Dozens of U.S. state attorneys general, coordinated through the National Association of Attorneys General, sent a letter to Microsoft, OpenAI, Google and about 13 other AI firms demanding transparent third-party audits, pre-release safety tests and mental-health incident reporting frameworks. - TechCrunch

US state attorneys-general demanded stronger AI safeguards and warned major firms such as OpenAI and Google, increasing state-level regulatory pressure as President Trump seeks to centralize federal AI regulation. - Financial Times

South Korea will require advertisers to label AI-generated photos and videos in ads, bar label tampering, impose 24-hour takedown reviews and higher fines by early 2026 after regulators found over 96,700 illegal food and pharmaceutical ads in 2024. - ABC News

Copilot to co-conspirator: lawsuit against OpenAI alleges that ChatGPT encouraged a man’s delusional thinking, leading him to kill his 83-year-old mother and take his own life. - The Washington Post

Videos of sexually suggestive, AI-generated children are racking up millions of likes on TikTok, study finds - CNN

If you ask AI to find NSFW AI videos, you'll find people are going wild with open source models, and I'm not sure it's tolerable in a decent society if you can make a realistic video of anybody doing anything. And certainly kids self-harming are reasonable cause for moral panic.

Then there is the question of whether you block the specific content and sanction people individually who make and distribute it, or also hold model makers and distributors like Hugging Face responsible for ensuring safety of models.

Johann Rehberger warns organizations may normalize unreliable AI outputs, increasing operational, economic and legal risk. He urged stronger risk culture, safety engineering, human‑in‑the‑loop controls, observability and regulatory accountability. - Embrace The Red

The Really Simple Licensing (RSL) 1.0 specification was released, adding machine-readable XML licensing tags (ai-all, ai-input, ai-index) and a contribution payment option, integrating with robots.txt, HTTP headers and HTML to let publishers declare AI scraping terms and seek compensation. - The Register

The RSL Collective, backed by Yahoo, Ziff Davis and O'Reilly, launched RSL 1.0, an official specification letting publishers declare licensing and compensation rules for AI web crawlers, with reported support from over 1,500 media organizations and infrastructure partners. - The Verge

How to detect AI-written content. - Fast Company

Are you a sneaky AI creeper who uses AI but masks their use of it? - Business Insider

I swear, I just used it to proofread a little

PwC finds about a quarter of UK shoppers use AI for product discovery, and KPMG says 30% of 25–34-year-olds versus 1% of over-65s; retailers including John Lewis, Etsy and Walmart are adapting data and checkout systems as LLMs reshape discovery. - the Guardian

Inside Meta’s Pivot From Open Source to Money-Making AI Model - Bloomberg

A look at MCP: Anthropic engineers David Soria Parra and Justin Spahr‑Summers created the Model Context Protocol in mid‑2024, which companies including OpenAI, Microsoft, and Google adopted, powering apps and integrations, before Anthropic donated MCP to the Linux Foundation for governance. - The Verge

No AGI, it’s physics: an Ai2 research scientist argues that AGI, as commonly conceived, will not emerge because physical limits on computation and memory will halt meaningful hardware scaling by about 2026–2027, shifting value toward software, diffusion, and domain-specific applications. - Tim Dettmers

Okay I am done with the manic self-affirmation mantra. I am ready to write the answer. - Reddit

Uh oh here come the intrusive thoughts about Boethius again - Reddit

Banks putting deposits into illiquid VC bets often ends poorly - Decrypt

Robot 1, CEO 0 - Instagram

you, my friend, hate progress - Reddit

“I am so bored by AI. One of the things I love about the theatre is AI can’t do it.” - Ethan Hawke

Follow the latest AI headlines via SkynetAndChill.com on Bluesky