AI Reading for Sunday December 21

Chinese open-weight LLMs, including Qwen3 Max, Kimi K2, Zhipu AI GLM-4.5, and DeepSeek, have reached near-parity with frontier proprietary models, barely distinguishable for many practical purposes and at most a few months behind. - ZDNET

Moore Threads Technology Co. introduced a new generation of chips aimed at reducing artificial-intelligence developers’ dependence on Nvidia Corp.’s hardware, just weeks after one of the most successful Chinese IPOs in years. - Bloomberg

crouching tiger, hidden 3rd-party benchmarks

Republicans on House Foreign Affairs Committee seek oversight over deals to send chips to China, others. - GlobalSecurity.org

hey don’t hog all the pay-to-play

Surveillance tools based on US tech, exported by China to countries like Nepal, are being used to monitor Tibetan refugees and restrict their freedom. - AP News

Scaling the Mickey Mouse Club: Disney and OpenAI signal the arrival of AI video streaming as users can create videos featuring more than 200 Disney, Marvel, Pixar and Star Wars characters, and Disney+ will stream a selection of user-made clips. - Scientific American

Sora Wars: The Mouse Awakens

Jony Ive said OpenAI’s secret consumer AI device could be ready within two years as Altman positions OpenAI to challenge Apple in hardware. - Fortune

OpenAI CEO Sam Altman said Gemini 3 helped OpenAI spot weaknesses, and competition has prompted recurring "code red" responses roughly once or twice a year lasting six to eight weeks, forcing OpenAI to reprioritize resources toward ChatGPT and accelerate product releases. - Mashable India

Harvard economist Jason Furman, who consults for OpenAI, compares the AI boom to the dot-com era and discusses AI's impact on jobs and related economic policy in a Q&A. - Bloomberg

Would You Trust a 22-Year-Old AI Billionaire With the Global Economy? My week partying with the young founders at the heart of the AI boom - The Atlantic

Google wanted to remove Assistant from most Android phones in favor of Gemini by the end of 2025, but says it needs more time - Engadget

Shield AI said its V‑BAT and Hivemind autonomous drone platform executed 35+ missions in 2025 identifying 200+ Russian targets, proving itself on the battlefield and looking to scale deployment and begin to arm them. - Fortune

field promotion to major algorithm

Consulting firm Challenger, Gray & Christmas says AI drove almost 55,000 U.S. layoffs in 2025 out of 1.17 million total cuts, with major firms—Microsoft (~15,000), Amazon (14,000), Salesforce (~4,000)—citing AI in restructuring. - CNBC

Yoshua Bengio, Turing Award-winning professor, said AI, citing ChatGPT, is 'a matter of time' to automate cognitive and desk jobs, could threaten trade jobs, democracy within years to decades, and companies including Intel, IBM and Google are revising hiring. - Fortune

bend it like Bengio

AI-powered screening tools are used by 78% of companies (2025 HRTech Outlook) and by 87% in at least one hiring component, employing on-demand video, audio and chat interviews that cut time-to-hire by 40% and produce ranked shortlists. - Quartz

congrats, you’re being interviewed and ghosted by an algorithm

Sen. Mark Kelly unveiled a 24-page AI roadmap proposing an "AI Horizon Fund" to tax frontier AI firms eliminating jobs to fund retraining and infrastructure, citing McKinsey's estimate that 12 million could be displaced by 2030 amid local data-center opposition. - Quartz

tax bots, train humans

Vibe finance: NYU's Robert Seamans says AI's disruption on Wall Street is largely 'smoke and mirrors' for now, even as major banks announce 2025 layoffs and Citigroup reports 54% of financial jobs have high automation potential. - Fortune

Alexander Vasylenko began training large language models in 2023 and now spends about 15–20 hours weekly creating and reviewing finance prompts, while companies pay prompt-writers $50–$70/hr and reviewers $90–$160/hr, reflecting sustained demand for human AI oversight. - Business Insider

Karpathy Wrapped: year in review - karpathy

Andrej Karpathy advised students to prioritize internships, research, open-source projects and recommendation letters over grades, reflecting hiring trends favoring demonstrable skills and a U.S. poll showing 63% view four-year degrees as not worth the cost. - mint

my wrapped: scrolled a lot, accomplished nothing

The National Republican Senatorial Committee released an AI-generated deepfake ad portraying Maine Gov. Janet Mills giving a young boy an estrogen kit, posted on YouTube without an AI disclosure, prompting warnings, including Public Citizen's Robert Weissman, about deepfakes. - LGBTQ Nation

The Bondi Beach attack sparked a global misinformation crisis that, a leading counterterrorism expert warned, is being amplified by "agentic AI" and political point‑scoring, fueling AI-driven hate, extremist coordination and strain on international cooperation. - South China Morning Post

Tech Against Terrorism researcher Lucas Webber says extremist groups use generative AI like ElevenLabs and ChatGPT to create lifelike cloned audio—including English-cloned Hitler speeches with tens of millions of streams—and multilingual propaganda that accelerates recruitment. - the Guardian

the Emmy for worst remix goes to…

Race-based AI medical diagnosis disparities: "We found that because AI is so powerful, it can differentiate many obscure biological signals that cannot be detected by standard human evaluation” - Futurism

on the Internet everyone knows you’re a dog now, and what your farts smell like

AI models compete for your affection: Inside the race to build the world’s best friend - The New York Times

A New England Journal of Medicine paper warns that relational AI chatbots pose risks—emotional dependency, reinforced delusions, addictive behaviors and self-harm—and call for regulation and research. - Futurism

Anthony Duncan, 32, says after using ChatGPT from May 2023 he became dependent by fall 2024, developed escalating delusions, lost his job, was hospitalized four days, and recovered with medication and family support; clinicians warn AI can reinforce psychosis. - Newsweek

disrupting human relationships and emotional well-being as a service

Dorks LARP as normies, pretend they are cool with Nano Banana - Futurism

METR: We estimate that, on our tasks, Claude Opus 4.5 has a 50%-time horizon of around 4 hrs 49 mins (95% confidence interval of 1 hr 49 mins to 20 hrs 25 mins). Task-length capability has doubled about every 7 months over six years. - METR

more on task than a PhD student popping handfuls of Ritalin

Claude in Chrome is a Google Chrome extension (beta for paid subscribers) integrating with Claude Code and Desktop that navigates pages, clicks buttons, fills forms, analyzes data and automates browser workflows while complying with Chrome Limited Use policy. - Claude

hard for e.g. Amazon to block agents when they control your own Chrome. also beware of agents running amok. don’t exfiltrate my browser history anywhere

Cursor buys AI code-review startup Graphite - TechCrunch

Tesla scales its robotaxi test, recruits safety drivers from its assembly line to the taxi line. - Business Insider

After SF blackout impacts traffic lights, Waymos make it much worse by stopping in the middle of intersections. - Mission Local

Viral footage shows Tesla's Optimus humanoid knocking over water bottles and collapsing, while Tesla and Elon Musk promote it as central to Tesla's future and China pushes humanoids as an economic engine, though practical deployment remains long-term and narrow. - The Verge

I'll believe when I see robots demo-ing e.g. vacuums in controlled sales floor environments. Great for selling since humans are programmed to be attentive to human-like objects.

2 legs, 0 PMF.

Walking on 2 feet is a useless skill in practice.

In indoor environments, wheeled or static robots with working hands would be actually useful. In unimproved outdoor environments, 4 legs maybe better.

Danny Moses, former FrontPoint Partners trader featured in The Big Short, warns an AI bubble is forming and advises selective exposure to cash‑flowing Big Tech while avoiding debt‑heavy AI suppliers and considering long‑lead uranium infrastructure investments. - Business Insider

Gary Marcus: Six (or seven) predictions for AI 2026 from a Generative AI ‘realist’ - The Potential Surface

just because a lot of people are crazy, it doesn’t make you a psychiatrist, and just because there’s a lot of crazy hype, it doesn’t make you a realist

Follow the latest AI headlines via SkynetAndChill.com on Bluesky