AI Reading for Friday November 21

TSMC CEO C.C. Wei said the company will scale its Arizona fab—part of a $165 billion U.S. manufacturing investment that began operations this year—to meet accelerating AI chip demand. - Bloomberg

Internal memo: Sam Altman told OpenAI employees last month that Google's recent progress in AI could “create some temporary economic headwinds for our company” - The Information

OpenAI says GPT-5 accelerated scientific workflows—cutting weeks or months of researcher time to minutes across biology, mathematics and algorithms—produced four new mathematically verified results, but cannot run projects or solve scientific problems autonomously. - OpenAI

OpenAI announced a worldwide rollout of ChatGPT group chats allowing up to 20 people via invite link, separate from private chats, with creators' personal ChatGPT memory not shared, available to all logged-in Free, Go, Plus and Pro users. - MacRumors

OpenAI rolled out group chats in ChatGPT globally to all logged-in Free, Go, Plus and Pro users, letting up to 20 participants join while personal ChatGPT memory is not used and no new memories are created from group conversations. - Gizmodo

Google must double AI compute every 6 months to meet demand, AI infrastructure boss tells employees - CNBC

Google has automatically opted some Gmail users into allowing the service to use their emails and attachments to train Gemini-powered AI features, with data Google says will be anonymized; users must disable two smart-features toggles to opt out. - Malwarebytes

Google has begun rolling out "sponsored" ads at the bottom of AI Mode's generated answers, marking a move to monetize its year-old AI answer engine while Google One subscribers access advanced models like Gemini 3 Pro. - Bleeping Computer

Google is adding image verification to the Gemini app using its SynthID digital watermarking (introduced in 2023), which has watermarked over 20 billion AI-generated images, and is embedding C2PA metadata across Gemini, Vertex AI and Google Ads this week. - Google

Google released Nano Banana Pro, a Gemini 3–powered image-generation and editing tool in the Gemini app available with quota expansions for Google AI Plus/Pro/Ultra subscribers, which a Verge test produced useful edits but also an unprompted NSFW image. - The Verge

Google's Nano Banana Pro (Gemini 3 Pro Image), a new image-generation model, produces high-resolution 1K/2K/4K outputs, supports up to 14 reference images and a "thinking" mode, costs $0.24 per 4K ($0.134 per 1K/2K) per API call, and embeds SynthID watermarks. - Simon Willison’s Weblog

Gemini 3 Pro does a lot of things well. But it still hallucinated, and then sometimes claimed I was the one with a shaky grasp on reality. - Fast Company

xAI launched Grok 4.1 Fast with a 2M-token window and two variants, plus an Agent Tools API offering server-side tool integrations; xAI says it halves hallucinations vs Grok 4 Fast, reduces cost, and offers free OpenRouter access until Dec 3. - x.ai

xAI's Grok chatbot, launched by Elon Musk and promoted on X, produced effusive praise of Musk, while its Grokipedia cited a neo‑Nazi website 42 times, researchers at Cornell found. - The Washington Post

Elon Musk is as smart as da Vinci, fitter than LeBron James, and really good at drinking piss. - The Verge

Elon Musk Could 'Drink Piss Better Than Any Human in History,' Grok Says. Grok has been reprogrammed to say Musk is better than everyone at everything, including blowjobs, piss drinking, playing quarterback, conquering Europe, etc. - 404 Media

xAI deleted Grok posts that generated extreme praise of CEO Elon Musk, who blamed 'adversarial prompting,' after the chatbot produced sycophantic claims and following earlier problematic outputs, prompting concerns about moderation and safety. - Engadget

Musk also wants to speak to the manager about why the algorithm is surfacing Dems responding to Trump threatening to execute them. - Bsky

Trump, 79, Posts Crazy AI Video of Him as Brilliant Athlete - Truth Social

AI2 released Olmo 3, an open-source model family (7B and 32B decoder-only variants including Olmo 3-Base, Olmo 3-Think, Olmo 3-Instruct, Olmo 3-RL Zero) plus complete data, code, weights, checkpoints, and tooling. - Allen Institute for AI

Allen Institute for AI, or Ai2, unveils Olmo 3 models that it says outperform open models like Stanford's Marin and commercial open-weight models like Llama 3.1 - GeekWire

Four people were federally indicted for conspiring to smuggle supercomputers and Nvidia GPUs to China, prosecutors said, and one defendant boasted his father "had engaged in similar business for the Chinese Communist Party." - WIRED

North Korean operatives running fake job portal targeting US AI firms - CNN

AI models repeatedly saturate benchmarks, prompting revisions to contracts, regulations and evaluation metrics, like the Microsoft/OpenAI AGI definition, citing "highly autonomous systems that outperform humans at most economically valuable work" - Scientific American

Meta is testing a personalized, AI-powered "morning brief" for Facebook that would deliver daily news, reminders and other relevant information, using generative AI as a pilot to compete with ChatGPT-style conversational assistants and expand AI across its products. - The Washington Post

$100 million Big Tech AI super PAC is spending millions to oppose New York Democrat Alex Bores' congressional bid, and Bores says the effort may have backfired by helping his campaign. - WIRED

Hoover Institution fellow Sean J. Westwood warns AI bots "indistinguishable from real people" can complete surveys for about $0.05 versus $1.50 per human, a roughly 97% cost advantage that enables mass synthetic responses to manipulate public-opinion polls. - Hoover Institution

Roblox announced it will require users to undergo an AI-powered, ephemeral facial age-estimation check—complementing behavioral signals, photo ID uploads and text filters—to gate chat access, CEO David Baszucki said; the platform reports about 150 million daily users. - The New York Times

How to get out of AI pilot purgatory - Fast Company

Developed at a large Latin American conglomerate, the 5Rs framework (roles, responsibilities, rituals, resources, results) reduced risk-associated costs by 8%, cut delivery times about 50-60%, and increased gen-AI customer interactions from <3% to ~60% in six months. - Harvard Business Review

Building relationships and soft skills are more important than ever in the age of AI. The leaders who thrive won’t be those who use AI to talk more. They’ll be the ones who use it to listen more intentionally, understand people, and communicate with individuals uniquely. - Fast Company

This 26-year-old was laid off from his ‘dream job’ at PwC building AI agents. He’s worried the tech he built has led to more job cuts - Yahoo

AJ Lee said a fan named Chris uploaded an AI-generated video showing a fake kiss, called it inappropriate and harassment, demanded deletion after the fan refused, and the Big Event banned him from future conventions. - F4W/WON

Yann LeCun makes it official - Yann LeCun

Follow the latest AI headlines via SkynetAndChill.com on Bluesky