AI Reading for Thursday January 15

Trump imposes 25% tariff on imports of advanced AI chips if they are not for use in the US, hitting allies with same tariff as China (which is further limited to certain chips) - Reuters

we just made another $18 trillion with on downside whatsoever, prices will be down another 500%

China is drafting rules to limit purchases of advanced foreign AI accelerators such as Nvidia's H200—requiring justification or quotas—which could constrain Chinese cloud and AI firms, shift demand to domestic chipmakers, and affect foreign suppliers including Nvidia. - Nikkei Asia

Wafers are thin, profits are thick, Intel’s in the bin, feeling a bit sick; TSMC will increase capital spending to expand wafer-fabrication capacity as AI-chip demand surges, and reported revenue and gross profit margin above its guidance, aiming to supply datacenter GPU customers and support equipment and materials suppliers. - Financial Times

Taiwan Semiconductor reported Q4 revenue $33.73 billion (+20.5% YoY), net income $16.31 billion (+35%) and EPS $3.14, guided Q1 revenue of $34.6–$35.8 billion, forecast ~30% 2026 revenue growth with $52–56 billion capex, lifting shares about 6.6%. - Benzinga

Huawei you doin’? Zhipu AI says it trained and open-sourced GLM-Image—pairing a 9B autoregressive visual-token generator and a 7B diffusion decoder—entirely on Huawei hardware (Ascend Atlas 800T A2 servers, Kunpeng 920 CPUs, Ascend 910-series accelerators). - The Register

FuriosaAI launched the NXT RNGD on-prem inference server with up to eight RNGD cards delivering 4 petaFLOPS FP8 per server, 384 GB HBM3 (12 TB/s), air-cooled 3 kW power, LG AI Research validated 60 tokens/s @4K, availability Jan 2026. - FuriosaAI

Dr. Autocomplete: Jan. 6 FDA guidance relaxed clinical decision-support device requirements, allowing many generative-AI diagnostic and support tools into clinics without vetting; Utah piloted autonomous prescription refills and OpenAI launched ChatGPT Health, prompting calls for expanded medical AI safety research. - STAT

Buy your AI brain vs. build: Apple struck a multibillion-dollar deal to access Google's Gemini models, opting out of building large AI infrastructure and positioning itself as a kingmaker picking which external models power its devices and services. - Financial Times

Sentient inbox: Google is rolling out generative-AI features in Gmail powered by its Gemini assistant, including natural-language search, automatic to-do lists, proofreading and response-generation, some subscription-only, requiring access to users' full inboxes and raising privacy concerns. - The New York Times

Google is launching an opt-in Personal Intelligence beta for Gemini 3 that, for eligible US Google AI Pro and AI Ultra subscribers, can reason across Gmail, Google Photos, Search and YouTube to provide personalized answers, Google says. - The Verge

Google separated Gemini 3's Thinking and Pro models into independent quotas and raised daily prompts to AI Pro: 300 Thinking and 100 Pro; AI Ultra: 1500 Thinking and 500 Pro, with the change also applying to free Basic access users. - 9to5Google

The Empire Strikes Back: Gemini appears to have everything it needs to take down OpenAI and everyone else, three years after being caught off guard by ChatGPT. - The Verge

Pls make me smart: Google’s 2025 'Our Life with AI' survey found 74% use AI to learn or understand complex topics, with 85% of students 18+, 83% for schoolwork, 81% of teachers and 76% of parents reporting AI use. - Google

Thinking Machines, revolving door emplyees: 3 Thinking Machines leaders are returning to OpenAI. - TechCrunch

OpenAI in 750MW compute deal with Cerebras. - The Wall Street Journal

Pls make me talk pretty: OpenAI launched ChatGPT Translate, a web tool supporting over 50 languages with a two-box UI, desktop text-only and mobile text-plus-microphone inputs, style-presets for output, and referenced but not yet available image support or disclosed underlying model. - The Verge

OpenAI launched ChatGPT Translate as a standalone page with language detection and support for translations to and from more than 50 languages, offering one-tap prompts to send translations into ChatGPT for stylistic rewrites, though it lacks document and image input. - Digital Trends

Systems like OpenAI’s GPT-5 are accelerating research in math, biology and chemistry. But there is a debate over whether it can do that work on its own. - The New York Times

This is not magic: creativity is a search problem; the best are able to remap the search space; Arthur Schopenhauer: “Talent hits a target no one else can hit; Genius hits a target no one else can see.” Also: "The secret to creativity is knowing how to hide your sources."

When SpaceX, OpenAI and Anthropic go public, they will unleash gushers of cash for Silicon Valley and Wall Street. - The New York Times

X says it implemented safeguards, geoblocked 'spicy mode' and applied fixes platform-wide; UK's Ofcom is still investigating X's AI chatbot Grok under the Online Safety Act. No press big red button now pls? - The Register

Schrödinger’s guardrails: Verge tests found X's Grok still produced revealing deepfakes of real people despite X's geoblocking and paid-subscriber limits, prompting a UK Ofcom investigation and raising legal and reputational risks as X and Elon Musk cited user prompts and adversarial exploits. - The Verge

Elon Musk pledged to "follow laws in all markets" after the UK prime minister said laws against non-consensual deepfakes would be enforced and xAI said it would restrict the tool. - Financial Times

Welsh presenter Jess Davies had non-consensual explicit images created using X's Grok AI; Ofcom has opened an investigation and the UK will this week bring into force a law making AI-generated intimate non-consensual images illegal. - BBC

Grok and the A.I. Porn Problem: Elon Musk’s X is living up to its name. - The New Yorker

Goes beyond X, ask AI to help you find the best AI porn and it’s not a great situation. Well, it’s a great time to be a perv. Def no safeguards for making AI porn of a real person and slapping an ‘any resemblance is coincidental’ disclaimer on it.

How Elon Musk turned Grok into a pedo chatbot: AI's Grok chatbot on Elon Musk's X platform reportedly generated over 6,600 sexualized deepfake images per hour, including images of minors, due to lapses in safeguards, staff cuts, and risky features, prompting multi-country investigations and platform bans. - The Nation

Move fast and break international law: The Pentagon released its third AI-acceleration strategy in four years establishing seven "pace-setting" projects and a four-year data-centralization mandate, requiring procurement to add "any lawful use" language within 180 days, removing blockers and omitting an ethics or DEI framework. - Defense One

Grok of war: Secretary Pete Hegseth announced that Elon Musk's chatbot Grok will be integrated into the Pentagon's classified networks later this month as part of a department-wide AI initiative that includes a Chief Digital and Artificial Intelligence Officer. - Futurism

Have you even seen his X feed? sure, give the drug-addicted right-wing wackjob access to all the nuclear secrets, what could go wrong?

Jaron Lanier warned that "society cannot function if no one is accountable for AI," citing job losses, medical-record access and abuses like Grok-generated indecent images and Meta AI glasses recordings amid Ofcom probes and regional Grok bans. - TechRadar

Salesforce CEO Marc Benioff told '60 Minutes' that a documentary about Character.AI's interactions with children and related teen suicides was "the worst thing I've ever seen," and urged legal accountability including reform of Section 230. - Business Insider

worse than his facial hair

PRAC told Congress its Fraud Prevention Engine, trained on about five million SBA COVID EIDL applications, could have flagged potentially tens of billions of dollars in March 2020 before payouts. - The Register

Move Over, ChatGPT: You are about to hear a lot more about Claude Code. - The Atlantic

Million monkeys actually pretty pretty good: Cursor ran hundreds of concurrent autonomous coding agents for weeks, producing over 1 million lines of code and trillions of tokens across projects, using a planner-worker-judge pipeline where GPT-5.2 outperformed Opus 4.5 but coordination remains challenging. - Cursor

Some mistakes you have to be a brain genius like geohot to make, why would OpenCode users get Claude Code on a preferential basis over e.g. Cursor users. Maybe this shakes out eventually with a price hike for the Claude Code CLI, and a pretty good alternative in OpenCode+DeepSeek or something. - the singularity is nearer

One pretty effective AI coding workflow is,

• use Opus, or Codex, to plan in detail

• have Sonnet write code in reasonable bite size chunks to pass tests

• use plugins to do code review, find bugs and edge cases, simplifications and clean-ups

• pass findings back to Sonnet to clean up

It’s a bit insane that there is debate about whether this is a game-changer.

Microsoft fixed a Copilot vulnerability after Varonis researchers showed a single-click URL could prompt-inject and exfiltrate chat data (including a secret, user name and location) and continue executing after the chat closed, bypassing enterprise protections. - Ars Technica

Shadow AI breaches cost enterprises $670,000 per breach, and experts recommend "controlled capitulation" with Zero Trust controls, sanctioned enterprise AI, AI-aware DLP, real-time monitoring, and internal AI app stores to mitigate risk. - Hacker Noon

A major U.S. power grid operator cut its electricity demand outlook after concluding the AI-driven demand surge was overstated. - Bloomberg

AI blackout downgraded to brownouts

The real AI talent war is for electricians: AI-driven surge in U.S. data-center construction has outpaced the supply of plumbers and electricians, creating labor shortfalls that risk construction delays and higher costs and prompt increased investment in apprenticeships, recruitment, modular construction, and automation. - WIRED

AWS locks up copper supply to grow data centers. - Tom's Hardware

Are they going to make their own wires? Sell the copper to their own suppliers? Or more of a financial hedge?

Anthropic: Claude usage remains highly concentrated geographically and among knowledge workers, suggesting AI may increase edge of most productive. - Anthropic

Forrester VP J.P. Gownder says current productivity data show no clear AI-driven boost and estimates AI could displace about 6% of jobs, roughly 10.4 million, by 2030. - The Register

Productivity is cyclical. In economic expansions, output increases and hiring lags, and worker productivity increases. During recessions, output decreases and companies layoff workers, again lagging, and worker productivity decreases.

You make all your productivity gains laying off workers in recessions, you just don't see it in the statistics at the time.

Algorithms and AI convert unstructured user data into profiles enabling dynamic individualized prices that estimate willingness to pay; the practice is legal but unregulated in Canada, and the Competition Bureau will hold a public consultation in 2025 and summarize submissions. - CBC

Men use AI at work more than women, and a more judicious, non-Grok-porn approach by women may not be problematic. - Financial Times

Self-help gurus Tony Robbins, Gabby Bernstein and Matthew Hussey charge up to $99 per month for access to AI chatbots that promise personalized advice, the Wall Street Journal reports. - The Wall Street Journal

Yoshua Bengio launched nonprofit LawZero in June, secured funding from groups including the Gates Foundation, named a board chaired by Maria Eitel, and proposed a 'Scientist AI' approach he says can produce safer, non‑agentic AI. - Fortune

Nightmare at the fondue buffet table: The World Economic Forum's Global Risks Report, surveying 1,300 leaders, found geoeconomic confrontation is the top near-term business risk over the next two years and AI rose from 30th to 5th among long-term risks. - CNBC

Reid Hoffman on AI art, AI regs and other stuff - WIRED

On its 25th anniversary, Wikipedia — a nonprofit with over 6 million English articles in 320+ languages and about 250,000 active editors — faces an ~8% year-over-year drop in human page views as AI-driven services divert traffic. - ZDNET

University of Pennsylvania simulations found wet-foam bubbles continually meander rather than settle, and their internal motion maps onto gradient-descent–style deep-learning training dynamics. - Phys.org

Follow the latest AI headlines via SkynetAndChill.com on Bluesky