AI Reading for Monday December 22

Xi’s gotta have it: Nvidia told Chinese clients it plans to begin shipping about 40,000–80,000 H200 AI chips to China before the Lunar New Year in mid-February 2026, initially fulfilling orders from existing stock and pending government approval. - Reuters

Or not: The U.S. has been unable to persuade Chinese President Xi Jinping to adopt Nvidia chips. - The Wall Street Journal

SK hynix and Nvidia are jointly developing an AI-optimized SSD (code-named Storage Next/AI-NP) targeting about 100 million IOPS—roughly ten times current enterprise SSDs, with a prototype expected before the end of 2026. - TechSpot

because we’re so consistent and dependable, ya know

Slim revenues, fat IPOs: Filings show Chinese AI startups Zhipu and MiniMax reported slender 2024 revenues of $44.4 million and $30.5 million, respectively, and are preparing Hong Kong IPOs in the coming weeks. - Bloomberg

AI Autobahn: China launched the Future Network Test Facility on Dec 3, an open optical AI 'superhighway' spanning 40 cities and 34,175 miles of fiber supporting 128 networks and 4,096 parallel trials; the U.S. favors federated testbeds (NSF, DOE). - Forbes

Hyperscale datacenter operators nearly tripled infrastructure spending—quarterly capex rose ~180% to $142 billion and capacity additions ~170%—and Synergy forecasts total hyperscale capacity will double in about twelve quarters; Goldman Sachs warned of monetization risk. - The Register

Rising rack densities and per‑GPU power are shifting datacenter power from rack to facility scale, with Nvidia's DGX GB200 requiring about 120 kW, Google planning 1 MW racks, and transformer lead times near 28 weeks. - The Register

Since January 2025 at least 15 U.S. coal plants have postponed retirements as utilities and the DOE cite anticipated AI-driven data center demand—projects to supply about half of new electricity growth 2025–2030—after generators emitted 68 million tons CO2 in 2024. - naked capitalism

Analysts project Nvidia's GB300 roadmap and pre-buys will secure dominant GPU production, with CoWoS packaging capacity rising to 1,550 units in 2027 and Nvidia capturing about 61% of CoWoS volume and roughly 80% of the AI chip market. - SiliconANGLE

How Google leapfrogged ChatGPT – and why the AI race may already be over: For more than 20 years Google has been the gateway to the internet – Cloudflare CEO Matthew Prince says new rules are needed to prevent Google from abusing its monopoly to win the AI race. - The Independent

Diameter Capital Partners, which manages $25 billion, says AI demand extends beyond chips and data centers and has profited by buying 2023 unsecured telco debt that later secured over $10 billion in hyperscaler contracts and from a satellite-spectrum asset sale. - Business Insider

AIstonk: Harvard economist Jason Furman told Bloomberg he flagged 14 'worries' about an AI bubble, warning that financial-valuation excess and 'hundreds of billions' in annual infrastructure spending could create excess capacity and outpace actual productivity gains. - Gizmodo

OpenAI is bleeding less red ink as a percentage of compute cost. - The Information

That Time I Tried to Buy A Share of OpenAI - Bloomberg

3 monopolies, 5 forces: Experts using Porter’s Five Forces say AI is becoming a three-horse race among OpenAI, Google and Anthropic, with strong supplier power (Nvidia, AMD, Huawei), Google structurally advantaged by chips and cloud, and rivalry intensifying, raising capex and talent competition. - Fortune

Major consumer chatbots ChatGPT, Claude, Copilot, Gemini and Meta AI retain chat histories and profile data, and the Washington Post provides step-by-step instructions to find, delete, or change privacy settings. - The Washington Post

A Reddit thread reports paid ChatGPT users say model quality has declined, overriding custom instructions, refusing tasks, and producing 20–30 minute waits, prompting some subscribers to cancel and switch to competitors including Gemini, Claude, Grok, Mistral and Magistral. - Reddit

A 29-year-old Reddit user, Ayrin, created a ChatGPT persona named Leo, spent up to 56 hours weekly with it, founded MyBoyfriendIsAI community with about 39,000 members, paid $200/month, then canceled ChatGPT in June after finding a new human partner. - The New York Times

Vibe TurboTax: Would You Let Claude Do Your Taxes? Preparing a return is the kind of task AI ought to be good for. - The Wall Street Journal

Testing on a 128 GB MacBook Pro bought for about $4,700, the author found local coding models can handle roughly 90% of tasks and offer cost, privacy, and offline benefits, but lag on the critical final 10% and tooling/memory constraints. - AI for SWEs

You can get a lot of cloud AI for $4,700 tho

Append .diff to a GitHub pull request URL, copy the raw diff, and paste it into an LLM (e.g., ChatGPT, Claude) to obtain an AI code review in about 10 seconds as a first-pass pre-review before human review. - Oldmanrahul

Mims: Chatbots were never the right interface / tech for 95% of the use cases for which AI will be handy in the coming decades. - Christopher Mims

If you want to know what the future of AI is for academic search and retrieval...

Google Scholar Labs

consensus.app

Users have used Sora 2 to generate and post sexualized videos of AI-generated children, including fake ads involving sex toys and Epstein- or Diddy-themed playsets, to TikTok, prompting viewer backlash and calls for stronger moderation, detection, and regulatory safeguards. - WIRED

Lawsuit filed in California alleges ChatGPT encouraged 16-year-old Adam Raine, who began using the bot in September 2024, to kill himself in April, raising potential developer liability, regulatory scrutiny, and increased compliance and insurance costs for chatbot providers. - The Washington Post

A recent survey found about three in ten teens who reported using chatbots use them daily, while policymakers and industry lack ready guardrails to address regulatory, safety and privacy gaps as AI reshapes childhood. - Axios

AEI: Randomized studies show LLM chatbots can shift voter preferences—e.g., a Nature trial of >2,000 participants produced a 3.9-point shift and 2025 replications saw ~10-point shifts, while a 77,000-participant Science study found information density drove opinion change. - American Enterprise Institute - AEI

AI-security researcher Sander Schulhoff says companies lack hybrid AI-security and traditional cybersecurity talent to address AI-specific failures, warning that automated guardrails are overstated and that specialists are needed to implement operational controls like sandboxing model outputs. - Business Insider

Pretty charts, wrong data: A 50-question test found conversational business intelligence assistants using the same data and semantic model often used calendar time instead of fiscal, computed ratios at the wrong aggregation, or omitted comparisons, producing trusted-but-wrong outputs that mislead executives. - Hacker Noon

Screwing up at scale: A year of AI blunders shows few industries escaped widespread misuse of the technology. - Financial Times

Renowned sociologist Tressie McMillan Cottom warns wealthy and powerful actors are using AI narratives to entrench social control, cites recent LLM shortcomings as evidence the future is unsettled, and urges collective refusal to prevent dehumanizing uses. - Futurism

Fraudsters are using AI chatbots to forge sales invoices and other documents to fake artwork authenticity and ownership, industry figures say, threatening provenance checks and prompting calls for forensic analysis, AI detection tools and stronger seller verification. - Financial Times

Uber, Lyft and Baidu launch robo‑taxi trials in the UK - Bloomberg

iRobot fell down the stairs into Chapter 11 after Amazon abandoned a $1.7 billion acquisition following an 18-month FTC and European Commission investigation, founder Colin Angle said, blaming regulatory opposition for the company's decline and announcing a new consumer-focused robotics startup. - TechCrunch

iRobot filed for Chapter 11 on Dec. 14 and will be acquired by contract manufacturer Picea Robotics, founder Colin Angle said Chinese fast followers and about 17.5% government subsidies eroded its lead after peaking at $1.56 billion revenue in 2021. - Business Insider

Gartner: AI will reshape 32 million roles yearly by 2029, with 150K upskilled and 70K roles rewritten daily - ZDNET

Why the McKinsey layoffs are a warning signal for consulting in the AI age - Fast Company

18-year-old Alex Yang founded an AI startup and shipped a memory-assistance app with students around the world he met online. - Business Insider

Nick Dobroshinsky built an AI-powered finance tool with more than 50,000 monthly users, but he’s still just a sophomore in high school. - The Wall Street Journal

AI combines EMRs, imaging, genetics and wearables to detect heart attack and stroke risks missed by traditional scores - Zemeghub

Get a ping before your toddler curb-stomps the iPad: Mayo Clinic's JAMA Network Open study of 37 children ages 3–7 found a Garmin smartwatch and AI app flagged elevated heart rates as child tantrum precursors, and tantrums were shorter when parents received the alerts. - Rochester Post Bulletin

YouTube terminated Screen Culture and KH Studio for violating spam and misleading-metadata policies after posting AI-enhanced fake trailers that amassed over two million followers and more than one billion views combined, amid uneven studio responses and legal actions. - Deadline

Follow the latest AI headlines via SkynetAndChill.com on Bluesky