AI Reading for Wednesday November 5

China foreign chip ban: Government guidance requires new state-funded data center projects to use only domestically made AI chips, forcing projects under about 30% completion to remove foreign chips and affecting suppliers such as Nvidia, AMD and Intel. - Reuters

The mind-boggling valuations of AI companies. Is Nvidia really going to deliver future earnings to shareholders equal to the annual output of Germany? - the Guardian

Analysts warn that foreseeable 'grey swan' shocks—rooted in heavy AI dependence (Nvidia's market cap over US$5 trillion and the 'Magnificent Seven' accounting for 40% of the S&P 500) and Taiwan's semiconductor concentration—could halt AI progress through security, legal, or innovation shocks. - The Conversation

the AWS us-east-1 saga demonstrates risk of concentration, interdependence, brittleness. There are probably many such dependencies on software and hardware infrastructure. Quantum computing is a Y2K scenario, in the near future basically all secure communications and computing will be obsolete.

Nvidia leads the AI chip market, and whether that dominance can last is a central question. - Bloomberg

basically Nvidia can price chips to extract the difference between its chips and its competitors’s total cost of ownership over their lifecycle. including training on a different stack, the more people get locked into their end to end solution the better. but Intel shows that the edge is not permanent. also 5 customers account for most of Nvidia’s revenue and are building their own chips and stacks.

How Much AI Spending Is Too Much? Investors are increasingly questioning corporate spending on artificial intelligence. - The Wall Street Journal

the limit does not exist tho? 

How Catastrophic Is It If the AI Bubble Bursts? An FAQ. The AI industry’s most important product is not a chatbot or a video generator; it’s the story the AI industry is telling about itself - The Ringer

The AI industry is running on FOMO - The Verge

OpenAI says it has over 1 million paying business customers, about 800 million weekly consumer users, ChatGPT for Work seats exceed 7 million (up 40% in two months), and ChatGPT Enterprise seats grew ninefold year‑over‑year. - OpenAI Help Center

Hyperscalers have announced about 46 GW of AI data-center capacity that Barclays estimates will cost roughly $2.5 trillion and, at 1.2 PUE, require about 55.2 GW of electricity—enough to power about 44.2 million U.S. homes. Numberwang! - Financial Times

Turner & Townsend surveyed 280 professionals on 300+ projects and found 48% cite power access as top scheduling constraint, with U.S. grid-connection waits up to seven years; Deloitte projects U.S. AI datacenter power needs could grow ~30x in a decade. - The Register

Caterpillar plans to double its gas turbine production capacity to meet rising demand from AI-related data centers. - Bloomberg

"China added 429 gigawatts of new power capacity in 2024, accounting for over half of global electricity growth. The US contributed just 51 gigawatts, roughly one-eighth of China’s increase” - OpenAI Help Center

Deutsche Bank is exploring hedges for its growing data-centre exposure as AI lending accelerates, including shorting a basket of artificial-intelligence stocks and using derivatives to transfer credit exposure off its balance sheet. - Financial Times

Michael Burry disclosed short positions against Palantir and Nvidia, triggering an immediate sell-off, prompting a public rebuke from Palantir CEO Alex Karp and expectations of heightened volatility and investor scrutiny in AI-related stocks. - The Washington Post

Amazon Demands Perplexity Stop AI From Making Purchases - Bloomberg

Amazon sent Perplexity a cease-and-desist demanding its Comet browser stop making automated purchases on Amazon, saying third-party purchasing agents must get permission, and Perplexity publicly pushed back, calling the demand a threat to internet users. - The Register

Agent wars: Perplexity says Amazon sent an "aggressive legal threat" demanding its Comet assistant stop making purchases on users' behalf, weeks after Comet's July launch and October global free rollout; Amazon alleges Perplexity is evading access and degrading shopping experience. - CNBC

it's a problem when something else becomes the front door to your store or SaaS. For one thing direct-to-consumer brands that lack distribution could sell direct through Perplexity or OpenAI. And ofc a company with 20 seat licenses for a SaaS could make an agent so everyone sends requests to 1 seat.

raising questions like, is the open Web doomed, if you have to ask permission before your browser can connect to a website, is it illegal to write a bot to access a site, or to put your bank credentials in a password manager that helps you log in?

Google preview AI agents can now book tickets and appointments - ZDNET

Talk to the map: Google Maps adds Gemini AI for planning and navigation, free for signed‑in Android and iOS users - The Verge

Google upgrades Vertex AI Agent Builder with ADK features, tool plugins, Go SDK and one-command deploy - Google Cloud Blog

Google launched Gemini for Home, an AI feature in the Google Home app that uses Nest camera footage and facial recognition to generate descriptions and alerts, which a 72-hour test found mostly accurate but occasionally incorrect, raising privacy concerns. - The Verge

Google DeepMind's weather AI ranked first among 11 models in this hurricane season, outperforming an aggregation-based forecast while the U.S. National Weather Service's supercomputer model ranked last, and it can run on a single computer in minutes. - Semafor

Google announced Project Suncatcher to deploy constellations of solar-powered satellites carrying TPUs in sun-synchronous orbit, example 81-satellite clusters at ~650 km with inter-satellite links of tens of terabits per second, and will fly two prototypes in 2027. - The Register

costs a lot to lift them, plus the solar panels and cooling fins or whatever, needs to be up there a long time to amortize that cost.

Researchers disclose seven GPT‑4o/GPT‑5 prompt‑injection vulnerabilities; OpenAI issues partial patches - The Hacker News

Microsoft in July uncovered a previously unseen backdoor dubbed 'SesameOp' that abuses OpenAI's Assistants API as a command-and-control channel, using .NET AppDomainManager injection to fetch encrypted commands, and said it shared findings with OpenAI, which disabled a related API key. - The Register

MIT Sloan posts, then withdraws, a paper with unsubstantiated claims about AI and cybersecurity, used to push a vendor that paid for the research. - doublepulsar.com

Despite AI companies' pledges not to influence voting choices, chatbots are recommending political parties, prompting concerns about their impact on democratic processes. - Financial Times

Developers are creating right-wing chatbots—such as Enoch, Arya and Grok—programmed to reflect creators' biases and used for ideological messaging, political persuasion and mobilization, raising disinformation risks, platform moderation costs and calls for regulation. - The New York Times

You can't post personal stuff online anymore, AI will abuse it: Rati Foundation/Tattle reported about 10% of helpline cases in India now involve AI-generated "nudify" images, prompting women to self-censor, lock accounts or avoid posting amid slow platform takedowns and repeated content recirculation. - the Guardian

France's equalities regulator Défenseur des Droits ruled Facebook's AI job‑ads algorithm constitutes indirect sex discrimination after tests found mechanic ads shown 9/10 to men and preschool‑teacher ads shown 9/10 to women, and ordered Meta to report measures within three months. - the Guardian

Future shock is real … everything is mediated in a virtual space, and you have to engineer it to provide basic human social needs and decency and not make us all crazy and hateful - The New York Times

METR benchmark analysis finds AI capability on long, complex software-engineering tasks doubles roughly every 6–7 months, with rapidly falling inference costs, but real-world robotics and messy tasks show an intercept "messiness tax" that slows applied performance. - Windows On Theory

Ninety percent of software engineers now use AI tools, Google Cloud says, and engineers report that reliance on LLMs and agents has reduced everyday peer interactions and informal mentorship while speeding coding and shifting collaboration toward higher-level decisions. - Business Insider

Common Crawl's petabyte-scale archive of billions of web pages funneled paywalled news into training data used by OpenAI, Google, Anthropic, Nvidia, Meta and Amazon, raising legal, financial and reputational risks for publishers and tech firms. - The Atlantic

seems like a problem if the thing everyone uses is full of unlicensed and sketchy stuff

Studio Ghibli, Bandai Namco, Square Enix demand OpenAI stop using their content to train AI - The Verge

Snowflake introduced AI agents, marking its entry into agent-driven assistance as enterprises increasingly budget for productivity AI like chatbots and coding tools despite frustration over reliability, intensifying competition among cloud and data platforms. - The Information

Microsoft released MAI-Image-1, its first in-house text-to-image model, making it available in Bing Image Creator and Copilot Labs (including Copilot Audio Expressions) outside the EU, with EU availability coming soon, according to Microsoft and Mustafa Suleyman. - The Verge

IBM will cut thousands of workers, affecting a low single-digit percentage of its roughly 270,000 global workforce, reallocating resources toward A.I. consulting and software while U.S. headcount stays roughly flat. - The New York Times

Apple ships Vision Pro Personas from beta; VisionOS 26 adds on‑device 3D faces and five‑person telepresence - CNET

Shopify reported AI-driven traffic to its stores rose 7x since January and purchases attributed to AI-powered search rose 11x, while Q3 revenue increased 32% to $2.84 billion and profit was $264 million. - TechCrunch

Zoe's AI meal tracker leverages Gemini and photo logging; claims >95% accuracy amid criticism it moralizes eating - CNET

Super-recognizers' advantage in facial recognitions appears to stem in part from how they scan faces and focus on the right things, AI eye-tracking suggests. - the Guardian

AI-inserted spreadsheet errors gonna crash the financial system - Simon Willison’s Weblog

OpenAI announced an "erotica for verified adults" policy as it enters the AI companion market, which generated about $120 million in revenue and roughly 220 million downloads this year amid growth, competition and rising regulatory and legal scrutiny. - Quartz

xAI required employees to submit biometric data to train Elon Musk’s AI girlfriend - The Verge

Musk wants to extract $1T from Tesla but spends all his time on xAI - The Wall Street Journal

Kim Kardashian said she used ChatGPT and blamed inaccurate chatbot answers for failing some law-school tests, while OpenAI Health AI lead Karan Singhal said terms didn't change and noted the clause prohibits legal or medical advice without a licensed professional. - TechRadar

Follow the latest AI headlines via SkynetAndChill.com on Bluesky