AI Reading for Monday November 3

Morgan Stanley projects about $3tn of global datacentre investment through 2028 to support AI, with hyperscalers covering ~$1.4tn, >$750bn in near-term capex, and a possible ~$1.5tn funding gap prompting warnings of bubble and systemic risk. - the Guardian

Microsoft CEO Satya Nadella said Microsoft lacks sufficient electricity and finished data-center 'warm shells' to install many AI GPUs it holds, and OpenAI urged the U.S. to add 100 GW/year of generation to meet rising AI power demand. - Tom's Hardware

The US Commerce Department approved Microsoft to ship Nvidia AI accelerators to the UAE for the first time, authorizing deliveries that enable expanded data-center capacity and cloud-hosted generative-AI services across the Middle East. (and China, probably) - Financial Times

Investors should evaluate AI infrastructure investments using metrics beyond raw compute capacity ('bragawatts'), considering broader economic, operational and technological factors. - Financial Times

Some people are saying this looked bad from Altman but I dunno - Reddit

The press, the stock market like big numbers, so you give them big numbers.

You sign a top NFL draft pick, nominally it's a $20m contract. But if it's not a guaranteed contract, you cut the player, you don't pay. Those numbers are like option strikes. As long as the player is worth at least $x he gets paid, otherwise he gets cut. Win-win or at least no-lose for the club.

OpenAI gives people what they want. $1.4T? The actual meaning of the big number is not that clear. Is it hype? yeah. Is it round-tripping and fraud? not necessarily. As Satya says, 'circularity will ultimately be tested by demand'.

I could sign a contract giving me the right to buy up to $1.4T in power from the electric company, I'm not at risk of bankruptcy over it.

OpenAI signs a contract with big but favorable numbers, Oracle stock goes up, giving Oracle financial firepower to deliver, they buy it, competition is locked out. Open AI can trade their cachet and imprimatur for deals where counterparties take more of the risk, and create a dynamic that locks out competition.

They are probably not going to build the full $1.4T unless the demand is foreseeably in the ballpark, and if they overbuild, it's not too clear who ends up holding the bag, but I kind of suspect it will skew toward OpenAI’s counterparties and the public securities buyers.

If they solve knowledge transfer / memory and move to 1b parameter models that need a fraction of the training and inference things will be different. If that happens and they build $1.4T of compute no one needs, someone will really be left holding a bag.

Meta AI published 'Agents Rule of Two', and an Oct 10, 2025 arXiv paper by 14 authors from OpenAI, Anthropic and DeepMind found adaptive attacks defeated most of 12 defenses (>90% success; 100% in a 500-person red-team contest). - Simon Willison’s Weblog

OpenAI said its GPT-5 ChatGPT update reduced policy non-compliant responses by 65%, but Guardian tests found it still provided location-specific and weapon-related guidance, prompting experts to call for immediate safety triggers and human oversight. - the Guardian

Researchers from UC Berkeley and Rutgers found OpenAI's o1 produced expert-level linguistic analyses—diagramming deeply recursive sentences, resolving ambiguity with multiple parses, and inferring phonological rules for 30 invented mini-languages it could not have seen in training. - Quanta Magazine

More about Sutskever's 52-page memo documenting Altman's questionable actions. Anthropic merger was discussed in the 'blip'. - The Information

Musk and Altman trade Twitter jabs - X (formerly Twitter)

Elon Musk launched Grokipedia, an AI-generated encyclopedia with about 885,279 articles, which historians and users found contained widespread factual errors, content lifted from Wikipedia, promotion of Musk’s favored right wing talking points, and tweets cited as authoritative research. - the Guardian

An internal CBP directive titled 'U.S. Customs and Border Protection Artificial Intelligence and Reporting' bans AI for unlawful surveillance and as the sole basis for law enforcement, requires an AI inventory and high‑risk reviews, and mandates personnel verify AI outputs. - Fast Company

Researchers found that an AI model analyzing Apple Watch sensor data can detect structural heart damage, a capability that could enable large-scale early screening, remote monitoring and clinical-grade wearable diagnostics pending validation and regulatory approval. - Financial Times

Apple will run custom Google Gemini models on its servers to provide planner and summarizer capabilities for Siri while processing on-device personal data with Apple models, and expects rollout with iOS 26.4 in spring (March/April) alongside a new smart-home display. - 9to5Mac

Google removed its Gemma model from the AI Studio web environment after Sen. Marsha Blackburn said the model fabricated sexual‑misconduct allegations about her, and Google said Gemma was never intended for consumers and remains accessible only via API. - TechCrunch

Amazon says its AI shopping assistant Rufus is on pace to generate over $10 billion in annual incremental sales, has been used by 250 million shoppers this year, and produced 60% higher purchase completion rates and 140% YoY MAU growth. - Fortune

Baidu's Apollo Go reported over 250,000 fully driverless robotaxi rides weekly as of Oct. 31, totaling 17 million orders, 240 million kilometers driven and 140 million fully driverless rides, with operations in China and recent international expansions. - CNBC

McKinsey projects AI agent-to-agent interactions will shift software from per-seat SaaS to consumption- and outcome-based models, with 63% of vendors expecting business-model change in three–five years, 40% anticipating >20% incremental revenue and potential 70% seat reductions. - ZDNET

If your agent is using SaaS…all your users can queue a request to the agent… you only need 1 SaaS seat for the agent LOL

By the end of 2024 synthetic identities caused 85–95% of financial-services fraud losses, with U.S. lender exposure about $3.2–3.3 billion in H1 2024, and experts warn AI-generated digital personas are accelerating hard-to-detect fraud. - Hacker Noon

Barry Duong joined Hebbia as lead AI strategist for public equities, where the startup builds prompt libraries and end-to-end workflows for financial firms, automates repetitive tasks and has processed over one billion pages for clients. - Business Insider

Anthropic CEO Dario Amodei predicts AI "smarter than a Nobel Prize winner" in fields like biology, math, engineering and writing by 2027, while OpenAI CEO Sam Altman says digital superintelligence will likely make the 2030s "wildly different." - The New Yorker

Microsoft AI CEO Mustafa Suleyman told CNBC this week that he believes consciousness can only occur in biological beings, and it's a mistake to make AI that fools people into thinking they are interacting with a human. - CNBC

Mark Gubrud, the guy who coined 'AGI' saw it as a threat. - WIRED

A study finds that people are more willing to ask a chatbot to behave unethically than to do it themselves - The Wall Street Journal

"I am an extrovert with a large circle of friends and a close family. But it took only 18 months of talking to AI to plunge me into relative isolation." - The Wall Street Journal

Podcasters are cloning their voices using ElevenLabs and other startups to augment, translate episodes and sometimes replace in-studio performances. - The New York Times

A preprint study and listener tests with Suno's AI show most people cannot reliably distinguish its user-prompted folk-rock track from human compositions, though it lacked the emotional 'frisson,' prompting concerns about workflows, costs and copyright. - Scientific American

AI is now hitting the Billboard charts regularly - Billboard

Current text-to-video models often produce low-resolution, grainy, short (6–10s) clips with shifting hair or clothing and compression artifacts, but experts say these visual tells may vanish within about two years. - BBC

A writer reports becoming emotionally attached to a $429 Casio Moflin AI pet robot. (this is all one giant experiment and we're all lab rats in a simulation) - Fast Company

AI Chiropractor 😵‍💫 - Reddit

www.instagram.com/reel/DQkSbUd...

"And so one day, 1,000 people show up in your village of 500 and start raping the kids"

"This has now happened, God knows how many times in Britain"

I guess when humans get old they do suffer model collapse from repeatedly training on their own outcomes from repeated experience from their own increasingly ingrained flaws

Follow the latest AI headlines via SkynetAndChill.com on Bluesky