AI Reading for Sunday March 22

At GTC, Jensen Huang announced new hardware and projected a $35 trillion AI-agent and $50 trillion physical-AI market, and forecasted $1 trillion in Blackwell/Vera Rubin chip orders by end-2027. But markets see risk around that story and the macro/political/energy situation. - TechCrunch

AWS has deployed 1.4 million Trainium chips, including over 1 million Trainium2 chips powering Anthropic's Claude, and agreed in its $50 billion OpenAI deal to supply 2 gigawatts of Trainium capacity while claiming up to 50% lower costs. - TechCrunch

OK Jensen, we’ll see, we’ve got ice cream and GPUs at home now

The entire semiconductor chip supply chain depends on energy and chemical imports from the Middle East, so disruptions to those flows could threaten manufacturing capacity, raise costs, and cause shortages and price volatility for chipmakers and AI firms. - Financial Times

Qatar halted about 30% of global helium production at Ras Laffan on March 2 after Iranian drone attacks, cutting annual exports by 14%, doubling spot prices and threatening semiconductor supply chains with shortages expected in a few weeks. - ABC News

the world rediscovers that everything runs on Middle East oil; talking in a squeaky voice gonna cost some coin

At Nvidia GTC, companies demonstrated multiple robots—IntBot, Humanoid's HMND 01 Alpha showing fleet control, Noble Machines' Moby 3 performing autonomous lifts up to 50 lbs, and OpenClaw AI on DGX Spark—while CEO Jensen Huang predicted broader AI-robot integration. - CNET

Softbank plans $33B, 10GW (gas equivalent of 9 big nukes) data center in Ohio - Tom's Hardware

Children of the Corn, but the kids are lost in a data center maze. post-apocalyptic data centers will make good scavenging.

Apple more than doubles many external storage prices amid AI-triggered shortages. - Apple Insider

Apple collected nearly $900 million in App Store fees from generative AI apps in 2025 and is projected to exceed $1 billion in 2026, with OpenAI's ChatGPT contributing about 75% and monthly AI-related fees rising from ~$35M to ~$101M. Probably more than they are paying for Gemini in Siri. - MacDailyNews

Reflect Orbital's FCC proposal for ~50,000 55‑m sun‑reflecting mirrors and SpaceX's plan for up to 1,000,000 data‑center satellites have prompted formal objections from major astronomy bodies, citing up to threefold night‑sky brightness increase and VLT image losses up to ~10–30%. - Space

koestler’s darkness at noon + literal sunlight at midnight. yay?

Anthropic says the Pentagon is lying. - TechCrunch

when you text your partner you are totally aligned right before the big breakup fight

Palantir is doubling down on a vision of AI built for battlefield advantage—and attracting customers who agree. - WIRED

more firepower, less guardrails and ethics

Deepfaking democracy: At least 15 political ads using AI-generated content have run since November, while 26 states now have deepfake rules and experts expect AI-created ads to increase ahead of the 2026 midterms. - NBC News

Federal bureau of infringing your privacy: FBI director Kash Patel told a Senate committee the FBI buys commercial location and consumer data from brokers, enabling warrant-free large-scale access, while the global data-broker industry—worth hundreds of billions—supplies detailed demographics, browsing and location records. - the Guardian

The Los Angeles Superior Court is pilot-testing Learned Hand, an AI that summarizes filings, organizes evidence and drafts rulings for a small group of judicial officers as court filings rose 49% year‑over‑year from 4,100 to 6,400. - Decrypt

KPMG has deployed a multifaceted governance framework for enterprise AI agents—unique identifiers and 'systems cards' for traceability, an AI operations center with human monitors, red‑teaming and least‑privilege controls. - Business Insider

pressing ‘kill switch’ in the SOC

Atlassian cut about 1,600 jobs—roughly 10% of its workforce, including nearly 500 Australian staff—CEO Mike Cannon-Brookes said the cuts were needed to become more agile and invest in AI, while affected employees reported abrupt email notifications and account lockouts. - the Guardian

Chief fine-tuning officer: an Durn turned Adobe's finance organization into an AI lab that auto-responded to about 300,000 emails across 19 inboxes (saving >5,000 hours), halved contract-review time, and helped AI-first product revenue more than triple YoY in Q1 FY2026. - Yahoo

Bank of America Institute found 'high propensity' new‑business interest rose 15.1% year‑over‑year in January while applications planning to hire fell 4.4%, small‑business tech spending rose 14%, and AI featured in about 8% of 2026 job‑cut announcements. - Yahoo

AI's revolution could mirror the Industrial Revolution by producing comparable economic, labor and technological transformations. - Bloomberg

Cursor has grown rapidly but faces a strong challenge from Claude Code. - Fortune

Their IDE is also highly complementary to CLIs, and an extremely attractive acquisition target for OpenAI, which has catching up to do, previously tried to buy Windsurf and has undoubtedly approached them.

The pangon/ai-sdlc-scaffold repository provides an Apache-2.0 template for Claude Code that structures the SDLC into four phases (Objectives, Design, Code, Deploy) and supplies agent skills that automate elicitation, design, task generation, implementation and deployment while recording decisions. - GitHub

Andrej Karpathy says he hasn't written a line of code since December, is getting into a state of psychosis, as agentic AI reversed his workflow from about 80% self-coded to agent-driven, built an agent called Dobby to control his home, and warns the shift is changing workflows. - Yahoo Tech

Is ChatGPT the Yahoo of AI? OpenAI says ChatGPT has over 800 million users, but users report Claude and Google's models outperform ChatGPT on many tasks. ChatGPT leads among free/consumer users while Claude is stronger in enterprise, and some subscribers have canceled paid plans. - Reddit

Thinking Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender - SSRN

they inserted errors into AI outputs. the more people used AI, the more they blindly followed the errors (cognitive surrender). Time pressure made it worse, incentives ($ per correct answer) and feedback on errors made it better, but, the more AI you use, the more you incorporate its errors.

I think in high stakes environments you need a very structured process with human in the loop to make sure it corrects more errors than it introduces. like, first radiologist reviews, then the AI evals radiologist input and suggests other areas to examine.

seems like if you don’t nerf the AI, the more you use it the better the results.