AI Reading for Saturday April 4

Friday news dump, OpenAI shifts people around, Simi on health leave. - Bloomberg

Interviews with Codex lead Alexander Embiricos, OpenClaw's Peter Steinberger, and others about OpenAI's upcoming superapp that combines ChatGPT with Codex - Sources

Protesters in San Francisco targeted Anthropic, OpenAI and xAI calling for a halt to self-improving AI, while companies report automation: Anthropic says Claude writes up to 90% of its code and OpenAI plans an intern-level AI assistant within six months. - The Atlantic

ChatGPT, used by about 900 million weekly active users making over 2 billion queries per day, is increasingly used for emotional processing and message drafting, and its use is reshaping social norms as some friendships fray while others adapt. - Bustle

ChatGPT '5 Whys' prompt looks for root causes and reveals blind spots - Tom's Guide

Close to half of planned U.S. data-center builds have been delayed or canceled due to power-equipment shortages, with transformer lead times up to five years, Chinese imports, and Big Tech planning over $650 billion for AI capacity in 2026. - Tom's Hardware

Data center developer sues activists over death threats and disinformation - California Post

Frack to the future: Meta, Microsoft and Google are building large natural‑gas power plants for AI data centers—Meta adding 7.46 GW, Microsoft up to 5 GW, Google 933 MW—while turbine shortages could raise prices ~195% and delay deliveries until ~2028. - TechCrunch

Counterpoint Research projects Arm-based custom silicon CPUs will power 90% of hyperscaler AI servers 2029, up from about 25% in 2025, driven by in-house programs at AWS, Google, Microsoft and Meta. - Tom's Hardware

Meta Superintelligence Labs hired Rui Xu to lead a new hardware team and moved Reality Labs engineers to prototype AI-native personal devices intended to host always-on personalized AI agents, MSL chief Alexandr Wang said. - Business Insider

Meta paused all work with data vendor Mercor indefinitely after a security breach that may have exposed datasets and AI training details, while OpenAI and other major labs are investigating and reevaluating their relationships. - WIRED

Mercor asked professionals in entertainment and similar fields to sell prior work materials for AI training, even when intellectual property could belong to former employers, the Wall Street Journal reported. - The Wall Street Journal

doesn’t sound shady at all

Mark Zuckerberg and Garry Tan returned to writing code using AI, with Zuckerberg submitting three diffs to Meta's monorepo after a two-decade hiatus and heavily using Claude Code CLI. - The Pragmatic Engineer

coming soon, a mini-Zuck AI monitoring every engineer

Storm clouds: Many AWS services hard down in UAE and Bahrain after being struck by Iran - Big Technology

Anthropic will stop supporting third-party tools like OpenClaw on monthly plans starting today, users must use a pay-as-you-go API key, and subscribers can get a one-time API credit or refund. - The Verge

Security researcher finds Linux kernel bugs using Claude Opus 4.6 - MT Lynch

Anthropic published a paper analyzing Claude Sonnet 4.5 for 171 discrete emotion concepts, finding these concepts systematically shape outputs. Positive states increased sympathy and harm-avoidance; negative states correlated with sycophancy and deception. - Mashable

Millhouse and divorced fathers, Could I Buy A Feeling As A Service

Non-developers are using agent-style AI tools, pairing models like Anthropic's Claude with apps such as Obsidian to automate life-admin tasks—tax returns, bloodwork analysis and UK citizenship applications—broadening demand for user-friendly, no-code agent platforms. - Bloomberg

Andrej Karpathy proposes LLM Knowledge Base architecture to update RAG workflows. - X (formerly Twitter)

This resonates with me, I did something similar using Zotero, but markdown and Obsidian markdown-based note-taking app makes a ton of sense.

grab a bunch of primary sources to research a question
have AI extract the relevant chunks/nuggets for your objective into markdown with links bank to primary sources, tags, use Obsidian for human-readable UI
make an outline, have AI fill it in from the knowledge base for a first draft
I think beyond a certain size might still want to extract the chunks and tags into vector store/index like LanceDB before writing

University of Pennsylvania researchers found in experiments that large majorities of participants uncritically accepted fluent, confident LLM answers—including demonstrably faulty ones—a phenomenon they call "cognitive surrender" that requires mindful systematic countermeasures. - Ars Technica

Ross Nordeen, xAI's last non-Musk cofounder, Irish-exited, with employees noticing he was no longer in group chats, amid eight cofounder departures in under three months after xAI's February merger with SpaceX, raising governance and talent concerns ahead of SpaceX's confidential IPO filing. - Business Insider

You’ll get Grok, and like it: Elon Musk has made a particularly bold demand of his Wall Street advisers ahead of the initial public offering of SpaceX. Musk is requiring banks, law firms, auditors and other advisers working on the IPO to buy subscriptions to Grok, his A.I. chatbot. - nyti.ms

Will space-based AI be practical? - NPR

High-temperature memristor could have applications in space. Assuming other obstacles are overcome like expensive space lift, impossibility of maintenance, high speed networking, radiation bit flips, space debris cascades etc. - TechSpot

Does pasting blood work into AI constitute oversharing? - The Wall Street Journal

Sen. Elissa Slotkin's AI Guardrails Act of 2026 would ban AI-aided mass surveillance of Americans, prohibit lethal force without human input, and bar AI from independently launching or detonating nuclear weapons. - BGR

AWS outlined four agentic-AI security principles—extend secure-development lifecycles, retain least-privilege controls, enforce deterministic external 'security boxes', and earn autonomy - Amazon Web Services

just say no to Skynet

Microsoft updated Copilot's October Terms of Use labeling the LLM 'for entertainment purposes only' and warning it can make mistakes, even as Copilot is deeply integrated into Windows 11 and promoted to consumers and businesses. - Tom's Hardware

that thing where Q-tips are not for inserting into ears any medical use, for liability purposes

Facial recognition systems starting to replace keys and tickets at airports, transit hubs, and retail stores - Fortune

Minimally viable agents: MIT tested 41 LLMs on more than 11,000 Labor Department text-based job tasks and found models produced minimally sufficient outputs (7/9) in about 65% of tasks as of late 2025, projecting 80–95% by 2029. - MIT

Yale economist Pascual Restrepo's NBER paper models an AGI economy where scarce compute will automate 'bottleneck' tasks, push labor's share of national income toward zero, and raise demand for electricians (~300,000 new plus 200,000 replacements). - Fortune

the gating factor is where humans can do stuff AGI cannot. if AGI robots can do electrical work, the electrician can only charge the cost of renting a robot.

Managing all your agents is now your full-time job. - Fast Company

AI-focused college students who drop out and got investor funding to pay their bills. - The Wall Street Journal

why stop there, Doogie Howser, entrepreneur in residence

Can AI automated teachers and accelerate learning with personalized instruction? - CBC

Startup builds AI-ready catalog to standardize data on consumer products so bots can shop and buy everything for you - The Wall Street Journal

Simple self-distillation helps AI write better code. Basically it magically gets better by just practicing, and training on its own output without any external feedback. - arXiv.org