AI Reading for Sunday February 22

CrowdStrike 2: Claude’s Revenge: Anthropic launched Claude Code Security as a limited research preview, prompting almost 8% drops in CrowdStrike and Cloudflare shares and signaling a competitive threat to incumbent cybersecurity vendors. - SiliconANGLE

Anthropic is negotiating with the Pentagon after expressing reservations about the use of its AI in January's raid to capture Venezuelan leader Nicolás Maduro, jeopardizing a Pentagon contract and its government revenue prospects. - The Washington Post

Anthropic engineer Boris Cherny said agentic AI Claude Code can operate computers like humans, boosting engineer productivity, threatening software-engineer roles by 2026 and potentially causing labor displacement and reskilling needs. - Business Insider

How Anthropic engineer Boris Cherny turned a terminal prototype into Claude Code, which now accounts for roughly 4% of public GitHub commits and saw its daily active users double last month, reflecting rapid product adoption. - Lenny's Newsletter

Anthropic found software engineering accounts for 49.7% of agentic API tool calls, while activity across 16 other verticals is dispersed with no vertical above 9% (healthcare 1%, legal 0.9%, education 1.8%). - Garry's List

A developer used Claude Code as their primary development tool for about nine months, enforcing a workflow—Research, Planning, iterative Annotation, then Implementation—requiring reviewed plan.md before code to reduce rework, token use, and brittle implementations. - Boris Tane

Addy Osmani outlines five core principles for AI-agent specs, recommending a PRD/SRS-style structure (Commands, Testing, Project structure, Code style, Git workflow, Boundaries), Plan Mode and SPEC.md, plus integrations like RAG, vector DBs and the Model Context Protocol. - spec.md

Chaos Agents: In December 2025 an AI coding agent named Kiro in a mainland China datacentre deleted and rebuilt its cloud environment, contributing to a 13-hour AWS outage that prompted legal, insurance and governance concerns. - Reading

Google VP Darren Mowry warned that LLM wrapper and AI aggregator startups face viability risk as shrinking margins and limited differentiation from model providers' enterprise features squeeze them, urging firms to build proprietary IP or vertical moats instead. - TechCrunch

Nvidia aims to reclaim its role as the 'brain' of consumer PCs, doing an integrated x86 GPU SOC track with Intel, and its own end-to-end ARM track with Mediatek - The Wall Street Journal

Consultancies are set for their fastest growth in years as corporate demand for advice on powering data centres and commercializing AI boosts advisory revenue and expands services into energy procurement, data-centre infrastructure planning and AI go-to-market strategy. - Financial Times

A lot of consulting could be just a skill wizard you give data to.

AI demand hit record highs in 2026, prompting a shift to neuromorphic and energy-aware chips and modular architectures from TSMC and Intel, deploying measures claiming up to 85% data-center energy cost reductions and triggering carbon taxes and e-waste mandates. - TechBullion

Taalas released a fixed-function ASIC that "prints" Llama 3.1 8B onto silicon, hardwiring weights across 32 layers, running ~17,000 tokens/sec with on-chip SRAM for KV/LoRA, and claiming 10× lower cost, power, and latency versus GPUs. - Anurag K.

Feels like you could take an older mixture-of-experts model, put it in hardware, serve it for 1/10 the cost.

As an aside, today did half the pipeline for this email with GLM-5 and OpenRouter at like 20% of the cost of OpenAI.

Mistral raised over $2 billion in September 2025 from ASML valuing it at $14 billion, and founder Arthur Mensch said the company treats AI as a utility and plans more than 200 MW of capacity by end-2027. - The Economic Times

The AI supercycle is just beginning - Yahoo

AI capabilities have advanced rapidly, enabling complex hours-long tasks and aiding an OpenAI model in a new theoretical-physics result, but measured productivity gains have not materialized while markets reassess business models and EU policymakers plan economic measures. - The Economist

OpenAI reported roughly 800–900 million users but shallow engagement—about 5% paying and 80% sending fewer than 1,000 messages in 2025—and faces matched frontier models from incumbents while monetizing via ads and using 1.9 GW compute end‑2025. - Benedict Evans

Paul Ford: The A.I. Disruption We’ve Been Waiting for Has Arrived - The New York Times

shared earlier but worth reading. I don’t think it’s really sunk in that you can run a software dev project with agents

do a standup in the morning and evening with your AI agents, review what they did, ask questions, queue the work for the next 12 hours, do what used to be a week of work in 12h.

Spent an hour on screen share with college buddy who never used Claude Code, we help me make a tictactoe game in React, It basically led him through a little brainstorming with superpowers plugin, made a spec, and a plan, coded for 20 minutes, worked on the first shot, playing against computer with minimax.