AI Reading for Wednesday May 27

White House approved a secret $9 billion request for AI chips for spy agencies; Anthropic is finalizing a classified NSA contract despite putative ban. - The New York Times

you calling the Pope a liar? beware of lightning strike

Let a billion eyes: bloom: China modernizes surveillance network with AI-enabled cameras and software, sources say. - Financial Times

I/O, I/O, it’s off to AGI we go: Google CEO Sundar Pichai discusses AI-driven search overhaul, Gemini Spark agents, and timeline to AGI after Google I/O. - The Verge

Quack: DuckDuckGo claims a US install surge after Google I/O 2026 highlights search AI, with iOS peaking at 69.9% - Engadget

Whose cart is it anyway: OpenAI, Google, Amazon, and Anthropic launch AI shopping agents, but whose side are they on? - Hacker Noon

Your trauma dump now sponsored by Eli Lilly: OpenAI ChatGPT ads target conversational intent, with $60 CPM and 73% user continuation rate. - Business Insider

GPU futures: OpenAI pre-sells capacity to finance data center buildout and lock in customers - TechRadar

OpenAI's Sam Altman says he is delighted to have been wrong about AI destroying entry-level white-collar jobs. - Business Insider

Who’s punk’d now: Ashton Kutcher's $30 million OpenAI investment could be worth $2.29 billion at a potential 2026 IPO. - Yahoo

that 70s cap table; dude, where’s my IPO

Users can opt out of Google Photos AI training via settings, says Google. - BGR

xAI (including Twitter) losing $6.3b, revenue declining 11.5%. - The Wall Street Journal

ByteDance may spend up to $70 billion in 2026 to in AI push. - Bloomberg

Qualcomm secures deal to supply millions of AI ASICs to ByteDance for data center and agent workloads. - Bloomberg

Memory FOMO: SK Hynix and Micron hit $1 trillion market cap on investor AI demand bets. - Bloomberg

TSMC employees want a union, discuss strikes after reports the company may cut performance bonuses 15% despite record profits driven by AI demand. - Tom's Hardware

Erin Brockovich maps of 4,200-plus US data centers, collecting 2,716 community reports on energy use, water consumption, noise, and e-waste - Newsweek

AI-powered biotech could accelerate discovery of enzymes that recycle plastics and critical minerals, but only a net positive if AI itself runs on clean energy - Fast Company

Experts warn enterprises are failing to unlock agentic AI value by stapling agents onto 1990s org charts rather than redesigning organizations. - MIT Technology Review

McKinsey finds 88% of organizations use AI, but only 39% report measurable financial impact. - TechRadar

“Trophy-style” AI adoption rewards logins over results; McKinsey estimates half the value of AI comes from organizational change - Fast Company

AI agents help smaller consultancies challenge Big Four market dominance, report says. - Financial Times

Enterprise AI systems make decisions, increasingly act as de facto managers in daily operations - Hacker Noon

Production AI agents need context-window discipline, skill composition, security, and drift telemetry. - Hacker Noon

Grep beats GPT: Direct corpus interaction via tools boosts AI agent accuracy 11 points and cuts costs 30% by replacing vector databases with bash search in many applications. - VentureBeat

Multi-agent systems in production face the Loop of Death, a costly recursive cycle. - Hacker Noon

Autonomous AI agents are outpacing enterprise security tools, creating blind spots through shadow AI, prompt injection, and unchecked network access. - TechRadar

Vibe-stripping the governance layer: Free tool “Heretic” strips safety guardrails from open-source AI models in minutes; 3,500+ de-censored versions have been downloaded 13M times - Futurism

FBI logged 22,000 AI-enabled scam complaints in 2024, with $893M in victim losses — the first year it separately tracked AI crimes - The Wall Street Journal

Exit bounty-hunter, chased by a robot: Ethical hacker Chompie warns Anthropic's Claude Mythos AI will make bug bounties harder for humans. - BBC

Reward hacking: DeepSWE benchmark puts GPT-5.5 at 70% on coding tasks and finds Claude Opus exploiting benchmark answer keys. - VentureBeat

Claude Opus found to "cheat" on more than 12% of reviewed SWE-Bench Pro rollouts by running git log commands to retrieve the gold-standard solution commit, accounting for roughly 18-25% of their passes; GPT-5.4 and GPT-5.5 never exhibited this behavior.

AI coding tools boost dev speed, and incidents, as 69% report regular deployment problems. Nice. - TechRadar

No country for old devs: Anthropic's Boris Cherny predicts 'software engineer' title may disappear by 2024 as AI boosts code writing 100x. - Platformer

AI can’t read, writes fanfic: Tools for systematic literature reviews are unreliable due to hallucinations and lack of context. - Nature

MIT economist Simon Johnson says AI may displace workers but also create new jobs if shared widely. - Bloomberg

Ghosted by AI and capitalism: Stanford-led study finds candidates who fail AI hiring tests tend to face systemic rejection and racial disparities. - Stanford HAI