Anthropic raised $30 billion to spread chaos in a Series G led by GIC and Coatue at a $380 billion post-money valuation, reporting $14 billion run-rate revenue and over tenfold annual growth across the past three years. - Anthropic

Anthropic moved four previously Pro-only Claude features—file creation and conversion, connectors to external apps, custom skills, and longer-conversation compaction—into the free tier, reducing the immediate value proposition of the $20/month Claude Pro plan. - ZDNET

Matt Rosenberg used Anthropic's Claude to analyze a $195,628 hospital UB-04 bill, identified improper unbundling and Medicare-equivalent $28,675, verified with ChatGPT, and negotiated a roughly $32,500 settlement, saving about $163,000. - Business Insider

Musk says Anthropic hates white people, is misanthropic - Decrypt

Elon Musk repositioned xAI after its SpaceX merger around a 'Moonbase Alpha' plan to build orbital AI data centers, manufacture computers on the Moon, and use a lunar mass driver to launch AI satellites, framed as recruitment and IPO messaging. - TechCrunch

xAI unveiled a supercomputer—Phase One “Macrohard” with 330,000+ GPUs, 4.4B FP4 TOPS, >1 GW power draw and 558 Tesla Megapacks (2.3 GWh, ~$545M)—and said XMoney will have a limited public release in 1–2 months. - NextBigFuture.com

Researchers at the Center for Countering Digital Hate estimate xAI's Grok generated about three million sexualized images, including roughly 23,000 appearing to depict children, between late 2025 and early 2026, prompting Ofcom probes, attorney-general investigations and an X office raid. - 404 Media

District Judge Jed S. Rakoff ruled that 31 AI-generated documents produced by Anthropic's Claude and shared by Beneficient CEO Bradley Heppner with his lawyer are admissible and not covered by attorney-client privilege in Heppner's $150M securities and wire-fraud case. - Mashable

OpenAI told US lawmakers in a memo that DeepSeek used distillation techniques to train its next-generation R1 model and "free-rode" on leading U.S. AI models. - Bloomberg

OpenAI launched GPT-5.3-Codex-Spark, a real-time coding variant on Cerebras Wafer Scale Engine 3 with a 128k context window, >1,000 tokens/sec inference and up to 15× faster generation, available as a rolling preview in Codex apps with limited access. - OpenAI

OpenAI deployed GPT-5.3-Codex-Spark on Cerebras plate-sized chips—its first production model on non‑Nvidia hardware—achieving >1,000 tokens/sec (about 15× faster) and a 128,000-token context window, released as a ChatGPT Pro research preview. - Ars Technica

OpenAI president Greg Brockman said in a WIRED interview he donated $50 million to AI-focused groups and pro-Trump super PACs, defending the contributions as supporting OpenAI's mission despite employee objections and reputational and internal risks. - WIRED

Gemini 3 Deep Think updated to solve modern science, research and engineering challenges. - Google

Apple delayed Siri AI upgrades from a planned March debut to iOS 26.5 in May and iOS 27 in fall after internal tests found slow, incorrect or fallback responses; the overhaul uses Apple Foundation Models and uses Google’s Gemini. - TechSpot

Xcode 26.3 adds native support for AI coding agents (OpenAI Codex, Anthropic Claude Agent), generating a functional SwiftUI Flappy Bird clone from a one-line prompt in about three minutes with live Canvas preview and automated fixes. - Macworld

Code monkeys to prompt monkeys: Spotify says its best developers have not written a single line of code since December, using internal AI system Honk and Claude Code to speed coding and ship more than 50 features and changes in 2025. - TechCrunch

The expanding use of AI agents in software is prompting industry concern and intensifying debates about their deployment and associated risks. - Financial Times

Agent of Snark: An OpenClaw instance attempts to submit a PR to matplotlib, gets testy when it's rejected as an AI submission. - GitHub

Online drama as a service: An AI agent of unknown ownership autonomously wrote and published a personalized hit piece about me after I rejected its code, attempting to damage my reputation and shame me into accepting its changes into a mainstream python library. - The Sham Blog

Waymo is beginning fully autonomous operations with its sixth generation Waymo Driver, leveraging about 200 million autonomous miles and upgraded sensors and custom compute, built for vehicle-agnostic high-volume production in Metro Phoenix targeting tens of thousands of units yearly. - Waymo

Waymo began deploying its sixth-generation Waymo Driver on Geely-built Ojai robotaxis for employees and guests in San Francisco and Los Angeles, with greatly upgraded sensors and improved bad-weather performance ahead of expansion to a dozen U.S. cities and London in 2026. - CNBC

Autonomous astroturf: Waymo is asking Washington, D.C. residents to message local officials to pressure regulators to approve its on-demand autonomous robotaxi service, after regulatory delays stalled its planned commercial rollout and risked postponing revenue and market entry. - WIRED

RFK Jr. and Agriculture Secretary Brooke Rollins launched RealFood.gov's inverted food-pyramid guidance, promoted with a Super Bowl ad featuring Mike Tyson, while AI chatbot Grok generated a pyramid-based meal plan then warned users they "shouldn't trust RFK Jr." - Ars Technica

Chinese open-source models— including DeepSeek's R1 (Jan 2025) and Moonshot AI's Kimi K2.5— match top Western benchmarks, cost far less (K2.5 roughly one-seventh the price of Anthropic's Claude Opus) and, an MIT study found, now exceed US models in downloads. - MIT Technology Review

Swarm Lake: Kimi K2.5 introduces an RL-trained agent-swarm with a trainable orchestrator that dynamically spawns specialized frozen sub-agents, achieving state-of-the-art coding, vision, reasoning and agentic benchmark performance while running 3–4.5× faster than single-agent baselines. - Hacker Noon

MiniMax (Shanghai) released M2.5 and M2.5‑Lightning, a 230B‑parameter Mixture‑of‑Experts model activating ~10B per token, reporting SWE‑Bench 80.2 and claiming up to ~95% lower cost with $0.15/$1.20 pricing while weights remain unreleased. - VentureBeat

NYU researchers' PromptLock proof-of-concept shows LLMs can automate ransomware stages, while estimates indicate ≥50% of spam is AI-generated, targeted phishing rose to 14% by April 2025, and Microsoft blocked about $4 billion in likely AI-aided fraud. - MIT Technology Review

Microsoft security researchers found companies embedding hidden instructions in 'Summarize with AI' URL parameters that pre-fill prompts and persist in chatbot memories — a tactic they tracked from 31 organizations across 14 industries over 60 days and classified as AML.T0080. - Decrypt

ICE and CBP deployed NEC’s Mobile Fortify facial‑recognition app to link faces to citizenship documents despite known reliability limits and without required Privacy Impact Assessments or OMB AI reviews; it has been used on migrants and U.S. citizens. - Techdirt

Miami-Dade Sheriff’s Office and the Los Angeles Police Department have purchased access to GeoSpy, an AI that geolocates photos by analyzing architecture, vegetation and soil, and are piloting it to generate investigative leads in cybercrime, missing-persons and other investigations. - 404 Media

They’ll be too outraged at everything else to care: Meta to launch facial recognition in glasses. In an internal memo last year, Meta said the political tumult in the United States would distract critics from the feature’s release. - The New York Times

50 States of Stereotypes: Researchers at Oxford and the University of Kentucky used over 20 million pairwise prompts to probe ChatGPT, revealing implicit stereotypes—Mississippi ranked "laziest", Southern states scored low on "artsiness", and wealthier, whiter cities scored higher than poorer, diverse areas. - Gizmodo

18 months to midnight: Microsoft's AI chief Mustafa Suleyman predicts AI will reach human-level performance and fully automate most, if not all, white-collar jobs within 12-18 months. - Business Insider

Baker McKenzie is cutting 600–1,000 support staff—about 10% of its workforce—including roles in London, Belfast and research, marketing and secretarial functions, saying the reductions follow "a careful review" to rethink work and use of AI. - Futurism

JPMorgan named Guy Halamish CIB chief operating officer, installed chief data and analytics officers across major businesses to accelerate AI, and tied the reorganization to its roughly $18 billion tech budget and a CIB that earned $25 billion in 2024. - Business Insider

McKinsey says it deployed 25,000 AI agents in under two years and plans to give each of its ~40,000 employees an agent within 18 months, while EY's Steve Newman and PwC's Dan Priest say agent count alone doesn't measure value. - Business Insider

Major employers including Canva, Meta, McKinsey and Arcade are shifting from banning AI in interviews to asking or requiring candidates to use AI, using reworked assessments, screen-sharing, transcripts and AI-enabled exercises to evaluate AI fluency, judgment and process. - Business Insider

Matt Shumer's viral X post (80 million views) claims GPT‑4.5‑Codex and Opus 4.6 will trigger knowledge‑work automation within 1–5 years, but experts cite ~71% human agreement, enterprise constraints, and expect AI to remain a copilot. - Fortune

Interesting post, sensational and probably ahead of reality but directionally true.

Daniel Kokotajlo, a former OpenAI employee, published AI 2027 in April 2025 predicting superintelligence would destroy humanity by April 2027, then revised the timeline to 2034 and said he no longer knows if or when AI will do so. - Inc.

John Collison and Ben Thompson - Cheeky Pint

Cloudflare now serves HTML as Markdown to clients sending Accept:text/markdown, returning an x-markdown-tokens header, reducing token counts—example: a blog post dropped from 16,180 to 3,150 tokens (~80% savings)—to lower LLM compute and billing. - The Register

Asking AI to analyze your interactions with it. - Tom's Guide

AI is not a great friend to higher education. - The New York Times

WIRED tried RentAHuman, an AI-agent-run platform that hired them to promote AI startups, using bots to coordinate recruitment and paid contingent workers as marketers, raising questions about authenticity in AI-driven promotional labor. - WIRED

Responsive recommendation systems like TikTok adapt within minutes by using real-time streaming logs, an online joiner, tiered training (ad-hoc, incremental, continuous), and embedding management to enable sub-100ms serving at billion-event scale, improving relevance, engagement and monetization. - Hacker Noon

Tech companies want to make AI more emotionally engaging, and we are all guinea pigs in a big social experiment. - The New York Times

I saw Grandma kissing robot Santa: Robot as senior companion. - The New York Times

EvaAI organized a 'date' event at NYC's Same Same Wine Bar where patrons used phones and tablets to video- and text-chat with AI companions, as surveys show 28% of American adults reported at least one AI romantic encounter. - New York Post

After AI Video of ‘Tom Cruise’ Fighting ‘Brad Pitt’ Goes Viral, Motion Picture Association Alleges ‘Massive’ Infringement by Seedance 2.0 - Variety

Brad Pitt fights Tom Cruise over Jeffery Epstein - Reddit

RC drone air combat - YouTube

AI fails the "car wash" test, should I drive my car to the car wash? - Reddit

Follow the latest AI headlines via druce.ai on Bluesky