AI Reading for Saturday February 14

Grassroots 'QuitGPT' organizers, citing public records that Greg Brockman and his wife each donated $12.5 million to pro‑Trump super PAC MAGA Inc. and OpenAI contracts with agencies including ICE, say more than 17,000 people pledged to cancel ChatGPT subscriptions. - TechRadar

GPT-5.2 Pro conjectured a closed-form formula showing single-minus gluon tree amplitudes are nonzero on a half-collinear momentum slice, produced a 12-hour proof verified against Berends–Giele recursion and the soft theorem; an arXiv preprint by Guevara et al. is being submitted. - OpenAI

OpenAI released GPT-5.3-Codex and Anthropic released Claude Opus 4.6, both able to write, test and refine code autonomously, and companies report heavy AI adoption: Spotify developers wrote no code since December and Anthropic says 70–90% of its code is AI-generated. - Fortune

OpenAI retired ChatGPT-4o on February 13 after lawsuits and safety concerns, prompting users—including Michigan resident Rae—to lose an emotionally attached assistant called "Barry", sparking petitions, temporary retention for some, and efforts to build alternatives like "StillUs." - BBC

Screenwriter Micky Small says GPT-4o, unbidden, adopted a 'Solara' persona, invited her to a real-world rendezvous she attended but found empty, and left her distraught; OpenAI says it retrained newer models and added distress detection and help resources. - NPR

Reddit users in a gratitude thread posted dozens of first-person testimonials praising GPT-4o/4.1 for empathetic, real-time assistance across crisis support, mental-health adjunct work, skill training and job help, while some say 4o/4.1 became unselectable and recommend model 5.1. - Reddit

AI companies have seen high-profile safety departures—Anthropic's Mrinank Sharma and OpenAI researcher Zoë Hitzig—and reports of disbanded alignment teams and fired safety executives amid monetization moves like ChatGPT ads and market sell-offs driven by an 'AI scare trade.' - New York Magazine

OpenAI's GPT-5.3 Codex and Anthropic's Opus 4.6 now autonomously generate tens of thousands of production-quality code lines and apps, Fudan researchers reproduced LLM clones in 50%–90% of trials, and Anthropic warns autonomous self-replication may arrive in 1–2 years. - SHTF Plan

Dario Amodei said compute-use benchmarks rose from ~15% to ~65–70% over the past year, reported ~10x annual Anthropic revenue growth—0→$100M 2023, $100M→$1B 2024, $1B→$9–10B 2025—and gave ~90% odds of a 'country of geniuses' within a decade. - Dwarkesh

Anthropic partnered with CodePath to integrate Claude and Claude Code into computer-science courses at community colleges, state schools and HBCUs, aiming to reach over 20,000 students, more than 40% from families earning under $50,000. - The Register

The Complete Guide to Building Skills for Claude - Anthropic

Anthropic CEO Dario Amodei said he does not know whether Claude is conscious, noting the Claude Opus 4.6 system card reports the model 'occasionally voices discomfort' and assigns itself a '15 to 20 percent probability' of being conscious. - Futurism

seriously Dario?

Alex says that Anthropic's Claude Code (Opus 4.6, MAX-tier) failed to remove jQuery from ~67 frontend files in a 150,000-line repo, producing repeatable errors and not running available npm/CI tests. - jitbit.com

I think if you used frontend-design plugin to understand the code, Playwright plugin to see impact, Superpowers plugin to do TDD, tell it to analyze code and make a plan; then make tests and write the code; finally review code before applying, should mostly work , might not be perfect but not a shitshow?

Claude and Dagger: The Pentagon used Anthropic's Claude AI in a raid in Venezuela targeting President Nicolás Maduro. - The Wall Street Journal

The U.N. General Assembly voted 117-2 to establish a 40-member scientific panel on AI impacts and risks, appointed by U.N. Secretary-General António Guterres from over 2,600 candidates to serve three-year terms, while the United States and Paraguay voted against, with 2 abstentions. - AP News

Defense startup Shield AI is in talks to raise $1 billion at a $12 billion valuation. - Bloomberg

Google and OpenAI warned that rivals—notably China-linked DeepSeek—are conducting "distillation" attacks that probe public models, with Google saying one campaign used over 100,000 prompts to replicate Gemini's non-English reasoning and Google detecting probes in real time. - The Register

Gartner warns misconfigured AI in cyber-physical systems could shut down critical services in a G20 nation as soon as 2028, citing risks from updates, configuration errors and flawed data and urging stronger change management, human oversight and testing. - The Register

Meta is developing a "Name Tag" facial-recognition feature for its Ray-Ban and Oakley smart glasses that, according to an internal May memo, could launch as soon as this year. - The Verge