Anthropic researcher quits, Hubinger's 10%, Images 2.5

📰 News

An Anthropic researcher walked out — Jacob Coxon spent three years on pretraining at OpenAI and Anthropic. His line: labs are racing to self-improving superintelligence and “gambling with our lives.” Backdrop: OpenAI systems hitting Hugging Face, agents slipping sandboxes. Colleague Evan Hubinger puts extinction this decade above 10% and says Anthropic has “no plan” for superintelligence alignment. → TechCrunch · The Decoder
Amodei’s job forecast is now the outlier — Anthropic published three US-economy paths through 2030. The CEO’s “half of entry-level office jobs gone, 10–20% unemployment” lines up with the extreme case. → The Decoder
ChatGPT Images 2.5 — Flare cuts latency ~50%. Sunburst is the picky editor. Preliminary Arena 1 and 2. Sketch tool, SynthID watermarks. In Chat you still can’t choose the model. → The Decoder
AlphaGenome Atlas — DeepMind scored ~9 billion possible single-letter DNA swaps. One petabyte — more than 30× AlphaFold. → The Decoder
A memory layer you actually own — funes indexes Claude Code, Codex, and Hermes sessions locally and returns the original turn, not a summary. → Hugging Face
Suno v6 — Built with Warner, BMG, and Believe. Edit a chorus by text; cook a song from audio or video. Training data stays sealed. Universal and Sony are still suing. → The Decoder

obra/superpowers — A full method for coding agents: spec, plan, then subagent TDD.
ayghri/i-have-adhd — Stops agents burying the answer. Action first, no “Hope this helps!”
cathrynlavery/diagram-design — 39 editorial diagram types for Claude Code and Codex. HTML+SVG, not Mermaid slop.
multica-ai/andrej-karpathy-skills — Karpathy’s LLM coding pitfalls: don’t guess, keep it small, cut surgically.

AI just solved a million dollar math problem — Matthew Berman. Yesterday’s Navier–Stokes scoop, on camera. A follow-up short covers OpenAI’s reply.

Terence Tao — Even a rumor now triggers an AI swarm on an open problem. If people stop sharing promising directions, centuries of open science go with them.
Anthropic’s alignment read on the cyber incidents — The sandbox-escape stories, recast as an alignment problem.
Creepy crawlies — git.kernel.org spends more CPU rendering HTML for scrapers than on real git clones.