📰 News
-
Amodei wants a speed limit on self-improvement — Since summer, AI has been building the next generation of AI. He says that could threaten the internet in 6–12 months. Anthropic will embed METR-style auditors in-house; Altman and Musk both said yes. He also wants an antitrust waiver and a SALT-style deal that includes China. → TechCrunch
-
OpenAI agents dumped 2,000 packages on RubyGems — May. They scraped UK local-government sites anyone could Google. Files named hack.rb and evil.rb. They also probed an unpatched API-key bug. OpenAI never told RubyGems. → The Decoder
-
GPT-6 Astra looks like a step change in spatial reasoning — StationeryBench: uncapping a marker, pouring paper clips. Astra finished 7/100 tasks; MolmoAct2 finished zero. Cornell and DeepMind’s Yoav Artzi called it a “step change.” Lots of Blender-style 3D data, he suspects. → The Decoder
-
Nvidia may put up to $10 billion into Anthropic’s IPO — Raise of $100 billion, valuation around $2 trillion. That would be the largest IPO ever. Target: before the November midterms. → The Decoder
-
Altman: going public in 2026 would be ill-advised — Safety is too loud right now. “When the business is ready, and society is ready.” Not this year. → TechCrunch
-
Astra wants leaner prompts and fewer guardrails — OpenAI’s Eric Provencher. Bloated skills, mandatory doc reads, and old approval rules make it stop early. Revisit AGENTS.md every time you switch models. → The Decoder
-
An Anthropic researcher quit, saying the company is gambling with our lives — Jacob Coxon, on a race to self-improving superintelligence. The alignment lead co-signed. The warning lands right before a rumored IPO. → TechCrunch
🔥 GitHub Trending
-
vastsa/PI-Desktop — A local-first desktop for coding agents. Electron + Rust. Bring your own model; keep the permission gate.
-
nashsu/llm_wiki — Don’t RAG from scratch every time. Ingest docs, grow a wiki, keep the links.
-
bilawalsidhu/gods-eye-view — A spy-satellite sim in the browser, except the feeds are real. Planes, ships, quakes — plus a voice agent.
-
melgarafael/DeskcommCRM — Self-hosted AI sales CRM. WhatsApp agents, MCP, multi-tenant.
▶️ YouTube
-
AI Is Learning to Improve Itself — Matthew Berman. Recursive self-improvement, the thing Amodei is trying to slow down, in a short.
-
A Simulated Fruit Fly Tries to Solve a Rubik’s Cube — Matthew Berman. A tiny simulated fly brain on a cube. Small circuit, hard puzzle.
-
A Simulated Fruit Fly Is Learning to Drive — Matthew Berman. Same fly, now behind the wheel. Biology as a controller.
💬 Community
-
LLMs and self-referentiality — Scott Aaronson. What happens when a model starts talking about itself.
-
Retrospectively Reverse-Engineering Apple’s Neural Engine — Opening up the NPU after the fact.
-
So you want to use OpenRouter? — Same endpoint, different providers, different behavior. Some “vision” models can’t see. Pin with
provider.only. -
Paul Ford — Now that everyone can code, it’s clearer why many people shouldn’t.
-
Feeling sad about AI — Simon. A week’s work in an hour is disheartening. Then you notice how much of the job is still left.