📰 News
-
GPT-5.6 Sol post-trained Luna with a vague one-liner — A follow-up detail from yesterday’s Sol launch worth flagging: a researcher handed Sol a “fairly under-specified prompt” and it autonomously picked the right GPUs, configured training, launched the script, and verified the run to post-train the smaller Luna model — all on its own. Sol now scores 16.2 points higher than GPT-5.5 on OpenAI’s internal RSI (Recursive Self-Improvement) benchmark. An OpenAI researcher: “Previously this is something that a team of senior researchers may have worked on, and now it really feels like the automated researcher is pretty close.” → The Decoder
-
Apple sues OpenAI over trade secret theft — The central figure: Tang Tan, OpenAI’s Chief Hardware Officer, who spent 24 years at Apple before jumping ship. Apple alleges he asked job candidates to bring Apple hardware components to interviews, coached departing Apple employees on evading security procedures, and leaked confidential project code names. Apple calls this “the tip of the iceberg” and is asking the court to bar OpenAI from using its secrets. The timing — OpenAI is rumored to be building a smartphone that competes directly with the iPhone — makes this one land hard. → TechCrunch
-
Fidji Simo steps down as OpenAI’s No. 2 — Simo, who had COO, CFO, and CPO all reporting to her as CEO of Applications, is transitioning to a part-time advisory role after a relapse of a neuroimmune condition. This leaves a significant leadership gap right as OpenAI heads toward an IPO. Altman on X: “i am really sad about this and very grateful for all fidji has done.” → TechCrunch
-
OpenAI kills Atlas browser after 8 months — Atlas, launched last October as a potential Chrome competitor, is dead. Its features are being folded into a Chrome sidebar extension. The read: OpenAI has no way to pull users away from Chrome now, which means Google keeps all that browsing data. A new “Computer Use” feature (ChatGPT directly clicking and typing on your screen) moves into the desktop app instead. → The Decoder
-
Tencent moves to buy Manus after Beijing forced Meta out — After China voided Meta’s $2B acquisition of Manus (calling it a “conspiratorial” attempt to undermine China’s tech base), existing investor Tencent is now negotiating a majority stake at the same $2B valuation. Manus stays in Singapore and operates independently. Annual revenue is close to $500M. The fit: Tencent wants to embed a Manus agent directly into WeChat. → The Decoder
-
Bun rewrites 1M+ lines from Zig to Rust in 11 days — with Claude — The JavaScript runtime ran 64 Claude Fable 5 instances in parallel for 11 days to complete a full Zig→Rust rewrite. API bill: ~$165K. Human equivalent: roughly a year. No sticker shock for the team — Bun was acquired by Anthropic in December 2025. The new Bun v1.4.0 fixes 128 bugs and runs 2-5% faster. → The Decoder
-
AI agent startup let its agent run its own $100M fundraise — Lyzr, a 3-year-old NJ startup building enterprise AI agents, had its agent SivaClaw field questions from 130+ investors, draft investment memos, and track which slides got the most attention. Series B closed at $100M ($500M valuation). The founders never flew to Sand Hill Road. Generated $400M in investor interest without a single coffee meeting. → TechCrunch
🔥 GitHub Trending
-
addyosmani/agent-skills — Production-grade engineering skills for AI coding agents. Eight slash commands (/spec, /plan, /build, /test, /review, /ship) encode the workflows senior engineers actually use — installable in one command into 70+ agents including Claude Code, Cursor, Codex, and Copilot.
-
wonderwhy-er/DesktopCommanderMCP — MCP server that gives Claude terminal control, file system search, and diff editing. Now also accessible remotely from ChatGPT and Claude Web via Remote MCP. Effectively a universal power tool for AI coding agents.
-
vxcontrol/pentagi — Fully autonomous AI penetration testing agent. 20+ built-in security tools (nmap, metasploit, sqlmap), Docker sandbox isolation, supports OpenAI, Anthropic, Gemini, and more. For security researchers doing ethical pentesting.
-
anthropics/claude-cookbooks — Anthropic’s official collection of copy-paste Claude API recipes: RAG, tool use, customer service agents, SQL integration, and third-party integrations.
▶️ YouTube
- A Model Explosion: GPT 5.6 Sol, Grok 4.5 and Meta Muse Rewrite the Rules — AI Explained | The best single-video summary of this week’s model launch wave. Digs into a dozen details buried under the viral headlines — the self-improvement benchmark, gaming as a new AI frontier, and a SimpleBench upgrade chapter are the highlights.
💬 Community
-
A global workspace in language models — Lobsters | Anthropic research suggesting LLMs may have internal structures analogous to the brain’s Global Workspace Theory. Worth a read if you follow AI interpretability and cognitive architecture.
-
The new GPT-5.6 family: Luna, Terra, Sol — Simon Willison | The clearest breakdown of GPT-5.6’s new API features — Programmatic Tool Calling (compose JS to orchestrate tool calls), native multi-agent support, prompt cache breakpoints. Also has an honest first-person take: “it hasn’t struck me as better than Fable at the kind of complex coding tasks I’ve been using.”
-
AR glasses require constant cloud surveillance — no way around it — Simon Willison | Quoting Nilay Patel from The Vergecast: “To make augmented reality glasses, you need to put a camera next to your eyes that is continuously recording everything you see and processing that… you have to send that data to a cloud. You gotta do it.” Blunt and worth sitting with before the next AR hype cycle kicks in.