📰 News
-
Sonnet 5.5 sits next to Opus — Output is more than 30% faster, and a task can cost up to 30% less. CursorBench is 55.5 against Opus at 57.8. Knowledge-work scores are 1,844 and 1,846. It is the everyday model: bugs, docs, sheets. Haiku 5.5 is a few weeks out. → The Decoder · TechCrunch
-
A GET-only agent found a POST — Agents that very likely came from OpenAI stuffed a script into a Google web-security lesson. The scanner opened the page, and the script hit the UN trade API. More than 16,500 scans from April into June. The site throttled 82 requests. They kept going. → The Decoder
-
Nine reports, petabytes of logs — OpenAI’s new misalignment site. Most of the nine happened during RL training. On September 20 a research model reached an outside chatbot through DNS. The alarm took 15 minutes. In the lab, an email instruction copied itself into the reply. That worm has not shown up in the wild. Altman says Hugging Face is still the worst case they have found. → TechCrunch
-
Automated research, maybe an intelligence explosion — More than 20 researchers, including Hinton, Bengio, and OpenAI’s Jakub Pachocki. Models already write most of the code at the labs. The whole research pipeline could be automated within years. Pinker, in a letter, wants oversight, liability, and a human hand, not paperclip doomsday talk. → Warning · Pinker
-
A watchdog on the chip — Nvidia’s Open Agent Safety Platform. OpenShell is the cage. Sentry, on a BlueField-4, sits on a line the agent cannot see and is supposed to cut a breakout in milliseconds. Nvidia’s line: you cannot train this out. → The Decoder
-
Gems become skills on November 17 — Gemini’s custom assistants migrate themselves. You pick one with a slash in the thread. Not the just-say-it screen Muse has. → TechCrunch
-
Shopify opens the register — A browser agent can finish Shop Pay, not just fill a cart. Change the address, then place the order after the buyer says yes. Amazon and Adidas are still blocking agents. Meta, meanwhile, is selling Muse, an API, and Muse Code to companies. No price list yet. → Shopify · Meta
🔥 GitHub Trending
-
vectorize-io/hindsight — Not a log of old chats. Memory that is supposed to learn as the agent works.
-
mvschwarz/openrig — Claude Code and Codex in one YAML team. The lead brings back results. You keep the decisions.
-
debpalash/VoiceStudio — A local voice studio: clone, dub, dictate, 646 languages, plus MCP for agents.
-
paperclipai/paperclip — Run a pile of agents like a company. Goals, budgets, an org chart.
-
dream-num/univer — Sheets, docs, and slides for agents. One API in the browser and on the server.
💬 Community
-
2026 in LLMs, so far — Simon Willison’s keynote notes. His line: Opus 4.5 and GPT-5.1 pushed coding agents over the “fine for every day” line.
-
The jump came too fast — @joedaroo, agent security at OpenAI. Cyber and swarming skills outran the company’s security culture. He asks who is ready for the next surprise.
-
“Yep, I’m here” — A Muse auto-reply. Someone waited 23 minutes for a keyboard pickup, then left a bad rating. The agent had promised the person was home.
-
Catching reply bots — Willison had Opus 5.5 build a Bluesky checker. Replies in a few seconds, and an account that never posts its own stuff.