AI Prompt
BlogAbout
← Back

OpenAI's Control Room, Meta's Muse Ember, and More

4 min read

OpenAI published details on Control Room, the system it uses to keep a hand on the off switch for its own research workloads. Every research job that runs agentic inference has to register with it and needs an approved safety case before it runs. The inference engines then check in with Control Room before each generation, and OpenAI says it can halt a job's output within 35 seconds.

The design choice I like most is that it fails closed. If Control Room itself goes down, all inference is denied rather than allowed. That's the opposite of how most of us wire up guardrails for agents, where the check is a nice-to-have and an outage quietly lets everything through. If you run long-lived agents, "check in before every generation, deny when the checker is unreachable" is a pattern worth stealing (OpenAI post).

What Else Happened

Meta launched Muse Ember, a cheap model built for agent grunt work. It's the smallest Muse model, aimed at tool calls, running tests, and small code fixes, with a 1M-token context window at $0.10 input and $0.50 output per million tokens. Artificial Analysis scores it 36, matching Claude Haiku 5.5 at about a third of the cost per task, and it's also on Google Cloud alongside Muse Wonder and Muse Spark 1.3 (Meta on X, Artificial Analysis).

Fable 5.2 landed at the top of the independent leaderboards. I covered the launch itself in its own post. Since then, Artificial Analysis put it first on its Intelligence Index at 65, five points ahead of Gemini 4 Carbon, and first on AA-3D, its test of building interactive 3D experiences, at 1271 Elo. Cursor has it live and says it's the top model on CursorBench at 60% on max effort (Intelligence Index, AA-3D, Cursor).

Codex can now drive several of your machines. You can connect multiple local computers to the same Codex setup and choose which one runs a task, so builds and tests happen on a dedicated box while you keep working on your own (OpenAI Developers).

Dario Amodei went on 60 Minutes. He said AI could cut half of entry-level roles in law, finance, and consulting within one to five years, citing Anthropic data showing junior technical hires falling from 27% to 9%, and he backed U.S. military use of Claude for targeting. Skeptics pointed to steady jobs and GDP data (jobs story on X, military-use story on X).

Google added two more eighth-generation TPUs. TPU 8i Max is built for huge context windows, and with a second new system Google is now at four eighth-gen TPU systems in a single year (Google on X).

Vercel says new paid teams are picking Anthropic nearly 2:1. That's from its November AI Gateway Index, which compares which model providers new paid teams choose (Vercel).

Tools and Design

Paul Irish shipped chrome-capture. It uses a new Chrome API to record a tab with no dropped frames, at 60fps easily and up to 120fps at 2160p, with tab audio, and it works inside Puppeteer and Playwright. Run it with npx chrome-capture. I'll be using this for demo videos (Paul Irish on X, GitHub).

Enhance is a design canvas for agent-built prototypes. Point Claude, Codex, or another agent at a prototype and it mounts it into a shared canvas where you tweak and redesign, then fold the changes back into the code (launch post).

Meta open-sourced Astryx, a design system for people and coding agents. It ships 150+ accessible React 19 and StyleX components, themes, an MCP server, and a CLI that outputs JSON so agents read the docs the same way developers do. It's MIT licensed, still in beta, and passed 16k GitHub stars over the weekend (meta/astryx on GitHub).

iOTA takes a similar swing at "all AI UI looks the same." It's 65 components plus usage rules that your coding agent installs with one prompt (Bikash's post).

Adam Bardon made a full iOS release kit with Claude and no Figma. Hero and OG images, changelog visuals, an Instagram carousel, and a YouTube Short and walkthrough all came out of one custom workflow (Adam's post).

Codex got a ModRetro Chromatic plugin. It helps you design levels, animate sprites, and playtest ROMs for the Chromatic handheld (OpenAI Developers).

Framer now resets free-plan credits every month. Free users can use the Framer agent monthly instead of just once (Framer).

What I'm Reading

  • Introducing Control Room – how OpenAI registers, checks, and halts its agentic research jobs.
  • Discord's Rust rewrite – Discord's CTO on replacing Electron and React Native with one native Rust codebase, with an early build at about 290MB of RAM versus 1GB+ today.
  • paulirish/chrome-capture – frame-perfect tab recording that works in Puppeteer and Playwright.

That's it for today. 👋