Introducing Quibbler MCP! Now you can use Quibbler to critique any coding agent, not just Claude Code. Here's a demo of Quibbler working on Cursor:

live monitoring of agents in real-world settings allows us to learn lessons about AI control that can later be applied to x-risk mitigation. hoping this open-source launch helps control researchers experiment with more realistic threat models!

Today, we’re excited to announce that we’re open-sourcing two tools: Quibbler is a background agent that critiques your coding agent’s actions. Orchestra is a multi-agent coding system: it uses a designer agent that spawns and coordinates your parallel coding agents.

i will be at COLM in montreal next week presenting the breakpoint eval with @KaivuHariharan Would love to meet and chat with people, some things I'm interested in right now: - human-in-the-loop training algorithms - scalable oversight - automated auditing

Going to COLM! Thinking about: 1. foundations of eval/RL env design (what's the design space? can we automatically measure eval quality?) 2. outer loop optimization for LLMs (GEPA, Alphaevolve, etc) 3. human preference learning in agentic settings (e.g., coding preferences)"

My friends @KaivuHariharan @uzpg_ are building a startup! I think they are highly capable, have good taste, and are motivated by the right reasons. Let them cook 🍳

Today, we’re announcing Fulcrum Research, a startup scaling human oversight. We are building debuggers that tell you why your agents fail, and what your rewards are truly testing for—the first step toward the inference-time infrastructure required to safely deploy agents.

Thought real machine unlearning was impossible? We show that distilling a conventionally “unlearned” model creates a model resistant to relearning attacks. 𝐃𝐢𝐬𝐭𝐢𝐥𝐥𝐚𝐭𝐢𝐨𝐧 𝐦𝐚𝐤𝐞𝐬 𝐮𝐧𝐥𝐞𝐚𝐫𝐧𝐢𝐧𝐠 𝐫𝐞𝐚𝐥.