Recommends

Crazy that Alexandr Wang is under 30. My takeaways: he has 200 direct reports in Meta Superintelligence Labs because he does not believe in bureaucracy. Some of those reports are technical leads, so they have technical management rather than people management. He said Muse needed a singular point of view, not a Frankenstein of features from different parties, or “PM hell.” I also found it really interesting that they are coupling the model and harness, with training built around personal agents. I am really loving Muse and really enjoyed this pod.

Read this and now mcp-pilled. What Mario and Armin say goes. Now I’m going to learn about code mode with MCP. From initial research, looks like a solved problem for all the issues with MCP. Composability through a JavaScript function, low latency from parallel tool calling, and context management by keeping tokens out of the window

Interesting to hear about Zuckerberg's obsession with the Roman Empire. The big takeaway for me is the dictator's dilemma. The more power you have, the more people are incentivized to tell you what you want to hear. You have to get close to the metal and go to the factory floor, rather than just believe the reports coming up to you.

Jev is super interesting. The example I like a lot is skill routing. Let Jev read the user's intent and choose the one skill that fits instead of having all the skill names and metadata in the context. I also see this as a fast alternative to the lazy loading of MCP server tools.

Loved hearing about Matt’s backstory and how he architects writing agents skills and relation to the fundamentals of software design. I bought all the books he recommended!

This is a masterclass in opinionated agentic coding. Love hearing their takes and I really want to research the grammar sampling used for RL.

the inspiration for this page