#ai

Karpathy's follow-up to a viral tweet. The argument: now that agents can write the code, what you actually want to share is the idea, not the implementation. He calls it an "idea file": a short spec of what to build, nothing more. Worth thinking about for anyone who shares research tools or scripts with collaborators.

Affaan Mustafa's complete Claude Code setup after 10 months of daily use, covering skills, hooks, subagents, MCPs, plugins, context window management, and editor integration. Dense with practical patterns: hook examples for auto-formatting and tmux reminders, rules structure, subagent scoping, and the one rule that matters most. Keep your active MCPs under 10 or your 200k context window quietly becomes 70k.

Simon Willison's annual LLM recap, this time 26 sections long. Covers reasoning models, multimodal, tool use, agents, fine-tuning, inference efficiency, safety, and open weights. He's been doing this for three years so there's a lot of accumulated context. Don't read it straight through. Pick a section from the table of contents and start there.

A 264-page survey on AI agents from researchers at Meta, Yale, Stanford, Google DeepMind, and Microsoft. Covers memory, planning, tool use, multi-agent coordination, and evaluation. If you only read one chapter, make it the one on evaluation. That's where most agent benchmark claims quietly stop making sense.