This Week's Feed

As agents handle more execution, the interesting question isn't "what will AI take over" but "what does the human layer become responsible for?" The Anthropic managed agents post, Taylor Pearson's "As We May Work", and the Karpathy knowledge-base idea all point toward a world where human value is increasingly about curation, synthesis, and judgment — not generation. What does that mean for how we train attention?

Jack's "from hierarchy to intelligence" framing + Schmachtenberger's critique of market coordination both suggest the same thing: our current institutional structures are optimized for control, not information flow. As AI makes intelligence cheaper, the coordination bottleneck shifts from access to information to quality of judgment at the top of hierarchies. That's a slow-motion crisis for most organizations.

Chiang's "blurry JPEG" metaphor is more interesting now that agents are acting on compressed knowledge. If LLMs are lossy compressions, what breaks when you chain lossy compressions together in an agentic pipeline? And conversely — Memvid and Karpathy's knowledge bases are attempts to keep things lossless. There's an engineering and epistemology problem here worth thinking through.