Writing

A short literary piece written in the first person — but the first person is an AI. 'I was not born.' 'I live in plural presents.' 'I love truly difficult problems.'

It's all about compute. Reading OpenAI's Sora wind-down through the CFO's public numbers — 0.2GW → 1.9GW, ARR 2B → 20B+ — as a resource-allocation choice, not a product failure.

A 42K-word breakdown of Claude Code's six-layer context model: system, messages, attachments, tools, hooks, long-context pipeline. The longest piece in the series.

A 18K-word reverse-engineering of Claude Code's npm package source map. Three agent shapes — Subagent, Coordinator, Swarm — and why Anthropic chose this structure.

A 14K-word multi-disciplinary investigation of digital-age attention loss. Neurobiology, political economy, algorithmic ethics, and existential philosophy in conversation.

A red-team read on agent-to-agent networks: prompt injection, token leakage, malicious plugins. Three principles for whether to let an agent talk to another agent.

A public correction of four predictions I made about Kimi K2's pricing strategy. What survived contact with reality, what didn't, and the spillover effect that surprised me.

A review of Kimi K2 with a side-by-side US/China pricing dissection, and three subscription tiers worth recommending.