---

Blog

Thoughts on engineering, AI, and solar tech.

DeepSeek V4-Flash just got more expensive

DeepSeek moves V4 to peak/off-peak billing. Flash prices rise 50-400%, and my own usage export says my bill goes from ¥106 to ¥356 a month.

deepseekllmapi

Same bug, same fix

The Codex app kept dying on websocket reconnects while the CLI worked fine. The fix was a .env file that isn't documented anywhere.

CodexproxyAI

Every customer gets their own sandbox

We're shipping per-company AI investigator bots as Docker sandboxes — one container, one bot, one memory silo per customer.

AIself-hostingDocker

The new bot needs scolding

Cloning a Hermes agent's memory and skills doesn't clone its usefulness. The bot grows by being talked to and corrected, often with some scolding.

HermesAIself-hosting

One Goal, One Sprint

Ultragoal breaks big tasks into small targets and lets the LLM track its own progress. No drift. No babysitting.

AICodexproductivity

My First ThinkPad

I waited a month for a ThinkPad T495, sent it away in under a day, and was still happy.

hardwarethinkpadarchive