570,000 Lines of LLM Code Compiled Fine. It Was 20,171x Slower Than SQLite.
Someone benchmarked an LLM-written Rust reimplementation of SQLite. The gap between code that looks right and code that is right turned out to be five orders of magnitude.
Simple thoughts on building, designing, and shipping.
Someone benchmarked an LLM-written Rust reimplementation of SQLite. The gap between code that looks right and code that is right turned out to be five orders of magnitude.
Four projects shipped in the last two months show what happens when AI agents handle not just coding but earning, orchestrating, and running entire companies.
After a year of agent-assisted development, I found that structured spec files fixed the inconsistency problem better than any prompt technique.
I reverse-engineered how Codex handles context overflow compared to Claude Code. The answer involves AES encryption, session handover patterns, and KV cache tricks.
Shopify CEO Tobias built QMD, an open-source search engine. Connect it to Claude Code and every session gets persistent memory.
Anthropic's Claude Code team rebuilt their tools three times. Fewer tools made the AI perform better. Here are four hard-won design principles.
Your AI isn't getting dumber. Your main session is overloaded. Sub-agents keep it lean and accurate for over an hour.
A race condition between Auto Memory and context compaction in Claude Code v2.1.59–v2.1.61 broke prompt caching and corrupted sessions. Anthropic reset all weekly limits as compensation.
Agentation gives AI agents pixel-perfect visual feedback via CSS selectors. Readout replays Claude Code sessions like video. Together they eliminate the two biggest friction points in AI-assisted frontend development.
After building a product with agents overnight, I finally get why Stripe Minions and Ramp Inspect both chose cloud-isolated environments over running everything locally.
Have a project in mind or just want to chat?
I'd love to hear from you.
I'm always open to a conversation.