

Distributed databases with Peter Mattis
Topic 13 of 13
Published Sep 30, 2026 · Clip 1:09:17–1:40:28 (31:11)
Peter traces his AI journey from GitHub Copilot autocomplete through Claude Code in the terminal to the Claude desktop app and occasionally the Codex desktop app for adversarial second opinions with models like Fable and Astra. He describes a Thanksgiving-weekend breakthrough where a long-postponed benchmarking matrix for CockroachDB materialized over four days with Opus, and building a B-tree of ~10,000 lines of optimized Rust in ~30 minutes with agents. He explains using models as sparring partners for design exploration (leading to a research breakthrough in autoscaling), running 5–10 concurrent sessions with many sub-agents, treating agents as a hands-on architect role rather than management, the need to be more ambitious, agents' laziness about testing and the fix via firm guidance and advanced testing techniques, designers writing production HTML/CSS directly, his view that code review scrutiny decreases as agents improve and may go the way of assembly, and his advice that AI is an amazing tutor for dissecting code and that curiosity and personal agency are essential.