Theo Browne describes experimental projects that keep coding agentsAn AI coding agent is a tool-using AI system that can inspect, modify, and validate software within a repository. running for extended periods, including a TypeScript compiler port, a Rust compiler and a skateboarding game. He cautions that the workflows are costly and does not recommend them as ordinary development practice.
He explains how objective testsDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. and performance targets can guide repeated changes. For the game experiment, he describes using model comparisons of screenshots as an imperfect feedback loopAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior..
Theo Browne discusses bottlenecks in builds and compute, adding machines and on-demand environmentsAI compute infrastructure is the hardware, facilities, networks, power, cooling, storage, and software used to train and run AI models., and asking agents what is slowing progress. He discloses a sponsorship relationship with Blacksmith and an investment in Boat.
Watch on YouTube
