THE WORKBENCH
Software I’m building, the problems behind it, and what I’ve learned along the way.
An operating system for computational organizations.
I’m building Syndicate so people can create institutions whose everyday work is done by agents. Its kernel tracks authority, commitments, completed work, and whether the work achieved its purpose. The system is pre-release.
An AI-assisted algorithm discovery platform for turning code search into a repeatable, auditable experiment.
I designed and built the search and evaluation system. It generates candidate programs, tests them in isolated workers, and keeps their results and ancestry. Separate checks have exposed candidates that improved the score by violating the intended arithmetic.
An AI tutor for studying on your own.
I lead the design and engineering of Autodidact. Learners work through reading and practice with Faraday, then return to what needs review. Saved Python exercises give the tutor their work to discuss.
An experimental foundation for a general computer agent: durable work, changing goals, and verifiable progress.
A workspace for running coding agents, following their work, and carrying changes through review and delivery.
A local toolkit that gives coding agents a place to begin in an unfamiliar repository.
Team reports that keep dependencies visible.
I’m building Sitrep to collect progress, blockers, and next steps through the web and email. It generates project, group, and organization summaries with the original updates available to read. Evaluations check whether important facts survive those summaries.
An archive of coding-agent sessions, so the reasons behind a change can survive alongside the code.
A research studio for institutions, connecting sources, shared writing, and the papers they publish.
A workspace for market research, with the data and calculations behind each explanation close at hand.
A Mandarin learning world for young children, with sounds, pictures, and room to explore.
No projects in this category yet.