Skip to main content

Self-improving cognitive systems

Research from the lab.

Governed recursive self-improvement, with receipts: systems that improve the lenses they think with, and leave a git trail of every improvement, with a human hand on the dial. Findings are published as they happen, and every entry links to the primary artifact: the paper, the talk, the repo.

Measurably self-improving multi-agent system — without retraining

Closed-loop forecasting: a multi-agent system that measures its own forecast error in production and improves on it, with no model retraining. Talk by Eugene Vyborov; the full method is in the paper.

Watch the talk ↗Read the paper (PDF) ↗