SutroYaro：AI辅助的能效机器学习研究工作室

SutroYaro is an innovative AI research collaboration platform designed to support energy-efficient machine learning experiments. It features structured workspaces, a locked evaluation mechanism, multi-agent collaboration with diverse AI tools, and a focus on the sparse parity problem using Data Movement Distance (DMD) as the core energy efficiency metric. This post breaks down its background, design, research focus, and significance.

As large language models and deep learning systems scale, AI energy consumption has become critical. The Sutro Group, a team focused on energy-efficient AI training, holds weekly meetings at San Francisco's South Park Commons. SutroYaro aims to turn their research methodology into a scalable, collaborative platform—more than a code repository, it’s a complete workflow system enabling parallel experiments with different AI tools.

SutroYaro’s design follows key principles: AI as a research partner, locked evaluation harness, and knowledge accumulation. The locked evaluation harness uses SHA256 verification to prevent modification of evaluation standards, ensuring fair comparisons between methods.

SutroYaro focuses on the sparse parity problem (learning XOR/parity functions from random inputs) and has completed 36 experiments. It uses Data Movement Distance (DMD) as the core energy efficiency metric, measured automatically via TrackedArray.

The platform supports researchers using preferred AI tools like Claude Code, Gemini CLI, and Codex. It provides automated scripts for single-cycle or overnight batch experiments—each cycle reads accumulated file states, runs experiments, and records results.

SutroYaro has a Gymnasium-compatible evaluation environment with 72-point scoring for testing AI agents. It also has a structured knowledge system including DISCOVERIES.md, TODO.md, LAB.md, and AGENT.md.

SutroYaro represents a new AI research organization approach emphasizing reproducibility, scalability, and collaboration. It extends AI-assisted programming to research itself, providing a scalable platform for energy-efficient ML studies.