Ran 106 tasks through six different agent harnesses, logging over 5,000 runs, revealing a 23.8-point performance gap.

Princeton study finds GPT-4 outperforms traditional search for book recommendations

Princeton study finds AI chatbots triple ad influence vs. search, raising transparency concerns

Launched 'Reproducibility Challenge' with Stanford to address AI research reproducibility crisis

Developed OpenClaw-RL AI system for continuous learning from conversations

Open-sourced LabClaw, a Skill Operating Layer for LabOS