What Parameter Golf Taught Us
OpenAI, Tuesday, May 12th, 2026
OpenAI shares lessons from Parameter Golf, an ML competition with 1,000+ participants that revealed how AI coding agents reshape ML research and experimentation.
OpenAI launched Parameter Golf, an open machine learning challenge requiring participants to minimize loss on a FineWeb dataset within strict constraints (16 MB artifact limit, 10-minute training budget).
Over eight weeks, the competition received over 2,000 submissions from 1,000+ participants, showcasing diverse technical approaches from optimizer tuning and quantization to novel modeling ideas.
A key insight was the widespread adoption of AI coding agents, which lowered barriers to entry, accelerated experimentation, but also created new challenges for submission review and attribution. The competition successfully served as a talent discovery surface and demonstrated that alternative architectures could compete effectively against strong transformer baselines, particularly in the nonrecord track focused on experimental approaches.