OpenAI Parameter Golf Retrospective: 2,000+ Submissions Reveal Coding Agents Dominated

OpenAI·medium signal

OpenAI published its retrospective on the Parameter Golf challenge (March-April 2026), where participants trained the best LLM fitting in 16MB on 8xH100s in 10 minutes. Over 2,000 submissions from 1,000+ participants revealed that the vast majority used coding agents as part of their workflow — a major shift from previous ML competitions. RunPod sponsored $1M in compute, making the challenge accessible beyond well-resourced labs.