What Parameter Golf Taught Us About AI Coding

The Parameter Golf challenge asked over 1,000 participants to minimize machine learning loss within tight limits: a 16 MB artifact (weights and code) and a 10-minute training budget on 8 H100 GPUs. What started as a technical puzzle revealed deeper insights. AI coding agents became a huge factor—they slashed the cost of experimentation, broadened participation, and sped up the pace of competition. But they also created headaches for scoring and attribution. The challenge surfaced remarkable creativity: from careful optimizer tuning and quantization tricks to test-time training and novel modeling ideas. Nine standout submissions showcased the range of winning strategies, including rule-bending evaluation approaches that required extra scrutiny. Organizers also found that open-ended challenges like this are powerful for discovering top machine learning talent. Why it matters: As AI agents become standard tools in research, competitions must evolve to reward genuine insight over agent-assisted brute force, reshaping how we evaluate technical creativity.