OpenAI opened Parameter Golf to test how researchers handle a tightly bounded ML task. Teams had to squeeze a model and training script into a 16 MB artifact, keep training under ten minutes on eight H100 GPUs, and minimize loss on the FineWeb benchmark. Over eight weeks more than 1,000 individuals submitted 2,000+ entries, ranging from optimizer tweaks to novel model ideas and scripts for reproducibility.

Most contestants leaned on AI coding agents, which slashed experiment setup time and let newcomers iterate quickly. RunPod’s $1,000,000 compute grant kept the barrier low, while OpenAI built a Codex‑based triage bot to flag out‑of‑rule submissions amid a flood of daily entries, and reduced overall cost for participants. Agent‑driven workflows spread strong ideas fast but also generated noisy, rule‑bending variants.

The contest surfaced unexpected talent; OpenAI says the open‑ended format highlighted participants with deep ML intuition and persistence. Judges reproduced each record‑breaking entry, confirming genuine advances in quantization, test‑time training and new modeling concepts. Parameter Golf proved that constrained, AI‑augmented challenges can drive both rapid prototyping and meaningful discovery in modern machine‑learning research, and set a benchmark for future contests.