Krux
May 14, 2026
OpenAI's Coding-Agent Contest Got 2,000 Submissions in 8 Weeks
Published: May 14, 2026 at 12:12 AM
Updated: May 14, 2026 at 12:12 AM
100-word summary
OpenAI ran an open machine learning challenge called Parameter Golf that drew 1,000+ participants racing to compress models under strict limits: 16 MB, 10 minutes on eight H100 GPUs. The twist? Coding agents dominated, lowering the barrier to entry but flooding organizers with submissions that needed their own AI triage bot to review. Half the entries beat the baseline; top performers hit 1.12 bits per byte. RunPod chipped in up to $1 million in compute credits to widen access. The contest surfaced novel techniques with names like SmearGate and BigramHash. The takeaway isn't the leaderboard. It's that AI-assisted research competitions can move fast and pull in non-experts, but only if...
What happened
OpenAI ran an open machine learning challenge called Parameter Golf that drew 1,000+ participants racing to compress models under strict limits: 16 MB, 10 minutes on eight H100 GPUs. The twist? Coding agents dominated, lowering the barrier to entry but flooding organizers with submissions that needed their own AI triage bot to review. Half the entries beat the baseline; top performers hit 1.12 bits per byte. RunPod chipped in up to $1 million in compute credits to widen access. The contest surfaced novel techniques with names like SmearGate and BigramHash. The takeaway isn't the leaderboard.
Why it matters
It's that AI-assisted research competitions can move fast and pull in non-experts, but only if you automate the refereeing as aggressively as the research itself.