Legends of the Kaggleverse

CPMP has every right to be upset

Nobody knows the final-days scramble like CPMP. Even in a winning year his team was still experimenting at the end: the NVARC writeup says an ensemble with Tiny Recursive Models first worked “the day before last,” with no time left to tune it. They won ARC Prize 2025 anyway.

So when his ARC-2 submissions started failing this week, he knew exactly what was at stake. His forum thread was titled “3rd day without working sub,” and his follow-up was the saddest line of the week: “I’m not stressed, I’m sad and frustrated.”

He pointed at the L4s early. ARC-2 runs on Kaggle’s L4x4 machines, four NVIDIA L4 chips apiece, and Kaggle’s own page calls that access limited. A month ago CPMP warned Kaggle staff the L4 queue was jammed, and Kaggle’s Dustin answered that demand had increased considerably and added capacity. Gemma’s competition page explicitly offers Kaggle L4x4 machines. CPMP pointed this out in his November 1 reply to our article, and he was right: our earlier wording gave the misleading impression that the hardware was exclusive to ARC-2. Both competitions offer that hardware.

Gemma’s discussion board supplies further evidence. On September 28, Ryan Holbrook paused rescoring because of L4x4 queues. On October 8, Dustin attributed long waits and erroneous failures to heavy L4 submission volume; Holbrook said the next day that doubling the machines had not contained demand. Dustin estimated a 33-hour queue on October 10, then about 14 hours on October 13. On October 16 he said the earlier queueing-error problem had been fixed. Those figures do not establish the queue length on October 31.

Late-October complaints continued in the same Gemma thread. On October 26, Jason Krijgsman reported a submission pending since the previous afternoon. On October 28, Starright reported scoring taking over a day and suspected renewed queue trouble. On October 31, Sawan Ruparel reported repeated submission timeouts after more than 14 hours or a day, asking whether his configuration was responsible. These participant reports establish continuing submission problems; they do not separate GPU queue time from execution time or establish the cause of each failure.

ARC-2 staff also acknowledged more sessions than machines. Both contests offer L4 hardware, and both had late-October complaints. Pressure on shared L4 capacity is a plausible explanation, consistent with CPMP’s concern. The available evidence does not establish how Kaggle divided that capacity between contests or how much of ARC’s delay came from Gemma.

Correction · November 1, 2026: We have corrected the implication that L4x4 access was exclusive to ARC-2. Thanks to CPMP’s reply; see Gemma’s Upgraded Accelerators section. We have added Gemma’s October 26, 28 and 31 participant reports alongside the earlier staff updates, including the reported improvement and fix. We distinguish the historical queue estimates from late-October reports and from our interpretation of shared capacity pressure.

Nobody here is a villain. Competitors are reporting long queues and failed submissions, and people who love this game are still trying to finish their entries. The legends haven’t gone anywhere. They’re still submitting.