Tufalabs wins ARC-AGI-3 Milestone #1 with a 1.21% score, open-sourcing its top-scoring solution
Two other winners built on the organizers' GPT-OSS template.
Story Brief
Two other winners built on the organizers' GPT-OSS template.
Two other winners built on the organizers' GPT-OSS template.
Two other winners built on the organizers' GPT-OSS template.
We just announced the winners of the ARC-AGI-3 milestone #1 prize. $37.5K in total prizes My notes: * 2nd and 3rd place both use a template we made (GPT-OSS) which is great. That means the tools are helpful * @tufalabs had the most novel solution of the 3. The best place to learn about their work is their write up, submission and @MLStreetTalk video * We're already seeing edits on top of Tufa's submission (a good thing!) * My guess is they have something up their sleeve they _didn't_ open source which we'll see in a few weeks. No inside information, just game-theory guessing * DevOps is still an issue on Kaggle I'd like to fix. Iteration times are slow, lack of error logs is very hard, and overall researchers are doing more engineering than research
