"Note on "tuned": OpenAI shared they trained the o3 we tested on 75% of the Public Training set. They have not shared more details. We have not yet tested the ARC-untrained model to understand how much of the performance is due to ARC-AGI data."

Really want to see the number of training pairs needed to achieve this socre. If it only takes a few pairs, say 100 pairs, I would say it is amazing!