A tiny recursive reasoning model achieves 45% on ARC-AGI-1 and 8% on ARC-AGI-2