The reinforcement learning system for Reflection AI's Beam model showed continuous improvement limited only by available compute