Part 7/9:
A 2021 academic paper affirmed a consistent relationship where increasing thinking time by tenfold yielded performance gains comparable to scaling model training. This relation underscores the potential for exploring not just data and computation but also the strategic timing of AI responses.
Future Directions for AI Scaling
Given the substantial costs of training today's frontier models—estimated in the hundreds of millions—the prospect of enhancing AI through System Two thinking appears promising. Improved performance can arise from scaling the time models spend "thinking", rather than merely amplifying data input or training time.