“If the current frontier is somewhere near DeepSeek-R1, and continues to grow by 10x every 3-5 months, then reasoning training compute will probably converge with the overall frontier by 2026 — after which this growth rate could slow down within a year or so.”
Said May 9, 2025Deadline Dec 31, 2026Epoch AI Gradient Updates, 'How far can reasoning models scale?', May 2025
Resolves on whether reported reinforcement-learning/reasoning training compute reaches the scale of total frontier training compute, after which growth would revert to ~4x per year.