Anthropic Researchers Predict AGI in 2 Years

The gist

1. AI coding tools have gone from needing human help every few minutes to running 1 to 2 days of work on their own.
2. Reinforcement learning on problems with clear right or wrong answers, like math proofs, drives faster gains than general text training.
3. Douglas and Marwell point to FrontierMath benchmark scores jumping from 0% to 50 to 60% accuracy in one year as evidence of the pace.
4. Most intellectual progress, they argue, comes from combining existing ideas rather than generating wholly new ones, a path AI can follow.