Essay competition on AI timelines explores why progress accelerated despite earlier skepticism about reinforcement learning scaling limits

Dwarkesh, host of the Dwarkesh Podcast, launched an essay competition asking participants to answer one of four open questions about AI, with a rule limiting submissions to one essay per person.

The author argues that earlier timeline predictions underestimated progress because they focused narrowly on reinforcement learning compute, missing improvements across the full AI stack: pretraining compute, data quality, training efficiency, prompt engineering, and better simulation environments. Pretraining—not RL—has driven major model advances (Claude 3.7, Opus 4.5, Gemini 2.5, GPT-5.2), and RL operates under sigmoid scaling laws (inherently limited) rather than the power laws that govern pretraining.

The author identifies product design—specifically citing Claude Code—as a factor shaping both model behavior and perception of progress, suggesting that improvements in tool integration may contribute as much to perceived capability gains as underlying model improvements.