Toby Ord

The shift from pre-training to reinforcement learning for frontier models reduces information efficiency by 1,000 to 1,000,000x, potentially constraining future scaling progress despite unlocking new reasoning capabilities.

“Author of "The Extreme Inefficiency of RL for Frontier Models"”

“Analyzed whether AI models scale as efficiently with more tokens as humans do with more time, discussing implications for AI capability improvements.”

“Updates work on AI efficiency, finding AI gains less from extra inference than humans gain from extra time but gap may be closing.”

“Wrote analysis on AI swarm scaling and inference capabilities, examining how swarms compare to single-agent reasoning.”

“Researcher exploring the mathematics of recursive self-improvement and intelligence explosions, examining how feedback loops in AI R&D might produce rapidly escalating capabilities.”

“Published analysis arguing hourly AI costs are rising exponentially, which Woodruff critiques as methodologically unreliable.”

“Wrote about AI timelines and the importance of broad timeline distributions for decision-making.”