Where is all the AI safety research uplift?

The author examines why external AI safety researchers appear to experience a much smaller productivity uplift from AI capabilities compared to reports from Anthropic. While reports suggest a 5x to 10x acceleration in AI-assisted work, the author notes that their own team at Redwood has not observed speedups of that magnitude. The piece explores potential reasons for this discrepancy, including limitations in research scaffolds, differences in how well-suited specific research tasks are to AI acceleration, and potential bottlenecks in research management and judgment. Open article →