METR's time horizon graph fuels AI acceleration debate among safety researchers

In a new LessWrong post, Lennart Finke challenges the prevailing assumption that measuring AI capabilities automatically reduces existential risk. Focusing on METR (Measuring and Evaluating AI Risk), he argues that its time horizon graph—which tracks AI's ability to complete longer tasks—has been widely cited by media and investors to justify accelerated AI development. The New York Times, The Atlantic, and Sequoia Capital have all used the graph to support narratives of rapid progress, with Sequoia explicitly calling it 'the one exponential curve to bet on.' Finke warns that such data, while intended for safety, now fuels a 'tragedy-of-the-commons' where AI companies and investors profit from automation externalities like job loss and existential risk.

Finke suggests that METR's work, though valuable, may inadvertently amplify the problem because its measurements serve as evidence for both safety advocates and accelerationists. He points to venture capital herd behavior and public statements by OpenAI employees claiming the graph is 'load-bearing' for global stock markets. The post concludes with a call to reweight AI safety efforts—moving from pure measurement toward strategies that directly prevent misuse or dangerous capability jumps, rather than relying solely on transparency metrics that can be co-opted.