Beliefs and position mid 2026

According to a new LessWrong post, AI researcher Russell Thor is revising his timeline for artificial superintelligence downward—and sees lower takeover risk than many in the safety community expect. He's now betting on two to ten years until transformative AI, with a fifty percent chance that current transformers will discover a better architecture by twenty thirty. Russell argues humans have powerful advantages: we can run AI in synthetic test environments, train multiple systems secretly, and use interpretability research to understand AI decision-making. He sketches an alternative scenario where hardware limits lead to only 'mild' superintelligence, fundamentally reshaping common assumptions about AI risk.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton