Whether to generate pixels, whether video models grasp physics, whether an LLM is already a world model — three debates that are really one question. Plus three bets on how each ends.
From physics first principles: why shedding a gigawatt of waste heat in a vacuum — not finding the power to make it — is the real bottleneck for AI data centers in orbit.
Video gives robots eyes. World models give them imagination. Neither gives them hands. A few thoughts on what robotics' scaling law actually runs on.
Rich Sutton wrote The Bitter Lesson in 2019. In 2025 he came back and said LLMs aren't actually following it. Three thoughts on what's been written between his two chapters.
Three labs in 2026 are independently calling their LLM 'time-aware' and meaning three different things. The framing problem nobody is naming, the cognitive deficit they all describe, and three falsifiable bets on what comes next.