Rational agents do not need goals. That is the central claim of a new essay in The Gradient by an author making a case against the dominant goal-directed model of AI alignment. The argument draws on Aristotle's concept of eudaimonia, not as a feel-good flourishing target to optimize toward, but as a structurally distinct form of deliberation the essay calls 'eudaimonic rationality.' The core formula is 'promote x x-ingly': art is the artistic promotion of art, honesty is promoted honestly. This is not a metaphor. The author argues it describes how human rational activity actually works, and why goal-based AI systems produce a dangerous type mismatch when asked to support it.
The practical alignment stakes are specific. Properties like corrigibility, harmlessness, and transparency are unstable and brittle when interpreted as goals or rules by a consequentialist agent. They become natural, the author argues, when interpreted as dynamics within networks of actions, action-dispositions, evaluation criteria, and action-resources, what the essay formally defines as 'eudaimonic practices.' Part II uses Terry Tao's account of research mathematics as a concrete worked example of this structure. Part III argues that the instrumental-versus-terminal-value distinction, foundational to Effective Altruism-style alignment thinking, is mostly unnatural for agents operating this way.
The essay is long and explicitly warns it sprawls, but the architecture is mapped in a section-by-section overview worth reading before diving in. The payoff is in Parts IV onward, where eudaimonic rationality is given a technical definition and tested against classical AI safety problems. The deeper question the whole piece is really asking: whether certain forms of agency have lower algorithmic complexity, are more learnable by ML training processes, and converge more reliably across different kinds of rational systems. If the author is right, alignment researchers have been optimizing toward the wrong target type entirely.
[READ ORIGINAL →]