Eliezer Yudkowsky

2004: Coherent extrapolated volition

In Coherent Extrapolated Volition (2004), he proposed that a superintelligence should act on what humanity would want if it knew more and had grown up farther together. A 2016 Arbital entry says this is not meant for the first AI. See Coherent extrapolated volition.

2022: Death with dignity

In MIRI announces new “Death With Dignity” strategy (2022-04-02), he wrote that humanity is unlikely to solve alignment in time and proposed measuring progress in log-odds of survival. See Death with dignity.

2022: List of Lethalities

AGI Ruin - A List of Lethalities (2022-06-05) lists 43 reasons he expects alignment attempts to fail in practice, while saying none is impossible in principle.

2026: Challenge to Against SURA

In a comment on Against SURA, he asked what policy Scott Alexander would recommend if labs cannot align superintelligence in the time they have. See Against SURA.

Pivotal use

In a post on X, he explained why a safe but limited AI does not by itself resolve the risk from other AI projects, and argued for a treaty instead. See Pivotal act.

Related: Superintelligence.