The p(doom) database · person
Samuel Marks’s p(doom)
15-20%
In their words
Samuel Marks (Scalable oversight lead, Anthropic) put their p(doom) at 15-20% (p(doom); <5% conditional on scalable oversight good enough to approximate a ground-truth reward signal), on X on Feb 16, 2024.
My P(doom) would probably go from like 15-20% to <5% if we developed scalable oversight good enough
Source · re-checked against the primary source.
In context
Among frontier labs in the database (20 people), the median is 12.5%. Across all 223 people it is 10%. How the medians and the index are built.
A p(doom) is a personal, subjective estimate of the probability that advanced AI ends in human extinction or a comparable, irreversible catastrophe. Scope and timeframe differ from person to person, so compare with care: why p(doom) numbers disagree.