What Do You Mean by a Two-Year AGI Timeline?

Koby Lewis, writing on LessWrong, addresses a subtle but critical confusion in discussions about AGI (Artificial General Intelligence) timelines. When someone says 'I have a two-year AGI timeline,' it's often unclear what statistical metric they're using. Lewis notes that many assume it refers to the expected value (mean) of the arrival time distribution. However, if there's any nonzero probability that AGI is never developed—due to a catastrophe, a ban, or other reasons—the unconditional expected value becomes infinite or undefined, making this metric meaningless.

Lewis outlines two common interpretations: the conditional mean (expected value only for scenarios where AGI does arrive) and the median or percentile (the time by which AGI is more likely than not to have arrived). He suggests the latter is more common, citing platforms like Metaculus, Manifold, and Epoch AI that explicitly use percentiles. The post encourages clearer communication, recommending that people explicitly state their metric and whether they assign non-negligible probability to AGI never arriving. This nuance matters for accurate risk assessment and planning in AI safety and development.