Scott Alexander kicked off a new round of an old AI argument by asking Claude for a literal dividing line between “fast” and “slow” takeoff. The chatbot put the boundary at about one year, describing fast takeoff as days to months and slow takeoff as a few years.
That answer was a generated synthesis, not a primary source for the views of Paul Christiano, Robin Hanson or Holden Karnofsky. Alexander nonetheless used it as a benchmark, writing that a takeoff lasting longer than one year still looked plausible even if the clock started now.
The argument moved beyond the calendar
Eliezer Yudkowsky responded with a more concrete economic marker. He recalled Christiano defining slow takeoff as a world in which gross domestic product first doubles over four years and then over one year. Yudkowsky argued that current conditions do not resemble the world Christiano had envisioned.
That interpretation quickly drew a challenge. The pseudonymous commentator Teortaxes argued that the present world could still fit the period before the first four-year doubling, and asked Christiano to clarify whether he disagreed. The attached discussion does not include a response from Christiano or his original writing, so both accounts remain attributed interpretations of his model.
Jan Kulveit then reframed the disagreement. In his view, “slow” and “fast” were always confusing labels because the more important distinction was continuity versus discontinuity. He argued that AI development so far has been mostly continuous, with stronger systems emerging from earlier, weaker ones instead of appearing in a single abrupt jump.
The exchange leaves no clean winner and no independently verified one-year rule. What it does reveal is a clash of yardsticks: elapsed time, economic acceleration and the smoothness of technical progress can each produce a different answer to the same question about whether takeoff is already underway.
