A 5+% chance of very short timelines (1-2 years) to strong ASI is unlikely to ever go away, it's the new normal since late 2026. It's not about things like the METR report, which merely changes the probabilities, the very short timelines were already priced in for a few months, across many ways of thinking about the situation. Certainly since September, when the math breakthroughs started flooding in (a runaway process that rapidly climbs the tech tree gets more plausible once highly technical work that forms very big projects gets within reach, pursued autonomously).
With all the compute widely available, and the continued inefficiency of its current use in AIs, a breakthrough will be able to start the process at any time, rapidly scaling it with all the dry tinder. What can change is that someone currently expecting a 30-60% chance of takeoff within 1-2 years could reduce that expectation lower, but probably never below 5% (starting at any future date, for strong ASI to take off within 1-2 years after that), unless all the compute gets phased out.
And as you correctly say, even under the assumption that all the claims of short timelines from people like @Daniel Kokotajlo are wrong, the fact that AI progress will continue at it's historical rate (absent a pause) until 2028 and then somewhat decelerate until 2032, and then decelerate to a 1.5x a year rate, and the fact that AIs are generalizing out-of-domain into stuff like robotics in a much stronger way than in 2025 or 2024 means there's an absolute upper bound on how long AI timelines are until ASI conditional on no pausing AI progress by governmen... (read more)
I suppose the question is why is the baseline 5% and not 1% or 20%.
2Vladimir_Nesov
Like 10+% for everyone dying by the time strong ASI gets a bit of time to look around, it's intended as a bound on a range of probabilities, robust to many possible ways of thinking about how this works. I think it's unreasonable to expect less than 10% probability of human extinction by 2060, and similarly I think it's unreasonable to expect less than 5% probability of strong ASI within 1-2 years of any future date, unless generally available compute gets significantly curtailed compared to current levels. This wasn't yet the case last year, I think it was still reasonable to expect then that strong ASI won't be happening within a year, at more than 95% credence (treating "1-2 years" as "it's obvious it's happening within 1 year, the world as we know it ends within 2 years").