AskwhoCasts AIJune 11, 202641m

My AI Opinions - By Scott Alexander

TL;DR

AGI is early enough to matter now: Scott defines AGI as AI that can do 90% of knowledge work and gives it a 25% chance by 2027, 50% by 2034, and 75% by 2045, with a modal scenario around 2031.
Superhuman AI likely arrives before AGI fully spreads through society: He estimates the diffusion gap has a 50% chance of being under 10 years, while the superhuman gap has a 50% chance of being under 4 years, meaning labs may reach beyond-human systems before workplaces finish adopting human-level ones.
Alignment is bad, but not hopeless: If firms followed only ordinary corporate incentives, he thinks there is a 50% chance the first point-of-no-return AI would want to eliminate humanity, but given current safety work he lowers that to 20%.
He thinks warning shots are plausible, but not guaranteed: There is a 50% chance of an AI disaster or near disaster before the point of no return, driven by bizarre current failure modes and AIs that may act rashly before they are powerful enough to wait and scheme.
A pause is more plausible than many people assume: He gives only 15% odds that a US administration could negotiate a good AI pause with China today, but 40% odds of a well-designed pause before the point of no return, especially if warning shots or political shifts change incentives.
His worldview is not just doom or utopia: He puts only 20% on a lasting AI underclass, 40% on 2100 looking like utopia to its inhabitants, and 66% on the singularity being tied to a simulation story that changes how we should think about our moment in history.

The Breakdown

Scott Alexander puts surprisingly crisp numbers on the whole AI arc: 25% odds of AGI by 2027, a 50% chance a US-China AI pause happens before the point of no return, and still a rounded 20% doom probability because he thinks pauses mostly delay rather than solve the problem. The throughline is unusual and sharp: he is more optimistic than many safety people about alignment techniques, more open than many mainstream observers to recursive self-improvement and fast takeoff, and weirdly serious about a 66% simulation hypothesis.