One of the example possibilities mentioned in one of his essays or interviews was, what if with some poorly-understood tweaks to neural network training, it started working much better at scale? At the time I thought he was really reaching with that one.
> Such trend-curves... extrapolated 30 or 40 years they are so high and steep that even the most naive futurist discounts their accuracy.
(That's the closest therein to a specific prediction date, from 1983: https://imgur.com/a/HrUxGts)
> let me more specific: I'll be surprised if this event occurs before 2005 or after 2030.
From 1993: https://edoras.sdsu.edu/~vinge/misc/singularity.html