A lot depends on how much uncertainty you have over that timeline, and how you expect the uncertainty to change as time passes. If it’s 10-200 years, with a peak at 20, you probably shouldn’t do anything particularly weird in response, except keeping a close eye on 5-10 year strong indicators, at which point you flip your savings/investment mechanisms from long-term optimization to short/medium-term utility maximization.
Do you know of any recent arguments from individuals with timelines like you mention, which are long/broad and emphasize uncertainty? At least the pool of people making a convincing case for the >100 year timelines has been drying up, from what I can see. Even Paul Christiano expects “40% on singularity by 2040”.
I don’t have great indicators. We’re still pretty far from human-level performance in any major complex industrial task (from cooking to warehouse work, automation exists to exist humans, but humans are executing the difficult/interesting part). Even driving is forever just a few years away.
I also don’t know of anyone arguing for >40 year timelines—most of the skeptics aren’t arguing at all, they’re just getting on with their lives. If you take everyone who’s NOT obsessing over AI safety as arguing for very-long timelines, that’s a massive weight of opinion to overcome.
A lot depends on how much uncertainty you have over that timeline, and how you expect the uncertainty to change as time passes. If it’s 10-200 years, with a peak at 20, you probably shouldn’t do anything particularly weird in response, except keeping a close eye on 5-10 year strong indicators, at which point you flip your savings/investment mechanisms from long-term optimization to short/medium-term utility maximization.
Do you know of any recent arguments from individuals with timelines like you mention, which are long/broad and emphasize uncertainty? At least the pool of people making a convincing case for the >100 year timelines has been drying up, from what I can see. Even Paul Christiano expects “40% on singularity by 2040”.
What in your opinion would a 5-10 year strong indicator look like?
I don’t have great indicators. We’re still pretty far from human-level performance in any major complex industrial task (from cooking to warehouse work, automation exists to exist humans, but humans are executing the difficult/interesting part). Even driving is forever just a few years away.
I also don’t know of anyone arguing for >40 year timelines—most of the skeptics aren’t arguing at all, they’re just getting on with their lives. If you take everyone who’s NOT obsessing over AI safety as arguing for very-long timelines, that’s a massive weight of opinion to overcome.
It might be 6 years.