(1). “Our forecasters’ aggregate estimate of when we will have “an AI system that is much better than the best human at every cognitive task” is 2028 to 2050, and with a median of medians of 2031.” Why was 2039 crossed out? Forecasters from of medians ranging from “2028 to 2050”. Would it be possible to give the different median predictions and list who made them?
(2). “The shortest 10th percentile includes 2027, and the longest 90th percentile includes “never”. When in 2027 is the 10th percentile?
(3). Maybe this can be question for next week but do forecasters think that the automated coder and TED-AI or AGI will be reached?
Hi Max, thanks for your comment and sorry I missed it. (1) 2039 was crossed out because we made an error/typo. The different median predictions were 2030, 2030, 2031, 2038 and 2050. (2) In October 2027. (3) We’ll consider looking at these in future weeks!
"Conditional on no government or self-imposed slowdown and precluding catastrophes, in which year will artificial superintelligence (ASI), an AI system that is much better than the best human at every cognitive task, be achieved? Our aggregate estimate lies at 2028 to 2050, with a median of medians of 2039."
Would it be better to have a more concrete operationalisation of ASI, like the one used by the AI Futures' team (gap between ASI and the best humans more than 2 times greater than the gap between the best humans and median professionals, at virtually all cognitive tasks)?
Possibly! One could also compare artificial systems against companies, countries and other teams of humans. But for a periodical like this one I don't think it makes that much sense to sweat the operationalization much.
"I don't think it makes that much sense to sweat the operationalization much"
Because you think reasonable operationalisations would lead to aggregate estimate covering a period similar to 2028 to 2050? I could see my median date for ASI being anything from a few years to a few centuries depending on the operationalisation.
Hi, a heads up that there is a mistake in your post in your calculation, assuming that $150 will be the price of oil if Trump escalates and $80, if he does not the current price would actually imply and escalation risk of 43%, not 13%. Not sure what went wrong here, unless there is a typing error.
Great post. I did have some questions.
(1). “Our forecasters’ aggregate estimate of when we will have “an AI system that is much better than the best human at every cognitive task” is 2028 to 2050, and with a median of medians of 2031.” Why was 2039 crossed out? Forecasters from of medians ranging from “2028 to 2050”. Would it be possible to give the different median predictions and list who made them?
(2). “The shortest 10th percentile includes 2027, and the longest 90th percentile includes “never”. When in 2027 is the 10th percentile?
(3). Maybe this can be question for next week but do forecasters think that the automated coder and TED-AI or AGI will be reached?
Hi Max, thanks for your comment and sorry I missed it. (1) 2039 was crossed out because we made an error/typo. The different median predictions were 2030, 2030, 2031, 2038 and 2050. (2) In October 2027. (3) We’ll consider looking at these in future weeks!
Thanks for the response. I was under the impression that at least one of the forecasters had a median of 2028. Is that not correct?
It seems the CIs are reversed, the one going till Sept should be for August.
```
Will the US place more than 1,000 troops on the ground in Iran before July 2026? We think there’s a 39% (20% to 55%) chance.
If so, when will the US end hostilities with Iran? Our aggregate estimate is August 12 2026 (80% CI: May 24 2026 - June 3 2030).
If not, when will the US end hostilities with Iran? Our aggregate is May 30 2026 (80% CI: April 20 2026 - Sept 13 2027).
```
Hey, thanks for your comment. It looks okay to me? The June 3 in the August one is in 2030.
Thanks for the update.
"Conditional on no government or self-imposed slowdown and precluding catastrophes, in which year will artificial superintelligence (ASI), an AI system that is much better than the best human at every cognitive task, be achieved? Our aggregate estimate lies at 2028 to 2050, with a median of medians of 2039."
Would it be better to have a more concrete operationalisation of ASI, like the one used by the AI Futures' team (gap between ASI and the best humans more than 2 times greater than the gap between the best humans and median professionals, at virtually all cognitive tasks)?
Possibly! One could also compare artificial systems against companies, countries and other teams of humans. But for a periodical like this one I don't think it makes that much sense to sweat the operationalization much.
"I don't think it makes that much sense to sweat the operationalization much"
Because you think reasonable operationalisations would lead to aggregate estimate covering a period similar to 2028 to 2050? I could see my median date for ASI being anything from a few years to a few centuries depending on the operationalisation.
Hi, a heads up that there is a mistake in your post in your calculation, assuming that $150 will be the price of oil if Trump escalates and $80, if he does not the current price would actually imply and escalation risk of 43%, not 13%. Not sure what went wrong here, unless there is a typing error.
Whoops, thanks, fixed!