"Metaculus is a community dedicated to generating accurate predictions about future real-world events by aggregating the collective wisdom, insight, and intelligence of its participants." In other words, it's a sort of betting website for smart people where they are asked to make predictions about future events. Any of you who read the "Astral Codex Ten" blog should have heard about it by now.
https://www.metaculus.com/questions/
I was browsing the site today, and thought I'd post some of the Metaculus community's predictions about AI.
Date ML model produces realistic 2-min video
lower 25% ... Jun 2022
median ... Dec 2022
upper 75% ... Sep 2023
https://www.metaculus.com/questions/305 ... or-better/
When will GPT-4 be announced?
Community Prediction: Aug 29, 2022
Metaculus Prediction: Jul 9, 2022
https://www.metaculus.com/questions/740 ... announced/
Billions of params of GPT-4 if released
Community Prediction 3.47k
Metaculus Prediction 3.48k
https://www.metaculus.com/questions/485 ... arameters/
When will AI achieve superhuman few-shot performance on SuperGLUE?
lower 25% ... Jul 2022
median ... Jul 2023
upper 75% ... Apr 2025
https://www.metaculus.com/questions/493 ... superglue/
Computer to pass Turing test by 2029
39% Median prediction
43% Mean prediction
https://www.metaculus.com/questions/364 ... ring-test/
How many years after the silver Turing Test is passed will an AGI system be developed?
lower 25% … 2.1
median … 7.0
upper 75% … 22
https://www.metaculus.com/questions/426 ... developed/
Date first AGI is publicly known.
lower 25% … 2028
median … 2038
upper 75% … 2063
< Mar 27, 2020 … 1%
> Jan 3, 2200 … 6%
https://www.metaculus.com/questions/347 ... -known-of/
Time Between AGI and Superintelligence (in months)
lower 25% … 0.80
median … 10
upper 75% … 75
< 0.033 … 8%
> 360 … 10%
https://www.metaculus.com/questions/412 ... elligence/
Metaculus' predictions about AI
Re: Metaculus' predictions about AI
Definitely examples there of only understanding linear and not exponential growth. I would not comment on anything achievable by an AI beyond the boundary of the next two to three years at maximum.
-
funkervogt
- Posts: 1178
- Joined: Mon May 17, 2021 3:03 pm
Re: Metaculus' predictions about AI
I don't know about that. Some pretty smart people contribute to Metaculus.
Re: Metaculus' predictions about AI
It is not a matter of intelligence as much as it is a matter of perspective, the sheer concept of exponential growth and thinking is not something that is easily understood but imagine trying to explain what GPT-4 or the equivalent may look like to someone in the year 2015.
It would be three entire years before Open AI even debuted the initial GPT model and the difference in what would be achievable with that software compared to what is conceptually possible this year is perhaps even greater than the difference between the chatbots of the early 2010's such as Cleverbot and GPT.
It is also a case of regardless of intelligence no-one really understands where the boundaries for intelligence lay, even the most intelligent human is only at best a few times smarter than the least intelligent human. AI models can and are able to think about things in ways we as limited biological creatures simply cannot, it is tempting to anthropomorphise and to assume that we as the most intelligent known species in the universe will mimic artificial intelligence or perhaps that it will mimic us but that simply is not the case.
I personally have always been amazed at the progression made in the fields of synthetic media ever since my attention was drawn to it initially I could not have conceived that as early as 2020 people would be able to make the admittedly false argument that GPT-3 is AGI. The fact that it can fool some people into believing that, and that the pace of the improvements made in modelling AI continues to increase, would make me doubt in the accuracy of anything beyond the near to immediate future regarding this technology and it's applications.
It would be three entire years before Open AI even debuted the initial GPT model and the difference in what would be achievable with that software compared to what is conceptually possible this year is perhaps even greater than the difference between the chatbots of the early 2010's such as Cleverbot and GPT.
It is also a case of regardless of intelligence no-one really understands where the boundaries for intelligence lay, even the most intelligent human is only at best a few times smarter than the least intelligent human. AI models can and are able to think about things in ways we as limited biological creatures simply cannot, it is tempting to anthropomorphise and to assume that we as the most intelligent known species in the universe will mimic artificial intelligence or perhaps that it will mimic us but that simply is not the case.
I personally have always been amazed at the progression made in the fields of synthetic media ever since my attention was drawn to it initially I could not have conceived that as early as 2020 people would be able to make the admittedly false argument that GPT-3 is AGI. The fact that it can fool some people into believing that, and that the pace of the improvements made in modelling AI continues to increase, would make me doubt in the accuracy of anything beyond the near to immediate future regarding this technology and it's applications.
