LESSWRONG
LW

AI TimelinesMachine Learning (ML)Updated Beliefs (examples thereof)AI
Frontpage

43

Alex Irpan: "My AI Timelines Have Sped Up"

by Vaniver
19th Aug 2020
AI Alignment Forum
1 min read
20

43

Ω 17

This is a linkpost for https://www.alexirpan.com/2020/08/18/ai-timelines.html
AI TimelinesMachine Learning (ML)Updated Beliefs (examples thereof)AI
Frontpage

43

Ω 17

Alex Irpan: "My AI Timelines Have Sped Up"
23Daniel Kokotajlo
24Pongo
12Daniel Kokotajlo
25jungofthewon
18Daniel Kokotajlo
13Ben Pace
5Daniel Kokotajlo
4Ben Pace
2Daniel Kokotajlo
1Amandango
2Daniel Kokotajlo
3jungofthewon
2Daniel Kokotajlo
2Daniel Kokotajlo
2ESRogs
2Amandango
23Raemon
4Vaniver
8sairjy
6ESRogs
New Comment
20 comments, sorted by
top scoring
Click to highlight new comments since: Today at 7:32 AM
[-]Daniel Kokotajlo5y230

Seems a bit weird to have 10% probability mass for the next 15 years followed by 40% probability mass over the subsequent 10 years. That seems to indicate a fairly strong view about how the next 20 years will go. IMO the probability of AGI in the next 15 years should be substantially higher than 10%.

Reply
[-]Pongo5y240

Whoa, I hadn’t noticed that. The old predictions put 40% probability on AGI being developed in a 5 year window

Reply
[-]Daniel Kokotajlo5y120

That was going to be my initial comment, then I noticed that the blog post addresses that problem. Not sufficiently IMO.

Reply
[-]jungofthewon5y250

Yea this was a lot more obvious to me when I plotted visually: https://elicit.ought.org/builder/om4oCj7jm

(NB: I work on Elicit and it's still a WIP tool) 

Reply
[-]Daniel Kokotajlo5y180

Ahhhh this is so nice! I suspect a substantial fraction of people would revise their timelines after seeing what they look like visually. I think encouraging people to plot out their timelines is probably a pretty cost-effective intervention.

Reply
[-]Ben Pace5y130

Sounds like we could have a thread for that. The image above looks great, would be interested in seeing more comparisons like it, if they're easy to generate using elicit.

Reply
[-]Daniel Kokotajlo5y50

Yes, let's please create a thread. If you don't I will. Here's mine: (Ignore the name, I can't figure out how to change it)

https://elicit.ought.org/builder/xt516PmHt

Reply
[-]Ben Pace5y40

Can you make the thread (make it a question post?) and share it with me? Then I'll suggest any rewrites and we can publish it?

Reply
[-]Daniel Kokotajlo5y20

I shared it with you and jungofthewon; maybe they could help me include some sort of instructions for how to use Elicit?

Reply
[-]Amandango5y10

I can help with this if you share the post with me!

Reply
[-]Daniel Kokotajlo5y20

Thanks so much!

Reply
[-]jungofthewon5y30

You want to change "Your Distribution" to something like "Daniel's 2020 distribution"? 

Reply
[-]Daniel Kokotajlo5y20

Could you help me by writing some instructions for how to use Elicit, to be put in my question on the topic? (See discussion with Ben above)

Reply
[-]Daniel Kokotajlo5y20

Yes, but I don't know how.

Reply
[-]ESRogs5y20

Is the "Your distribution" one Alex's updated estimates, or is it your (jungofthewon's) distribution?

Reply
[-]Amandango5y20

The blue distribution labeled "Your distribution" in this snapshot is Alex's updated 2020 prediction.

Reply
[-]Raemon5y230

I think the title is particularly confusing because it includes an "I" statement. Maybe change to:

Alex Ipran: "My AI Timelines Have Sped Up" 

Reply
[-]Vaniver5y40

Changed.

Reply
[-]sairjy5y80

If 65% of the AI improvements will come from compute alone, I find quite surprising that the post author assigns only 10% probability of AGI by 2035. By that time, we should have between 20x to 100x compute per $. And we can also easily forecast that AI training budgets will increase 1000x easily over that time, as a shot to AGI justifies the ROI. I think he is putting way too much credit on the computational performance of the human brain.

Reply
[-]ESRogs5y60

For this post, I’m going to take artificial general intelligence (AGI) to mean an AI system that matches or exceeds humans at almost all (95%+) economically valuable work.

I'm not sure this is such a good operationalization. I believe that if you looked at the economically valuable work that humans were doing 200 hundred years ago (mostly farming, as I understand), more than 95% of it is automated today. And we don't spend 95% of GDP on farming today.

So I'm not quite sure what the above means. Does it mean 95% of GDP spent on compute? Or unemployment at 95%? Or 95% of jobs that are done today by people are done then by computers? If that last one, then how do you measure it if jobs have morphed s.t. there's neither a human nor a computer clearly doing a job that today is done by a human?

I think that productivity is going to increase. And humans will continue to do jobs where they have a comparative advantage relative to computers. And what those comparative advantages are will morph over time. (And in the limit, if I'm feeling speculative, I think being a producer and a consumer might merge, as one of the last areas where you'll have a comparative advantage is knowing what your own wants are.)

But given that prices will be set based on supply and demand it's not quite obvious to me how to measure when 95% of economically valuable work is done by computers. Because, for a given task that involves both humans and computers, even if computers are doing "more" of the work, you won't necessarily spend more on the computers than the people, if the supply of compute is plentiful. So, in some hard-to-define sense, computers may be doing most of the work, but just measured in dollars they might not be. And one could argue that that is already the case (since we've automated so much of farming and other things that people used to do).

Alternatively, you could operationalize 95% of economically valuable work being done by computers as the total dollars spent on compute being 20x all wages. That's clear enough I think, but I suspect is not exactly what Alex had in mind. And also, I think it may just be a condition that never holds, even when AI is strongly superhuman, depending on how we end up distributing the spoils of AI, and what kind of economic system we end up with at that point.

Reply
Moderation Log
Curated and popular this week
20Comments

Blog post by Alex Irpan. The basic summary:

In 2015, I made the following forecasts about when AGI could happen.

  • 10% chance by 2045
  • 50% chance by 2050
  • 90% chance by 2070

Now that it’s 2020, I’m updating my forecast to:

  • 10% chance by 2035
  • 50% chance by 2045
  • 90% chance by 2070

The main underlying shifts: more focus on improvements in tools, compute, and unsupervised learning.

Mentioned in
135Forecasting Thread: AI Timelines
53Reflections on AI Timelines Forecasting Thread