Surprising Things AGI Forecasting Experts Agree On:
I hesitate to say this because it’s putting words in other people’s mouths, and thus I may be misrepresenting them. I beg forgiveness if so and hope to be corrected. (I’m thinking especially of Paul Christiano and Ajeya Cotra here, but also maybe Rohin and Buck and Richard and some other people)
1. Slow takeoff means things accelerate and go crazy before we get to human-level AGI. It does not mean that after we get to human-level AGI, we still have some non-negligible period where they are gradually getting smarter and available for humans to study and interact with. In other words, people seem to agree that once we get human-level AGI, there’ll be a FOOM of incredibly fast recursive self-improvement.
2. The people with 30-year timelines (as suggested by the Cotra report) tend to agree with the 10-year timelines people that by 2030ish there will exist human-brain-sized artificial neural nets that are superhuman at pretty much all short-horizon tasks. This will have all sorts of crazy effects on the world. The disagreement is over whether this will lead to world GDP doubling in four years or less, whether this will lead to strategically aware agentic AGI (e.g. Carlsmith’s notion of APS-AI), etc.
I’ve begun to doubt (1) recently, would be interested in seeing the arguments in favor of it. My model is something like “well, I’m human-level, and I sure don’t feel like I could foom if I were an AI.”
I’ve also been bothered recently by a blurring of lines between “when AGI becomes as intelligent as humans” and “when AGI starts being able to recursively self-improve.” It’s not a priori obvious that these should happen at around the same capabilities level, yet I feel like it’s common to equivocate between them.
In any case, my world model says that an AGI should actually be able to recursively self-improve before reaching human-level intelligence. Just as you mentioned, I think the relevant intuition pump is “could I FOOM if I were an AI?” Considering the ability to tinker with my own source code and make lots of copies of myself to experiment on, I feel like the answer is “yes.”
That said, I think this intuition isn’t worth much for the following reasons:
The first AGIs will probably have their capabilities distributed very differently than humans—i.e. they will probably be worse than humans at some tasks and much better at other tasks. What really matters is how good they are the task “do ML research” (or whatever paradigm we’re using to make AI’s at the time). I think there are reasons to expect them to be especially good at ML research (relative to their general level of intelligence), but also reasons to expect them to be or especially bad, and I don’t know which reasons to trust more. Note that modern narrow AIs are already have some trivial ability to “do” ML research (e.g. OpenAI’s copilot).
Part of my above story about FOOMing involves making lots of copies of myself, but will it actually be easy for the first AGI (which might not be a generally intelligent as a human) to get the resources it needs to make lots of copies? This seems like it depends on a lot of stuff which I don’t have strong expectations about, e.g. how abundant are the relevant resources, how large is the AGI, etc.
Even if you think “AGI is human-level” and “AGI is able to recursively self-improve” represent very different capabilities levels, they might happen at very similar times, depending on what else you think about takeoff speeds.
In any case, my world model says that an AGI should actually be able to recursively self-improve before reaching human-level intelligence. Just as you mentioned, I think the relevant intuition pump is “could I FOOM if I were an AI?” Considering the ability to tinker with my own source code and make lots of copies of myself to experiment on, I feel like the answer is “yes.”
Counter-anecdote: compilers have gotten ~2x better in 20 years[1], at substantially worse compile time. This is nowhere near FOOM.
Proebsting’s Law gives an 18-year doubling time. The 2001 reproduction suggested more like 20 years under optimistic assumptions, and a 2022 informal test showed a 10-15% improvement on average in the last 10 years (or a 50-year doubling time...)
1. an human-level AGI would be running on hardware making human constraints in memory or speed mostly go away by ~10 orders of magnitude
2. if you could store 10 orders of magnitude more information and read 10 orders of magnitude faster, and if you were able to copy your own code somewhere else, and the kind of AI research and code generation tools available online were good enough to have created you, wouldn’t you be able to FOOM?
The more you accelerate something, the slower and more limiting all it’s other hidden dependencies become.
So by the time we get to AGI, regular ML research will have rapidly diminishing returns (and cuda low level software or hardware optimization will also have diminishing returns), general hardware improvement will be facing the end of moore’s law, etc etc.
I don’t see why that last sentence follows from the previous sentences. In fact I don’t think it does. What if we get to AGI next year? Then returns won’t have diminished as much & there’ll be lots of overhang to exploit.
Sure - if we got to AGI next year—but for that to actually occur you’d have to exploit most of the remaining optimization slack in both high level ML and low level algorithms. Then beyond that Moore’s law is already mostly ended or nearly so depending on who you ask, and most of the easy obvious hardware arch optimizations are now behind us.
Well I would assume a “human-level AI” is an AI which performs as well as a human when it has the extra memory and running speed? I think I could FOOM eventually under those conditions but it would take a lot of thought. Being able to read the AI research that generated me would be nice but I’d ultimately need to somehow make sense of the inscrutable matrices that contain my utility function.
Is it really true that everyone (who is an expert) agrees that FOOM is inevitable? I was under the impression that a lot of people feel that FOOM might be impossible. I personally think FOOM is far from inevitable, even for superhuman intelligences. Consider that human civilization has a collective intelligence is that is strongly superhuman, and we are expending great effort to e.g. push Moore’s law forward. There’s Eroom’s law, which suggests that the aggregate costs of each new process node doubles in step with Moore’s law. So if FOOM depends on faster hardware, ASI might not be able to push forward much faster than Intel, TSMC, ASML, IBM and NVidia already are. Of course this all depends on AI being hardware constrained, which is far from certain. I just think it’s surprising that FOOM is seen as a certainty.
Re 1: that’s not what slow takeoff means, and experts don’t agree on FOOM after AGI. Slow takeoff applies to AGI specifically, not to pre-AGI AIs. And I’m pretty sure at least Christiano and Hanson don’t expect FOOM, but like you am open to be corrected.
What do you think slow takeoff means? Or, perhaps the better question is, what does it mean to you?
Christiano expects things to be going insanely fast by the time we get to AGI, which I take to imply that things are also going extremely fast (presumably, even faster) immediately after AGI: https://sideways-view.com/2018/02/24/takeoff-speeds/
I don’t know what Hanson thinks on this subject. I know he did a paper on AI automation takeoff at some point decades ago; I forget what it looked like quantitatively.
Slow or fast takeoff, in my understanding, refers to how fast an AGI can/will improve itself to (wildly) superintelligent levels. Discontinuity seems to be a key differentiator here.
In the post you link, Christiano is arguing against discontinuity. He may expect quick RSI after AGI is here, though, so I could be mistaken.
Christiano is indeed arguing against discontinuity, but nevertheless he is arguing for an extremely rapid pace of technnological progress—far faster than today. And in particular, he seems to expect quick RSI not only after AGI is here, but before!
Whoa, what? That very much surprises me, I would have thought weeks or months at most. Did you talk to him? What precisely did he say? (My prediction is that he’d say that by the time we have human-level AGI, things will be moving very fast and we’ll have superintelligence a few weeks later.)
Not sure exactly what the claim is, but happy to give my own view.
I think “AGI” is pretty meaningless as a threshold, and at any rate it’s way too imprecise to be useful for this kind of quantitative forecast (I would intuitively describe GPT-3 as a general AI, and beyond that I’m honestly unclear on what distinction people are pointing at when they say “AGI”).
My intuition is that by the time that you have an AI which is superhuman at every task (e.g. for $10/h of hardware it strictly dominates hiring a remote human for any task) then you are likely weeks rather than months from the singularity.
But mostly this is because I think “strictly dominates” is a very hard standard which we will only meet long after AI systems are driving the large majority of technical progress in computer software, computer hardware, robotics, etc. (Also note that we can fail to meet that standard by computing costs rising based on demand for AI.)
My views on this topic are particularly poorly-developed because I think that the relevant action (both technological transformation and catastrophic risk) mostly happens before this point, so I usually don’t think this far ahead.
Thanks! That’s what I thought you’d say. By “AGI” I did mean something like “for $10/h of hardware it strictly dominates hiring a remote human for any task” though I’d maybe restrict it to strategically relevant tasks like AI R&D, and also people might not actually hire AIs to do stuff because they might be afraid / understand that they haven’t solved alignment yet, but it still counts since the AIs could do the job. Also there may be some funny business around the price of the hardware—I feel like it should still count as AGI if a company is running millions of AIs that each individually are better than a typical tech company remote worker in every way, even if there is an ongoing bidding war and technically the price of GPUs is now so high that it’s costing $1,000/hr on the open market for each AGI. We still get FOOM if the AGIs are doing the research, regardless of what the on-paper price is. (I definitely feel like I might be missing something here, I don’t think in economic terms like this nearly as often as you do so)
But mostly this is because I think “strictly dominates” is a very hard standard which we will only meet long after AI systems are driving the large majority of technical progress in computer software, computer hardware, robotics, etc.
My timelines are too short to agree with this part alas. Well, what do you mean by “long after?” Six months? Three years? Twelve years?
Less relevant now, but I got the “few years” from the post you linked. There Christiano talked about another gap than AGI → ASI, but since overall he seems to expect linear progress, I thought my conclusion was reasonable. In retrospect, I shouldn’t have made that comment.
I disagree with the first one. I think that the spectrum of human-level AGI is actually quite wide, and that for most tasks we’ll get AGIs that are better than most humans significantly before we get AGIs that are better than all humans. But the latter is much more relevant for recursive self-improvement, because it’s bottlenecked by innovation, which is driven primarily by the best human researchers. E.g. I think it’d be pretty difficult to speed up AI progress dramatically using millions of copies of an average human.
Also, by default I think people talk about FOOM in a way that ignores regulations, governance, etc. Whereas in fact I expect these to put significant constraints on the pace of progress after human-level AGI.
If we have millions of copies of the best human researchers, without governance constraints on the pace of progress… Then compute constraints become the biggest thing. It seems plausible that you get a software-only singularity, but it also seems plausible that you need to wait for AI innovation of new chip manufacturing to actually cash out in the real world.
I broadly agree with the second one, though I don’t know how many people there are left with 30-year timelines. But 20 years to superintelligence doesn’t seem unreasonable to me (though it’s above my median). In general I’ve updated lately that Kurzweil was more right than I used to think about there being a significant gap between AGI and ASI. Part of this is because I expect the problem of multi-agent credit assignment over long time horizons to be difficult.
Surprising Things AGI Forecasting Experts Agree On:
I hesitate to say this because it’s putting words in other people’s mouths, and thus I may be misrepresenting them. I beg forgiveness if so and hope to be corrected. (I’m thinking especially of Paul Christiano and Ajeya Cotra here, but also maybe Rohin and Buck and Richard and some other people)
1. Slow takeoff means things accelerate and go crazy before we get to human-level AGI. It does not mean that after we get to human-level AGI, we still have some non-negligible period where they are gradually getting smarter and available for humans to study and interact with. In other words, people seem to agree that once we get human-level AGI, there’ll be a FOOM of incredibly fast recursive self-improvement.
2. The people with 30-year timelines (as suggested by the Cotra report) tend to agree with the 10-year timelines people that by 2030ish there will exist human-brain-sized artificial neural nets that are superhuman at pretty much all short-horizon tasks. This will have all sorts of crazy effects on the world. The disagreement is over whether this will lead to world GDP doubling in four years or less, whether this will lead to strategically aware agentic AGI (e.g. Carlsmith’s notion of APS-AI), etc.
I’m doubtful whether the notion of human level AGI makes much sense.
In it’s progression of getting more and more capability there’s likely no point where it’s comparable to a human.
Why?
I’ve begun to doubt (1) recently, would be interested in seeing the arguments in favor of it. My model is something like “well, I’m human-level, and I sure don’t feel like I could foom if I were an AI.”
I’ve also been bothered recently by a blurring of lines between “when AGI becomes as intelligent as humans” and “when AGI starts being able to recursively self-improve.” It’s not a priori obvious that these should happen at around the same capabilities level, yet I feel like it’s common to equivocate between them.
In any case, my world model says that an AGI should actually be able to recursively self-improve before reaching human-level intelligence. Just as you mentioned, I think the relevant intuition pump is “could I FOOM if I were an AI?” Considering the ability to tinker with my own source code and make lots of copies of myself to experiment on, I feel like the answer is “yes.”
That said, I think this intuition isn’t worth much for the following reasons:
The first AGIs will probably have their capabilities distributed very differently than humans—i.e. they will probably be worse than humans at some tasks and much better at other tasks. What really matters is how good they are the task “do ML research” (or whatever paradigm we’re using to make AI’s at the time). I think there are reasons to expect them to be especially good at ML research (relative to their general level of intelligence), but also reasons to expect them to be or especially bad, and I don’t know which reasons to trust more. Note that modern narrow AIs are already have some trivial ability to “do” ML research (e.g. OpenAI’s copilot).
Part of my above story about FOOMing involves making lots of copies of myself, but will it actually be easy for the first AGI (which might not be a generally intelligent as a human) to get the resources it needs to make lots of copies? This seems like it depends on a lot of stuff which I don’t have strong expectations about, e.g. how abundant are the relevant resources, how large is the AGI, etc.
Even if you think “AGI is human-level” and “AGI is able to recursively self-improve” represent very different capabilities levels, they might happen at very similar times, depending on what else you think about takeoff speeds.
Counter-anecdote: compilers have gotten ~2x better in 20 years[1], at substantially worse compile time. This is nowhere near FOOM.
Proebsting’s Law gives an 18-year doubling time. The 2001 reproduction suggested more like 20 years under optimistic assumptions, and a 2022 informal test showed a 10-15% improvement on average in the last 10 years (or a 50-year doubling time...)
The straightforward argument goes like this:
1. an human-level AGI would be running on hardware making human constraints in memory or speed mostly go away by ~10 orders of magnitude
2. if you could store 10 orders of magnitude more information and read 10 orders of magnitude faster, and if you were able to copy your own code somewhere else, and the kind of AI research and code generation tools available online were good enough to have created you, wouldn’t you be able to FOOM?
No because of the generalized version of Amdhal’s law, which I explored in “Fast Minds and Slow Computers”.
The more you accelerate something, the slower and more limiting all it’s other hidden dependencies become.
So by the time we get to AGI, regular ML research will have rapidly diminishing returns (and cuda low level software or hardware optimization will also have diminishing returns), general hardware improvement will be facing the end of moore’s law, etc etc.
I don’t see why that last sentence follows from the previous sentences. In fact I don’t think it does. What if we get to AGI next year? Then returns won’t have diminished as much & there’ll be lots of overhang to exploit.
Sure - if we got to AGI next year—but for that to actually occur you’d have to exploit most of the remaining optimization slack in both high level ML and low level algorithms. Then beyond that Moore’s law is already mostly ended or nearly so depending on who you ask, and most of the easy obvious hardware arch optimizations are now behind us.
Well I would assume a “human-level AI” is an AI which performs as well as a human when it has the extra memory and running speed? I think I could FOOM eventually under those conditions but it would take a lot of thought. Being able to read the AI research that generated me would be nice but I’d ultimately need to somehow make sense of the inscrutable matrices that contain my utility function.
Is it really true that everyone (who is an expert) agrees that FOOM is inevitable? I was under the impression that a lot of people feel that FOOM might be impossible. I personally think FOOM is far from inevitable, even for superhuman intelligences. Consider that human civilization has a collective intelligence is that is strongly superhuman, and we are expending great effort to e.g. push Moore’s law forward. There’s Eroom’s law, which suggests that the aggregate costs of each new process node doubles in step with Moore’s law. So if FOOM depends on faster hardware, ASI might not be able to push forward much faster than Intel, TSMC, ASML, IBM and NVidia already are. Of course this all depends on AI being hardware constrained, which is far from certain. I just think it’s surprising that FOOM is seen as a certainty.
Depends on who you count as an expert. That’s a judgment call since there isn’t an Official Board of AGI Timelines Experts.
Re 1: that’s not what slow takeoff means, and experts don’t agree on FOOM after AGI. Slow takeoff applies to AGI specifically, not to pre-AGI AIs. And I’m pretty sure at least Christiano and Hanson don’t expect FOOM, but like you am open to be corrected.
What do you think slow takeoff means? Or, perhaps the better question is, what does it mean to you?
Christiano expects things to be going insanely fast by the time we get to AGI, which I take to imply that things are also going extremely fast (presumably, even faster) immediately after AGI: https://sideways-view.com/2018/02/24/takeoff-speeds/
I don’t know what Hanson thinks on this subject. I know he did a paper on AI automation takeoff at some point decades ago; I forget what it looked like quantitatively.
Thanks for responding!
Slow or fast takeoff, in my understanding, refers to how fast an AGI can/will improve itself to (wildly) superintelligent levels. Discontinuity seems to be a key differentiator here.
In the post you link, Christiano is arguing against discontinuity. He may expect quick RSI after AGI is here, though, so I could be mistaken.
Likewise!
Christiano is indeed arguing against discontinuity, but nevertheless he is arguing for an extremely rapid pace of technnological progress—far faster than today. And in particular, he seems to expect quick RSI not only after AGI is here, but before!
I’d question the “quick” of “quick RSI”, but yes, he expects AI to make better AI before AGI.
I’m pretty sure he means really really quick, by any normal standard of quick. But we can take it up with him sometime. :)
He’s talking about a gap of years :) Which is probably faster than ideal, but not FOOMy, as I understand FOOM to mean days or hours.
Whoa, what? That very much surprises me, I would have thought weeks or months at most. Did you talk to him? What precisely did he say? (My prediction is that he’d say that by the time we have human-level AGI, things will be moving very fast and we’ll have superintelligence a few weeks later.)
Not sure exactly what the claim is, but happy to give my own view.
I think “AGI” is pretty meaningless as a threshold, and at any rate it’s way too imprecise to be useful for this kind of quantitative forecast (I would intuitively describe GPT-3 as a general AI, and beyond that I’m honestly unclear on what distinction people are pointing at when they say “AGI”).
My intuition is that by the time that you have an AI which is superhuman at every task (e.g. for $10/h of hardware it strictly dominates hiring a remote human for any task) then you are likely weeks rather than months from the singularity.
But mostly this is because I think “strictly dominates” is a very hard standard which we will only meet long after AI systems are driving the large majority of technical progress in computer software, computer hardware, robotics, etc. (Also note that we can fail to meet that standard by computing costs rising based on demand for AI.)
My views on this topic are particularly poorly-developed because I think that the relevant action (both technological transformation and catastrophic risk) mostly happens before this point, so I usually don’t think this far ahead.
Thanks! That’s what I thought you’d say. By “AGI” I did mean something like “for $10/h of hardware it strictly dominates hiring a remote human for any task” though I’d maybe restrict it to strategically relevant tasks like AI R&D, and also people might not actually hire AIs to do stuff because they might be afraid / understand that they haven’t solved alignment yet, but it still counts since the AIs could do the job. Also there may be some funny business around the price of the hardware—I feel like it should still count as AGI if a company is running millions of AIs that each individually are better than a typical tech company remote worker in every way, even if there is an ongoing bidding war and technically the price of GPUs is now so high that it’s costing $1,000/hr on the open market for each AGI. We still get FOOM if the AGIs are doing the research, regardless of what the on-paper price is. (I definitely feel like I might be missing something here, I don’t think in economic terms like this nearly as often as you do so)
My timelines are too short to agree with this part alas. Well, what do you mean by “long after?” Six months? Three years? Twelve years?
Thanks for offering your view Paul, and I apologize if I misrepresented your view.
Less relevant now, but I got the “few years” from the post you linked. There Christiano talked about another gap than AGI → ASI, but since overall he seems to expect linear progress, I thought my conclusion was reasonable. In retrospect, I shouldn’t have made that comment.
But yes, Christiano is the authority here;)
I disagree with the first one. I think that the spectrum of human-level AGI is actually quite wide, and that for most tasks we’ll get AGIs that are better than most humans significantly before we get AGIs that are better than all humans. But the latter is much more relevant for recursive self-improvement, because it’s bottlenecked by innovation, which is driven primarily by the best human researchers. E.g. I think it’d be pretty difficult to speed up AI progress dramatically using millions of copies of an average human.
Also, by default I think people talk about FOOM in a way that ignores regulations, governance, etc. Whereas in fact I expect these to put significant constraints on the pace of progress after human-level AGI.
If we have millions of copies of the best human researchers, without governance constraints on the pace of progress… Then compute constraints become the biggest thing. It seems plausible that you get a software-only singularity, but it also seems plausible that you need to wait for AI innovation of new chip manufacturing to actually cash out in the real world.
I broadly agree with the second one, though I don’t know how many people there are left with 30-year timelines. But 20 years to superintelligence doesn’t seem unreasonable to me (though it’s above my median). In general I’ve updated lately that Kurzweil was more right than I used to think about there being a significant gap between AGI and ASI. Part of this is because I expect the problem of multi-agent credit assignment over long time horizons to be difficult.