Forecasting transformative AI: what’s the burden of proof?
Audio version available at Cold Takes (or search Stitcher, Spotify, Google Podcasts, etc. for “Cold Takes Audio”)
This is one of 4 posts summarizing hundreds of pages of technical reports focused almost entirely on forecasting one number: the year by which transformative AI will be developed.1
By “transformative AI,” I mean “AI powerful enough to bring us into a new, qualitatively different future.” I specifically focus on what I’m calling PASTA: AI systems that can essentially automate all of the human activities needed to speed up scientific and technological advancement.
The sooner PASTA might be developed, the sooner the world could change radically, and the more important it seems to be thinking today about how to make that change go well vs. poorly.
In future pieces, I’m going to lay out two methods of making a “best guess” at when we can expect transformative AI to be developed. But first, in this piece, I’m going to address the question: how good do these forecasting methods need to be in order for us to take them seriously? In other words, what is the “burden of proof” for forecasting transformative AI timelines?
When someone forecasts transformative AI in the 21st century—especially when they are clear about the full consequences it would bring—a common intuitive response is something like: “It’s really out-there and wild to claim that transformative AI is coming this century. So your arguments had better be really good.”
I think this is a very reasonable first reaction to forecasts about transformative AI (and it matches my own initial reaction). But I’ve tried to examine what’s driving the reaction and how it might be justified, and having done so, I ultimately don’t agree with the reaction.
I think there are a number of reasons to think that transformative AI—or something equally momentous—is somewhat likely this century, even before we examine details of AI research, AI progress, etc.
I also think that on the kinds of multi-decade timelines I’m talking about, we should generally be quite open to very wacky, disruptive, even revolutionary changes. With this backdrop, I think that specific well-researched estimates of when transformative AI is coming can be credible, even if they involve a lot of guesswork and aren’t rock-solid.
This post tries to explain where I’m coming from.
Below, I will (a) get a bit more specific about which transformative AI forecasts I’m defending; then (b) discuss how to formalize the “That’s too wild” reaction to such forecasts; then (c) go through each of the rows below, each of which is a different way of formalizing it.
|“Burden of proof” angle||Key in-depth pieces (abbreviated titles)||My takeaways|
|It’s unlikely that any given century would be the “most important” one. (More)||Hinge; Response to Hinge||We have many reasons to think this century is a “special” one before looking at the details of AI. Many have been covered in previous pieces; another is covered in the next row.|
|What would you forecast about transformative AI timelines, based only on basic information about (a) how many years people have been trying to build transformative AI; (b) how much they’ve “invested” in it (in terms of the number of AI researchers and the amount of computation used by them); (c) whether they’ve done it yet (so far, they haven’t)? (More)||Semi-informative Priors||Central estimates: 8% by 2036; 13% by 2060; 20% by 2100.2 In my view, this report highlights that the history of AI is short, investment in AI is increasing rapidly, and so we shouldn’t be too surprised if transformative AI is developed soon.|
|Based on analysis of economic models and economic history, how likely is ‘explosive growth’ - defined as >30% annual growth in the world economy—by 2100? Is this far enough outside of what’s “normal” that we should doubt the conclusion? (More)||Explosive Growth, Human Trajectory||Human Trajectory projects the past forward, implying explosive growth by 2043-2065.
|“How have people predicted AI … in the past, and should we adjust our own views today to correct for patterns we can observe in earlier predictions? … We’ve encountered the view that AI has been prone to repeated over-hype in the past, and that we should therefore expect that today’s projections are likely to be over-optimistic.” (More)||Past AI Forecasts||“The peak of AI hype seems to have been from 1956-1973. Still, the hype implied by some of the best-known AI predictions from this period is commonly exaggerated.”|
Some rough probabilities
Here are some things I believe about transformative AI, which I’ll be trying to defend:
I think there’s more than a 10% chance we’ll see something PASTA-like enough to qualify as “transformative AI” within 15 years (by 2036); a ~50% chance we’ll see it within 40 years (by 2060); and a ~2/3 chance we’ll see it this century (by 2100).
Conditional on the above, I think there’s at least a 50% chance that we’ll soon afterward see a world run by digital people or misaligned AI or something else that would make it fair to say we have “transitioned to a state in which humans as we know them are no longer the main force in world events.” (This corresponds to point #1 in my “most important century” definition in the roadmap.)
And conditional on the above, I think there’s at least a 50% chance that whatever is the main force in world events will be able to create a stable galaxy-wide civilization for billions of years to come. (This corresponds to point #2 in my “most important century” definition in the roadmap.)
I’ve also put a bit more detail on what I mean by the “most important century” here.
Formalizing the “That’s too wild” reaction
Often, someone states a view that I can’t immediately find a concrete flaw in, but that I instinctively think is “just too wild” to be likely. For example, “My startup is going to be the next Google” or “College is going to be obsolete in 10 years” or “As President, I would bring both sides together rather than just being partisan.”
I hypothesize that the “This is too wild” reaction to statements like these can usually be formalized along the following lines: “Whatever your arguments for X being likely, there is some salient way of looking at things (often oversimplified, but relevant) that makes X look very unlikely.”
For the examples I just gave:
“My startup is going to be the next Google.” There are large numbers of startups (millions?), and the vast majority of them don’t end up anything like Google. (Even when their founders think they will!)
“College is going to be obsolete in 10 years.” College has been very non-obsolete for hundreds of years.
“As President, I would bring both sides together rather than just being partisan.” This is a common thing for would-be US Presidents to say, but partisanship seems to have been getting worse for at least a couple of decades nonetheless.
Each of these cases establishes a sort of starting point (or “prior” probability) and “burden of proof,” and we can then consider further evidence that might overcome the burden. That is, we can ask things like: what makes this startup different from the many other startups that think they can be the next Google? What makes the coming decade different from all the previous decades that saw college stay important? What’s different about this Presidential candidate from the last few?
There are a number of different ways to think about the burden of proof for my claims above: a number of ways of getting a prior (“starting point”) probability, that can then be updated by further evidence.
Many of these capture different aspects of the “That’s too wild” intuition, by generating prior probabilities that (at least initially) make the probabilities I’ve given look too high.
Below, I will go through a number of these “prior probabilities,” and examine what they mean for the “burden of proof” on forecasting methods I’ll be discussing in later posts.
Different angles on the burden of proof
“Most important century” skepticism
One angle on the burden of proof is along these lines:
Holden claims a 15-30% chance that this is the “most important century” in one sense or another.3
But there are a lot of centuries, and by definition most of them can’t be the most important. Specifically:
Humans have been around for 50,000 to ~5 million years, depending on how you define “humans.”4 That’s 500 to 50,000 centuries.
If we assume that our future is about as long as our past, then there are 1,000 to 100,000 total centuries.
So the prior (starting-point) probability for the “most important century” is 1⁄100,000 to 1⁄1,000.
It’s actually worse than that: Holden has talked about civilization lasting for billions of years. That’s tens of millions of centuries, so the prior probability of “most important century” is less than 1⁄10,000,000.
This argument feels like it is pretty close to capturing my biggest source of past hesitation about the “most important century” hypothesis. However, I think there are plenty of markers that this is not an average century, even before we consider specific arguments about AI.
One key point is emphasized in my earlier post, All possible views about humanity’s future are wild. If you think humans (or our descendants) have billions of years ahead of us, you should think that we are among the very earliest humans, which makes it much more plausible that our time is among the most important. (This point is also emphasized in Thoughts on whether we’re living at the most influential time in history as well as the comments on an earlier version of “Are We Living at the Hinge of History?”.)
Additionally, while humanity has existed for a few million years, for most of that time we had extremely low populations and very little in the way of compounding technological progress. Human civilization started about 10,000 years ago, and since then we’ve already gotten to the point of building digital programmable computers and exploring our solar system.
With these points in mind, it seems reasonable to think we will eventually launch a stable galaxy-wide civilization, sometime in the next 100,000 years (1000 centuries). Or to think there’s a 10% chance we will do so sometime in the next 10,000 years (100 centuries). Either way, this implies that a given century has a ~1/1,000 chance of being the most important century for the launch of that civilization—much higher than the figures given earlier in this section. It’s still ~100x off from the numbers I gave above, so there’s still a burden of proof.
There are further reasons to think this particular century is unusual. For example, see This Can’t Go On:
The total size of the world economy has grown more in the last 2 centuries than in all of the rest of history combined.
The current economic growth rate can’t be sustained for more than another 80 centuries or so. (And as discussed below, if its past accelerating trend resumed, it would imply explosive growth and hitting the limits of what’s possible this century.)
It’s plausible that science has advanced more in the last 5 centuries than in the rest of history combined.
A final point that makes our time special: we’re talking about when to expect transformative AI, and we’re living very close in time to the very beginnings of efforts on AI. In well under 1 century, we’ve gone from the first programmable electronic general-purpose computer to AI models that can compete with humans at speech recognition,6 image classification and much more.
More on the implications of this in the next section.
Report on Semi-informative Priors (abbreviated in this piece as “Semi-informative Priors”) is an extensive attempt to forecast transformative AI timelines while using as little information about the specifics of AI as possible. So it is one way of providing an angle on the “burden of proof”—that is, establishing a prior (starting-point) set of probabilities for when transformative AI will be developed, before we look at the detailed evidence.
The central information it uses is about how much effort has gone into developing AI so far. The basic idea:
If we had been trying and failing at developing transformative AI for thousands of years, the odds of succeeding in the coming decades would be low.
But if we’ve only been trying to develop AI systems for a few decades so far, this means the coming decades could contain a large fraction of all the effort that has ever been put in. The odds of developing it in that time are not all that low.
One way of thinking about this is that before we look at the details of AI progress, we should be somewhat agnostic about whether developing transformative AI is relatively “easy” (can be done in a few decades) or “hard” (takes thousands of years). Since things are still early, the possibility that it’s “easy” is still open.
A bit more on the report’s approach and conclusions:
Angle of analysis. The report poses the following question (paraphrased): “Suppose you had gone into isolation on the day that people started investing in building AI systems. And now suppose that you’ve received annual updates on (a) how many years people have been trying to build transformative AI; (b) how much they’ve ‘invested’ in it (in terms of time and money); (c) whether they’ve succeeded yet (so far, they haven’t). What can you forecast about transformative AI timelines, having only that information, as of 2021?”
Its methods take inspiration from the Sunrise Problem: “Suppose you knew nothing about the universe except whether, on each day, the sun has risen. Suppose there have been N days so far, and the sun has risen on all of them. What is the probability that the sun will rise tomorrow?” You don’t need to know anything about astronomy in order to get a decent answer to this question—there are simple mathematical methods for estimating the probability that X will happen tomorrow, based on the fact that X has happened each day in the past. “Semi-informative Priors” extends these mathematical methods in order to adapt them to transformative AI timelines. (In this case, “X” is “Failing to develop transformative AI, as we have in the past.”)
Conclusions. I’m not going to go heavily into the details of how the analysis works (see the blog post summarizing the report for more detail), but the report’s conclusions include the following:
It puts the probability of artificial general intelligence (AGI, which would include PASTA) by 2036 between 1-18%, with a best guess of 8%.
It puts the probability of AGI by 2060 at around 3-25% (best guess ~13%), and the probability of AGI by 2100 at around 5-35%, best guess 20%.
These are lower than the probabilities I give above, but not much lower. This implies that there isn’t an enormous burden of proof when bringing in additional evidence about the specifics of AI investment and progress.
Notes on regime start date. Something interesting here is that the report is less sensitive than one might think about how we define the “start date” for trying to develop AGI. (See this section of the full report.) That is:
By default, “Semi-informative Priors” models the situation as if humanity started “trying” to build AGI in 1956.7 This implies that efforts are only ~65 years old, so the coming decades will represent a large fraction of the effort.
But the report also looks at other measures of “effort to build AGI”—notably, researcher-time and “compute” (processing power). Even if you want to say that we’ve been implicitly trying to build AGI since the beginning of human civilization ~10,000 years ago, the coming decades will contain a large chunk of the research effort and computation invested in trying to do so.
Bottom line on this section.
Occasionally I’ll hear someone say something along the lines of “We’ve been trying to build transformative AI for decades, and we haven’t yet—why do you think the future will be different?” At a minimum, this report reinforces what I see as the common-sense position that a few decades of “no transformative AI yet, despite efforts to build it” doesn’t do much to argue against the possibility that transformative AI will arrive in the next decade or few.
In fact, in the scheme of things, we live extraordinarily close in time to the beginnings of attempts at AI development—another way in which our century is “special,” such that we shouldn’t be too surprised if it turns out to be the key one for AI development.
Another angle on the burden of proof is along these lines:
If PASTA were to be developed anytime soon, and if it were to have the consequences outlined in this series of posts, this would be a massive change in the world—and the world simply doesn’t change that fast.
To quantify this: the world economy has grown at a few percent per year for the last 200+ years, and PASTA would imply a much faster growth rate, possibly 100% per year or above.
If we were moving toward a world of explosive economic growth, economic growth should be speeding up today. It’s not—it’s stagnating, at least in the most developed economies. If AI were really going to revolutionize everything, the least it could be doing now is creating enough value—enough new products, transactions and companies—to make overall US economic growth speed up.
AI may lead to cool new technologies, but there’s no sign of anything nearly as momentous as PASTA would be. Going from where we are to where PASTA would take us is the kind of sudden change that hasn’t happened in the past, and is unlikely to happen in the future.
(If you aren’t familiar with economic growth, you may want to read my brief explainer before continuing.)
I think this is a reasonable perspective, and it especially makes me skeptical of very imminent forecasts for transformative AI (2036 and earlier).
My main response is that the picture of steady growth—“the world economy growing at a few percent per year”—gets a lot more complicated when we pull back and look at all of economic history, as opposed to just the last couple of centuries. From that perspective, economic growth has mostly been accelerating,8 and projecting the acceleration forward could lead to very rapid economic growth in the coming decades.
Could Advanced AI Drive Explosive Economic Growth? explicitly asks the question, “How likely is ‘explosive growth’ - defined as >30% annual growth in the world economy—by 2100?” It considers arguments on both sides, including both (a) the long view of history that shows accelerating growth; (b) the fact that growth has been remarkably stable over the last ~200 years, implying that something may have changed.
It concludes: “the possibilities for long-run growth are wide open. Both explosive growth and stagnation are plausible.”
Modeling the Human Trajectory asks what future we can expect if we extrapolate out existing trends over the course of economic history. The answer is explosive growth by 2043-2065 - not too far from what my probabilities above suggest. This implies to me that the lack of economic acceleration over the last ~200 years could be a “blip”—soon to be resolved by technology development that restores the feedback loop (discussed in The Duplicator) that can cause acceleration to continue.
To be clear, there are also good reasons not to put too much weight on this as a projection,9 and I am presenting it more as a perspective on the “burden of proof” than as a mainline forecast for when PASTA will be developed.
History of “AI hype”
Another angle on the burden of proof: I sometimes hear comments along the lines of “AI has been overhyped many times in the past, and transformative AI10 is constantly ‘just around the corner’ according to excited technologists. Your estimates are just the latest in this tradition. Since past estimates were wrong, yours probably are too.”
However, I don’t think the history of “AI hype” bears out this sort of claim. What should we learn from past AI forecasts? reviewed histories of AI to try to understand what the actual historical pattern of “AI hype” has been.
Its summary gives the following impressions (note that “HLMI,” or “human-level machine intelligence,” is a fairly similar idea to PASTA):
The peak of AI hype seems to have been from 1956-1973. Still, the hype implied by some of the best-known AI predictions from this period is commonly exaggerated.
After ~1973, few experts seemed to discuss HLMI (or something similar) as a medium-term possibility, in part because many experts learned from the failure of the field’s earlier excessive optimism.
The second major period of AI hype, in the early 1980s, seems to have been more about the possibility of commercially useful, narrow-purpose “expert systems,” not about HLMI (or something similar) …
It’s unclear to me whether I would have been persuaded by contemporary critiques of early AI optimism, or whether I would have thought to ask the right kinds of skeptical questions at the time. The most substantive critique during the early years was by Hubert Dreyfus, and my guess is that I would have found it persuasive at the time, but I can’t be confident of that.
My summary is that it isn’t particularly fair to say that there have been many waves of separate, over-aggressive forecasts about transformative AI. Expectations were probably too high in the 1956-1973 period, but I don’t think there is much reason here to impose a massive “burden of proof” on well-researched estimates today.
Other angles on the burden of proof
Here are some other possible ways of capturing the “That’s too wild” reaction:
“My cause is very important” claims. Many people—throughout the world today, and throughout history—claim or have claimed that whatever issue they’re working on is hugely important, often that it could have global or even galaxy-wide stakes. Most of them have to be wrong.
Here I think the key question is whether this claim is supported by better arguments, and/or more trustworthy people, than other “My cause is very important” claims. If you’re this deep into reading about the “most important century” hypothesis, I think you’re putting yourself in a good position to answer this question for yourself.
Expert opinion will be covered extensively in future posts. For now, my main position is that the claims I’m making neither contradict a particular expert consensus, nor are supported by one. They are, rather, claims about topics that simply have no “field” of experts devoted to studying them. Some people might choose to ignore any claims that aren’t actively supported by a robust expert consensus; but given the stakes, I don’t think that is what we should be doing in this case.
(That said, the best available survey of AI researchers has conclusions that seem broadly consistent with mine, as I’ll discuss in the next post.)
Uncaptured “That’s too wild” reactions. I’m sure this piece hasn’t captured every possible angle that could be underlying a “That’s too wild” reaction. (Though not for lack of trying!) Some people will simply have irreducible intuitions that the claims in this series are too wild to take seriously.
A general take on these angles. Something that bugs me about most of the angles in this section is that they seem too general. If you simply refuse (absent overwhelming evidence) to believe any claim that fits a “my cause is very important” pattern, or isn’t already backed by a robust expert consensus, or simply sounds wild, that seems like a dangerous reasoning pattern. Presumably some people, sometimes, will live in the most important century; we should be suspicious of any reasoning patterns that would reliably11 make these people conclude that they don’t.
Of course, the answer could be “A kajillion years from now” or “Never.” ↩
Technically, these probabilities are for “artificial general intelligence”, not transformative AI. The probabilities for transformative AI could be higher if it’s possible to have transformative AI without artificial general intelligence, e.g. by via something like PASTA. ↩
From Wikipedia: “Genetic measurements indicate that the ape lineage which would lead to Homo sapiens diverged from the lineage that would lead to chimpanzees and bonobos, the closest living relatives of modern humans, around 4.6 to 6.2 million years ago. Anatomically modern humans arose in Africa about 300,000 years ago, and reached behavioural modernity about 50,000 years ago.” ↩
E.g., it emphasizes the odds of being among the most important “people” instead of “centuries.” ↩
I don’t have a great single source for this, although you can see this paper. My informal impression from talking to people in the field is that AI speech recognition is at least quite close to human-level, if not better. ↩
“The field of AI is largely held to have begun in Dartmouth in 1956” ↩
There is an open debate on whether past economic data actually shows sustained acceleration, as opposed to a series of very different time periods with increasing growth rates. I discuss how the debate could change my conclusions here. ↩
“Modeling the Human Trajectory” emphasizes that the model that generates these numbers “is not flexible enough to fully accommodate events as large and sudden as the industrial revolution.” The author adds: “Especially since it imperfectly matches the past, its projection for the future should be read loosely, as merely adding plausibility to an upswing in the next century. Davidson (2021) [“Could Advanced AI Drive Explosive Economic Growth?”] points at one important way the projections could continue to be off for many decades: while the model’s dynamics are dominated by a spiraling economic acceleration, people are still an important input to production, and, if anything becoming wealthy has led to people having fewer children. In the coming decades, that could hamper the predicted acceleration, to the degree we can’t or don’t substitute robots for workers.” ↩
(Absent overwhelming evidence, which I don’t think we should generally assume will always be present when it is “needed.”) ↩