My primary aim in this post is to present two basic models of the development and future of “intelligence”, and to highlight the differences between these models. I believe that people’s beliefs about the future of AI and “AI takeoff scenarios” may in large part depend on which of these two simple models they favor most strongly, and hence it seems worth making these models more explicit, so that we can better evaluate and critique them.
Among the two models I present, I myself happen to consider one of them significantly more plausible, and I will outline some of the reasons why I believe that.
The models I present may feel painfully basic, but I think it can be helpful to visit these most basic issues, as it seems to me that much disagreement springs from there.
It might be tempting to view the discussion of the contrasting models below as a clash between the “AI priority” camp and the “AI skepticism” camp. But I think this would be inaccurate. Neither of the two models I present imply that we should be unconcerned about AI, or indeed that avoiding catastrophic AI outcomes should not be a top priority. Where the models will tend to disagree is more when it comes to what kinds of AI outcomes are most likely, and, as a consequence, how we can best address risks of bad AI outcomes. (More on this below.)
Two contrasting definitions of “intelligence”
Before outlining the two models of the development and future of “intelligence”, it is worth first specifying two distinct definitions of “intelligence”. These definitions are important, as the two contrasting models that I outline below see the relationship between these definitions of “intelligence” in very different ways.
The two definitions of intelligence are the following:
Intelligence 1: Individual cognitive abilities.
Intelligence 2: The ability to achieve a wide range of goals.
The first definition is arguably the common-sense definition of “intelligence”, and is often associated with constructs such as IQ and the g factor. The second definition is more abstract, and is inspired by attempts to provide a broad definition of “intelligence” (see e.g. Legg & Hutter, 2007).
At a first glance, the difference between these two definitions may not be all that clear. After all, individual cognitive abilities can surely be classified as “abilities to achieve a wide range of goals”, meaning that Intelligence 1 can be seen as a subset of Intelligence 2. This seems fairly uncontroversial to say, and both of the models outlined below would agree with this claim.
Where substantive disagreement begins to enter the picture is when we explore the reverse relation. Is Intelligence 2 likewise a subset of Intelligence 1? In other words, are the two notions of “intelligence” virtually identical?
This is hardly the case. After all, abilities such as constructing a large building, or sending a spaceship to the moon, are not purely a product of individual cognitive abilities, even if cognitive abilities play crucial parts in such achievements.
The distance between Intelligence 1 and Intelligence 2 — or rather, how small or large of a subset Intelligence 1 is within Intelligence 2 — is a key point of disagreement between the two models outlined below, as will hopefully become clear shortly.
Two contrasting models of the development and future of “intelligence”
Simplified models at opposite ends of a spectrum
The two models I present below are extremely simple and coarse-grained, but I still think they capture some key aspects of how people tend to diverge in their thinking about the development and future of “intelligence”.
The models I present exist at opposite ends of a spectrum, where this spectrum can be seen as representing contrasting positions on the following issues:
- How large of a subset is Intelligence 1 within Intelligence 2? (Model 1: “Very large” vs. Model 2: “Rather modest”).
- How much of a bottleneck are individual human cognitive abilities to economic and technological growth? (Model 1: “By far the main bottleneck” vs. Model 2: “A significant bottleneck among many others”).
- How suddenly will the most growth-relevant human abilities be surpassed by machines? (Model 1: “Quite suddenly, perhaps in months or weeks” vs. Model 2: “Quite gradually, in many decades or even millennia”).
These three issues are not strictly identical, but they are nevertheless closely related, in that beliefs that lie at one end of the spectrum on one of these issues will tend to push beliefs toward the corresponding end of the spectrum on the other two issues.
Model 1 says that individual human cognitive abilities are by far the main driver of, and bottleneck to, economic and technological progress. It can perhaps best be described with a few illustrations.
The following figures are meant to illustrate how Model 1 sees the respective contributions of humanity and humanity’s external tools to economic and technological progress over time. That is, the figures are meant to depict how much of a driver humanity is in creating technological and economic growth compared to how much of a driver our external tools and technologies are. (The illustrations are not to scale in any sense.)
The bar to the right tracks how large a proportion of human cognitive abilities that is surpassed by machines at the time in question, since this is the key measure to keep track of for Model 1.
In 20.000 BC, we have relatively few tools and virtually no technological or economic progress (meant to be illustrated by the thin arrows). Machines do not surpass individual human cognitive abilities in any domain.
In 1500, we have more tools and far more humans — thousands of times more humans, ostensibly — and we have more technological and economic growth (though the global GDP per capita is still only estimated to be 50 percent higher than in 20.000 BC). Machines still do not surpass individual human cognitive abilities in any domain.
In 1950, we have considerably more tools and humans (more than five times as many humans as in 1500). We also have considerably more growth and wealth — global GDP per capita is estimated to be more than ten times as high as in 1500. Machines are beginning to surpass human cognitive abilities in a narrow range of domains, such as in calculations of the decimals of pi.
In 2020, we again have more tools, including advanced AI programs that surpass or are close to surpassing humans in many domains, e.g. AlphaZero, AlphaStar, and GPT-3. The human population roughly tripled since 1950, and global GDP is more than ten times higher than in 1950.
In 2040 (or within a few decades from now; the exact year is not too important), Model 1 predicts that all or virtually all human cognitive abilities have likely been surpassed by AI systems. Importantly, and most distinctively, Model 1 predicts that once AI systems surpass humans on a sufficient set of cognitive abilities, human cognitive abilities will no longer be the main limiting factor to technological progress. The limiting factor will then be the AI systems themselves, which are continually growing themselves in ever more competent ways, in effect doubling the economy on a timescale of months or weeks, if not faster. The picture we started with has thus been reversed, and humans, if they are around anymore, are now but a tiny speck compared to the mushrooming Jupiter of AI.
Model 2 disagrees with Model 1’s claim that individual human cognitive abilities are by far the main driver of, and bottleneck to, economic and technological progress. According to Model 2, there are many bottlenecks to economic and technological progress, of which human cognitive abilities are but one — an important one, to be sure, but not necessarily one that is more important than others.
Broadly speaking, Model 2 focuses more on Intelligence 2 (the ability to achieve goals in general) than on Intelligence 1 (individual human cognitive abilities), and it sees a large difference between the two.
Since Model 2 does not see progress as being chiefly bottlenecked by Intelligence 1, it is also less interested in tracking the extent to which machines surpass humans along the dimension of Intelligence 1. Instead, Model 2 is more interested in tracking how well our machines and tools do in terms of the broader notion of Intelligence 2. In particular, Model 2 considers it more important to track the relative contribution of our external tools to humanity’s ability to achieve goals in general, including the goal of creating technological progress.
Thus, the following figures are meant to illustrate how Model 2 sees the respective contributions of humanity and external human tools to economic and technological progress over time. And the bar to the right now tracks the relative contribution of external tools in humanity’s technological progress (operationalized and tracked in accordance with Model 2).
As we can see, Model 2 is largely in agreement with Model 1 in the year 20.000 BC. The only difference between the respective figures at this stage is that Model 2 highlights how, in the very limited growth that occurred back then, humanity’s external tools already played a meaningful role. (Model 1 would obviously not disagree that there were contributions of this kind, but it sees these contributions as much less important than the, for Model 1, all-important question of how well our external tools match human cognitive abilities.)
In 1500, the differences between Model 1 and Model 2 begin to emerge. Model 2 includes many distinct classes of external tools, as it sees each of them as significant bottlenecks to progress. (Model 2 sees these bottlenecks as markedly more significant than does Model 1, and Model 2 therefore represents them more prominently.)
Examples of these tools include scientific instruments, tools for food harvesting, tools for mining, etc. Model 2 highlights the complex interplay of these many different classes of tools, and how the existing set of tools in large part determines what new tools can be built — hence the large relative contribution in the progress bar.
Furthermore, Model 2 stresses the point that the latest, most advanced technologies are, already at this stage, crucial components in the creation of the next generation of technology. There is thus already significant recursive self-improvement at this stage, only it is a recursive self-improvement that occurs gradually at the level of the larger set of tools, rather than at the level of a single tool that directly improves itself.
In 1950, we see a continuation of the preceding trend: our set of external tools has been expanded further, and has become an even greater factor in economic and technological growth. Unlike in 1500, humanity has now made inventions such as computers, telephones, and combustion engines, which in turn unlock new avenues of innovation.
Humans gradually begin to emerge as a node in a network of factors — a network of tools, broadly construed — that are driving technological growth. Humans are, of course, a critically important node in this network, yet the same is increasingly true of the other nodes in the network. For example, if we instantly removed combustion engines or telecommunication in the world of 1950, economic and technological growth would have grinded to a halt — indeed turned negative — and it could not have resumed until these technologies (or their analogues) had been reinvented and widely redeployed.
In 2020, our interconnected web of tools has grown even larger, and our external tools likewise play a still greater role in the creation of new technologies. Humans now use computers and other machines on a large scale, including when they create the next generation of technology. Machines thus effectively do a large share of the work in tech development — in the design of new technologies, in calculations of optimal solutions to technical problems, in transportation, in manufacturing, etc.
Of course, machines still do not perform this work wholly independently (even as many manufacturing systems are now practically fully automated). But this is a symmetric point, in that humans by no means perform the work independently either. Both humans and machines are necessary nodes in the network of tools that drive technological growth. And the core point on which Model 1 and Model 2 disagree is how large of a node — i.e. how large of a contributor and bottleneck — humans are compared to other nodes in this network. In 2020, Model 2 sees humans as a considerably more modest node in this network compared to Model 1 (even if Model 2 may still see humans as the single most significant node).
In 2040 (again, the exact year is not important), Model 2 predicts a very different outcome compared to Model 1. Where Model 1 predicts a growth explosion, Model 2 predicts roughly continuous technological progress along existing trend lines.
The reason for this divergence is that — unlike Model 1 — Model 2 does not see human cognitive abilities as the overriding bottleneck to technological progress, and hence it does not expect future progress in surpassing human cognitive abilities to be a master key that will unlock the floodgates to explosive technological growth.
To reiterate, Model 2 does expect continued growth, partly as a result of further progress in software that enables machines to surpass human cognitive abilities in ever more domains. But in terms of growth rates, Model 2 expects this future growth to largely resemble the technological growth that we have seen in past decades, in which machines have already been gradually surpassing humans in an ever wider range of tasks.
Model 2 sees many bottlenecks to technological growth (e.g. economic and technological bottlenecks in manufacturing and energy harvesting), and it expects these many bottlenecks to continue to exist — and indeed to become increasingly significant — if or when AI systems surpass 80, 90, or even 100 percent of human cognitive abilities.
Contrasting narratives centered on contrasting notions of “intelligence”
Broadly speaking, the two models outlined above can be seen as representing two different narratives of the development of “intelligence”. Model 1 is essentially the story of how machines came to surpass humans in terms of their individual cognitive abilities — i.e. “Intelligence 1”, the chief driver of growth, according to Model 1 — and thereby rapidly took off on its own.
Model 2, in contrast, is essentially the story of how humans created an ever larger and more intricate set of tools that in turn expanded technological civilization’s ability to achieve a wide range of goals, i.e. “Intelligence 2”, including the goal of creating further economic and technological growth. It is a story that sees the ability to create technological progress as not being chiefly determined by Intelligence 1, but as instead being determined and constrained by many factors of similar significance, such as the capabilities of specialized hardware tools found across a wide range of domains.
Examples of views that are loosely approximated by these models
Of course, neither of the models described above are a perfect illustration of anyone’s views. But my impression is that these models nevertheless do roughly approximate two contrasting views that one can find in discussions and projections of the future of AI.
Examples of people whose views mostly seem to resemble Model 1 include I. J. Good, Ray Solomonoff, Vernor Vinge, Eliezer Yudkowsky, Nick Bostrom and Anna Salamon. These authors all appear to have argued that explosive technological growth is likely to occur once machines can surpass human cognitive abilities in a sufficient range of tasks — that is, an “intelligence explosion”.
There are also views that are more difficult to categorize, such as those of Robin Hanson and Eric Drexler. Hanson and Drexler have both argued that an intelligence explosion is unlikely, at least in the sense of a single recursively self-improving AI system that ignites a growth explosion, as they argue that growth in AI capabilities is more likely to be distributed among a broader economy of AI systems. In these respects, their views resemble Model 2. Yet their views resemble Model 1 in that they still believe that AI progress is quite likely to result in explosively fast growth, even if the nature of this AI progress will mostly resemble Model 2 in terms of consisting of a larger conglomerate of disparate AI systems and industries that co-develop.
Why I consider Model 2 most plausible
What follows is an overview of some of the reasons why I consider Model 2 most plausible.
A picture akin to Model 2 is arguably what we have observed historically
When looking at the history of economic and technological growth, it seems to me that Model 2 better describes the pattern that we observe. For example, in terms of individual cognitive abilities — or rather, cognitive talents — humans today do not seem to have significantly bigger or more advanced brains than did humans 20.000 years ago. In fact, some groups of humans who lived more than 20.000 years ago may have had brains that were 15-20 percent larger than the average human brain of today.
In other words, our individual cognitive abilities do not seem to have increased all that much over the last thousands of years, whereas our collective abilities and our cultural and technological tools have undergone a drastic evolution, especially in the last few centuries. Hence, it seems that the key bottlenecks that we have overcome in the last 20.000 years — from our stable ~no-growth condition to our current state of technological advancement — have been unlocked by our collection of external tools and our collective cultural development. It is in large part by overcoming bottlenecks in these domains that humanity has created the economic and technological growth that we have seen in recent millennia (see also Muthukrishna & Henrich, 2016). This recent cultural and technological development is arguably what most distinguishes humans from chimpanzees, more so than our individual cognitive abilities in isolation (Henrich, 2015, ch. 2; Vinding, 2016; 2020; Grace, 2022).
Superhuman software capabilities increasingly advance technological development — yet growth rates appear to be declining
Model 1 emphasizes that computers could rapidly become not just a little better than humans, but far better than humans across virtually all cognitive domains. This would be a momentous event, according to Model 1, since computers themselves would then be driving scientific and technological progress at an unprecedented pace.
Model 2, in contrast, tends to emphasize how computers have already been superhuman in various domains for decades — and not just slightly superhuman, but vastly and increasingly superhuman (e.g. in multiplying astronomical numbers and in memory storage). Additionally, Model 2 highlights how superhuman software capabilities have already been pushing the frontier of scientific and technological development for decades, and how they continue to do so today.
For instance, supercomputers already help advance cutting-edge mathematics (see e.g. experimental mathematics and computer-assisted proofs), and they are likewise among the main tools pushing physics forward, along with software in general. A recent example is how bootstrapping methods run on supercomputers help to clarify the conditions that our physical theories (e.g. of gravity) must satisfy — what physicist Claudia de Rham called “one of the most exciting research developments at the moment” in fundamental physics.
Compared to Model 1, Model 2 tends to see the gains from progress in software as being more uniformly distributed over time, as something that is reaped and translated into further technological progress in a rather gradual manner starting already decades ago. Model 2 expects a continuation of this trend, and hence it sees little reason to predict the kind of growth explosion that is predicted by Model 1.
The respective predictions that Model 1 and Model 2 make about what we should observe in our current condition are thus another reason why I consider Model 2 more plausible. For if Model 1 sees human cognitive abilities as the key bottleneck to technological growth, a bottleneck whose unwinding would effectively give rise to a growth explosion, and if computers increasingly surpass human cognitive abilities and increasingly sit at the forefront of technological development, then Model 1 would seem to predict that growth rates should now be steadily rising.
Yet what we observe is arguably the opposite. Economic and technological growth rates appear to be steadily decreasing, across various measures, despite superhuman cognitive abilities — and software in general — playing an increasingly large role in economic and technological development. This weakly suggests that progress in creating and employing superhuman software capabilities has diminishing returns in terms of its effects on economic and technological growth rates. Model 2 fits better with this putative observation than does Model 1.
Declining growth rates
The observation that growth rates appear to be declining across a wide range of measures deserves further elaboration, as it to my mind represents a key component of the evidence against Model 1. Below are some specific examples. None of these examples represent strong evidence against a future growth explosion in themselves; they each merely count as weak evidence against it. But in combination, I think these observations do give us considerable reason to be skeptical.
Global economic growth
The growth rate of the global economy has, in terms of its underlying trend, seen a steady decline since 1960.
In the 1960s, the average growth rate of the global economy was above 5 percent, and it has since gradually declined such that, in the 2010s, the average growth rate was just around 3 percent (excluding the recession year of 2020, and including the post-recession recovery year of 2010).
The following graphs show the growth rates of the world economy from 1961 to 2021 (purple), along with a linear fit and a 10-year moving average, respectively (pink).
10-year moving average:
It is likewise worth noting that most economists do not appear to expect a future growth explosion.
The fastest supercomputers and other hardware measures
A similar trend of decline can be observed in the performance development of the top 500 supercomputers worldwide, except the decline started occurring sometime around 2010-2015. As the graph below illustrates, the performance of the fastest supercomputer is now a full order of magnitude below what one would have expected based on an extrapolation of the trend that we saw in 1993-2013.
This decline is arguably what one would expect given the predicted end of Moore’s law, projected to hit the ceiling around 2025. In particular, in approaching the end of Moore’s law, we should arguably expect progress to steadily decline as we get closer to the ceiling, rather than expecting a sudden slowdown that only kicks in once we reach 2025 (or thereabout).
These declining trends all count somewhat against projections of imminent explosive growth driven by computers. After all, growth in hardware and software seems to have gone in lockstep so far, with greater computation being a key driver of software progress. This tentatively suggests that the observed slowing trends in hardware will lead to similar trends in software.
Many key technologies only have modest room for further growth
It appears that many technologies cannot be improved much further due to physical limits. Moore’s law mentioned above is a good example: we have strong reasons to think that we cannot keep on shrinking transistors much further, as we are approaching the theoretical limit (in the ultimate limit, we cannot make silicon transistors smaller than one atom, and limits are expected to be reached before that, due to quantum tunneling).
Physicist Tom Murphy gives additional examples in the domain of other machines and energy harvesting technologies: “Electric motors, pumps, battery charging, hydroelectric power, electricity transmission — among many other things — operate at near perfect efficiency (often around 90%).”
This means that many of our key technologies can no longer be made significantly more efficient, which in turn implies that, at least in many domains, sheer scaling is now pretty much the only way to see further growth (as opposed to scaling combined with efficiency gains).
This relates to one of the main theses explored by Robert J. Gordon in The Rise and Fall of American Growth, namely that many of the key innovations that happened in the period 1870 to 1970 — the period in which human civilization saw the highest growth rates in history — were in some sense zero-to-one innovations that cannot be repeated (e.g. various efficiency gains and light-speed communication).
To briefly zoom out and make a more general point, I believe that Model 2 is right in its strong emphasis on the many-faceted nature of growth. As a famous economics essay explains, nobody knows how to make a pencil — it involves work and tools across countless industries. And this point is even more true of the manufacture of more advanced technologies. There are countless processes and tools involved, and improvements in a wide range of this set of tools — like what we saw around the 1960s, according to Gordon — will tend to lead to greater growth. Conversely, when an increasing number of these key tools can no longer be improved much further, such that only a comparatively limited set of tools in our collective toolbox see significant improvements, the growth of the whole system will likewise tend to be more modest, as it is partly constrained by the components that are no longer improving much.
In other words, this picture suggests that the various areas in which we are now approaching the theoretical ceilings will be, if they are not indeed already, considerable bottlenecks to growth.
An unprecedented decline in economic growth rates across doublings
A final observation that I wish to draw attention to is that the decline in economic growth rates that we have witnessed in recent decades is in some sense unprecedented in the entire history of human growth over the last million years.
The world economy has seen roughly three doublings since 1960. And as noted above, the average growth rate today is considerably lower than it was in 1960 — around 3 percent today versus an average growth rate above 5.5 percent in the period 1950-1970 (DeLong, 1998, “Total World Real GDP”).
This decline in growth rates across doublings is unprecedented in the history of human economic growth. Throughout economic history, three doublings of the global economy appears to have resulted in a clear increase in the growth rate of the economy. For example, the economy doubled roughly three times from 400 BC to 1750, the latter being at the beginning of the Industrial Revolution where growth rates rose to unprecedented levels (DeLong, 1998, “Total World Real GDP”).
To be clear, I am not saying that this observation is in itself strong evidence against a future growth explosion. But it does serve as counter-evidence to a claim that is sometimes made in favor of a future growth explosion, namely that the history of economic growth suggests that we are headed for a new, even faster growth mode (see e.g. Hanson, 2000; note that Hanson likewise mostly bases his claims on DeLong’s estimates).
The unprecedented growth decline that we have seen over the past three doublings may be taken as weak inductive evidence that we are generally headed for slower rather than faster growth, and that we might now find ourselves past a long era of increasing growth rates (see also Modis, 2012, pp. 20-22).
This wouldn’t be extremely surprising, given that we have good reasons to think that current growth rates can at most continue for a couple of thousand years, and perhaps only a few hundred years. For example, estimates suggest that a continuation of current growth rates in information processing would hit ultimate theoretical limits in less than 250 years. Hence, if future economic growth will be strongly tied to growth in computation, we have some reason to think that current economic growth rates cannot be sustained for much longer than a couple of centuries either.
In light of such strikingly imminent limits — whether they lie a few thousand or a few hundred years ahead — it would not be that surprising if we are already now witnessing a gradual slowdown in growth rates. After all, as mentioned earlier, we should arguably expect a gradual decline in growth rates to set in a good while before we hit ultimate limits.
Again, none of the considerations I have outlined above are in themselves strong reasons to doubt a future growth explosion. But I find that they do begin to paint a fairly compelling picture when put together.
Below are a couple of counterpoints that may be raised in favor of Model 1, along with my comments on them.
Some versions of Model 1 are consistent with the evidence
There are versions of Model 1 that do not predict gradually rising growth rates when just any set of human cognitive abilities get surpassed, yet which do predict explosive growth rates once AI systems exceed humans in the right set of cognitive skills. Specifically, there may well be a certain set of cognitive skills relevant to growth such that, once AI systems surpass humans along those dimensions, then we will see explosive growth. These skills may include strategic planning or the ability to write code.
I think it is worth taking this hypothesis seriously, and to carefully scrutinize its plausibility. The following are some of the reasons why I am skeptical.
Probably the main reason I consider this picture unlikely is that it still seems to overlook the multitude of bottlenecks that work against explosive growth. In particular, it still appears to ignore the increasingly challenging constraints found in the realm of hardware and hardware production. It is difficult to see how greatly improving a given set of cognitive abilities would vastly improve the process of hardware production, since hardware production in large part seems bottlenecked by the sophistication of our hardware itself, not just information.
To be sure, improvements in domains such as strategic planning and coding may produce considerable gains. Yet would these gains be enough to not only get us back to something that resembles the growth in computational efficiency, memory, processing speed, and performance that we saw more than a decade ago, but also enough to positively exceed those growth rates by a considerable margin?
I think we have good reasons to doubt that, and it is worth reiterating that the reasons to be skeptical are partly a matter of hard physical limits, many of which we have practically already reached or are close to reaching (as mentioned above).
It is also worth asking why progress in surpassing a particular set of human cognitive abilities, such as strategic planning, should give rise to explosive growth, when vastly surpassing other human cognitive abilities — e.g. in various calculation tasks since around 1950 — has not resulted in explosive growth (even as it has helped power recent growth).
After all, it is not clear why future progress in surpassing human cognitive abilities should boost growth rates more than did past progress of this kind. For example, one could argue that past progress likely reaped the low-hanging fruit, particularly in the domains where computers have the clearest advantages compared to humans, and that software progress is now facing diminishing returns in terms of how much it will boost growth rates going forward. Again, this is arguably consistent with the patterns that we currently observe, with the growth rates of the global economy and growth in computer hardware steadily declining.
In other words, expectations of greater increases in growth rates from future progress compared to past progress seem to call for an explanation, especially since past trends and our current trajectory can seem to suggest the opposite.
We should chiefly focus on Model 1 regardless
For any set of plausible credences in Model 1 and Model 2, we should still give far more priority to scenarios that conform to Model 1, as they seem far more consequential and worrying.
Model 2 also implies many serious risks, and it is not clear that these are significantly less consequential.
After all, near-term growth rates do not seem critical to the size of the future. If one assumes an eventual colonization wave that occurs near the speed of light, there is virtually no difference between whether such a wave is approached through something like 3 percent annual growth or whether it is approached at the fastest pace allowed by the laws of physics. These two scenarios would likely only differ by a few hundred — or at most a few thousand — years in terms of when they start expanding at the speed of light (cf. Krauss, 2011; Vinding, 2017a). That difference becomes proportionally minuscule as time progresses, implying that a slight difference in the quality of the respective colonization waves would be far more consequential than their respective starting times.
Second, it seems that many specific risks that one might worry about in the context of Model 1 would have analogues in Model 2, including risks of lock-in scenarios.
Third, one could argue that we have greater reason to focus mostly on Model 2 scenarios. For example, one could argue that Model 2 and its basic predictions seem significantly more plausible (as I have tried to do above). Additionally, one may argue that we have a better chance of having a beneficial impact under Model 2 — e.g. because the effects of interventions might generally be more predictable in Model 2 scenarios, and because such scenarios give us more time to understand and implement genuinely helpful interventions. Likewise, one might think that Model 2 implies a significantly lower risk of accidentally making things much worse. For instance, it seems plausible that Model 1 scenarios come with a significantly greater risk of unwittingly creating suffering on a vast scale, which may be an additional reason to focus more on Model 2 scenarios (cf. Trammell’s “Which World Gets Saved” and Tomasik’s “Astronomical suffering from slightly misaligned artificial intelligence”).
Fourth, even if we were to grant that Model 1 scenarios deserve greater priority, one could still argue that such scenarios are currently overprioritized among people concerned about creating beneficial AI outcomes in the long term, and hence that it would be better to give greater priority to Model 2 scenarios on the current margin. For example, one might think that the current focus is roughly 80/20 in favor of Model 1 (among people working to create beneficial AI outcomes), whereas it should ideally only be, say, 60/40 in favor of Model 1.
Lastly, I think it is worth flagging how priority-related assumptions of the kind expressed in the objection above can potentially distort our views and priorities. For one, such assumptions about how some scenarios and worldviews deserve far more priority than others might be risky in that they may reduce our motivation to carefully explore which outcomes are in fact most likely. Moreover, to the extent that we do seriously explore which outcomes are most likely, there is a risk that priority-related assumptions of this kind will exert an undue influence on our descriptive picture — e.g. if we are inclined to downplay a line of argument for priority-related reasons, we might confuse ourselves as to whether we also disagree for purely descriptive reasons. It may be difficult to keep these classes of reasons separate in our minds, and to not let priority-related reasons pollute our descriptive worldview.
Worrying, too, is the empirically documented tendency of “belief digitization”, namely the tendency to wholly dismiss a less plausible hypothesis when making inferences and decisions, even when one acknowledges that the hypothesis in question still deserves considerable weight. This tendency could lead us to assign unduly low credences to scenarios that might not seem most likely or most worrisome, yet which still deserve considerable weight and priority (in line with the previous point about the possibility that we underprioritize Model 2 scenarios on the current margin, even if we believe that Model 1 scenarios deserve the most weight in absolute terms).
Potential implications for altruistic priorities
In closing, I will briefly outline what I see as some of the main implications of updating more toward Model 2 compared to Model 1 (if one were to make such an update).
There is obviously much uncertainty as to what the exact implications would be, but it seems that a greater credence in Model 2 likely pushes toward:
- Comparatively less focus on individual AI systems.
- Individual AI systems in themselves are less likely to be the critical determinants of future outcomes on Model 2. This is not to say that the design of individual AI systems is unimportant on Model 2, but merely that it becomes less important.
- A greater focus on improving institutions and other “broader” factors, e.g. norms and values.
- This may be the most promising lever with which to steer civilization in better directions in scenarios with more distributed growth and power, as it has the potential to simultaneously influence a vast number of actors, creating better equilibria for how they cooperate and make decisions.
- Note that a focus on institutions would not necessarily just mean a focus on AI governance in particular, but also better political institutions and practices in a broader sense.
- More research on what could go wrong in Model 2 scenarios.
- For example, what are the most worrying lock-in dynamics that might occur in these scenarios, and how might they come about? How can they best be prevented?
Updating more toward Model 2 would by no means change everything, but it would still imply significant changes in altruistic priorities.
Thanks to Tobias Baumann for suggesting that I write this post. I am also grateful to the people who gave feedback on these ideas at a recent retreat and in writing.
Appendix: Critical questions regarding the models
It seems worth briefly reviewing a couple of critical points and questions that one might raise in response to the two basic models reviewed in this post.
Just qualitative models?
One critique might be that the two models presented here are merely qualitative, and hence they are less useful than proper quantitative models of future growth. I would make two points in response to this sentiment.
First, it seems that qualitative models are a good place to start if we want to think straight about a given issue. In particular, if we want to make sure that the more elaborate models we construct are plausible, we have good reason to first make sure that we base these models on plausible assumptions about the nature of technological growth, especially as far as the key drivers and bottlenecks are concerned.
This brings me to the second point, which is that quantitative models are only as good as the underlying assumptions on which they rest. A qualitative model can thus be more useful than a quantitative model, if the qualitative model better reflects reality in terms of the core assumptions and predictions it makes. That being said, it would, of course, be ideal to have quantitative models that do rest on accurate assumptions.
The two models presented here are not meant to provide precise quantitative predictions (even as they do provide some rough predictions), but rather to illustrate an underlying source of disagreement in a simple and intuitive way.
How do the models presented here relate to the economic models presented in Aghion et al., 2019?
The models presented by Aghion et al. relate quite closely to the models presented in this post. The first model they present is an automation model in which a set of tasks that are difficult to automate serve as a bottleneck to more rapid economic growth — a model that seems to describe recent growth trends reasonably well (Aghion et al., 2019, sec. 9.2).
The authors then proceed to describe various models that are similar to Model 1, in which growth becomes ever faster due to greater automation or self-improving AI (Aghion et al., 2019, sec. 9.4). Yet it is worth noting that the authors do not claim that any of these models are plausible. Indeed, these abstract models are in some sense trivially implausible, at least in the limit, since they either imply that economic output becomes infinite in finite time, or that growth rates will increase exponentially without bounds.
Furthermore, concerning the equations that the authors use to formalize an intelligence explosion, they note that there are reasons to believe that those equations “would not be the correct specifications”, and they go on to discuss these reasons, “which can broadly be characterized as ‘bottlenecks’ that AI cannot resolve” (Aghion et al., 2019, p. 258). These reasons relate to Model 2, in that Model 2 can be seen as an intuitive illustration of some of the counterpoints that Aghion et al. raise against the plausibility of automation/AI-driven explosive growth (in Section 9.4.2).
Overall, Aghion et al. do not appear to take a strong stance against a future growth explosion driven by AI, but my impression is that they are tentatively skeptical.
How do the points made in favor of Model 2 in this post relate to the conclusions made in Davidson 2021?
Davidson concludes that it is plausible (>10 percent likely) that ‘explosive growth’ — by which he means >30 percent annual global economic growth — will occur before 2100. Yet it is worth noting that 30 percent annual growth is far closer to current growth rates than to the growth rates found in an economy that doubles on the order of months, weeks, or shorter. For example, an economy that doubles every other month would see 6300 percent annual growth(!), whereas an economy that doubles every week would see more than (i.e. >4 quadrillion) percent annual growth(!!).
So we should be clear that Davidson is evaluating a much milder growth explosion than the extreme growth explosions typically implied by Model 1, or by, say, Robin Hanson’s projections (Hanson, 2000; 2016). And assigning >10 percent probability to >30 percent annual growth is obviously far less radical than is assigning >10 percent probability to >6300 percent annual growth, let alone to >4 quadrillion percent annual growth. Worth noting, too, is that Davidson seems to assign an even higher probability to a growth slowdown: “I place at least 25% probability on the global economy only doubling every ~50 years by 2100”; “my central estimate is not stable, but it is currently around 40%”. This seems broadly consistent with the arguments I have presented in favor of Model 2 (although I would place strictly less than 10 percent probability on >30 percent growth rates in light of these arguments).
How does one operationalize the quantities measured by the respective “progress bars” in the figures above?
This is a fair question. There is obviously no single right way to operationalize “the proportion of individual human cognitive abilities surpassed by machines” or “the relative contribution of external tools to human technological progress”. But I also don’t think much hinges on the exact operationalization. The main point is that Model 1 and Model 2 disagree on what kinds of progress are most informative to track. The exact positions of the progress bars are thus not critical, in contrast to the underlying points that the respective bars are meant to illustrate.
Aaronson, S. (2008). The limits of quantum. Scientific American, March 2008, pp. 62-69. Ungated
Aghion, P. et al. (2019). Artificial Intelligence and Economic Growth. In Agrawal, A. et al. (eds.), The Economics of Artificial Intelligence: An Agenda. The University of Chicago Press. Ungated
Cowen, T. (2011). The Great Stagnation: How America Ate All the Low-Hanging Fruit of Modern History, Got Sick, and Will (Eventually) Feel Better. Dutton Adult.
Davidson, T. (2021). Could Advanced AI Drive Explosive Economic Growth? Ungated
DeLong, B. (1998). Estimating World GDP, One Million B.C. - Present. Ungated
Gordon, R. (2012). Is U.S. Economic Growth Over? Faltering Innovation Confronts the Six Headwinds. Ungated
Gordon, R. (2016). The Rise and Fall of American Growth: The U.S. Standard of Living since the Civil War. Princeton University Press.
Grace, K. (2022). Counterarguments to the basic AI x-risk case. Ungated
Hanson, R. (2000). Long-Term Growth As A Sequence of Exponential Modes. Ungated
Hanson, R. (2009). Limits To Growth. Ungated
Hanson, R. (2016). The Age of Em: Work, Love and Life When Robots Rule the Earth. Oxford University Press.
Henrich, J. (2015). The Secret of Our Success: How Culture Is Driving Human Evolution, Domesticating Our Species, and Making Us Smarter. Princeton University Press.
Jones, C. (2021). The past and future of economic growth: a semi-endogenous perspective. Ungated
Jones, C. (2022). The Past and Future of Economic Growth: A Semi-Endogenous Perspective. Annu. Rev. Econ., 14, pp. 125-152. Ungated
Krauss, L. (2011). The Future of Life in the Universe. Ungated
Krauss, L. & Starkman, G. (2004). Universal Limits on Computation. Ungated
Legg, S. & Hutter, M. (2007). Universal Intelligence: A Definition of Machine Intelligence. Ungated
Modis, T. (2012). Why the Singularity Cannot Happen. In Eden, A. et al. (eds.), Singularity Hypotheses: A Scientific and Philosophical Assessment. Springer. Ungated
Murphy, T. (2011). Can Economic Growth Last? Ungated
Muthukrishna, M. & Henrich, J. (2016). Innovation in the collective brain. Philosophical Transactions of the Royal Society B: Biological Sciences, 371: 20150192. Ungated
Sotala, K. (2017). Disjunctive AI scenarios: Individual or collective takeoff? Ungated
Tomasik, B. (2018). Astronomical suffering from slightly misaligned artificial intelligence. Ungated
Trammell, P. (2018). Which World Gets Saved. Ungated
Vinding, M. (2017a/2020). The future of growth: Near-zero growth rates. Ungated
Vinding, M. (2017b/2022). A Contra AI FOOM Reading List. Ungated
Vinding, M. (2020). Chimps, Humans, and AI: A Deceptive Analogy. Ungated
Vinding, M. (2021). Some reasons not to expect a growth explosion. Ungated
Vinding, M. (2022). What does a future dominated by AI imply? Ungated
Other examples of people whose views conform reasonably well to Model 2 may be found in the following Contra AI FOOM Reading List. Though as I note shortly in the text, the views of some people on this list also conform well to Model 1 in some significant respects.
Cognitive abilities are arguably strongly tied to culture, so it seems worth distinguishing cognitive talents vs. abilities.
Likewise, Model 2 is more skeptical as to whether it is helpful to talk about “a point” at which humanity achieves AGI, as opposed to there being a continued, decade-long process of AI systems surpassing humans in ever more cognitive domains, where many of the most important superhuman cognitive abilities may be achieved early in that process, e.g. superhuman calculation and memory abilities that serve critical functions in the development of technology.
This would at least be the prediction of Model 1 in its most straightforward version; I discuss an alternative view below.
Note that diminishing returns in terms of growth rates need not imply diminishing returns in terms of absolute growth. For example, an economy whose growth rate steadily declines from above 5 percent to around 3 percent would obviously see diminishing growth rates, i.e. it would grow sub-exponentially, but the size of the economy may still be growing superlinearly, since absolute growth can increase despite decreasing growth rates (as the development of the global economy since the 1960s illustrates).
The following subsections expand on some of the points presented in Vinding, 2021.
Note that the trend shown below extends further back than the 1960s, as the average annual growth rate in the 1950s was likewise above 5 percent.
The linear regression is obviously not a good fit (), but it still helps give some sense of the overall direction of economic growth rates over the last 60 years.
But can’t we still just scale our way to massive growth? This might be possible in theory, but a problem is that hitting limits in many kinds of efficiency gains and hardware progress (as described in the previous section) also means that further scaling, e.g. in terms of getting more computing power, will become increasingly expensive compared to what we are used to. Efficiency gains appear to have been an important driver of economic growth to date (see e.g. Gordon, 2012), and hence as efficiency gains begin to decline, we should not only expect to get less relative growth for the money that we invest into further growth, but we should arguably also expect the relative growth of our investment capital itself to begin to decline (compared to what we have been used to). So simply scaling up at unprecedented rates may be a lot easier said than done, as the costs may be prohibitive.
A similar thesis is defended in Cowen, 2011. Both Gordon’s and Cowen’s books focus on American growth, yet they still capture much of the global picture, since the US economy accounted for roughly 40 percent of the global economy in 1960, and hence it drove much of the global growth in that period.
Charles Jones likewise argues that most of the growth that the US economy has seen since the 1950s has been due to zero-to-one improvements that cannot be repeated, e.g. misallocations of resources that cannot be corrected much further. A smaller fraction of growth, around 20 percent, has been due to an increase in population, which is the only factor in the growth model presented by Jones that could keep on driving growth in the long term (Jones, 2022).
Yet note that even population growth cannot continue at a high rate for long. Indeed, a maximum growth scenario that involves a colonization sphere that expands outward from Earth at the speed of light could only expand for around 100 years before the annual growth rate of its population would fall below 3 percent, assuming a constant population density in the colonized sphere. Constant 3 percent growth could in principle go on for somewhat longer, though not much longer; if the human population were to grow by 3 percent every year from now, there would be more humans in 5000 years than there are atoms in our galaxy, cf. Hanson, 2009.
That is, in terms of the underlying trendline, disregarding short-lasting recessions.
Estimates of global GDP in historical and prehistoric times are surely far from perfectly reliable. But a closer look at these estimates renders it unlikely that the estimates are so unreliable that it would turn out to be false that three doublings have historically led to consistently higher growth rates (prior to 1960), in terms of the average growth trend across doublings. The main reason is that it is clear that, prior to the modern age, it took a very long time to double the economy three times — so long that global civilization appears to have increased its ability to grow the economy considerably in the meantime.
Specifically, at any point prior to 1960 in DeLong’s estimates (under “Total World Real GDP”), one can try to work three doublings ahead and then see whether it is plausible that the underlying growth rate should be lower than at the starting point. A further test might be to compare the time it took for another doubling to occur from the three-doublings-later point relative to how long a single doubling took at the starting point.
Note also how declining or relatively stable growth rates need not imply that absolute growth will be modest compared to today. For example, if we imagine that economic growth were to continue at roughly the current pace (~3 percent per year) for another 200 years, then the 3 percent growth occurring in a single year 200 years from now would, in absolute terms, be greater than 11 times our entire economy today. (I should note that I do not believe that the most likely scenario is that growth will remain at around 3 percent in the coming centuries, but rather that growth rates will continue to gradually decline along the downtrend we have seen since the 1960s, as also projected by this review of long-term GDP projections and this PWC report.)
As for the possibility that quantum computers might overcome these constraints, physicist Scott Aaronson argues that quantum computers could solve certain problems much faster, but they would likely only do slightly better than conventional computers on most tasks.