This is a crosspost for Benjamin Jones' review of Coefficient Giving's (CG's) report Could Advanced AI Drive Explosive Economic Growth? by Tom Davidson, which was published on 25 June 2021. It includes Ben's initial comment, and a discussion of it between Ben and Tom. I remain open to bets against short AI timelines, or what they supposedly imply, up to 10 k$.

Initial review

Comments on “Could Advanced AI Drive Explosive Economic Growth?”

This is a thought-provoking report and an enjoyable read.  It is balanced, engaging a wide set of viewpoints and acknowledging debates and uncertainties.  The balance occurs empirically (including differing viewpoints on historical growth patterns) and theoretically, engaging various growth models that have been used to meet historical patterns and think about future ones. The report is also admirably clear in its arguments and in digesting the literature.  I particularly valued the clarity around feedback loops, including the role of accumulable inputs in driving growth, and how AI may substitute for labor in this vein and thereby accelerate growth.  Overall, to my read the report engages key ideas in a transparent way, integrating perspectives and developing its analysis clearly and coherently.

Making predictions is of course challenging, and one has to approach one’s predictions with appropriate modesty, as this report does, and I am going to shy away from assessing specific probabilities. Nonetheless, while this report suggests that a rapid growth acceleration is substantially less likely than singularity-oriented commentators sometimes advocate, to my mind this report still sees 30% growth by 2100 as substantially likelier than my intuitions would suggest.  Without picking numbers, and acknowledging that my views may prove wrong, I will just say that achieving 30% growth strikes me as very unlikely.  Here I will articulate some reasons why, to provoke further discussion.

A. Growth at 30% per annum

Part of the argument in the report in favor of a massive growth acceleration is that there is some historical precedent for a massive growth acceleration.  And it does seem clear historically that the sustained growth since the 19^th century in frontier economies like the United States has been much faster than what was seen further in the past.  Data before the 19^th century is of low quality, but I think that Pritchett (“Divergence, Big Time” in the Journal of Economic Perspectives, 1997) has a quite definitive argument for a large acceleration.  Namely, below a certain level of income per person, humans cannot survive.  Most simply, people would not have enough food and the population cannot sustain itself.  This implies a consumption floor.  On this basis, one must infer that growth rates since 1870 are far higher than they were through human history. Specifically, as an exercise in backward extrapolation, if growth rates prior to 1870 were anywhere near as good as we know they are after 1870, then people further in the past would have been impossibly poor.  Growth in per-capita income had to be much closer to 0% for most of human history, and then after 1870 it is much higher.  

What was the process of this acceleration? The report asks whether there has been “super-exponential” growth with a slow/smooth acceleration of growth through human history, prior to the 19^th century.  Another view is that growth was close to 0% through history and then there was a step function increase in the 19^th century. Whether or not you believe in a more-or-less smooth acceleration (as presented in Section 4.1 of the report) or a step function can matter for how you model growth and population dynamics and thus the right conceptual model.  But to me the main point is the historical precedent for acceleration.  Either way (smooth acceleration or a more discrete jump), it is clear that growth accelerated a lot.

Where I become less in line with the report is the idea that such a historical growth acceleration makes 30% growth in the future plausible.  This strikes me as tenuous. The argument in the report is that, since growth in the past has gone from 0.3% to 3% in the past, a 10x increase in growth rates is historically possible.  Therefore, another 10x increase in growth rates, from 3% to 30% has some historical precedent.  I have trouble with this argument on several grounds. 

First, ratios of growth rates seem misleading.  I think what happened historically was that growth rates accelerated to 3% per year.  And growth rates started somewhere close to 0%.  If we say growth was once 0.3% than you get a 10x increase, as emphasized in the report.  But if you say that growth was 0.1% or 0.03% for periods of human history, then the ratio is radically increasing, and if you say historical growth rates were often 0% then we have experienced an infinite increase in the ratio.  Would we then infer that a 300% growth rate or an infinite growth rate is also possible based on historical evidence?  The ratio is extremely sensitive.  But the fact that growth accelerated by about 3% is not very sensitive. So I would say we have a robust sense in which adding 3% growth has been demonstrated historically.  We have never seen anything like a 30% jump in the growth rate.

Second, it’s important to realize what a 30% growth rate means.  This is not like going from 0.3% to 3%.  This is something else entirely, and it’s really, really hard to imagine.  To be clear, at a 30% annual growth in income per capita, productivity and the standard of living double every 2.5 years.  And this compounds.  After just 25 years, we would be 1,000 times richer and more productive than we are now.  Keep in mind that, through all of human history to the present, per-capita income in the U.S. is now about 100 times the consumption floor (that is, the starvation threshold).  So, in other words, if at 30% growth we will be 1,000 times where we are now, then we would need to experience all the technological and productivity gains in all of human history, not just once but ten times over, in just 25 years. That seems hard to imagine in its own terms.  Now think about the technologies and advantages in which the historical gains are embedded – airplanes, skyscrapers, MRI machines, open-heart surgery, industrial robots, smartphones, GPS satellites, on-demand video and music.  Many things we take for granted today would be hard to imagine for people just 100 years ago, let alone 10,000 years ago.  Yet at 30% growth we are imagining that in 25 years we could create and deploy some set of advances that raise standards of living by 10 times more than all the things we have ever figured out and deployed through all of history.  All in all, while 30% growth is not technically a singularity, it would sure feel like one, and this is way outside any historical precedent.

Third, there are adjustment costs, both technological and political in nature, that would seem to undermine the possibility of 30% growth.  This kind of growth would require enormous “creative destruction.”  To double productivity every 2.5 years, we are almost surely talking about brand new capital equipment being implemented.  This is likely very hard to do.  Concretely, history took us from the horse to the airplane. Now there are lots of airports and supporting infrastructure and air-transport supply chains and skills, and few stables and coach drivers.  We would presumably need an even bigger change in infrastructure and capital equipment to get from airplanes to whatever the 1,000 times more productive technology would be, and in the blink of an historical eye. Is the new system low-cost underground hypersonic railways that move you across the U.S. in 1 minute?  Or take you through space, quickly and cheaply?  Even if that were technologically possible, it would be a lot of new infrastructure, and a lot of defunct airplanes, businesses, and workers.

The related creative destruction point is in political economy.  The owners of the vintage capital stock (airplanes), as well as those with relevant vintage human capital (pilots), won’t want to see their capital assets or skills become worthless.  Such capital owners and workers will then likely attempt to block such advances.  Governments face large challenges with the churn of creative destruction at ordinary growth rates, where displaced workers or business owners seek protection.  It’s a little hard to imagine the political implications of rates of change and churn beyond all historical precedent.  These political economy obstacles are a different sort from the technological aspects of AI but ones we should take seriously.  The majority of the world’s population today lives in nations that have not been able to use and deploy the technologies we already have, but rather live at 5-20x or even 100x lower consumption levels behind the frontier.  Governance often obstructs growth in the world.  Even in the most advanced economies, workers and citizens struggle with job displacement (whether through technology like automation or globalization).  This is a very live issue today in advanced economies, even at ordinary rates of worker displacement.

B. Technology

My main reasoning about the potential technological obstacles to formal singularities are articulated in the paper “Artificial Intelligence and Economic Growth” (with Philippe Aghion and Chad Jones).  Because those arguments are laid out there, and because this report already engages them in depth, I won’t repeat them in general here.  But one main theme worth emphasizing is “bottlenecks.”  The Baumol cost-disease or “bottleneck” view of the world, and related recent experience in an advanced economy like the United States, to my mind poses big challenges in connecting even very-advanced AI to a large structural acceleration in growth.  Here I will add some ideas and examples along the bottleneck line to simulate further these debates.

The point of Baumol’s cost disease is to direct one’s attention away from what we are getting good at and recognize that it is the bottlenecks that ultimately govern the economy.  For example, we have successfully automated an amazing amount of agricultural production (in advanced economies) since the 19^th century.  One fact I like:  In 2018, a farmer using a single combine harvester in Illinois set a record by harvesting 3.5 million pounds of corn in just 12 hours.  That is really amazing.  But the result is that corn is far cheaper than it used to be, and the GDP implications are modest.  As productivity advances and prices fall, these amazing technologies tend to become rounding errors in GDP and labor productivity overall.  Indeed, agricultural output used to be about half of all GDP but now it is down to just a couple percent of GDP.  The things you get good at tend to disappear as their prices plummet.  Another example is Moore’s Law.  The progress here is even more mind-boggling – with growth rates in calculations per unit of resource cost going up by over 30% per year.  But the price of calculations has plummeted in response.  Meanwhile, very many things that we want but don’t make rapid progress in – generating electricity; traveling across town; extracting resources from mines; fixing a broken window; fixing a broken limb; vacation services – see sustained high prices and come to take over the economy.  In fact, despite the amazing results of Moore’s Law and all the quite general-purpose advances it enables – from the Internet, to smartphones, to machine learning – the productivity growth in the U.S. economy if anything appears to be slowing down.  One common and related way to look at the U.S. economy is that we are in a structural transformation away from agriculture and manufacturing (where we have become extremely productive) and are left with a rising share of the economy in services (where productivity advances appear harder to propel).  

This “bottleneck” orientation strikes me as a very credible view of the world.  What we advance quickly is exciting and attracts a lot of attention.  But in the Baumol perspective it’s actually the things we want but are bad at that matter, and increasingly so with time.  The world is full of amazing technologies, including those with wide ranging applications.  But the better they get, the more we are left struggling with what those technologies don’t do. 

If this is right, then a growth acceleration via AI requires something without historical precedent.  Yes, machine learning already does some amazing things, and maybe it will do a widening range of amazing things, and maybe more than any technology before.  But changing the growth path is a very tall order.  On the production side, AI might theoretically change the growth game if it can automate everything.  Then it might seem you can simply scale the AI to get more and more of that thing.  But if AI can’t do everything, then bottlenecks will prevail.  Otherwise, on the innovation side, AI could theoretically change the growth game by substantially advancing our ability to innovate – e.g. by scaling human equivalents in R&D processes.  So there are two kinds of pathways here where we could theoretically get a growth acceleration.  

Nevertheless, call me skeptical.  Moore’s Law and computing are already amazing tools on both the production and innovation sides.  Computers have automated a widening range of tasks in production, often with great efficiency.  And computers are great tools for research, and again with a widening range of applications and efficiency.  But economic growth has not been accelerating over this digital revolution period.  The gulf between the amazing progress of computing and anemic overall economic growth is naturally bridged due to bottlenecks, and it seems to me like there are going to be myriad ongoing bottlenecks.  Take the production tasks side.  Here the key question for AI is whether it can automate all the tasks, as discussed in the report. But to push this further, even if you could automate all the tasks (which strikes me as doubtful), you would further need to be able to scale output at each task by applying more capital inputs.  But this also seems doubtful, as many task outputs may not scale much through more machines.  For example, let’s talk about traveling.  Let’s say AI advances do give us self-driving vehicles.  That replaces labor in driving.  But scaling the number of self-driving cars won’t get you somewhere increasingly fast.  The productivity bottleneck here is really the speed of travel (which also depends on how much traffic there is).  So adding more self-driving machinery won’t actually improve output much in transportation (scaling the number of cars could actually make travel worse through congestion).  To achieve 30% per annum productivity gains you actually need to change the method of travel – maybe flying point-to-point vehicles that go really fast and don’t get in each others’ way?  For many wants (e.g., travel, fixing a broken bone, eating a nice meal – and to be clear transportation, health services, and restaurant services are all large sectors of the economy) I don’t think scaling machinery would solve bottlenecks in consumption and hence radically advance GDP per-capita and the standard of living.

Returning to self-driving vehicles, transportation bottlenecks then aren’t overcome through automation of production itself.  Rather, the root bottleneck is a creative one– inventing 1,000 times faster vehicles.  We then have to imagine that AI can do a lot to advance the rate of innovation.  The Aghion et al. (2017) paper lays out reasons to be skeptical on this dimension, which the report also engages, including search limits and fundamental physical laws.  But to add a bit here:  On the inputs to creativity side, I’m not seeing where machine learning is anywhere close to taking over general cognitive functions, so it seems like we need big advances of a different kind of AI.  And creative advances require many inputs, and not just theoretical ones. Scientific and technological progress currently requires the gathering of empirical information, often in experiments directed by new theories. This whole process is an iterative interplay between theoretical and empirical developments. If we need to gather data, then time of experimentation becomes its own bottleneck.  One way to say this is:  if a theory can be thought of as a mathematical function that intersects a bunch of points (the empirical facts), then there are an infinite set of theories for any set of evidence.  Refining theories then requires new empirical facts.  (Occam’s razor says pick the simplest theory, but nature may have more complex ideas.  And the progress of science is often a story of surprising new facts rejecting prior theories.)  Ultimately, if the AI says:  we need a bigger Large Hadron Collider to adjudicate my theories, we would have to wait quite a while for the result.  

Ultimately, I think bottlenecks are where the action is.  An interesting descriptive exercise would be to consider (a) the current array of goods and services that humans consume to see which ones seem both essential to the standard of living and least amenable to a scalable automation solution and (b) the array of activities in scientific and technological advance that are amenable to a scalable intelligence.  It doesn’t take much in the way of bottlenecks to severely undermine the growth implications of AI, even if AI is really fantastic at very many things.

Author response

Hi Ben,

First off, thanks again for your thoughtful comments - I found thinking about them and writing this response very helpful. 

I’ll jump right into the substance. 

Does the historical acceleration of growth make 30% per annum plausible?

As you nicely summarise, the report argues that since growth in the past has gone from 0.3% to 3% in the past, another 10x increase in growth rates might be possible. 

I agree with your objection that the exact size of the historically observed ratio is sensitive to the ancient growth rate, and this makes me less inclined to lean strongly on the argument. So I agree with the direction of your critique. 

But I still find the 10x argument somewhat persuasive, for two reasons. 

First, imagine economists in 1600 discussing the possibility of 3% annual growth. One says:

Well annual growth already increased from 0.03% to 0.3%, so we shouldn’t rule out another 10X increase out of hand.

The other replies 

Hmmm, that 10X ratio is sensitive to the exact the speed of ancient growth. I’d just say we have evidence that growth can increase by 0.3%, and that doesn’t tell us that a 3% increase in growth is possible. Also, keep in mind how crazy 3% growth would be. The economy would have to double in size in a mere 25 years. That would involve all the tech progress we’ve made over centuries happening in just a decade.

In this case at least, the person making the 10X argument was right. To be clear, I don’t think the 10X argument proves 30% growth is possible -- far from it -- but I think it’s a good response to someone who wants to dismiss it out of hand because it seems crazy. The way to move forward is then to consider more carefully whether it might in fact be possible, without a strong bias in either direction.

Second, I don’t think the ‘sensitivity of the ratio’ objection is decisive. We could say that the ancient growth rate you use has to be one that human societies have had for multiple consecutive doublings of GWP. This criterion prevents us using arbitrarily low values for the ancient growth rate and proving too much. 

A 1000X increase in GWP/capita might be possible without big infrastructure changes

A few of your comments relate to just how hard it would be to actually achieve a 1000X increase in GWP/capita in practice. You highlight frictions relating to embedding new technologies in physical infrastructure, resistance to innovation by incumbents, and bottlenecks.  

I wanted to push back on this a little bit, by suggesting some relatively low-friction paths to very large increases in GWP/capita.

Increasing average global consumption to the level of rich people

Some people alive today enjoy annual consumption in excess of $1 million per year. If everyone alive consumed the same basket of goods as these people, average GWP/capita would be $1 million per year. Today, average GWP/capita is $10,000 per year. So if advanced AI allows everyone to enjoy the consumption of the richest alive today, it could drive a 100X increase in GWP/capita. If it did this in 18 years, this would be 30% annual growth. This wouldn’t require lots of new physical infrastructure.

And you could push the argument further. Perhaps some people consume baskets of goods worth $10 million or $100 million each year. If average global income rose to this level, there’d be a 1000X or 10,000X increase in GWP/capita.

Providing services and goods that don’t require infrastructure changes

There are many GDP-enhancing services that very advanced AI could provide without large infrastructure changes. Suppose disembodied AI’s general intelligence and learning abilities match that of the best humans. Then AIs could learn to be world class experts in many domains of knowledge; e.g. coding, medicine, law, therapy, life-coaching, teaching, lobbying, theoretical research, entertainment creation, logistics, transportation and business consulting. They could learn to become experts by studying the multitude of online materials related to these topics. AIs can be copied, so once there is one AI expert in the domain we can run very many copies. These world class experts could dramatically increase output in the relevant industries by creating much higher quantities of output (e.g. everyone can have as much therapy, life-coaching, and medical advice as they want) and by significantly improving the average quality of the service (e.g. better business ideas and more personalised entertainment and medical advice). 

Of course, regulation will prevent AIs from directly replacing the relevant human practitioners in many locations. Despite this, AIs might still do most of the work while formally qualified humans just give the legal stamp of approval. Although some may not choose to use these world-class advisors, those who do would be very successful so others would probably copy. Similarly, if some areas have more lax regulations allowing AIs to have greater economic impact, those areas will prosper and others would be incentivized to copy.

It’s true that as these knowledge-based services are supplied in greater quantities, Baumol effects will diminish their importance to GDP. But before this happens there could be very large increases in GWP/capita. It seems like that top quality cognitive labour is in relatively short supply; so if it became plentiful the increase to GWP/capita could be very large before Baumol effects really kick in.

As an extension of this argument, AIs might learn to be experts in more practical jobs, again by learning from publicly available information online. (E.g. youtube videos explain how to be a plumber and there are online books on mechanical engineering.) AIs could then advise individual human workers in real-time, significantly increasing their productivity. If factories gave AIs access to their private data, AIs could give their workers real-time guidance to boost their productivity. Today there are huge skill differences in wages, suggesting that giving top-quality cognitive guidance would fairly directly apply a multiple to average worker productivity.

To extend the argument further still, suppose AI experts can direct the movements of versatile robots. Then they could enhance the output of manufacturing and construction by literally directing the robots to do the work. 

In summary, if we grant the (very strong!) assumption that disembodied AIs develop general intelligence and learning abilities similar to the best humans, it seems like they could significantly boost output in many industries. Even if bottlenecks and Baumol effects eventually come into play, there could be very large increases in GWP/capita before this happens. If the development of this kind of AI happens sufficiently quickly, it seems like it could drive 30% growth of GWP. 

Would 30% growth really involve 10X humanity’s tech progress in just 25 years?

You point out that 30% growth of GWP/capita for 25 years corresponds to a 1000X, while US GDP/capita is only 100X that needed to survive. So it seems like those 25 years would involve 10X as much tech progress as the whole of human history!

This point is minor, but I think a lot of the technological progress over long-run history went into increasing the population size rather than the standard of living. E.g. between 10,000BC and 1500 AD GWP/capita increased very little but population increased by orders of magnitude. Tech progress allowed us to support more people on a fixed quantity of land, rather than increasing income.

The way I’d put it is that 10 years of 30% annual growth would involve all the tech progress made in the last 100 years (when growth was 3%). This is still an extreme situation, but somewhat less extreme!

Reviewer response

Hi Tom,

Thanks very much for this and sending your note.

Here are some further thoughts, responding to your notes.  Free disposal of course, but hopefully these help advance the thinking and debates a bit further.

1) Increasing average global consumption to the level of rich people

This is an interesting perspective but also strikes me as a very different take from the conceptual orientations of the usual AI-growth literature.  The AI-growth literature and its models focus on the frontier level of technology in the world. I had interpreted the 30% growth feature in your report as happening in the US or other technologically advanced nation.

When it comes to the world income distribution, US mean income is on the order of 100x the mean income in the very poorest countries.  So if one wants to argue that a lot of the 30% growth is coming through convergence to existing income levels, then perhaps we no longer need to imagine 100x or 1000x in the US per se.  But this strikes me as a very different type of case and set of considerations, and requires engaging different areas of economics.  In particular, the main fact about economic development is that most countries / economies in the world are nowhere near the current technological frontier.  That is, they do not adopt and implement cutting edge technology. The modern explanations for this typically emphasize governance challenges (weak property rights, contracts, corruption, political instability, civil conflict, etc.) among other forces.  The story one needs is then one that revolves around so-called “convergence” – something where AI solves the problems that prevent economies from failing to grow.  But, for example, I don’t see an articulated argument where AI solves issues around property rights, corruption, conflict, etc.  So I think a convergence / catch-up type of story for AI needs a quite different set of conceptual considerations.  But I grant you that, if AI somehow solves the challenges of development economics, then faster growth is more within reach.

Somewhat different considerations would also bear on the reduction of inequality within an economy, including advanced economies.  Income distributions are very durable.  So, for example, why is there a large wage and income distribution within the US, why is there enduring poverty in the US?  Then:  how does AI solve those problems?  These again seem like very different conceptual and theoretical considerations (the sources of inequality) than the theories currently emphasized in your report or that I have seen (drivers of growth in mean income).

2) Providing services and goods that don’t require infrastructure changes

My main concern here would be twofold. 

First, and not to harp on bottlenecks, but I don’t see how a services story gets one past a bottleneck problem.  That is, AI-based services could be great but we would still have the Baumol price effect, where the GDP share of such services will go to zero.  Your comment suggests there is a delay in the price decline but I don’t see why there is a delay between quantity and price movements.  (Or how that can happen at scale because markets wouldn’t clear.) Related, I was recently making a calculation for another paper which suggests that effects of even infinite productivity and quantity in a substantial share of sectors will have modest effects on GDP.  For example, taking the usual bottleneck parameter you can let 50% of the tasks in the economy be done by AI and let that AI become infinitely productive, and per capita income goes up by only 50%.

Second, the service advances (medicine, law, etc.) will create political economy problems – doctors and lawyers who are being displaced will try to stop it.  Or, if you believe the doctors or lawyers are still needed for some sub-task, as in one version you lay out, then we have the bottlenecks again and the AI doesn’t get you very far.  Or take government services which are a big part of the economy – it seems unlikely over the short run that the government would adopt such AI and fire all or nearly all of its workers, who are typically organized in powerful public unions.  In general, I think the political economy problem challenges around worker displacement are real – and an active concern around AI advances -- and would meaningfully slow things down.  But the first point above – the Baumol one – is sufficient on its own to greatly limit a growth surge.

3) Would 30% growth really involve 10X humanity’s tech progress in just 25 years?

That’s a fair point that to some extent early progress went in a Malthusian direction.  So my point may have been overstated when comparing 25 years of growth to all of human history.  But also my 25 years was arbitrary and I think it really matters how long you are imagining the 30% growth to go on for.  So long as 30% continues, then we are getting another 10x every 8.5 years.  Therefore, if one means sustained 30% growth we are really talking about something extraordinarily radical.  For example, if 30% continued for 100 years, than we would be 250 billion times the initial income per capita, which is not a mathematical singularity but is, I think, practically the same.  So perhaps one can have in mind that AI drives growth rate to 30% for a modest period of time but sustained 30% to me seems very extreme.  If what you mean is that there could be a limited period of 30% growth, like a decade, then things seem more plausible.

Author response

Thank you for these comments! 

Some thoughts in response:

1) Increasing average global consumption to the level of rich people

I take the point that many countries may be behind the technological frontier due to governance issues, and the AI-growth literature doesn’t address this. 

But one factor explaining cross-country income is differences in human capital, and advanced AI could address that issue fairly directly. 

It’s a complex topic, but I believe that some evidence suggests that performance on cognitive tests can explain a lot of the variation in income between countries. And this is borne out by the fact that when high-skilled immigrants/expats move to low income countries, they typically earn high salaries. Not as high as if they worked in a rich country, but not too far off. This suggests that if there was a sufficient quantity of high-skilled labour in low income countries, their GDP/capita would rise significantly.

This possibility does tie in with the AI-growth literature, where AI substituting for humans in goods and services production plays a comparably central role as AI substituting in R&D. (E.g. in this Nordaus paper AI only substitutes in goods and services, and tech progress is exogenous.) So I do think there’s a link between the AI-growth literature and the possibility that advanced AI rapidly increases GDP/per-capita in low-income countries.

2) Providing services that don’t require infrastructure changes

Bottlenecks

I agree that while AI is limited to services with no infrastructure requirements, bottlenecks will eventually slow growth. 

I think GDP could increase a bit more than you said though, before the bottleneck kicks in. My understanding is that a standard value for the substitutability parameter is 𝜌 = -½. (This corresponds to an elasticity of substitution of ⅔.) When I do the math, this implies that if AI made half of economic tasks infinitely productive, output would increase by 4X (rather than the 50% you mentioned). So there could be fast growth while this 4X increase in output is obtained before the Baumol effect kick in.

Once it does kick in, the question is: will our fast-growing population of AIs be able to pivot to improve a new area of the economy that people still value? Perhaps people have no further willingness to pay for services, but AIs can do R&D to invent radical life extension. How much would people pay to live for >1000 years? Or perhaps AIs can design robots and control them remotely to do physical construction tasks. Or invent other technologies we haven’t thought of yet, but would value highly.

I agree that the longer we imagine 30%/year growth of GWP/capita lasting, the less plausible it is. One distinction I want to draw out more in future is between growth of human GDP vs growth of AI GDP, that can include consumption of AIs and robots. I think the latter could plausibly grow at 30%/year for longer. This possibility is highlighted by the review of Anton Korinek.

Political economy

I agree this could delay things significantly. 

But if we grant that advanced AI and robotics would drive explosive growth absent this issue, I don’t think it would prevent explosive growth permanently. For then if just one region allowed advanced AI and robotics to automate the economy, its output would grow explosively. Then other regions would probably follow suite, or the initial region’s economy would eventually come to dominate GWP.