Have you considered the brain as a possible example, though evolved instead of engineered? Gradual increase in volume across the last six million years, but with technological stagnation for long stretches (for example, Homo erectus’ stone tools barely changed for well over a million years). Then some time in the last hundred thousand years or so, we got accelerating technological progress despite steady or decreasing brain size. Is it better algorithms (via genetics)? Better data (cultural)? Feedback loops from these leading to larger population? Unclear, but the last, rate-limiting step wasn’t larger or faster brains.
And I think the idea of the rate-limiting step, rather than overhang, is exactly the key here. In your post you talk about S-curves, but why do you think s-curves happen at all? My understanding is that it’s because there’s some hard problem that takes multiple steps to solve, and when the last step falls (or a solution is in sight), it’s finally worthwhile to toss increasing amounts of investment to actually realize and implement the solution. Then, we see diminishing returns as we approach the next wall that requires a different, harder set of solutions. In time each one is overcome, often for unrelated reasons, and we have overhang in those areas until the last, rate-limiting step falls and we get a new s-curve.
Consider the steam engine, first given a toy demo about two and a half millennia ago by Archytas. Totally impractical, ignored, forgotten. Then we developed resource and compute overhang (more material wealth, more minds in the population, greater fraction of minds focused on things other than survival, better communications from the printing press and shared Latin language among scholars). We developed better algorithms (algebra, calculus, metallurgical recipes for iron and steel, the scientific method, physics). Then, and only then, did James Watt and his contemporaries overcome the last hurdle to make it practical enough to kickstart the s-curve of the industrial revolution that we’re still riding to this day.
Your post reads, to me, as saying, “Better algorithms in AI may add new s-curves, but won’t jump all the way to infinity, they’ll level off after a while.” Which is a reasonable observation and true enough for the effects of each advance. But at some level that’s almost the same as saying, “There is compute overhang,” because both mean, “Available compute is not currently the rate-limiting step in AI development.”
Now, you can go on from there to debate where or when a given s-curve will level off. You can debate whether the fact that each AI s-curve increases available intelligence for problem solving makes AI’s expected trajectory different than other technologies. Faster planes don’t invent their successors, and we can only produce so many aeronautical engineers to address the successively-harder problems, but if AI hits a point of “We can make a new mind as good as the best human inventor ever at near-zero cost and point it at the next AI problem AND the next hardware problem AND the next energy problem AND etc., all at 1000x the speed of neurons” it’s not quite the same thing. Regardless, you don’t need to address this to discuss the idea of compute overhang.
why do you think s-curves happen at all? My understanding is that it’s because there’s some hard problem that takes multiple steps to solve, and when the last step falls (or a solution is in sight), it’s finally worthwhile to toss increasing amounts of investment to actually realize and implement the solution.
I think S-curves are not, in general, caused by increases in investment. They’re mainly the result of how the performance of a technology changes in response to changes in the design/methods/principles behind it. For example, with particle accelerators, switching from Van der Graaff generators to cyclotrons might give you a few orders of magnitude once the new method is mature. But it takes several iterations to actually squeeze out all the benefits of the improved approach, and the first few and last few iterations give less of an improvement than the ones in the middle.
This isn’t to say that the marginal return on investment doesn’t factor in. Once you’ve worked out some of the kinks with the first couple cyclotrons, it makes more sense to invest in a larger one. This probably makes S-curves more S-like (or more step like). But I think you’ll get them even with steadily increasing investment that’s independent of the marginal return.
So Swanson’s law is an observation for solar panel cost, where each increase in production volume results in lower cost and it is driving an S curve.
It seems like there would be 2 separate effects running here: the S curve like technology improvement to photovoltaic cells, and as production volume increases, greater and greater automation is justified.
Note also you would expect that for silicon PV we are well into the diminishing returns area of the curve, yet costs continue to decline.
Eyeballing the plot..well....it actually looks kinda flat. Like increased solar cell volume is leading to scaling r&d investment and leading to almost linear efficiency improvements with time.
I would argue that AI inference hardware production would be an example of something that should benefit from the learning effect and lead to a similar S curve adoption of ai, totally decoupled from the r&d effort for the model capabilities.
Investment scaling with volume looks like an important effect.
You’re right, I was switching between performance s-curves and market size s-curves in my thinking without realizing it. I do think the general point holds that there’s a pattern of hit minimum viability --> get some adoption --> adoption accelerates learning, iteration, and innovation --> performance and cost improve --> viability increases --> repeat until you hit a wall or saturate the market.
Your post reads, to me, as saying, “Better algorithms in AI may add new s-curves, but won’t jump all the way to infinity, they’ll level off after a while.”
The post is mostly not about either performance s-curves or market size s-curves. It’s about regulation imposing a pause on AI development, and whether this would cause catch-up growth if the pause is ended.
Stacked s-curves can look like a pause + catch-up growth, but they are a different mechanism.
I think I was thrown by the number of times I’ve read things about us already being in hardware overhang, which a pause would make larger but not necessarily different-in-kind. I don’t know if (or realistically, how much) larger overhangs lead to faster change when the obstacle holding us back goes away. But I would say in this proposed scenario that the underlying dynamics of how growth happens don’t seem like they should depend on whether the overhang comes from regulatory sources specifically.
The reason I got into the whole s-curve thing is largely because I was trying to say that overhangs are not some novel thing, but rather a part of the development path of technology and industry generally. In some sense, every technology we know is possible is in some form(s) of overhang, from the moment we meet any of the prerequisites for developing it, right up until we develop and implement it. We just don’t bother saying things like “Flying cars are in aluminum overhang.”
Have you considered the brain as a possible example, though evolved instead of engineered? Gradual increase in volume across the last six million years, but with technological stagnation for long stretches (for example, Homo erectus’ stone tools barely changed for well over a million years). Then some time in the last hundred thousand years or so, we got accelerating technological progress despite steady or decreasing brain size. Is it better algorithms (via genetics)? Better data (cultural)? Feedback loops from these leading to larger population? Unclear, but the last, rate-limiting step wasn’t larger or faster brains.
And I think the idea of the rate-limiting step, rather than overhang, is exactly the key here. In your post you talk about S-curves, but why do you think s-curves happen at all? My understanding is that it’s because there’s some hard problem that takes multiple steps to solve, and when the last step falls (or a solution is in sight), it’s finally worthwhile to toss increasing amounts of investment to actually realize and implement the solution. Then, we see diminishing returns as we approach the next wall that requires a different, harder set of solutions. In time each one is overcome, often for unrelated reasons, and we have overhang in those areas until the last, rate-limiting step falls and we get a new s-curve.
Consider the steam engine, first given a toy demo about two and a half millennia ago by Archytas. Totally impractical, ignored, forgotten. Then we developed resource and compute overhang (more material wealth, more minds in the population, greater fraction of minds focused on things other than survival, better communications from the printing press and shared Latin language among scholars). We developed better algorithms (algebra, calculus, metallurgical recipes for iron and steel, the scientific method, physics). Then, and only then, did James Watt and his contemporaries overcome the last hurdle to make it practical enough to kickstart the s-curve of the industrial revolution that we’re still riding to this day.
Your post reads, to me, as saying, “Better algorithms in AI may add new s-curves, but won’t jump all the way to infinity, they’ll level off after a while.” Which is a reasonable observation and true enough for the effects of each advance. But at some level that’s almost the same as saying, “There is compute overhang,” because both mean, “Available compute is not currently the rate-limiting step in AI development.”
Now, you can go on from there to debate where or when a given s-curve will level off. You can debate whether the fact that each AI s-curve increases available intelligence for problem solving makes AI’s expected trajectory different than other technologies. Faster planes don’t invent their successors, and we can only produce so many aeronautical engineers to address the successively-harder problems, but if AI hits a point of “We can make a new mind as good as the best human inventor ever at near-zero cost and point it at the next AI problem AND the next hardware problem AND the next energy problem AND etc., all at 1000x the speed of neurons” it’s not quite the same thing. Regardless, you don’t need to address this to discuss the idea of compute overhang.
I think S-curves are not, in general, caused by increases in investment. They’re mainly the result of how the performance of a technology changes in response to changes in the design/methods/principles behind it. For example, with particle accelerators, switching from Van der Graaff generators to cyclotrons might give you a few orders of magnitude once the new method is mature. But it takes several iterations to actually squeeze out all the benefits of the improved approach, and the first few and last few iterations give less of an improvement than the ones in the middle.
This isn’t to say that the marginal return on investment doesn’t factor in. Once you’ve worked out some of the kinks with the first couple cyclotrons, it makes more sense to invest in a larger one. This probably makes S-curves more S-like (or more step like). But I think you’ll get them even with steadily increasing investment that’s independent of the marginal return.
https://en.m.wikipedia.org/wiki/Swanson’s_law
So Swanson’s law is an observation for solar panel cost, where each increase in production volume results in lower cost and it is driving an S curve.
It seems like there would be 2 separate effects running here: the S curve like technology improvement to photovoltaic cells, and as production volume increases, greater and greater automation is justified.
Note also you would expect that for silicon PV we are well into the diminishing returns area of the curve, yet costs continue to decline.
https://en.m.wikipedia.org/wiki/Solar-cell_efficiency
Eyeballing the plot..well....it actually looks kinda flat. Like increased solar cell volume is leading to scaling r&d investment and leading to almost linear efficiency improvements with time.
I would argue that AI inference hardware production would be an example of something that should benefit from the learning effect and lead to a similar S curve adoption of ai, totally decoupled from the r&d effort for the model capabilities.
Investment scaling with volume looks like an important effect.
You’re right, I was switching between performance s-curves and market size s-curves in my thinking without realizing it. I do think the general point holds that there’s a pattern of hit minimum viability --> get some adoption --> adoption accelerates learning, iteration, and innovation --> performance and cost improve --> viability increases --> repeat until you hit a wall or saturate the market.
The post is mostly not about either performance s-curves or market size s-curves. It’s about regulation imposing a pause on AI development, and whether this would cause catch-up growth if the pause is ended.
Stacked s-curves can look like a pause + catch-up growth, but they are a different mechanism.
True, that was poor framing on my part.
I think I was thrown by the number of times I’ve read things about us already being in hardware overhang, which a pause would make larger but not necessarily different-in-kind. I don’t know if (or realistically, how much) larger overhangs lead to faster change when the obstacle holding us back goes away. But I would say in this proposed scenario that the underlying dynamics of how growth happens don’t seem like they should depend on whether the overhang comes from regulatory sources specifically.
The reason I got into the whole s-curve thing is largely because I was trying to say that overhangs are not some novel thing, but rather a part of the development path of technology and industry generally. In some sense, every technology we know is possible is in some form(s) of overhang, from the moment we meet any of the prerequisites for developing it, right up until we develop and implement it. We just don’t bother saying things like “Flying cars are in aluminum overhang.”