My cached view was that doubling cutting-edge semiconductor production was plausibly going to be the main bottleneck to doubling physical capital, and that doing this doubling annually with just current tech would potentially be tricky. Do you disagree?
My all considered view is that this is unlikely to prevent a pretty fast industrial explosion because:
You can improve AI efficiency to make a limited chip supply go further, and this could suffice for a massive quantity of robotics and cognitive work.
Probably it’s doable to double semiconductor production quickly enough even with current tech, given full automation and a huge fraction of output focused on this.
We won’t just have current tech.
But, I still think analyzing this possible bottleneck would be pretty informative, and it seems that many people (e.g. maybe Dario?) think slow chip production scaling will be the bottleneck to a rapid industrial explosion.
It’s worth distinguishing “how fast do things grow given the stuff around today?” vs “how fast could things grow with today’s technology but the stuff optimized to grow fast?” This post is about the latter, and there I don’t really see how semiconductor production would be the bottleneck; it’s just not that large a fraction of the physical capital required even with pessimistic assumptions about cost, and you can compensate for longer lead times by increasing sector production.
In the next post I look at the transition to rapid growth, planning to post sometime in the next week or so. Given my results there I’m skeptical of a semiconductor bottleneck, but I haven’t investigated in detail so it’s possible I’m missing them because they are hidden within a larger sector (this should be easy to fix though, I’ll add to my to do list).
Note I am assuming that the R&D required to create AI and robotic workers has been done and costs are comparable to the human brain wrt compute requirements. Obviously if compute costs are 1000x higher because AI systems aren’t very efficient that could be a strong drag on economic growth, but I expect that won’t last very long as inference costs continue to fall and the R&D needed for the transition gets done.
I don’t really see how semiconductor production would be the bottleneck
I mean, it’s really just going to depend on whether massively increased spending can compensate for doing a insanely large scale up of a very complex sector with long lead times? I don’t understand your intuitions for why it’s clearly not going to be the bottleneck, but maybe I’m missing something. Like, maybe it’s the bottleneck after the 4th doubling rather than as of the first, but it seems super plausible it’s the bottleneck to me.
Notably, it’s pretty plausible that semi will soon be a serious bottleneck in practice in AI in the short term. (Rather than being capex or energy.)
Are you distinguishing leading-node chips from chips in general? It might be hard to rapidly scale up EUV lithography in particular, but very doable to scale up chips at larger nodes.
My cached view was that doubling cutting-edge semiconductor production was plausibly going to be the main bottleneck to doubling physical capital, and that doing this doubling annually with just current tech would potentially be tricky. Do you disagree?
My all considered view is that this is unlikely to prevent a pretty fast industrial explosion because:
You can improve AI efficiency to make a limited chip supply go further, and this could suffice for a massive quantity of robotics and cognitive work.
Probably it’s doable to double semiconductor production quickly enough even with current tech, given full automation and a huge fraction of output focused on this.
We won’t just have current tech.
But, I still think analyzing this possible bottleneck would be pretty informative, and it seems that many people (e.g. maybe Dario?) think slow chip production scaling will be the bottleneck to a rapid industrial explosion.
It’s worth distinguishing “how fast do things grow given the stuff around today?” vs “how fast could things grow with today’s technology but the stuff optimized to grow fast?” This post is about the latter, and there I don’t really see how semiconductor production would be the bottleneck; it’s just not that large a fraction of the physical capital required even with pessimistic assumptions about cost, and you can compensate for longer lead times by increasing sector production.
In the next post I look at the transition to rapid growth, planning to post sometime in the next week or so. Given my results there I’m skeptical of a semiconductor bottleneck, but I haven’t investigated in detail so it’s possible I’m missing them because they are hidden within a larger sector (this should be easy to fix though, I’ll add to my to do list).
Note I am assuming that the R&D required to create AI and robotic workers has been done and costs are comparable to the human brain wrt compute requirements. Obviously if compute costs are 1000x higher because AI systems aren’t very efficient that could be a strong drag on economic growth, but I expect that won’t last very long as inference costs continue to fall and the R&D needed for the transition gets done.
I mean, it’s really just going to depend on whether massively increased spending can compensate for doing a insanely large scale up of a very complex sector with long lead times? I don’t understand your intuitions for why it’s clearly not going to be the bottleneck, but maybe I’m missing something. Like, maybe it’s the bottleneck after the 4th doubling rather than as of the first, but it seems super plausible it’s the bottleneck to me.
Notably, it’s pretty plausible that semi will soon be a serious bottleneck in practice in AI in the short term. (Rather than being capex or energy.)
Are you distinguishing leading-node chips from chips in general? It might be hard to rapidly scale up EUV lithography in particular, but very doable to scale up chips at larger nodes.