I think an AI bubble burst may genuinely slow down the underlying technology much more compared to something like the dot com bubble burst slowing down its underlying technology. I think this is probably an important dynamic to consider especially as an AI bubble burst start to look more plausible.
The basic argument here is that a bunch of AI progress is driven by exponentials in compute and algorithmic progress. These exponentials are in turn driven by exponentials in investment and people joining the field. A bubble burst would break the investment exponential and probably the people joining the field and would therefore slow down compute and algorithmic progress, which would slow down AI.
In some sense this is obvious (which is why your timelines would lengthen conditional on a bubble burst), but it's worth noting just how sensitive AI progress is to investment. For example, a bubble burst probably would not hurt the underlying technology to build iphone's as much. (one piece of evidence for this was apple holding >250 billion dollars of cash in 2017 and not having a good way of investing it).
I agree if investment in compute declines significantly and stays down, that will slow down AI capabilities a lot for that time.
It's possible, though, for the stocks to crash, but without investment falliing much as well - especially since the hyperscalers tend to have non-cyclical revenues. (This seemed more plausible when I first wrote the post, today a lot more compute investment seems debt funded, and hyperscaler capex was a lot lower relative to revenues.)
Another factor is that a lot of compute growth is already 'baked in' by past investment. If capex were cut back now, the effects might only kick in in a couple of years. And by that time further advances might have spurred more investment.
Oh I see, I didn't consider the "baked in" point. I changed my mind about burst dynamics because of this. (In particular I now think it's less likely that "lower investments <-> lower capabilities" will form a feedback loop that significantly decreases the rate of AI progress.)
Similar energy:
https://www.vox.com/future-perfect/367435/artificial-intelligence-openai-chatgpt-boom-bust-safety-superintelligence-google
See discussion of this post on the EA Forum:
https://forum.effectivealtruism.org/posts/LJzvCWnwnSSxrjXzi/ai-stocks-could-crash-and-that-could-have-implications-for
I think an AI bubble burst may genuinely slow down the underlying technology much more compared to something like the dot com bubble burst slowing down its underlying technology. I think this is probably an important dynamic to consider especially as an AI bubble burst start to look more plausible.
The basic argument here is that a bunch of AI progress is driven by exponentials in compute and algorithmic progress. These exponentials are in turn driven by exponentials in investment and people joining the field. A bubble burst would break the investment exponential and probably the people joining the field and would therefore slow down compute and algorithmic progress, which would slow down AI.
In some sense this is obvious (which is why your timelines would lengthen conditional on a bubble burst), but it's worth noting just how sensitive AI progress is to investment. For example, a bubble burst probably would not hurt the underlying technology to build iphone's as much. (one piece of evidence for this was apple holding >250 billion dollars of cash in 2017 and not having a good way of investing it).
I agree if investment in compute declines significantly and stays down, that will slow down AI capabilities a lot for that time.
It's possible, though, for the stocks to crash, but without investment falliing much as well - especially since the hyperscalers tend to have non-cyclical revenues. (This seemed more plausible when I first wrote the post, today a lot more compute investment seems debt funded, and hyperscaler capex was a lot lower relative to revenues.)
Another factor is that a lot of compute growth is already 'baked in' by past investment. If capex were cut back now, the effects might only kick in in a couple of years. And by that time further advances might have spurred more investment.
Oh I see, I didn't consider the "baked in" point. I changed my mind about burst dynamics because of this. (In particular I now think it's less likely that "lower investments <-> lower capabilities" will form a feedback loop that significantly decreases the rate of AI progress.)
Thanks for your response!