I think the AI as normal technology people are imagining diffusion etc bottlenecks to limit a lot of economic growth from AI. If you run the logical implications of their predictions forwards (which tbc I think is kinda insane), the direct implication of their worldview is that there’s a lot of economic growth left from current generation models, so Anthropic’s current 50B ARR is an extreme underestimate of future value (though much of it might not be captured by the labs, similar to how the impact of electricity is mostly not captured by the utility companies).
I on the other hand think the main economic gains are much more limited by capabilities rather than diffusion, at least as of today’s models (I’m even more confident about the models of a year back, naturally[1]). Tbc 10x or even 30x the diffusion of today’s models is plausible to me, tho probably still smaller than the internet’s sustained effect.
Reasons that I might be underestimating the potential economic impact of today’s models, that mostly won’t apply to those of a year ago:
The labs have access to private models that the public have not seen or evaluated
Publicly available models like Fable 5 and 5.6 Sol might be massively nerfed by Anthropic and OpenAI, there’s suggestive evidence that the nerfs hit hardest on the most economically valuable tasks. Presumably they’d let up in the future, even if we otherwise freeze capabilities.
The latest models have only been released for a few weeks, maybe there are industries where much of the job can be automated but we haven’t figured out the scaffolding or institutional processes to do so.
Hmm, I think I think AI is pretty diffusion bottlenecked? I think if Claude Code was properly leveraged everywhere the world would look pretty different. I’m not sure how it’d translate into GDP (I don’t have great intuitions about how real world benefit translates into GDP)
Yeah tbc think of my views here as “strong opinions weakly held,” I could definitely be wrong!
In my opinion, the prior that things that look like a big deal barely shows up if at all in the aggregate growth statistics has generally held up. Obviously at some point this is no longer true (if we have a billion remote-only superintelligences running around we’d likely see either massive economic growth or death).
I’m not sure where the crossover points are (and obviously in some sense the underlying distribution is kinda smooth), an intuitive model to me is when AIs cross over from being better modeled as “high-skilled immigration” than as “tools.” [1]Imo they are still very tool-shaped for most remote-only jobs that aren’t programming.
this isn’t necessarily true because of jaggedness. Eg if we get sustained across-the-board superhuman R&D but we freeze further AI R&D specifically, the non-AI scientific advances will probably pay off in GDP...eventually.
and obviously in some sense the underlying distribution is kinda smooth
I actually disagree that the output distribution has to be smooth, even assuming the inputs are smoothly increasing.
The reason for this is phase changes, and an example is when you heat up water from say 20 degrees Celsius to 100 degrees Celsius in increments of say 0.01 degrees per minute, and while it is true that increasing the heat by 0.01 Celsius per minute matters in and of itself, at 100 degrees Celsius, water heating behaves discontinuously and boiling starts to happen.
And the impact of AI is also likely to be a phase change, because of long-tail effects meaning that 99% automation of something is vastly less valuable than 100% automation of something.
More generally, even assuming the inputs are continously increasing, you cannot assume that downstream consequences/output will continuously increase.
This likely has important implications for how to think about AI.
I definitely buy this argument within industries, but not across the whole economy. Mathematically, you can have a kinda smooth distribution even if the underlying components are very jagged (cf Central Limit Theorem). Mechanically, this is because you should expect the tipping point of different jobs and industries to be at different points in the capabilities growth curve, so even if (eg) customer support agents go from losing 10% of jobs to losing 99% of jobs over night, you still shouldn’t expect the 10%-99% job loss transition to be at approximately the same time as other jobs.
There are a few ways this assumption breaks. The most obvious break is automated AI R&D/RSI. It’s easy to see how a smooth growth in AI capabilities there can potentially translate into a sharp increase in the pace of AI capabilities overall, which leads to ripple effects across the economy.
If I squint, I can also maybe buy a story here with superhuman non-AI R&D, where smooth increases in R&D ability of AIs nonetheless translates to non-smooth increases in scientific advancements and thus total factor productivity across the whole economy.
But I don’t think there are many things quite like that.
What do you imagine the differences between “AI is normal tech” and this (Plan S) hypothetical?
I think the AI as normal technology people are imagining diffusion etc bottlenecks to limit a lot of economic growth from AI. If you run the logical implications of their predictions forwards (which tbc I think is kinda insane), the direct implication of their worldview is that there’s a lot of economic growth left from current generation models, so Anthropic’s current 50B ARR is an extreme underestimate of future value (though much of it might not be captured by the labs, similar to how the impact of electricity is mostly not captured by the utility companies).
I on the other hand think the main economic gains are much more limited by capabilities rather than diffusion, at least as of today’s models (I’m even more confident about the models of a year back, naturally[1]). Tbc 10x or even 30x the diffusion of today’s models is plausible to me, tho probably still smaller than the internet’s sustained effect.
Reasons that I might be underestimating the potential economic impact of today’s models, that mostly won’t apply to those of a year ago:
The labs have access to private models that the public have not seen or evaluated
Publicly available models like Fable 5 and 5.6 Sol might be massively nerfed by Anthropic and OpenAI, there’s suggestive evidence that the nerfs hit hardest on the most economically valuable tasks. Presumably they’d let up in the future, even if we otherwise freeze capabilities.
The latest models have only been released for a few weeks, maybe there are industries where much of the job can be automated but we haven’t figured out the scaffolding or institutional processes to do so.
Hmm, I think I think AI is pretty diffusion bottlenecked? I think if Claude Code was properly leveraged everywhere the world would look pretty different. I’m not sure how it’d translate into GDP (I don’t have great intuitions about how real world benefit translates into GDP)
Yeah tbc think of my views here as “strong opinions weakly held,” I could definitely be wrong!
In my opinion, the prior that things that look like a big deal barely shows up if at all in the aggregate growth statistics has generally held up. Obviously at some point this is no longer true (if we have a billion remote-only superintelligences running around we’d likely see either massive economic growth or death).
I’m not sure where the crossover points are (and obviously in some sense the underlying distribution is kinda smooth), an intuitive model to me is when AIs cross over from being better modeled as “high-skilled immigration” than as “tools.” [1]Imo they are still very tool-shaped for most remote-only jobs that aren’t programming.
this isn’t necessarily true because of jaggedness. Eg if we get sustained across-the-board superhuman R&D but we freeze further AI R&D specifically, the non-AI scientific advances will probably pay off in GDP...eventually.
I actually disagree that the output distribution has to be smooth, even assuming the inputs are smoothly increasing.
The reason for this is phase changes, and an example is when you heat up water from say 20 degrees Celsius to 100 degrees Celsius in increments of say 0.01 degrees per minute, and while it is true that increasing the heat by 0.01 Celsius per minute matters in and of itself, at 100 degrees Celsius, water heating behaves discontinuously and boiling starts to happen.
And the impact of AI is also likely to be a phase change, because of long-tail effects meaning that 99% automation of something is vastly less valuable than 100% automation of something.
More generally, even assuming the inputs are continously increasing, you cannot assume that downstream consequences/output will continuously increase.
This likely has important implications for how to think about AI.
I definitely buy this argument within industries, but not across the whole economy. Mathematically, you can have a kinda smooth distribution even if the underlying components are very jagged (cf Central Limit Theorem). Mechanically, this is because you should expect the tipping point of different jobs and industries to be at different points in the capabilities growth curve, so even if (eg) customer support agents go from losing 10% of jobs to losing 99% of jobs over night, you still shouldn’t expect the 10%-99% job loss transition to be at approximately the same time as other jobs.
There are a few ways this assumption breaks. The most obvious break is automated AI R&D/RSI. It’s easy to see how a smooth growth in AI capabilities there can potentially translate into a sharp increase in the pace of AI capabilities overall, which leads to ripple effects across the economy.
If I squint, I can also maybe buy a story here with superhuman non-AI R&D, where smooth increases in R&D ability of AIs nonetheless translates to non-smooth increases in scientific advancements and thus total factor productivity across the whole economy.
But I don’t think there are many things quite like that.