Non-obvious considerations in superhuman AI diffusion/adoption rates across various institutions/society
People sometimes want to use the track record of technology adoption and diffusion to benchmark/estimate AI diffusion rates. Below are some considerations that make this trickier than other “technologies.”
Endogeneous dynamics can be a bit of a wild-card: Unlike most past technologies (eg electricity, computers), AI is a technology that would likely do a lot to aid in its own adoption. So unlike dynamics with past technological adoptions, where you primarily look at the relative benefit of the technology against the adoption costs and various social factors, you also need to look at how much the technology itself can aid in reducing the relevant adoption costs and social factors. This is in contrast to historical technologies, where the tech itself can create stronger reason to adopt the technology, or reduce practical and economic costs, or make adoption more socially salient/permissible due to network effects, but not suggest/implement (superhumanly good!) plans to make its own adoption easier.
This is similar perhaps to railways or the internet? (TODO: investigate)
Gated access: Both the AI companies making the technology and the national security institutional environments might limit broad adoption of the latest models.
Trust: Nation-state actors and competing organizations may have justified(!) reasons why they don’t want to install scary/agentic software of a presumed adversary/supply-chain risk in their servers, and/or integrate it with highly important and sensitive contexts. The Anthropic-DoD dispute is an early hint of this.
[I then did a first-pass guess at relative adoption speeds at various institutions. My guess is that this isn’t very interesting to other people].
tl;dr I basically expect superintelligence diffusion/adoption in important institutions to be faster than pretty much every major historical technology, in the absence of strong policy breaks. So there aren’t relevant entities that feels like they’d be “slow,” in absolute terms.
Overall takes:
I feel less intellectually excited about this idea than some others, at least in the absence of a foil/someone to fight and argue against
I feel like my handles on this topic feel somewhat flimsy/slippery
“Diffusion” in some key ways feels like the wrong frame since a lot of it is more questions about structured access and trust, rather than diffusion per se.
I’d be somewhat excited to work more on this if there’s a natural “customer” for this work, I have a better angle for improving people’s thinking on it, but by default won’t want to hammer further on this by myself
Further questions:
The endogeneous dynamics here seems pretty central to my model. Can we model this better…
…theoretically? Like write down a toy mathematical model and explore the relevant implications
…empirically? Like try to come up with historical examples of similar technologies (Railways? Printing press? internet?) and see if there are things we can learn from them in terms of adoption speeds
Think through/model diffusion dynamics in China in particular in more detail.
Most of this was generated in a horizon-scanning sprint at FT where I looked into a half-dozen areas to work on. At the time I wrote it I didn’t feel super-excited about further work here personally (and I’m still not super excited), but since then I’ve been encountering more “technology diffusion” arguments that seem obviously wrong to me so in the spirit of “somebody’s wrong on the internet”, decided to share it here.
Non-obvious considerations in superhuman AI diffusion/adoption rates across various institutions/society
People sometimes want to use the track record of technology adoption and diffusion to benchmark/estimate AI diffusion rates. Below are some considerations that make this trickier than other “technologies.”
Endogeneous dynamics can be a bit of a wild-card: Unlike most past technologies (eg electricity, computers), AI is a technology that would likely do a lot to aid in its own adoption. So unlike dynamics with past technological adoptions, where you primarily look at the relative benefit of the technology against the adoption costs and various social factors, you also need to look at how much the technology itself can aid in reducing the relevant adoption costs and social factors. This is in contrast to historical technologies, where the tech itself can create stronger reason to adopt the technology, or reduce practical and economic costs, or make adoption more socially salient/permissible due to network effects, but not suggest/implement (superhumanly good!) plans to make its own adoption easier.
This is similar perhaps to railways or the internet? (TODO: investigate)
Gated access: Both the AI companies making the technology and the national security institutional environments might limit broad adoption of the latest models.
Trust: Nation-state actors and competing organizations may have justified(!) reasons why they don’t want to install scary/agentic software of a presumed adversary/supply-chain risk in their servers, and/or integrate it with highly important and sensitive contexts. The Anthropic-DoD dispute is an early hint of this.
[I then did a first-pass guess at relative adoption speeds at various institutions. My guess is that this isn’t very interesting to other people].
tl;dr I basically expect superintelligence diffusion/adoption in important institutions to be faster than pretty much every major historical technology, in the absence of strong policy breaks. So there aren’t relevant entities that feels like they’d be “slow,” in absolute terms.
Overall takes:
I feel less intellectually excited about this idea than some others, at least in the absence of a foil/someone to fight and argue against
I feel like my handles on this topic feel somewhat flimsy/slippery
“Diffusion” in some key ways feels like the wrong frame since a lot of it is more questions about structured access and trust, rather than diffusion per se.
I’d be somewhat excited to work more on this if there’s a natural “customer” for this work, I have a better angle for improving people’s thinking on it, but by default won’t want to hammer further on this by myself
Further questions:
The endogeneous dynamics here seems pretty central to my model. Can we model this better…
…theoretically? Like write down a toy mathematical model and explore the relevant implications
…empirically? Like try to come up with historical examples of similar technologies (Railways? Printing press? internet?) and see if there are things we can learn from them in terms of adoption speeds
Think through/model diffusion dynamics in China in particular in more detail.
Most of this was generated in a horizon-scanning sprint at FT where I looked into a half-dozen areas to work on. At the time I wrote it I didn’t feel super-excited about further work here personally (and I’m still not super excited), but since then I’ve been encountering more “technology diffusion” arguments that seem obviously wrong to me so in the spirit of “somebody’s wrong on the internet”, decided to share it here.