Does anyone have good resources on where value might accrue in worlds where AI can do most (or all) cognitive labour (this Q also applies to physical labour, but is harder)?
Context is I’m interested in learning more about the question of whether “all/most value accrues to AI companies”. My current guess is that value probably just accrues to whatever remains hard to replicate, not necessarily to the model layer.
If you follow this line of reasoning, here are some candidates for where value might end up:
Distribution and customer trust (I’d assume there’s a decent chance users stick with institutions they already trust, especially in high-stakes domains?)
Proprietary workflows/data (i.e not just static datasets, but also embedded org context, feedback loops, tacit knowledge etc)
Regulatory permissions (e.g. medicine, law, finance, government procurement stuff)
Physical-world access (robots, factories, human relationships, and so on)
Compute/energy/inference (but TBC only insofar as these remain structurally scarce, not just temporarily supply-constrained).
Frontier model capability (only if a small number of AI companies maintain a persistent lead, or if integration costs are really high for some reason—which seems unikely due to strong commercial incentives for them to be easy)
(to be clear I think a world where value concentrates in frontier AI companies would be a worse world compared to one where it is more diffuse, due to power concentration concerns + generally more dynamism and resilience in that world).
Does anyone have good resources on where value might accrue in worlds where AI can do most (or all) cognitive labour (this Q also applies to physical labour, but is harder)?
Context is I’m interested in learning more about the question of whether “all/most value accrues to AI companies”. My current guess is that value probably just accrues to whatever remains hard to replicate, not necessarily to the model layer.
If you follow this line of reasoning, here are some candidates for where value might end up:
Distribution and customer trust (I’d assume there’s a decent chance users stick with institutions they already trust, especially in high-stakes domains?)
Proprietary workflows/data (i.e not just static datasets, but also embedded org context, feedback loops, tacit knowledge etc)
Regulatory permissions (e.g. medicine, law, finance, government procurement stuff)
Physical-world access (robots, factories, human relationships, and so on)
Compute/energy/inference (but TBC only insofar as these remain structurally scarce, not just temporarily supply-constrained).
Frontier model capability (only if a small number of AI companies maintain a persistent lead, or if integration costs are really high for some reason—which seems unikely due to strong commercial incentives for them to be easy)
(to be clear I think a world where value concentrates in frontier AI companies would be a worse world compared to one where it is more diffuse, due to power concentration concerns + generally more dynamism and resilience in that world).