Excellent distinction between capability and deployment. We would add that even a genuine frontier slowdown may change the composition, not the size, of capex: fewer giant training runs, but more inference, networking, memory and power infrastructure. Investors may therefore be asking the wrong binary question. The key is not whether AI spending stops, but where each marginal dollar migrates.
I agree with you that "we are nowhere near peak compute". It feels though that you downplay the risks associated with RSI a bit too much.
More importantly the risk/reward calculation points towards allocating resources to AI adoption, inference rather than pushing frontier models (which are already scary capable) further. I believe we can easily capture 90% of potential economic benefit and a large chunk of discovery benefits over the next 5 years even with 0 progress in frontier.
China argument is fair, but they have 5-8x less compute - which would make them receptive to any reasonable deal proposal. Also an argument can be made that US is doing much more to turn AI effort into a race.
Good thoughts. Thank you for this. It is entirely likely I don’t understand the import of RSI.
“we can easily capture 90% of potential economic benefit and a large chunk of discovery benefits over the next 5 years even with 0 progress in frontier.”
I am pretty sure you know way more about RSI than me - my concern is entirely based on 2 assumptions
- there is no low ceiling (subhuman or near human) for model performance
- it is very rare that a less intelligent entity controls the more intelligent one - other than a child and a mother (where child achieved perfect alignment :)) there aren't a ton of examples.
Btw wrt China in AI race matter, found a couple of podcasts that give a good overview:
I agree, because it is just a matter of time, until Mike Jhonson is replaced. After that there will be a pause first, regulatory framework next. before any further progress. This was already expressed by Democrats and few Republicans.
Excellent distinction between capability and deployment. We would add that even a genuine frontier slowdown may change the composition, not the size, of capex: fewer giant training runs, but more inference, networking, memory and power infrastructure. Investors may therefore be asking the wrong binary question. The key is not whether AI spending stops, but where each marginal dollar migrates.
It’s the proverbial capex mix shift that we should be focusing on. Good point.
I agree with you that "we are nowhere near peak compute". It feels though that you downplay the risks associated with RSI a bit too much.
More importantly the risk/reward calculation points towards allocating resources to AI adoption, inference rather than pushing frontier models (which are already scary capable) further. I believe we can easily capture 90% of potential economic benefit and a large chunk of discovery benefits over the next 5 years even with 0 progress in frontier.
China argument is fair, but they have 5-8x less compute - which would make them receptive to any reasonable deal proposal. Also an argument can be made that US is doing much more to turn AI effort into a race.
I liked this podcast on the topic:
https://podcasts.apple.com/gb/podcast/the-but-china-dilemma-driving-the-a-i-race/id1548604447?i=1000789743497
Good thoughts. Thank you for this. It is entirely likely I don’t understand the import of RSI.
“we can easily capture 90% of potential economic benefit and a large chunk of discovery benefits over the next 5 years even with 0 progress in frontier.”
I 100% agree with this though.
I am pretty sure you know way more about RSI than me - my concern is entirely based on 2 assumptions
- there is no low ceiling (subhuman or near human) for model performance
- it is very rare that a less intelligent entity controls the more intelligent one - other than a child and a mother (where child achieved perfect alignment :)) there aren't a ton of examples.
Btw wrt China in AI race matter, found a couple of podcasts that give a good overview:
https://x.com/ALikhodedov/status/2100941675380043930
(there is also a summary of key points made by Fable).
This reads like someone who is heavily invested in Ai companies trying to save their bags.
This last post reads like someone who would like to stop Trump.
I agree, because it is just a matter of time, until Mike Jhonson is replaced. After that there will be a pause first, regulatory framework next. before any further progress. This was already expressed by Democrats and few Republicans.
One reason they all three want to pace AI could be that they really feared that AI could take them. A genuine concern after earning so much wealth.