“If you take the Epoch AI scaling curve seriously, which I think we should, these models are going to get so expensive that it gets to a point where there's only kind of one or two states that can continue to maintain a sovereign frontier. … What I think is a really obvious idea is that you will not soon be able to buy the best models, that labs will become something much closer to the defense industrial complex, that they will have models that are available to nation states for national security purposes. You will not be using them.”
sputnikintelligence.com
The hosts asked for the short version of his new 140-page essay, 'Commanding Heights'. Manidis says the debate should move past bottlenecks to who is advantaged, and introduces one of the essay's central ideas, the 'sovereign frontier'.
…to say about the coming scaling laws that are going to happen? What is the finance take on everything that's going to happen? And we set out to write this manuscript. Joe Roberts and I, Farrell Gregory, sat down and spent the last year drafting this thing and published this week. Very exciting.
And what is the TLDR on the knock-on effects? Because everyone knows the bottleneck constraints. The situational awareness thesis has been digested. You're saying AI is real, but what are some of the knock-on effects?
Yeah, I mean, I think we need to move past bottlenecks and start to think about who is uniquely advantaged as a result of this technology, right? It is staggering to me, permanently. It's staggering to me that you can use instinct every day at work and not process that millions of people's jobs are within the capabilities cone of the models we have today. And if you take the epoch AI scaling curve seriously, which I think we should, these models are going to get so expensive that it gets to a point where there's only kind of one or two states that can continue to maintain a sovereign frontier, meaning models that are kind of world-class and can continue to be trained. So if we have a world where the best models are, you know, held inside of American skiffs, where the national security community is given, you know, Fable 5.5, right? Maybe some unreleased TPPN episodes, really kind of high-security stuff, and the Chinese are able to maintain their own sovereign frontier in the East, then what does this mean for the populations that are within the diffusion net? And we introduced a couple new ideas as part of this PDF. The sovereign frontier is the big one. What I think is a really obvious idea that you will not soon be able to buy the best models, that labs will become something much closer to the defense industrial…
So, how do you take that thesis into some of the more low-risk, mundane AI-enabled technologies? I'm thinking of just like website builders. Like, that's something, like, are you going to have to go, does Lovable have to go to the government for their inference, or can they just use current models?…
cases are like pretty saturated. They
are saturated, but do they continue? Okay.