grand theft eigenvalue 🔆

Value plus discounted

🙋 I'll stay on haiku 4

consistent with "it's natsec not collusion" stories

virtual compute market solves* this

*rations by price

US rural to urban migration accelerates beyond counterfactual pre-2025 trend, visible in IRS tax stats or equivalent no later than 2029

anyway there's a version of the story where coordinating to restrict public model quality isn't collusion, it's a reasonable response to national security concerns

i obviously have no idea if this is what's going on but my folk story of the narrative here is ~"people looked at this kind of graph long enough to convince themselves reasoning models (i.e., variable cost constant output) were an economic improvement on vanilla (constant cost variable output)"

good article but the issue isn't tokens getting more expensive, it's a rebound effect big enough to call a backfire

tokens are getting more expensive "language models will get cheaper by 10x" will not save ai subscriptions from the short squeeze

i think in some ways this is broadly correct (SOTA models are remarkably inflexible in terms of $/mtok) but I also think it misses the forest for the trees. inference drives today’s dollars, training drives tomorrows.

ethanding.substack.com/p/ai-subscri...

1909/11000

i don't think it'll be on pricing. doesn't feel right. maybe they divide the market or some segments but that doesn't feel stable either

i am actually very curious to see what the collusion looks like here. i suspect it'll be on models/training data/workers if the industry is booming and on GPUs if it's constrained or lackluster

sure, but that's why very few of the big companies are selling models as premium products anymore

no smart businessperson making big capital outlays wants to compete on price bc even if you win you lose. far better to compete on some quality margin with lock-in. hence, agents, toolchains, etc

that's not to say it won't push wages down. it's just that the usual stories of everything getting cheaper because of Technological Progress and Amazing Innovation are probably too simple

the people demand slop or things apparently indistinguishable from it

at a basic level all the fixed costs mean they can't price at marginal cost anytime soon -- they need profits to recoup the up-front investment. they'll try to stay cheap with the VC money for a while but people will start wanting exits by 2030

they're not competing just against other AI firms, they're competing against human workers too. the human wage creates margin

but even otherwise... I don't think the AI market is or will be pricing tokens or agents at marginal cost. they'll find ways to make switching hard, or they'll go bust

tapping the sign

yeah probably

anyway i think we're gonna get here at scale soon enough

"we're spending almost as much on office buildings for agents as much as we are on office buildings for humans" is a funny sentence to think on

1848:

"is it a crime when the state does it," the greatest thread in the history of etc etc

tangent but there's a case to be made that the most sinister technologies SV surveillance and behavior mod firms use is actually causal inference and GSP auctions

like, imagine you are zuck. how do you sell this story to people in SV/NYC/...?

obviously idk but there's a track record here

my basic issue here is that yes, information has to be collected by people, unrelated has anyone noticed that the profitable business models score high on surveillance and behavior modification

hey i'll have you know i have been described as a producer of high information density text

Value plus discounted

one of the harder truths of this time imo is that the unique position of the US was unique for many reasons and -- shocker -- not all of them were about the US

doing the work tomorrow for a better today