Google DeepMind's Alex Imas argues retail users will eventually be priced out of frontier AI models by token costs
Alex Peysakhovich says high costs stem from multi-agent developer workflows.
@alexolegimas I think satisficing hits hard here
At a certain capability level normal models just “work” & you need to be pretty smart to get the most out of advanced ones
We are almost at that level
I think choice will be restricted by cost rather than prohibition. Same # of tokens on frontier model costs a lot more than cheaper model. Retail will eventually just be priced out.
@alexolegimas important to point out that even the "mythos class" models you're looking at $10 for ~400 pages of text.
the people you're seeing complaining about things being expensive are swe-types running 50 agents in the background constantly writing code and running it.
I think choice will be restricted by cost rather than prohibition. Same # of tokens on frontier model costs a lot more than cheaper model. Retail will eventually just be priced out.
@alexolegimas the race is between increasing flops used per query and increasing flops delivered per unit of electricity, i think current stuff (both hw and sw) has so much room to be made more efficient that i am bullish on the intelligence/price ratio going way way up in the new few years
@alex_peys Yeah right now the prices for retail are very low, my point was on future models. You don't think they'll start raising prices given how early we are with coding agents as far as demand?
@alex_peys Yeah right now the prices for retail are very low, my point was on future models. You don't think they'll start raising prices given how early we are with coding agents as far as demand?
@alexolegimas important to point out that even the "mythos class" models you're looking at $10 for ~400 pages of text. the people you're seeing complaining about things being expensive are swe-types running 50 agents in the background constantly writing code and running it.