Agree that this flip may take 10 years. There's a few reasons why:
1) If frontier scaling persists and models get better/more capable, building sustainable/differentiated applications is very difficult as a model update and/or frontier labs building vertical products (Claude Code vs Cursor) can wipe out your business.
2) Adoption speed across sectors in the economy will likely be a slow/choppy process with tons of bottlenecks, so even if model progress stalls, and sustainable apps are built, it will take time to gain mass adoption. To be clear, these are unique applications that drive productivity increases. Swapping OpenEvidence for Google likely does drive better marginal outcomes but it's not the type of application that's so differentiated that it flips the pyramid.
More generally it seems like the path forward will be some combination of frontier models hitting some bottlenecks in intelligence scaling (I see this a gap period between till algo techniques/unhobblings occur that get to AGI). Inference scaling massively as frontier training slows down, making triangle turn into a diamond for a period of time.
Most consumer downloads are used for Chat, which morphs into search. What this doesn’t factor in is that every iPhone user has a device with a utility installed (safari) that now defaults to Gemini when they ask a question… so if distribution is king, google still has the edge.. no doubt they'll soon release a stat that talks to their DAU's or MAU's which includes every Gemini response served to the top of a google page.
Well structured, thoughtful and excellent work. Looking forward to your thought line as we “all” are on this ai journey
Great article, thanks!
Agree that this flip may take 10 years. There's a few reasons why:
1) If frontier scaling persists and models get better/more capable, building sustainable/differentiated applications is very difficult as a model update and/or frontier labs building vertical products (Claude Code vs Cursor) can wipe out your business.
2) Adoption speed across sectors in the economy will likely be a slow/choppy process with tons of bottlenecks, so even if model progress stalls, and sustainable apps are built, it will take time to gain mass adoption. To be clear, these are unique applications that drive productivity increases. Swapping OpenEvidence for Google likely does drive better marginal outcomes but it's not the type of application that's so differentiated that it flips the pyramid.
More generally it seems like the path forward will be some combination of frontier models hitting some bottlenecks in intelligence scaling (I see this a gap period between till algo techniques/unhobblings occur that get to AGI). Inference scaling massively as frontier training slows down, making triangle turn into a diamond for a period of time.
Most consumer downloads are used for Chat, which morphs into search. What this doesn’t factor in is that every iPhone user has a device with a utility installed (safari) that now defaults to Gemini when they ask a question… so if distribution is king, google still has the edge.. no doubt they'll soon release a stat that talks to their DAU's or MAU's which includes every Gemini response served to the top of a google page.