You say "if it takes you 6 months to pay back Model A’s training cost on a $1B revenue base, and Model B costs 4x more to train but you’re on a $3B revenue base… the payback period actually shrinks."
On a $1B revenue base a 6 months payback period assumes $500M investment. On a $3B revenue base the 4 x $500M investment would take 8 months to pay back.
I challenge your argument favoring lock-in vs switching costs. Already today many applied AI companies have developed multi-model architectures with well-defined abstraction layer (separate business logic) and dynamic routing to optimize model selection based on task. As model capabilities continue to improve, for many tasks (but not all) the incremental differences will not be material
Spurred along by what has felt like usurous token churning by the Big Labs, I've been actively tinkering with the Master / Sub agent approach, in my case specifically for coding. Three directional observations:
Not only are the open models quite good (Qwen as example), the Master oversight approach seems pretty effective, with many fewer Master tokens.
I'm running the open model locally on my Mac Studio. There definitely seems to be something to the edge potentially having a large advantage in many use cases.
As much as I know use case demand will expand with model capabilities, it is that open model are more than sufficient (i.e. much better than good enough) for many of use cases, leaving so much work do being available to being done much less expensively.
Hi, Jamin. Your math for payback does not add up.
You say "if it takes you 6 months to pay back Model A’s training cost on a $1B revenue base, and Model B costs 4x more to train but you’re on a $3B revenue base… the payback period actually shrinks."
On a $1B revenue base a 6 months payback period assumes $500M investment. On a $3B revenue base the 4 x $500M investment would take 8 months to pay back.
I challenge your argument favoring lock-in vs switching costs. Already today many applied AI companies have developed multi-model architectures with well-defined abstraction layer (separate business logic) and dynamic routing to optimize model selection based on task. As model capabilities continue to improve, for many tasks (but not all) the incremental differences will not be material
The biggest bear case for the models is the macroeconomic picture. Come to think of it, that's the biggest bear case for all private capital.
Spurred along by what has felt like usurous token churning by the Big Labs, I've been actively tinkering with the Master / Sub agent approach, in my case specifically for coding. Three directional observations:
Not only are the open models quite good (Qwen as example), the Master oversight approach seems pretty effective, with many fewer Master tokens.
I'm running the open model locally on my Mac Studio. There definitely seems to be something to the edge potentially having a large advantage in many use cases.
As much as I know use case demand will expand with model capabilities, it is that open model are more than sufficient (i.e. much better than good enough) for many of use cases, leaving so much work do being available to being done much less expensively.