An alternative path is to find a domain-specific application that outperforms the general model by an order of magnitude (if you can’t, don’t bother), scale quickly, and build a moat with network effects or by means of the legal frameworks (copyright (still an open question), IP laws, etc).
A lab is still probably a better bet, but perhaps not the only way.
Something like this *might* work. There's so many unsettled things in the supply chain: what's going to have relative pricing power, for example? Maybe you make the domain-specific application built on an LLM under the hood. And the LLM learns some new ability in 1 year that means a new entrant can outcompete you. Maybe there's a network effect in the product design that outcompetes the new entrant, but can't know for sure that new capabilities won't come out that outcompete the network effect.
Great post. Very similar to the conclusion I reached thinking about this over the month.
An alternative path is to find a domain-specific application that outperforms the general model by an order of magnitude (if you can’t, don’t bother), scale quickly, and build a moat with network effects or by means of the legal frameworks (copyright (still an open question), IP laws, etc).
A lab is still probably a better bet, but perhaps not the only way.
Something like this *might* work. There's so many unsettled things in the supply chain: what's going to have relative pricing power, for example? Maybe you make the domain-specific application built on an LLM under the hood. And the LLM learns some new ability in 1 year that means a new entrant can outcompete you. Maybe there's a network effect in the product design that outcompetes the new entrant, but can't know for sure that new capabilities won't come out that outcompete the network effect.