Alex Imas, Google DeepMind's director of AGI economics, argues enterprise lock-in is the primary moat for frontier AI companies

@alexolegimas 🎯

@intellectronica But that would limit the benefits of the models. That's almost like running 3 separate operating systems with a firm (the "it just works" is due to single model).

@intellectronica I don't think it's a question on whether alternatives exist. It's that once a firm adopts a "cloud AI", either through a single provider or an openclaw/hermes alternative, it will be difficult to switch.

@alexolegimas hopefully, enough enterprises realise that they can have the exact same functionality with swappable models. it's hard to imagine them not figuring this out sooner or later.

Yes I absolutely agree with that, that's the first comment on architecture. But I think it's difficult to have the type of AI cloud Andrej is describing across multiple models. It's more about incorporating the models into the entire org, multiple models don't play together as seamlessly as a single one across the whole thing.

@alexolegimas the models are surprisingly similar at their weight classes. for some reason the labs are always tied, with negligible margin. for "just works" to be model-dependent, one of them would have to release a model with unique capabilities. that hasn't happened yet.

@alexolegimas hoards of openclaw/hermes users would beg to disagree. it isn't perfect, but if the experience with young community projects is already acceptable, how good can it get when cloud software providers start catering to the needs of rich enterprises?

@alexolegimas Moats are bad tho, for social welfare. Theory suggests they shouldn’t exist.

@alexolegimas yes, that's definitely true. but these are still early days, and in my experience IT departments are nowadays rather mindful about lock-in risks, having been burnt with windows and similar platforms. many, will not tolerate a single-provider solution.

@alexolegimas The enterprise lock-in path is the realistic one. Architectural moats erode fast because talent moves between firms. But tacit knowledge accumulated across a company's workflows can't be copied or poached. That's the only moat that compounds instead of leaking.

@alexolegimas the third is network effects