Comment by api
1 day ago
The biggest exception is cloud. Big cloud carries an insane markup (bandwidth is like 10000X!) and everyone runs on it.
The strategy there is false openness where deployment complexity is the real proprietary moat. Sure Linux, Docker, Kubernetes, Postgres, and all the other standard tools in the box are open source and free, but they're also arcane and complex to run and hard to make fault tolerant. So you're lured in by "open" and then locked in via a kind of "death by a thousand cuts" complexity moat.
(Personally I hold the view that complexity and arcane-ness beyond a certain point is indistinguishable from closed in practice. Open source that's really complex and hard to run is not open in any meaningful sense.)
AI may not admit that kind of moat though, because AI is very good at slicing through that kind of thing. You can prompt a model to make itself compatible with another model or to change code to make it compatible. There's no moat because the moat bridges itself.
I think the whole idea of people running models is wrong. Models will run models, training will become distributed, and people will ask ModelNet to do whatever.
Computing tends to oscillate between centralised and decentralised models. It also oscillates between batch and timesharing.
Currently training is batched and centralised, access is timeshared and centralised.
But eventually a previous generation of computing turns into transparent networked infrastructure, and then you get another layer of new kinds of applications on top of it.
That's what happened with the Internet, and it will happen again with AI.
If Linux doesn't qualify as open, then what does?