Size is absolutely not going to become the moat unless there's some hardware revolution that makes running big models very very cheap, but that requires a very large up-front capital cost to deploy. Big models are inefficient, and as smaller models improve there will be very few use cases where the big models are worth the compute.
I imagine that going forward, the typical approach would be a multi-level LLM, such that there's a relatively small and quick model in front of the user, which can in turn decide to consult an "expert" larger model as part of its "system 2".
Absolutely, that is 100% the way things are going to go. What's going to happen is that eventually there will be an online model directory that a local agent knows how to query to identify other models to call in order to build up an answer. Local agents will be empowered with online learning since it won't be possible to pre-train on the model catalog.
> as smaller models improve there will be very few use cases where the big models are worth the compute
I see very little evidence of this so far. The use cases I'm interested in just barely works on GPT-4 and lesser models give mostly garbage. I.e. function calling and inferring stuff like SQL queries. If there are smaller models that can do passable work on such use cases I'd be very interested to know.
Claude Haiku can do a LOT of the things you'd think you need GPT4 for. It's not as good at complex code and really tricky language use/abstractions, but it's very close for more superficial things, and you can call haiku like 60 times for each gpt4 call.
I bet you could do multiple prompt variations with haiku and then do answer combining to compete with GPT4-T/Opus at a fraction of the price.
Interesting! I just discovered that Anthropic indeed officially support commercial API access in (at least) some EU countries. They just don't support GUI access in all those countries:
