Comment by captainbland
5 days ago
I think the most likely thing is the cheap self hostable copies will broadly stop improving significantly. It'll be too costly for a community project to distill a bleeding edge cloud model and companies will stop releasing them. What's free now will remain free, we might even get another gen or 2 of improvements (possibly with diminishing returns) on free/cheap local models but those days are numbered.
What's the force that you see keeping those days numbered? Many ollama models are runnable on a local laptop, and models like deepseek even more so. Aside from open source interest, there are many large state and corporate actors who would like to see models commoditized.
The only real wild card to me is state sponsorship. Everywhere else I expect such a capital intensive tech industry to approach winner (or maybe duopoly) takes all over time.
My view is that right now, because of the willingness of corporations and other investors to swallow short term (but massive) losses on this, we're basically in AI fiscal fantasy land.
The question we should be asking is how do we get access to these local models in the first place? It's all based on the work of these hyper expensive base models as the best small models are quantisations and distills. None of this works as soon as the profit motive comes into play and companies start gatekeeping effectively, which they will.