Will the era of using these to generate code end? Is the assume that the inference problem will be solved?
Personally, I've found that with guardrails, local 8-14B models can match frontier models on agentic tasks. The key is simple tasks with volume. For very complex things, the big models win. But a simple HR agent auditing 30,000 employee records to make sure all your info is filled out correctly, one at a time? You don't need frontier size for that.
1. Quantization: bigger models can be compressed and retain most of their quality, but run with a far smaller compute footprint
2. Moore's law: chips still are scaling so you can run more compute cheaper as it improves
3. Open source competition: if models from Anthropic or OpenAI get too expensive people will opt to use open source Chinese models, which would reduce demand and thus reduce prices down to a more reasonable equilibrium
They for sure are testing LLMs and checking the performance of local models. Once they reach a performance and quality enough for some tasks they will announce Apple AI or some variation of the name.
All of this is speculation, but I think is obvious the right way.
That's what Sam Altman said we should do.
> Will we shift towards slower/worse LLMs running locally?
Bet on faster/better LLMs running locally and invest/brace yourself for recession accordingly.
My long term expectation is that the big labs will build big models primarily for training and distilling to economically viable models. Even if the model capabilities don't plateau so soon, I think the economics of this will.
But that would lead to another competition on prices again.
So: when the money runs out and the bubble pops, we'll still get cheap existing models, what we lose is the race for new models.
We'd probably even keep free models: I forget where I saw it, but back in the early days someone noticed that models were so cheap that you could generate a decent sized blog post about any topic for about the same as the expected revenue from putting a few adverts on it and having it viewed *exactly once*.
That said, when (/if) these businesses stop chasing new models, it can make sense to burn the weights of the best at that date into a fixed (and analog, given how well they work with only a few bits of precision) circuit, making them more efficient. Not my field, so I'm not sure exactly how much more efficient analog can be; one or two orders of magnitude from what I've heard, but don't hold me to that, not my field.