LLMs are inherently unreliable. We do already work with unreliable technologies for many things but there's this engineering principle called "use the right tool for the job".
And it's a lot more likely that LLMs can't be changed to become unreliable, it's just how they work. So we would need more basic research, that doesn't grow on trees and for which the timelines are basically open ended. Maybe tomorrow, maybe right after cold fusion hits mass adoption.
They’ve become a lot more reliable. It would be strange if that progress stopped now. It’s possible. My main point is that they won’t get worse, as the other commenter suggested.
This isn’t a theoretical or technical problem; in fact, it’s not about the models at all: it’s about the available services getting worse. When they preview a new model and people are benchmarking it, they afford them a shitload of compute and they work great. Then, they get worse. That extra compute is surely in the billions of dollars OpenAI allotted to marketing, and especially in higher-volume periods, people talk about Anthropic‘s responses being lower quality all the time. The finances don’t indicate these companies’ current MOs are sustainable. Since self-hosted models don’t stand a chance of competing with frontier model capabilities, there is a very real chance that the available services will get much worse, and where the rubber hits the road, that’s really all that matters in the foreseeable future.
And it's a lot more likely that LLMs can't be changed to become unreliable, it's just how they work. So we would need more basic research, that doesn't grow on trees and for which the timelines are basically open ended. Maybe tomorrow, maybe right after cold fusion hits mass adoption.