Same, I think there are genuinely useful applications for this tech, but it's far more limited than what's being marketed. The sooner we get past the hype phase the better, because then we can start focusing on figuring out what this tech is actually good for.

The aspect that's driving the hype currently is that we don't know what the plateau is. And people are figuring out ways to improve both accuracy and performance of LLMs, so it's difficult to say definitively what will be possible in the near future. For example, using a router with a set of LLMs can drastically reduce energy consumption while maintaining quality of the output https://lmsys.org/blog/2024-07-01-routellm/

Another example, is using a consensus model leads to more consistent outputs https://www.wired.com/story/game-theory-can-make-ai-more-correct-and-efficient/

So, it's possible that some of the more glaring limitations could be addressed, but it's also possible that we'll hit a wall because there are inherent problems with the approach. The context issue is one example where you start getting diminishing returns as the model gets bigger.

It's going to be interesting to watch how this all develops.