It’s a tool with some interesting capabilities. It’s very much in a hype phase right now, but legitimate uses are also emerging. Automatically generating subtitles is one good example of that. We also don’t know what the plateau for this tech will be. Right now there are a lot of advancements happening at rapid pace, and it’s hard to say how far people can push this tech before we start hitting diminishing returns.

For non generative uses, using neural networks to look for cancer tumors is a great use case https://pmc.ncbi.nlm.nih.gov/articles/PMC9904903/

Another use case is using neural nets to monitor infrastructure the way China is doing with their high speed rail network https://interestingengineering.com/transportation/china-now-using-ai-to-manage-worlds-largest-high-speed-railway-system

DeepSeek R1 appears to be good at analyzing code and suggesting potential optimizations, so it’s possible that these tools could work as profilers https://simonwillison.net/2025/Jan/27/llamacpp-pr/

I do think it’s likely that LLMs will become a part of more complex systems using different techniques in complimentary ways. For example, neurosymbolics seems like a very promising approach. It uses deep neural nets to parse and classify noisy input data, and then uses a symbolic logic engine to operate on the classified data internally. This addresses a key limitation of LLMs which is the ability to do reasoning in a reliable way and to explain how it arrives at a solution.

Personally, I generally feel positively about this tech and I think it will have a lot of interesting uses down the road.