Oh I'm not one of those "LLMs are good for nothing" people, I use them often at work and at home and they do have some real world use cases that are very compelling. However they are a far very from the "solve everything machine" they are being pitched as, and they do seem to be plateauing in their improvements, seems like there's not much further they are going to be able to go even with a billion people using them everyday, there just isn't any new training data that is actually useful and they don't make new data, only associations between data they have already ingested. My qwen 3.8 model running on my 3090 at home is often better than pay for play enterprise Claude at work, if not just slower. I expect local models will explode in the future due to the cost / benefit of running them.