LLMs can automate some tasks (especially basic coding), but with some major caveats:

- They have rapidly diminishing returns on higher contexts. They are very easy to overload with user information. You can post-train them, but that’s too much into actual computer science to be useful to the average office worker. They can work well as a co-worker assistant in a few cases, but they really can’t replace humans long term.

- LLMs only really work well when given the ability to call on-device tooling. Which, with a cloud system like Copilot, is going to be super tame and underwhelming because of lawsuits and various business deals. Tool calling is only really agnostic when you’re running the model yourself on your own device.

- On the topic of on-device models, there’s also the fact that AI provider companies have been caught evading things like robots.txt and causing so much crawling activity that it effectively becomes a DDoS attack. On-device AI doesn’t solve this totally, but most people likely won’t be pushing their gaming GPUs to the absolute limit 24/7 to constantly hit websites with crawl requests.