Earlier quoted context omitted.
The staggeringly effective compression of LLMs is still under appreciated, I think. 2 years ago you had downloaded onto your laptop an effective and useful summary of all of the information on the Internet, that could be used to generate computer programs in an arbitrarily selected programming language.
I got excited about that, until I actually tried to download a model and run it locally and ask it questions. A current gen local LLM which is small enough to live on disk and fit in my laptop's RAM is very prone to hallucination of facts. Which makes it kind of useless. Ask your local model a verifiable question - for example a list of tallest buildings in Europe. I did it with Gemma on my laptop, and after the top…
But what you can do with local models is give them actual data and tools to search it. Download a copy of Wikipedia locally, give the agent a way to search it and BOOM accurate information without an internet connection.
Also "small enough to live on disk" is a bit vague, especially when models get super stupid super fast when you get to the smaller size. At that point they're just basically 40k servitors that can use tools and nothing much.