I'm a hobbiest at heart; I really enjoy just building and breaking things. For me this means a completely underpowered (and probably over priced) setup with a bunch of raspberry pis, and a few other boards.
I went in that direction because it's just a hobby and for fun. Maybe slightly ironic, I absolutely love tech.. but my career has nothing to do with it (relatively speaking) so everything I do is just out of curiosity and enjoyment... And most of what I do is just dumb things like "hey, what if I made this weird workflow"..
Anyways.. so because I'm always building things I'm constantly looking up and using tutorials. But I keep having issues. Probably partly because I'm dumb, but also because the web is filled with depreciated instructions, instructions that assume a very specific setup, or an OS, or a tool.. or what have you.
And sure, I could probably refine my results by refining my search.. but it's difficult to tell when I should be like "how to install x on RasPi" vs "how do I install x on RasPi with a, b, c, d, e, f ,g .... Configuration"
So I started cheating a bit.. I added some memories to Chat that maintain my system configuration and can help to more quickly find the info I need.
And then, over time I expanded that a bit. Now I have some local models with RAG, and can work through processes determine what works what doesn't.. and save that for easy retrieval.
Its not perfect, but saves a ton of time with trouble shooting because I often do things once, without documenting (or poorly document) and then am like WTF.
But then it occurred to me, if I can get these sort of results from an AI and manually perform tasks.. there really isn't a reason why I can't have the AI just do it automatically. And I've played with it a bit. Spin up a VM, have the AI attempt to do whatever install.. have it search for solutions when it encounters an error, etc.. it's not 100% reliable (not even 60%) especially with local models...
But this sparked an idea, and this isn't about a personal project because it's well beyond me.
But, if a local model can even halfway do something like this (find instructions, implement, verify if it works).. you could potentially have hundreds or thousands of machines constantly working through information verification and issues.
And with that information you can begin to dynamicly adjust relevance of information on the web... For someone really smart, maybe even constantly test and update that information.
I'm using tutorials as an example, because it's something I commonly encounter... But I think we are really just at the beginning of AI usage.
Sure there are a few things already changing, perplexity, web summaries,etc.. but most of these are just grabbing multiple top rank items, comparing and summarizing (I'm probably overly simplifying this).
Anyways, I don't think it's a far stretch to say that soon, relatively speaking, AI agents will be not just summarizing, but verifying information through actual implementation and that the "search providers" will begin dynamically ranking information based on accuracy, relevance, etc.
And it's kind of neat to think about because if you think about it it kind of creates this hive mind or decentralized information store that is continually checked and updated... And, while not perfect, we are at a point where just about anyone's machine can contribute to that.
[link] [comments]