Large language models, AIs as you know them, are getting more and more powerful by the day. They are also getting more and more complicated too.
A small model (Gemma 3, for example) is like a Casio watch - light and cheap, and tells time, maybe has a calculator. Meanwhile, a large model (like LLAMA 4 from Meta) is like a mainframe from the 80s, it takes the whole room.
Which ones should we use, and can we actually use them?
With great power, comes great power consumption!
It’s a wild west in the LLM space right now. We are all racing forward, but not sure we know where to. We are just racing, hoping that along the way someone will discover something, and it will change everythign.
Are we going to get one super model that can do everything and is only accessible to the super rich? Are we going to stop with AI, and go back to doing things by hand? Are we going to put giant solar arrays into space and crunch numbers in floating containers, cooled by dark matter?
Well, the latter actually might be the case, but …
Libraries, I think the future will be full of libraries.
We had libraries in the past, and while they were tedious, they also worked quite well for some use cases. You could go to a Library and get any book that you wanted, on any topic that you could find. When finished, you could go get another book, from that, or another library.
Library is an infinite source of knowledge, capped only by time.
There’s a very high chance that just like we used to have books, we will now have libraries full of models. In fact, using actual libraries as data centers might be an interesting concept. You go there, and you get to have access to powerful models, but only while you are physically tethered to the fiber cable. Perhaps directly through a Neuralink.
These libraries of the future will be “books”, but instead of physical paper, you will be connecting to different models. Some models will be good for coding, others for cooking, for history and for mathematics…etc. In fact, there will be libraries dedicated to those subjects, with books on different sub-topics.
In the future, you will connect to a provider that serves physics models, for example, and then you will choose whoever is the best in nuclear physics, and chemistry.
Supervising it all there will be a librarian, if you well, who will be able to reroute your information request to just the right place. You won’t have to talk to her, it will be just an AI with a specific set of reasoning skills, quickly guiding their audience to the right data.
Some of these ebooks you will take home for free, some of them, you will access remotely for a reasonable price, based on time and value.
I think this will solve a lot of problems with these below the bottles that are too large to effectively run on laptop. Furthermore, just like most people don’t read books on ALL the topics, most people will not need all the models. There will be cheap publicly available models serving AI slop, news, and some basic conversations. Paid models will do that much, much better. Other models you’ll be able to buy, or check out from those libraries and bring home. Math and chemistry libraries for the nerds, gaming libraries for others.
— Kirill.
Addendum:
Why does this even matter? Can’t we have local models and be happy with them?
Existing AIs have a few constrains. Models like Grok or Claude Code that run via the web API (aka, from their servers to my home computer) are very-very good. They however require an internet connection, and they do cost quite a bit. Models that can run locally on my MacBook Pro, thus only constrained by cost of a powerful laptop, those models have issues:
What issues, you ask?
Local models are capped in their training data at the time of release, which means if I ask a question about something that happened this morning, the will not have a clue. These models need to be updated with new content if you want them to access to that content, and while there are multiple ways of doing so, it doesn't work out of the box, it require rather a lot of effort.
On top of that, the small local models are only great a few things, not everything. A programming model might do well at programming, but it won’t be able to do project management, or creative writing. This requires switching models around, and who wants to do that?
And last, but not least is reasoning. A good model is no longer just a predictive text generator, it knows how to think. Not in the way you and I would think quite yet, but it’s getting there. It know how to figure out what I actually want to know, and give me that precise bit as the answer.
So yeah, we need more than just local models. A lot more, and for a lot less, if we want to truly make this technology the revolution that it wants to be. This has to be the new internet, available for all, with added benefit to those with money.