The above is, after all, the whole genesis of the word 'hacker'. We should celebrate that.
I tried to modify the embedding output of bert to make it generate box embeddings instead of point ones. At the time I had access to university provided A100 gpus but even with all that a training run took half a day. Models these days I don't think I can train it in any reasonable time with that much compute.
(As a TML person, I'm obviously biased, but I couldn't resist because of "tinkering").
TBF it's hard to imagine a real architecture change that wouldn't require a ton of compute, but you could certainly fine tune and play with different recipes, loss functions, etc. And Claude can carry you a lot of the way through doing this.
One fun task is to invent a tool and then train a small model to use it. You could export that small model and run it locally for free forever to do your thing. I think this is what a lot of Software Engineering will look like later.
There are a lot of other high level abstractions here to look at. Prime Intellect has one.
The other thing to play with is self-hosting small models, but IMO most of the interesting stuff is actually related to multi-gpu or multi-node inference so there's not necessarily a ton to learn here.
So, you build one from scratch.
The best analogy is strip mining (big labs) vs cave exploration (solitary/small teams). I think this is how science progresses at the boundaries by smart/curious/hardworking individuals because depth is a requisite for finding the right questions and then the answer. It is not for everyone and it does not always work. But you learn a ton even if it doesn't pan out to be a big breakthrough.
https://ravinkumar.com/GenAiGuidebook/book_intro.html
This guidebook covers pretraining, post training (SFT, RL) and a couple other topics. And others authors have also written books that fit on single node reasonable hardware.
If you want to start with a pretrained base I built Gemma 270m and released it last year. This fits on a raspberry pi.
https://developers.googleblog.com/en/introducing-gemma-3-270...
The fundamentals of AI don't require industrial amounts of large scale. Think of it like this, when I was learning how a plane worked when I was a kid I didn't build a 747 at home, I started with scale sized model planes. Same idea here.
And FWIW I'm a staff researcher at Deepmind (opinions here are my own) so I want to specifically encourage all people out there, you can learn a lot about how these LLMs work at home, for (mostly free), using resources like colabs or spot pricing on accelerator providers. There's many great resources out there and I encourage anyone willing to learn to go for it!