I really like this article. I hadn't thought that an RTX 3090 would be capable of generating a sort-of decent small LLM from scratch in a reasonable time, but he shows how in detail.