Fine tuning. For training from scratch, you want a dataset of at least 20GB gathered from all corners of the internet. I think OpenAI used around 160GB.

Even if you can't process that much data, merely having it available forces the model to learn a diverse variety of knowledge.

The difficulty of training from scratch (and generating a quality model) vs the difficulty of fine tuning is like the difficulty of becoming fluent in emacs vs using notepad. It's doable, but quality results take focused effort.

It's fun! Definitely within reach of lots of people who wouldn't normally consider themselves data scientists / ML engineers. (I'm one of 'em.)