How LLMs Actually Work Under the Hood

I've spent the last few weeks unpacking how LLMs actually work under the hood. The mechanics run deeper than the common "next word predictor" explanation.

Text becomes vectors, vectors become embeddings, and embeddings enable the model to weigh meaning and context before generating a response. It builds a representation of intent from the full context.

Fine tuning changes the game by adjusting the model’s internal parameters directly. You are pretty much adding specific context into the model weights, kind of like prompting but on a deeper level.

Most tools like ChatGPT rarely use RAG for general response generation, apart from personalizing context.

This means the following are retrieved with RAG: • Your conversation history and thread context • Decision patterns from past interactions   • Preference signals and response history

This surfaces personalization without touching the base model.

It's been fun learning about all of this, wanted to keep you updated.