BuffetGPT - Finetuning vs RAG vs MemGPT vs SRT

Hi everyone, I am getting a little bit confused with the concept, the purpose and the "difference in outcome" between a couple of techniques that afaik all try to "customize" an LLM.

As the title says it is Finetuning, RAG, MemGPT, SRT

Can anyone clarify?

Yesterday I have watched some videos about tools used for fine-tuning. They were presented on World of AI and Matthew Berman, namely MonsterAPI, AutoTrain, Fast-GPT-LLM-Trainer.

It made me thinking, say I want to create a GPT called "BuffetGPT" like Warren Buffet, the master investor, that has all the knowledge about value investing.

What would be the difference between finetuning Llama with a dataset of a dozen of value investing books (fine-tuning) and using vanilla Llama creating a vectordb (RAG) with langchain that includes the PDF information of all these books? What is better or would you do both?

(my hypothesis: fine-tuning is like sub-conciousness, and RAG is like working memory)

And how would MemGPT and SRT help?

Anyone knows a comprehensive video / blog post on the subject?

12 comments