Here is the evolutionary tree of the existing LLMs (as of July 2023)
CREDIT | JingfengYang
The best place to learn how LLMs actually work and the main principles behind them:
LLM University (LLMU) | Cohere
When you want an LLM to produce desired answers, you can:
1 - Rely on it’s knowledge (training data)
That is what you do when you use ChatGPT and write instructions there.
Some language models, like models from Open AI (GPT-3.5 and GPT 4) allow you to set a system prompt in addition to the instructions you provide in every message. That is important when you want all the answers to follow the same rules.
The problem? - model knows everything on a superficial level, but lack narrow specific knowledge - like knowledge about some tools, businesses, individuals, etc.
2 - Fine-tune the model
You can read more about fine-tuning here:
3 - Augment the model, i.e. give it your information in addition to the model’s internal knowledge
Here you can find a brief explanation of how it works:
Tutorial: ChatGPT Over Your Data
When building LLM-driven apps, it’s crucial to understand how data processing works - it’s based on vector embeddings. Here you can find a good overview of vector embeddings and databases: