How does the Large language models (LLMs) work under the hood?

✅ The Basics:
- In a nutshell, LLMs like ChatGPT are autocomplete systems.
- It gets the input as a text and uses deep learning to predict the most likely next word (token), the next word and so on.

✅ The Training :
To understand and predict the next token, LLMs are put through rigorous pre-training with loads of datasets (think of it as many thousand gigabits of data)

✅ The Architecture:
LLM applications use the traditional client-server model.
➡ The client (a website or an app) sends the user’s text input to the server.
➡ The server sends it through the LLM to generate a response.
➡ The response gets sent back to the client to display to the user.

But the real magic behind these LLMs is, how these models are adapted and fine-tuned for a specific use case. It's like giving the artist direction - "Hey, we need you to paint something inspirational for a children's hospital." With some guidance and fine-tuning, that same talent can be channelled into something truly impactful.

The competent but unrefined LLM models are refined with the below few techniques to specific needs and goals:

➡ Prompting: Carefully crafting the input to steer the model’s output to a desired direction.
➡ Constitutional AI: Instilling the ethics and principles via training
➡ RLHF: Reinforcement learning is like to optimise for quality. The model is given feedback iteratively on a good or bad output.

- Large language models (LLMs) like GPT-4 are incredibly impressive feats of deep learning like a raw diamond with brilliant potential.
- But it's the meticulous process of cutting, polishing, and setting that gemstone into a custom AI assistant that allows its true brilliance to shine. 🔥

#ai #artificialintelligence #llm #finetuning #rlhf #promptengineering #prompts