How chatGPT works

Fastest growing app in history, 10 million users in just 2 months.

Base of (free version) chatGPT is GPT-3.5 Large Language Model (LLM)

LLM: Neural Networks based model trained on large amounts of text data to understand and generate human languages. It uses training data to learn statistical patterns and then predicts word after word. Input and output words are organized by tokens (numerical representation of words), to process more efficiently. GPT-3.5 has 175 billion parameters with 96 layers in deep learning neural networks, in 500 billion tokens.

Without proper guidance, the model can generate untruthful statements. Fine tuning aligns the model with human values, to become chatGPT. For finetuning, a new prompt is sampled from the dataset, a reward model calculates the reward for the output, which, based on cumulative feedback, finetunes, or updates/aligns the policy for the output.

For every input in a chat conversation, the entire chat context and prompt tone of the past is fed along with the current input to understand the conversation.

Detailed video: youtu.be/bSvTVREwSNw?si=pJ4LBSkvg8LpZ8JR