Prompt and completion flow simplified

There are already countless transformer models, such as GPT, Llama 2, Dolly, BERT, BART, T5, and so on. These are essentially LLMs and, as you already know from Chapter 1, they are trained on vast quantities of unstructured text in a self-supervised manner. In this self-supervised learning, the training objective is automatically derived from the model’s inputs, eliminating the need for human-annotated labels or input (more on this later in this section). This allowed the transformer models or LLMs to be massive in terms of their parameters. GPT-4 has more than 1.75 trillion parameters alone. Sam Altman stated that the cost of training GPT-4 alone was more than $100 million (https://www.wired.com/story/openai-ceo-sam-altman-the-age-of-giant-ai-models-is-already-over/)!

Such models gain a statistical comprehension of the language they are trained on. However, they are not particularly useful for specific practical tasks. To overcome this...