Introduction to LLMs

By the end of this lesson, we will be able to speak the same language about the most important aspects of generative AI. I will explain what large language models actually are and clarify the key concepts, including those related to the pitfalls we are likely to encounter when working with LLMs.

• ChatGPT is not synonymous with artificial intelligence. AI covers a much broader range of technologies, with generative AI being only one part of it, and large language models (LLMs) representing an even narrower category.

• A prompt is the content sent to a model, such as an instruction, question, or input material. It may also include a negative prompt, which specifies what the model should avoid.

• Modern models are becoming increasingly multimodal. They can process not only text, but also files and other types of data, and can generate different types of output, such as text and images.

• AI agents can operate with a certain degree of autonomy, collaborate with one another, and carry out non-linear processes. For example, they can evaluate their own results and return to earlier stages if the outcome is not satisfactory.

• Models can be specialized in several ways: through costly fine-tuning, through RAG using external knowledge sources, or through simpler specialized solutions based on a fixed prompt and additional files, which can be used for basic automation.

• Tokens are used to describe the amount of information processed by a model, including the capacity of its context window.

• Models can be used either as conversational partners or as assistants with a defined, persistent role. However, it is important to remember their limitations, including hallucinations (plausible-sounding information that is not actually true), bias, excessive confidence in their answers, and knowledge cut-off dates. Important information should always be verified.

You will also find explanations of the most important concepts in THE GLOSSARY, which you can return to at any time.