At a time when large language models such as OpenAI's GPT series are becoming increasingly important, developers and software architects are faced with the exciting challenge of making the best use of these technologies in chat applications. The quality of interaction with such models depends largely on how precisely and thoughtfully our queries are formulated.

In this talk we will take a look at LLMs like chatGPT from the user's perspective. We will learn how we as developers can use modern AI to get better in what we do.

The first part of the talk dives deep into the art of "prompt engineering". Here we look at how targeted "priming" can be used to steer the AI in certain answer or thought directions. This subtle control of the AI makes it possible to obtain answers that are precisely tailored to the context and the needs of the user. We also look at the differences and areas of application of few-shots, zero-shots and many-shots. Each of these approaches has its own strengths and it is crucial to understand when and how they can best be used in chat applications. The challenge of limited token counts and strategies to achieve clear, coherent and helpful responses within this limitation will be highlighted.

As a special bonus of the talk, we will briefly discuss the use of embeddings, which enable topic-specific communications with AI and thus further increase the relevance of the answers.

W-JAX 2023