Large Language Models can seem almost magical in their capabilities - brilliantly insightful one moment, frustratingly obtuse the next. This apparent cognitive asymmetry stems from how these models navigate their latent space: a high-dimensional realm where knowledge is encoded not as discrete facts, but as geometric relationships between concepts.

In this talk, we’ll demystify this abstract space and provide practical tools for exploring it. You’ll learn why the same model can give wildly different responses based on how you structure your prompts, and how to craft them effectively. We’ll examine why models sometimes display deep insight into topics not explicitly present in their training data, through their ability to traverse novel paths through latent space.

Using Python code examples and visual analogies, we’ll build an intuitive understanding of latent space navigation. You’ll discover why providing a thoughtfully diverse set of “coordinate references” in your prompts often works better than hyper-focused queries, much like how dropout prevents neural networks from overfitting. We’ll explore practical techniques for prompting that treat the model not as a simple question-answering system, but as a geometric navigator responding to carefully placed waypoints.

The talk will provide both theoretical foundations and hands-on techniques, including:

• A practical framework for understanding latent space navigation
• Visual tools for conceptualizing high-dimensional concept spaces
• Concrete Python examples for systematic latent space exploration
• Best practices for prompt engineering based on geometric intuition
• Common pitfalls and how to avoid them

Whether you’re building AI applications or just curious about how these models really work, you’ll come away with a deeper understanding of the geometric nature of machine learning and practical tools for more effective model interaction. Join us for this journey into the hidden dimensions of artificial minds.