Rather than manually iterating on prompt phrasing, why not evolve them?

This talk utilizes a live coding environment where LLMs generate, evaluate, and refine their own prompts in pursuit of a specific goal. Using genetic algorithms and test-driven evaluation, we explore how an LLM can teach itself to:

- Write code to spec
- Red-team itself into break LLM guardrails
- Produce the best image-to-text pipeline
- Write summaries or sentiments that pass human-alignment benchmarks

Choose your own adventure: The audience chooses the challenge at the start. We’ll show how each prompt is scored, mutated, and selected in real time—offering insight into both model behavior and prompt design strategies. The result is part performance, part education, and all open-source!