Ann Arbor AI/ML Meetup

Why was Claude able to produce a candidate counterexample to one of mathematics' longest-standing open problems? And what does that have in common with robot vacuums, packing a car trunk, AI coding assistants, and quantum computers?
This talk explores a common pattern that appears across all of these examples: finding or creating a solution is often the expensive part, while verifying a proposed solution can be surprisingly quick. Once you recognize this pattern, it becomes easier to understand why AI can be remarkably effective on some problems—and where human judgment still matters most.
We'll explore this idea through six stories:

• Everyday Examples: Robot vacuums, bricklaying, and packing problems introduce the intuition behind problems involving delegating difficult work and simultaneously keeping verification simple.
• Mathematics & AI: We'll examine Claude's reported Jacobian conjecture counterexample as a case study in combining human insight, reduced search spaces, and AI-assisted discovery.
• Modern Computing: See how the same principle appears in AI-generated TensorFlow code and quantum algorithms for integer factorization.

Our speaker, Patrick Sherry, is a software engineer pursuing a career in machine learning for healthcare applications, focusing on areas such as computer vision, time series analysis, natural language processing, and full-stack development. He currently co-organizes the Ann Arbor AI/ML Meetup event and holds a master's in Computer Science from the University of Michigan. When not building models, he enjoys mountain biking, regular biking, cross-stitch, crochet, board games, and video games.
Matters for Dialogue
When Is Verification Easier Than Discovery?

• AI as a Search Engine for Ideas: When can AI effectively search enormous solution spaces while humans validate the results?
• Human-in-the-Loop AI: Which tasks benefit from delegating the expensive work while keeping people responsible for verification?
• Limits and Risks: When is quick verification sufficient, and when do problems require deeper analysis or stronger guarantees?