“Unlikely Models: How the next AI breakthroughs may come from unexpected places”
David Gurzick, Ph,D.
Everyone expects LLMs to dominate coding, planning, and reasoning. But what if the real breakthroughs come from unexpected model types in unconventional domains?
This talk explores a bold new trend in AI: the use of atypical models in domains not designed to perform in—think diffusion models used for text generation, image models adapted for tabular reasoning, and small-scale models outperforming giants through clever structure. We’ll examine real-world cases of this purposeful anomaly, like Apple’s use of diffusion models for programming tasks, and highlight surprising places where VAEs, graph neural nets, or symbolic logic models quietly outperform transformer-based systems.
Together, we will learn how shifting the architecture—not just the size—can lead to more efficient, safer, and task-optimized AI. For national security and mission-critical work, this architectural agility could be the difference between performance, failure, and, maybe, the “wrong” model—that happens to outperform.