Can AI Really Reason? Inside AI Research, Generalization & LLM Evaluation | Michelle Yi

Can today’s AI models actually reason, or are they simply getting better at producing answers that look like reasoning? In this episode of AI Chronicles: Season 1 — AI Research, I speak with Michelle Yi, an experienced AI research and technology leader whose career has spanned IBM Research, Slalom, and senior roles across applied AI. We explore how AI research has evolved from the earlier days of systems like IBM Watson to today’s large language models, and dig into some of the harder questions researchers are still trying to answer. Michelle shares her perspective on what reasoning should really mean for an AI system, why generalization and robustness matter, and how even small changes to a problem can reveal weaknesses in models that otherwise appear highly capable. We also discuss the limitations of current AI benchmarks, how researchers should evaluate increasingly capable models, and what gets lost when AI moves from the research lab into real-world applications. Topics include: • Can LLMs truly reason? • Reasoning vs pattern recognition • Generalization and robustness in AI models • Adversarial attacks and model weaknesses • How AI evaluation needs to evolve • The changing nature of AI research • Moving AI from research into production If you're interested in AI research, LLM reasoning, model evaluation, trustworthy AI, or the future of artificial intelligence, this conversation goes beyond the hype and into the questions that still don’t have easy answers. Subscribe to AI Chronicles for more conversations with researchers, builders, and leaders shaping AI.