Voices is a new mini-series from Humanitarian AI Today. In daily five-minute flashpods we pass the mic to humanitarian experts and technology pioneers, to hear about new projects, events, and perspectives on topics of importance to the humanitarian community. In this flashpod, Cole Leng, an AI Researcher at Harvard and former Project Manager with Nexa AI, sits down with Humanitarian AI Today producer Brent Phillips to discuss the state-of-the-art and future trajectory of on-device machine learning. The discussion provides researchers and staff from humanitarian organizations with important insights into where the state-of-the-art in on-device machine learning stands today and where the cutting-edge is heading. Cole examines the critical trade-offs between on-device and cloud models, analyzing their respective workflows, performance limitations, and implementation considerations to help listeners evaluate whether on-device ML applications are suitable for their specific needs. He also offers insight into choosing LLMs, the impact of new specialized hardware on performance and capability, and how current advances in ML are shaping the next generation of applications. This episode highlights a core goal of the Humanitarian AI Today podcast: fostering dialogue between technology researchers and humanitarian practitioners. As AI and ML capabilities scale rapidly, this cross-sector engagement is crucial for mapping new technical advances to the unique operational, privacy, and resource constraints of the humanitarian field. Substack notes: https://humanitarianaitoday.substack.com/p/cole-leng-on-the-state-and-future