SE Radio 715: Sahaj Garg on Designing for Ambiguity in Human Input

In this episode of Software Engineering Radio, host Amay Ambadeh interviews Sahaj Garg, co-founder and CTO of Whisper, a voice-to-text AI company focused on designing systems that handle ambiguity in human input. Garg discusses how human communication is inherently ambiguous due to lack of context, tone, accent, and intent—challenges that machine learning models struggle with because they typically process inputs in isolation without retaining context. Whisper's mission is to build a voice-first interface that adapts to users' preferences, styles, and histories by leveraging context, user behavior, and personalization. The conversation explores how ambiguity can be reduced through additional context, such as voice characteristics, prior corrections, and conversational history, and how models can be trained using synthetic data, instruction tuning, and user feedback. Garg emphasizes that the key to solving ambiguity lies in understanding what users truly want, not just what they say, and that systems should learn from 'revealed preferences'—like repeated corrections—rather than relying on explicit instructions. He also highlights the importance of balancing personalization with consistency, especially when users are inconsistent, and warns against AI regressing toward generic, unoriginal content unless users provide clear intent and narrative structure. The episode concludes with broader lessons applicable across AI applications: more context leads to better decisions, and the most successful systems mirror natural human interaction.