Unscripted with Manas Talukdar

Manas Talukdar has spent 19 years building the infrastructure underneath AI systems: the data backbone of the process industry, the platform behind one of the largest enterprise AI companies, and the training data systems behind modern language models.

 

In this conversation with Jeff Pedowitz, he argues that the model, the part everyone argues about, is actually the easy part. The hard part is the data engineering, context management, and system architecture surrounding it, and that's where most enterprise AI initiatives quietly fail, particularly on long horizon problems where accuracy degrades as workflows get longer regardless of how large the context window gets. Talukdar walks through why reliable, production-grade agent fleets barely exist yet despite how good the demos look, why the industry's shift from token maxing to token optimization is really an economics problem in disguise, and why intellectual property exposure, not just cost, is quietly shaping how enterprises think about feeding their proprietary knowledge into third-party models.

 

He closes with a concrete picture of what an AI-native company actually looks like in practice, and lays out where he believes durable competitive advantage will live once the models themselves are fully commoditized: proprietary data, deeply integrated workflows, and operational discipline.