Descriptions:
Ed Elson’s “First Time Founders” series features an in-depth interview with Nick Frosst, co-founder of Cohere — the enterprise-focused AI company founded in 2019 by three former Google engineers, now valued at approximately $7 billion. Cohere’s clients include Dell, SAP, Salesforce, and the Canadian federal government, positioning it as one of roughly ten companies worldwide capable of building and deploying foundation models at scale.
Frosst explains Cohere’s core differentiator: a singular focus on enterprise deployment with private, secure infrastructure where Cohere itself cannot access customer data. He articulates why only a handful of companies can build foundation models at all — likening it to building a rocket rather than a software product, requiring massive coordinated teams, enormous compute, and iterative data pipelines that most organizations cannot assemble.
A technically rich portion of the conversation walks through the modern LLM training stack: initial training on web-scale text (volumes representing thousands of person-years of reading), followed by supervised fine-tuning with human-annotated data, reinforcement learning from human feedback, and increasingly, reinforcement learning on synthetic data generated by the model itself. Frosst also addresses the AGI question directly — arguing that embodiment and real-world physical interaction represent genuine, not just rhetorical, blockers to human-level general intelligence, and that text-trained models remain bounded by the limits of language as a representation of the world.
📺 Source: The Prof G Pod – Scott Galloway · Published March 01, 2026
🏷️ Format: Interview
