What I Learned Watching OpenAI's Early Days & How to Build Reliable AI w/ Kyle Corbitt | E132

Imagine AI Podcast

What if you could fully trust your AI agent—especially in high-stakes environments like customer support or voice automation? Kyle Corbitt, co-founder and CEO of OpenPipe, joined us live at Imagine AI Live 25 to explain how his company is solving exactly that. Kyle shares how OpenPipe uses open-source LLMs and reinforcement learning to make AI agents dramatically more reliable—without your proprietary data ever leaving your system. From voice agents taking restaurant orders to compliance-critical customer support tools, OpenPipe enables companies to deploy agents trained directly on their data, inside their firewalls. We also get personal—Kyle reflects on his time at Y Combinator, reporting directly to Sam Altman during OpenAI’s formative years, and what it was like to watch the AGI conversation begin from the inside. If you’re scaling with AI and want long-term control and security, this episode is a must-watch. (0:00) Welcome & Kyle's high-energy first impression (0:36) Use cases: customer support & voice agents (1:42) How OpenPipe handles sensitive data securely (2:40) What models OpenPipe uses (LLaMA, Gemma, Qwen) (3:23) Letting clients choose their own models (4:04) Team size and client breakdown (4:58) How OpenPipe fits into your existing agent stack (5:45) Kyle’s origin story and background at Y Combinator (6:58) What it was like working with Sam Altman (8:04) Predicting AGI timelines in 2019 (9:39) Still collaborating with OpenAI today (10:13) OpenPipe’s long-term vision (11:09) Fundraising and growth status (11:53) Why switching providers is hard (and sticky)