Builders Unscripted: Ep. 2 - Ashe Magalhaes, Founder of Hearth AI

By OpenAI

Categories: AI, Product

Summary

Early AI product builders in 2023 had to rebuild their entire agentic stack every few weeks because user behavior around AI agents didn't exist yet—requiring founders to educate users on new interaction patterns while simultaneously inventing the technology itself.

Key Takeaways

  1. Agentic CRM required rebuilding the stack every few weeks in 2023 when using GPT-3.5, unlike traditional ML which takes a year to lay data pipes before seeing predictive results. This forced rapid iteration and user education simultaneously.
  2. Consumer behavior patterns didn't exist for AI agents in 2023—users didn't understand that Slack bots could respond interactively vs. just send outbound pings. Founders must actively educate users on new interaction models, not assume adoption.
  3. Early AI builders were trying to implement structured outputs (JSON schemas with retrieval) before OpenAI had native support for them. Necessity drove workarounds and API feedback that shaped platform development.
  4. Building relationships with other builders and AI lab insiders during frontier tech moments is critical—peer networks provide shared learning, transparent trick-sharing, and institutional knowledge that makes products work.
  5. Frontier tech building requires high rigor and accepting real risk (literal body-on-the-line decisions like racing a solar car). This mindset transfers to startup decision-making and technical resilience under uncertainty.

Topics

Transcript Excerpt

Hey, Ashe. So great to see you. Thank you for doing this. Thank you for having me it’s awesome to be here. I actually remember the very first time we met was in this OpenAI office. And at the time, you were building Hearth, right? Can you tell us a bit more about what it was? Yes. Hearth AI was founded in 2022, and it was the first agentic CRM. So 2023, we're using GPT 3.5, and we were trying to kick off all these agentic workflows when the underlying systems were still pretty fragile. So I was ...