Dark matter. Two miles underground. Next to a dam that caused an earthquake.

That's where Deepgram actually started. Today Faraz Siddiqi finally kicked me out of my own car long enough to get Scott Stephenson to tell the whole story from the passenger seat.

Before Deepgram, Scott was hunting dark matter two miles underground — and building custom compute from scratch because nothing off the shelf was precise enough. On this episode of Skywatch, Bluejay's podcast on conversational AI, he and Faraz dig into how that same obsession with getting it right from first principles is what let him see that end-to-end deep learning was going to rewrite speech recognition entirely.

A few things from this episode we can't stop thinking about:

🔹 The real AI debate isn't small vs. large models — it's real-time vs. not. Scott reframes the whole conversation here, and it's clarifying.

🔹 The laws of physics will determine voice AI architecture. Latency is geography. You can't route around it.

🔹 Scott's framework for when to launch a product — wait until the market is pulling, not just the technology is ready — is something every founder building in voice should think about.

If you're building in voice, real-time infra, or agentic systems — this one's a must.

Spotify: https://open.spotify.com/episode/0Y9lYoExz3PtXJzExx1OJq?si=dGVK3HEGQ1y2EJWSPnn-RA
YouTube: https://youtu.be/mIm1jIezSGU