This episode moves from AI in life-critical settings to the technical details of making voice feel genuinely natural. Justin Uberti walks through the design constraints behind real-time speech systems, while Jessica Jensen and Jeremy Greenberg examine how emergency-response tools are actually being used and where they still fall short.
The conversation also touches on biology, AI infrastructure, model switching, and how product teams are adapting to voice-first workflows and usage-based AI pricing.
Show Notes
Justin Uberti joins Prakash Narayanan and Nathan Labenz to break down the OpenAI Realtime API, including natural turn-taking, latency, asynchronous reasoning, telephony, SIP, voice safety, accent coverage, and speech training data. Earlier in the episode, Jessica Jensen and Jeremy Greenberg discuss the current state of emergency AI, from predictive warnings and damage assessment to connectivity, privacy, preparedness, and the limits of automation in unique disasters.
Chapters
(0:00) A vaccine built for your tumor.
(1:23) Eight seconds can mean safety.
(2:56) AI answers when humans sleep.
(4:13) PMs can build features instantly.
(5:00) Opening and morning news
(5:51) Claude protein binders
(8:18) Specialist model pipelines
(11:24) Real-world AI testing
(13:31) Anthropic’s safety prompt
(15:22) Moderna-Merck cancer combo
(17:15) AI’s role in treatment
(18:48) Personalized cancer vaccines
(23:59) Cancer vaccine manufacturing
(25:34) The value of prevention
(27:49) Genetic screening tradeoffs
(31:21) Healthcare spending and value
(32:05) Structured biology models
(33:16) AI and self-experimentation
(34:13) Guest introductions
(39:45) Dual-use emergency tools
(40:56) Human control in disaster response
(43:32) Real-time damage assessment
(47:07) Predictive disaster warnings
(50:07) Connectivity and offline AI
(54:40) 1,179 emergency AI products
(55:59) Integrated emergency tools
(1:05:47) Automating preparedness work
(1:07:36) Privacy and life safety
(1:11:12) AI limits in unique disasters
(1:16:41) Robots and situational awareness
(1:20:16) Justin Uberti and Realtime AI
(1:23:40) Natural voice turn-taking
(1:24:31) Voice latency and gaps
(1:28:37) Real-time voice reasoning
(1:32:33) AI agent interoperability
(1:35:26) Voice AI telephony
(1:38:29) Realtime API and SIP
(1:41:34) Asynchronous reasoning
(1:44:51) Voice safety boundaries
(1:47:20) Accent and dialect coverage
(1:51:17) Voice agents on desktop
(1:52:38) Speech training data
(1:53:53) Why voice AI struggles to sing
(1:55:12) Etched inference hardware
(1:58:02) Voice mode and mind dumps
(1:59:26) Claude, Codex, and voice
(2:02:00) AI agents and deep work
(2:07:20) OpenRouter and model switching
(2:09:42) Capital and intelligence flows
(2:13:59) AI labs beyond token prices
(2:15:25) From tokens to digital employees
(2:17:01) Why AI models differ
(2:27:19) Usage-based AI pricing
(2:34:17) Replit for product managers
(2:35:39) Replit for hobbyists
(2:36:49) Closing thoughts
Guests:
Jessica Jensen — Senior Policy Researcher (RAND), AIDE Initiative
Justin Uberti — OpenAI Realtime Lead, OpenAI (𝕏)
Jeremy Greenberg — Senior Advisor (Aspen Digital), AIDE Initiative
