Agents Have Turned Podcasts Inside Out—Old Interviews with AI Founders Are About to Be Fact-Checked

Exa has turned podcasts into a searchable database. 119,000 shows, 3.2 million episodes, roughly 2.2 million hours of audio, 23.7 billion words, with complete transcripts, updated daily. Exa's tweet says 3M+ podcasts; broken down, the original text is 119,000 shows and 3.2 million episodes. For Agents, podcasts have changed from audio that could only be listened to manually into a text library searchable by person, time, and topic. The search API is already available; coverage and benchmarks are on the Exa blog.

Podcasts have long been an information source in the tech world with high information density but low utilization. In one- to two-hour interviews, founders, investors, and executives say many details: why they cut a certain business line, how they assess competitors, when their attitude toward a certain technology changed. This content is often more interesting than press releases and earnings calls. In the past, to find a particular thing someone said among hundreds of episodes, you could only listen episode by episode manually.

Now Agents can search directly. For example, compile an investor's public judgments on AI Agents over the past three years, arrange them chronologically, and see how their views changed. You can also track how frequently an AI lab is mentioned across different podcasts, find out when it entered discussions in Silicon Valley investment circles, and who was the earliest to repeatedly bring it up. For in-depth reporting, research on individuals, and investment research, this kind of capability is very practical. Details scattered across different shows and never separately reported by the media have a chance to be systematically unearthed.

The line chart released by Exa counts the number of episodes that mention an AI lab in the full podcast text, showing changes by month.

Benchmarks were provided too. Exa tested on 250 active shows whether the latest episode ranked in the top five: Exa Auto 98.4%, Perplexity 93.6%, Parallel Advanced 89.6%, Brave 67.6%, Tavily Advanced 63.6%, You.com 49.2%, Linkup 22.4%. On 667 questions about specific episodes, Exa answered 85.0% correctly, with the second best at 80.7%.

Retrieval quality is not just about recall. User Aleksandar Janca reminded that for Agents, speaker labels are as important as the text; a quote attributed to the wrong guest is worse than no result. User Marvin mentioned that retrieval results need timestamps and surrounding context, and when comparing changes in views, one must distinguish recording time from publication time—don't treat an old interview as a new statement. User Jason Garoutte is already using it to track brand and competitor mentions in podcasts; what was missed before can now be caught. User Byron Xiao put it more directly: for a one-person company doing competitor research, prioritizing Agents to extract changes in views by person and timeline is better than scrolling through headlines.

One foreseeable consequence is that many people in the tech world will change their attitude toward public expression. In the past, an offhand remark in a two-hour podcast that wasn't clipped into a short video might slowly drown in the audio. In the future, any sentence could be found by an Agent and compared side by side with what you said three or five years ago. The original poster mentioned that this will change many people's attitudes toward public expression in the tech world. The internet has already recorded what people have written; now it is beginning to systematically organize what people have said.

Reference: Exa tweet