PaperDive: who makes it, and how.
PaperDive publishes one long-form audio episode per new AI research paper — the podcast is listed as “AI Papers: A Deep Dive” — with a full transcript, key takeaways, the chapters, and our reservations about the paper on every episode page.
Who makes it
Garrett Casey — Editor and publisher. PaperDive started as a private feed: the deep-dive segments at the end of Last Week in AI were the most useful part of a long commute, there were never enough of them, and building a pipeline to make more was a good excuse to get serious about coding agents. The feed went public once the episodes were good enough to share. Garrett Casey picks the paper every day, reviews what ships, and answers for the corrections.
How an episode is made
Every step below is automated except the two marked human.
- Candidate mining. New arXiv papers are scored on relevance, novelty and importance for an audience that follows AI but does not do research for a living.
- Deep read of the shortlist. The top candidates are read in full and diagnosed: the finding, the honest caveat, how strong the evidence is.
- Paper pick (human). Garrett Casey chooses the day's paper from that report. Nothing is produced without that pick.
- Close reading. A language model writes a long brief from the paper itself — motivating problem, method, the load-bearing math, concrete numbers, a steelman critique, and the limitations the authors acknowledge.
- Script. A two-host script is written from that brief, then checked by a separate fact-review pass that flags claims which do not match the paper (up to three fix rounds), then rewritten once more for listening rather than reading.
- Voices. The hosts are synthetic voices from ElevenLabs. The audio is transcribed back with speech recognition and compared with the script; an episode with too many misreads is not published.
- Notes, transcript, video. Show notes, chapter timestamps, the transcript and the YouTube video are generated from the same script.
- Review before publishing (human). Garrett Casey listens and reads before an episode goes to the feed, the site and YouTube.
What is AI and what is not
The scripts are written by an AI language model and the host voices are synthesized by ElevenLabs; every episode says so on air and in its notes. The producer is not affiliated with Anthropic or ElevenLabs. The paper selection, the final review, and this page are human. Please cite the paper and its authors, not the episode, for the research itself — every episode page links the paper.
Corrections
Mistakes happen, and the model can misread a table. If you spot an error, reply in the pinned comment under the YouTube video or write to hello@paperdive.ai. Confirmed corrections are added to the video description's correction block and to the episode page, dated, with the timestamp they apply to. We do not silently edit audio that has already been published.
Contact and where to listen
hello@paperdive.ai · YouTube · Apple Podcasts · Spotify · RSS