Cairn
A static site generator for podcasts
Cairn turns a feed into a site where every episode has a digest, topical sections with specific points, and the verbatim words behind each one, with a play button that plays exactly that stretch of audio. Search the full transcript, ask about anything you hear, and save an episode for offline listening.
Every episode page works at three levels of zoom. Start at the top, or drop straight into the words.
Five to seven sentences. Each one a claim the episode makes, so you can decide whether to go further.
Topical sections with a short summary, and the specific points made inside each one, attributed to who said it.
Every point has a play button behind it, and the exact quote it comes from, over a speaker-labelled transcript you can search.
From the newest episode · Latent Space: The AI Engineer Podcast
John Platt says many scientific problems can be recast as scorable tasks, where you want a piece of code that maximizes a score, and ERA works because the thing mutating that code is an AI with broad prior knowledge.
The episode opens with a clip from later in the conversation about predictive versus descriptive models. Brandon then introduces co-host RJ and guest John Platt: college at 14, a Caltech PhD co-advised by John Hopfield, Platt scaling and SMO, two asteroids, an Oscar, and a decade of Google work on fusion, quantum and climate.
Brandon welcomes listeners to Latent Space Science with co-host RJ and introduces John Platt, a Google Fellow and head of applied science at Google Research, who calls himself a "giga nerd".
Brandon1:14Welcome to Latent Space Science. Um, I'm Brandon, uh joined by my co-host RJ. Uh it's a pleasure to have John Platt with us today. John is a Google fellow and head of applied science uh Google Research. He has really a…
Digests and section summaries above are written by a language model and paraphrase what was said. Quotes and the transcript are verbatim: the page assembles them from the transcript by segment index, so a quote can never drift from what was actually said.
Every quote and every line of the transcript has its own play button. Press it and hear exactly that moment, not the whole episode.
Every line is tagged with who said it. One search box finds it, in the points and in the full transcript.
Claude never writes the words inside a quote. It cites a range of transcript segments, and the page joins them, so a quote is always what was actually said.
Pause the audio and ask a companion that already knows the episode and the last minute of what was said. It answers in text and, if you want, out loud.
The page keeps your position and what you have heard. The show page shows how far along you are on every episode.
A service worker keeps the audio and the page on your device, so you can keep listening and reading without a connection.
The newest episodes across every show.
John Platt says many scientific problems can be recast as scorable tasks, where you want a piece of code that maximizes a score, and ERA works because the thing mutating that code is an AI with broad prior knowledge.
Nathaniel Whittemore says Meta's Muse is the first personal agent to catch on with ordinary consumers outside tech circles, and that it climbed past ChatGPT to number one on Apple's free app chart.
JS Denain says the acceleration evidence OpenAI and Anthropic have published, such as researchers' Codex spending roughly doubling each month, shows real value but is not strong evidence of a software intelligence explosion within six months.
Diogo Almeida argues that the field collapsed onto one model shape — autoregressive, chat-tuned, human-facing — and that Jev is a different class of model whose intended consumer is code rather than a person.
Alexander Mattick tells the history of generative modelling as a question of where the cost of inference is paid, from rigid variational families and slow sampling to diffusion and flow matching, which move that cost into training.
Zvi Mowshowitz says the real story behind the pacing debate is that people at the labs genuinely see dramatic improvement in internal models and are freaking out about it.
The shows above are everything processed so far. If there is an episode or a show you want, say so below and it'll be queued up.
Thanks, got it. I will email you when it is up.
Each episode goes through six stages, running locally on a GPU with NVIDIA Parakeet for transcription and NVIDIA Sortformer for speaker diarization, both through NeMo, and Claude for the outline.
Read the show's RSS feed, save the episode metadata, and download the audio.
Sentence-level segments with timestamps and speaker turns, encoded down to a small web-size MP3.
When the show publishes its own transcript, align it word by word and take its wording and names, keeping our timestamps.
Claude reads the indexed, speaker-labelled transcript and writes a headline, a digest, sections, and points that cite transcript ranges.
Verify that every section and point cites a real, non-overlapping stretch of the transcript.
Render the site: episode, show, and home pages, ready to deploy.