Summary

Snipd found its direction by disproving its first thesis: users enjoyed a TikTok-like discovery feed, but they kept listening to full episodes and created Snips “like crazy.” The four-person, all-technical team shifted from social clips to a learning system that captures, summarizes, and preserves podcast knowledge; a triple tap on headphones converts the just-heard moment into a note. That behavioral evidence gives the business a precise job: stop listeners forgetting “99%” before reflexively starting another episode.
The wedge now resembles a structured knowledge layer over audio, with more than a million podcasts already processed. Episodes can receive transcripts, diarized and named speakers, guest bios, AI chapters, extracted books, author-linked appearances, timestamp-sourced chat, and Snips retaining the original text and audio. Dynamic advertising breaks fixed timestamps, so Snipd built fuzzy audio matching—“basically Shazam for podcasts”—to resynchronize whichever file reaches the listener.
Snipd’s model strategy is margin- and latency-aware: self-host some cost-sensitive transcription and diarization models, then route higher-level work among OpenAI, Gemini, and Perplexity according to the minimum intelligence required. Perplexity search was quoted at roughly $5 per 1,000 queries, materially more than an ordinary LLM call, while Claude 3.5 Sonnet is admired for its formulation and personality but its cost leads the team to use other models for many workloads. Predictable preprocessing runs continuously; bursty user requests stay on APIs, which Kevin compares to “the AWS of AI.”
The defensibility lies in production reliability rather than the ease of a demo: episode chat can be prototyped in 15 minutes, but reaching “99% of the time” requires prompt formats, frontend rendering, heuristics, and countless regexes. Snipd supplements startup-style “vibe evals” with an LLM-as-judge pattern—generate five candidates cheaply, then let a stronger model choose—used for quotes, books, and speakers. Hosted closed models currently win because iteration velocity matters more than uncertain open-model savings.
The consumer opportunity is AI that disappears into the workflow, because “you can already ask the chat” is not equivalent to opening an app and finding the desired personalized summary already rendered. Kevin keeps returning to the job to be done: a simple interaction such as clicking a transcript word and jumping to the audio can matter more than model novelty. His endpoint is that “AI is the electricity of the future”—eventually Snipd should simply be a podcast app whose intelligence is assumed.
Voice is Snipd’s proposed learning loop: when an episode ends, a two- or three-minute conversation could force the listener to choose one takeaway and connect it to action. More than half a billion monthly podcast listeners already possess the listening habit, so Snipd need not manufacture a new one with Duolingo-style owls, streaks, and notifications. Kevin believes this might “10x” the value of an episode, while explicitly leaving the detailed interaction open for experimentation.
Expansion spans content, discovery, video, and creators, but distribution incentives will determine how much of the stack Snipd can own. Kevin wants native audiobooks, YouTube, AI-generated material, recommendation systems users can talk to, and “backgroundable video,” whose principal advantage is discovery even when roughly 90% of consumption accompanies another activity. swyx’s counterproposal is creator tooling—record, press one button, finish—because YouTube is already the strongest podcast platform; hard migrations and Substack-hosted Apple Watch playback failure expose both the opening and the platform risk.

Deep dive

1. Demand, not the hackathon trophy, created Snipd

Kevin studied at ETH Zurich in mathematics and economics, specializing in quantitative finance, but his retrospective signal of weak fit was revealing: “I never read an academic paper about the subject in my free time.” A friend sent him machine-learning lecture slides; after one weekend, his reaction was, “Freaking hell, like this is it. I’m in love.”
He hunted for excuses to apply machine learning at his bank, then concluded that pursuing it required “a real cut.” Kevin quit and spent five years building an AI team at an early-stage Zurich startup, delivering bank-sales models and systems that converted cryptic transaction booking text into readable merchant descriptions.
The same friend later joined him for HackZurich, where they built natural-language search within podcasts. Their stage demo searched “smoking weed” inside the 2½-hour Joe Rogan episode and found the exact Elon Musk moment. Winning supplied “activation energy,” but the stronger validation came when other participants immediately asked, “Can I use this?” and described adjacent problems.
Snipd remains unusually lean: four technical people, split between two backend/AI engineers and two frontend engineers. That team supports iOS, Android, and an Apple Watch app. The product is freemium, with Kevin offering a link for a free month of premium access.

2. A failed TikTok thesis exposed the real job

Snipd’s first release arrived roughly 3½ years earlier, before ChatGPT and Whisper. The initial concept was a social, TikTok-like feed: some users would listen to full episodes, clip the best moments into Snips, and let everyone else consume those clips for discovery or as a means to an end.
The founders expected attracting Snip viewers to be easy and persuading people to create them to be hard, so they optimized aggressively for creation. Reality inverted the hypothesis: users liked discovering podcasts through clips but still wanted long-form audio—and “they were creating Snips like crazy.” Snipd consequently doubled down on capture, retention, and learning.
swyx had lived the underlying pain through an earlier “personal mixtape podcast”: note a timestamp and URL, later download the MP3, manually cut a five-to-10-minute clip, record commentary, and republish it. Kevin’s reframing was that podcast apps remain “repurposed music players” despite podcasts being “one of the largest sources of knowledge in the world.”

3. Every episode becomes a navigable knowledge object

swyx’s baseline requirement was uncompromising: Snipd still had to be a full podcast player, effectively a superset of Overcast or Apple Podcasts. He initially disliked its opt-in queue and download behavior, but found that visible Snip counts and their positions inside an episode became a useful proxy for audience interest.
For enabled episodes, Snipd transcribes the audio, diarizes and names speakers, generates a mini biography and image for each guest, creates titled chapters with descriptions, and exposes a clickable transcript. Selecting transcript text jumps directly to its timestamp—a simple interaction Kevin uses to illustrate why consumer value is not synonymous with AI sophistication.
Book extraction goes beyond detecting a title. An LLM reads the transcript, while Perplexity and other orchestration retrieve the cover, author, and description; Snipd then shows other episodes featuring that author. Earlier models over-selected frequently mentioned people such as Sam Altman and Elon Musk, so the team tightened the product toward genuine guest appearances.
A listener can keyword-search or chat with an episode, ask when a topic appears, request takeaways, and follow timestamp citations back to audio. Snips preserve the summarized insight alongside transcript and sound; runners can create them from the Watch, although Kevin said Substack-hosted podcasts cannot play there—even through Apple Podcasts—and that Substack did not seem to care despite outreach.

4. Dynamic advertising forced Snipd to build “Shazam for podcasts”

The only processing-scale number Kevin disclosed was more than one million podcasts. Some shows are handled automatically; others enter the pipeline after a paying user requests them. Transcription and diarization produce speech blocks, after which LLM orchestration identifies guests and assigns names to the appropriate blocks.
Dynamic ads make ordinary timestamps unreliable. Each playback request can return a different MP3 with advertising inserted according to factors such as IP address; if Snipd transcribed another version, every word-level timestamp after the insertion shifts, breaking chapters, transcript navigation, chat citations, and Snips.
Snipd therefore resynchronizes the listener’s audio against its transcript on the fly. The matching happens near the audio-byte level rather than through partial transcription, and it is fuzzy rather than dependent on exact equality. Kevin’s concise description: “We basically built Shazam for podcasts” as a side project needed to make the main product work.

5. The stack buys only the intelligence each job requires

Roughly 90% of the backend is Python on Google Cloud Platform. The mobile application uses Flutter and Dart from one codebase, with native work where Flutter does not reach, such as Apple Watch. Asked whether Flutter was a good decision, Kevin’s bounded answer was “up until now, yes”; the team has no plan to leave it.
Snipd began before GPT-3.5 Turbo and initially ran and fine-tuned open models itself. For transcription it used wav2vec 2.0, whose combination of transformers, continuous audio, and self-supervised learning convinced Kevin that audio would follow text’s trajectory—even if the desired product capabilities did not exist yet.
Today it uses Whisper and still self-hosts some models for cost-sensitive transcription and diarization, while most downstream work uses OpenAI and Gemini APIs. Perplexity handles tasks needing web search; Kevin cited about $5 per 1,000 queries and said Google search grounding was not cheaper enough to justify replacing a system that already worked.
The governing rule is not “take the best model,” but identify the intelligence level a task needs and buy the best price at that level. Claude 3.5 Sonnet is Kevin’s favorite for formulation, personality, coding, and brainstorming, yet its cost is still high, so the team opts for different models for many workloads. “Given that we do work with a lot of content, price is actually something that we do look at.”

6. Podcast structure makes diarization tractable, not solved

Snipd benefits from a bounded domain: podcasts generally provide high-quality, controlled, studio-recorded audio and consistent background noise. Kevin contrasted that with meeting-oriented voice products that must handle arbitrary rooms and conditions, where usable structural heuristics and in-domain data are far scarcer.
One heuristic is airtime: in a one-hour episode, a voice heard for only 30 seconds is probably neither host nor guest, but an inserted advertiser. The underlying system embeds speech segments and clusters them by speaker; Snipd has modified the clustering and applies podcast-specific rules unavailable to a generic diarization service.
LLMs combine transcript meaning with acoustic clusters to identify speakers and recalibrate transition points. swyx objected that LLMs can introduce errors and are unsafe as precise authorities; Kevin conceded plainly that Snipd’s diarization “is also not perfect,” though he said recent episodes had improved substantially over those processed a year earlier.
swyx nevertheless judged Snipd’s presentation and identification better than Descript’s in a difficult, multi-guest TechMeme Ride Home episode. Both speakers expect today’s specialized pipeline eventually to collapse into a multimodal model ingesting raw audio; Kevin tested Gemini 1.5 Flash, but said its cost remained far above the self-hosted pipeline.

7. Production quality lives in regexes, judges, and “vibe evals”

Episode chat demonstrates the distance from prototype to product. Pasting a transcript into ChatGPT and requesting timestamped answers can yield a demo in “15 minutes,” Kevin said; making it work correctly “99% of the time” requires carefully formatted context, prompt engineering, citation behavior, frontend rendering, and recovery from malformed responses.
Snipd uses “countless regexes” to repair formatting mistakes before they become ugly UI. A second LLM might sound more AI-native, but chat must stream immediately; the available APIs cannot conveniently stream incomplete output into another live corrective stream. swyx’s verdict was that this supposedly inelegant machinery is ordinary “real-world engineering.”
Kevin characterized much of the company’s evaluation culture as “vibe evals.” A startup can tolerate occasional roughness because users receive rapid iteration in exchange, whereas a Spotify-scale organization might subject half the same features to six months of legal and organizational review. The trade is speed, not a claim that failures have disappeared.
Snipd Wrapped exposed the tail problem: a model usually selected an interesting representative quote, but occasionally returned something inexplicably dull. The fix was to generate five candidates with a cheaper model, then have a much stronger LLM judge choose one. Snipd applies the same challenger pattern when validating detected books and speakers.

8. Consumer AI wins when the chat box disappears

swyx’s formal feature request was customizable summarization through a user-supplied prompt. Kevin noted that long-context personalization remains expensive if every summary must be regenerated for every user, although falling model costs should eventually make “everything” personalized.
Kevin’s immediate answer was technically correct: open episode chat and ask for a summary in the desired style. But he then accepted the deeper request—swyx listens while moving and does not want to initiate a chat. The real product should open with the preferred information already generated, structured, and visually consumable.
That distinction underpins Kevin’s consumer-AI thesis: builders must “move beyond the chat box.” Intelligence may already be available, but the opportunity lies in identifying the natural interface through which a person engages with it. His recurring discipline for AI-oriented founders is to ask, “What’s actually the job to be done here?”
The endpoint is invisible infrastructure. Kevin’s analogy was that “AI is the electricity of the future”: nobody markets a microphone or phone as electricity-enabled, and eventually AI-enabled will sound equally redundant. Snipd uses the phrase now because novelty attracts attention, but its desired destination is simply a better podcast product.

9. Voice can turn an episode ending into a learning trigger

Kevin evaluates consumer products by the real-life trigger that causes someone to open them: travel prompts Airbnb, but language study supplies no natural daily moment. Duolingo is the exceptional company that manufactured such a habit through notifications, streaks, leaderboards, and “the owl memes”; Kevin called it “the GOAT” of that game.
Snipd already inherits a trigger—being alone, walking, driving, exercising, or otherwise starting a podcast—but lacks one for processing what was learned. Kevin’s own forcing function is to choose exactly one takeaway after every episode, even when ten ideas were valuable, because the prioritization tests relevance and makes action more likely.
His proposed voice product starts when playback ends: rather than rolling into another episode, an AI companion holds a two- or three-minute reflection. More than half a billion monthly podcast listeners already have the preceding habit. If users feel that this tiny investment materially improves retention, application, and thinking, Kevin believes Snipd may “10x the value” they receive.
The implementation remains deliberately unsettled, and swyx initially resisted the framing as another “chat with your podcast” feature before recognizing the post-listening use case. Kevin also expects voice cloning to normalize: capabilities that provoked grave ethical debate at an earlier NeurIPS workshop now exist in mass-market tools, and an authorized “AI swyx” could eventually discuss Latent Space episodes.

10. YouTube turns discovery into a platform question

Kevin’s content roadmap extends beyond RSS podcasts: users can already add YouTube videos and upload audiobooks, while he wants native audiobooks and AI-generated programming that combines Deep Research with NotebookLM-style podcast generation. swyx pushed him to prioritize video podcasts over audiobooks because, in his view, YouTube is already “the best podcasting platform.”
Discovery should also move from opaque personalization to an algorithm the user can address. Kevin’s example was telling a feed: “I know I freaking love” cat videos, but for the next two hours show AI material instead. TikTok may lack the incentive to surrender engagement optimization; a learning product aligned with user goals does not.
Kevin adopted Spotify CPO Gustav Söderström’s phrase “backgroundable video”: roughly 90% of podcast consumption still accompanies another activity, even on YouTube, but visuals help with clips, slides, demonstrations, and connection to hosts. His largest reason for adding video is discovery—video recommendations are more engaging, while full attention remains optional.
swyx identified podcast creators as a third constituency beyond listeners and Snip makers. Riverside comes closest and Descript owns editing, yet neither delivers his desired workflow: “sit down, record, press the button, done.” The opportunity is paired with real switching friction—despite loving Snipd, he waited four to six months to migrate because OPML could not carry half-listened episodes, rankings, and accumulated state.

27. 315 Special: The Past and Present of Deal Hunters Apollo: Connoisseurs of Complexity - [Business Breakdowns, EP.208]