Apple Podcasts Was Never Built for Podcasts  🔗 **Source:** [Apple Podcasts Was Never Built for Podcasts](https://share.snipd.com/episode/e31eb597-451a-4fd0-a644-eb81baf920ee) ## Highlights - **Bike Commute Sparked SNIPD** - Kevin Smith discovered podcasts during a Zurich bike commute when Sam Altman's How to Start a Startup kept him learning without losing time. - He repeatedly struggled to capture or revisit insights mid-ride, which motivated building a one-tap bookmark feature that became SNIPD's seed idea. Transcript: Kevin Smith So really the moment was in 2016. I had just joined an early stage tech startup in Zurich, Switzerland. And it was my very first experience at a startup. So I had no idea how startups work at all. So what the founder did is he recommended this podcast to me called How to Start a Startup by Sam Altman back when he was at Y Combinator. And here I was in Zurich, Switzerland, on the other side of the planet of where the epicenter of startups is, that is San Francisco. Every morning I'm riding my bike to work and I'm listening to this podcast, getting the best advice from some of the greatest entrepreneurs that we've seen. Like people like, you know, Peter Thiel, Dustin Moskovitz, Paul Graham. People that if you would want to get their advice, you would have to move to the other side of the planet, get into the right room with them, get their attention, and then hopefully you Get some advice from them. Kevin Rose Next to impossible. Yeah. Kevin Smith And I'm doing all of this while not even losing time because I'm going to ride my bike to work anyway. So this was the moment when I realized like, oh my God, there's like podcasts. It's such an incredible, valuable source of knowledge that you're not getting anywhere else in this shape and form. But I then also started to encounter all of these problems. You know, you're riding your bike, you get this amazing insight from Paul Graham, and you're like, oh my God, I need to remember that. Yeah. So what do you do? You know, you get off your bike, you take out your- Pause it, yeah. Pause the podcast, open up the notes app, write it down. You know, all of this goes on for like a couple of minutes, and like, okay, I'm good. Get back on your bike, you know, keep riding. And it's like, oh, that's another great idea. Yeah, so your commute's going from 15 minutes to an hour and a half basically and your boss is like Kevin you're late again whoa why um and it actually continues then right like at some Point you know you get to work um of course you didn't write down all of the insights because you know as you said like you're not getting to work otherwise um at work you maybe want to share Some of these insights. ([Time 0:02:55](https://share.snipd.com/snip/b9551c9e-2214-491a-91b0-d4115b32d65f)) - **Podcast Apps Are Just Music Players** - Podcast apps are mostly repurposed music players and lack features needed for knowledge work like clipping, indexing, and discovery. - Treating spoken audio as a knowledge library enables new UX: searchable moments, snips, and AI-guided discovery instead of simple play/pause. Transcript: Kevin Smith Because with music, you don't need more. You know, you just want to listen to right to it but with spoken audio it's once you look at it as a knowledge library you start wanting to do all of these other kinds of things and now you Have these two three hour conversations you know like swedman goes to eight hours sometimes um where you you want to uh in with certain parts of it. You want to take that knowledge and do something else with it. And discovery becomes a whole new area that you can tackle differently than in music. So our vision was always how can we create the next generation spoken audio app. ([Time 0:14:01](https://share.snipd.com/snip/c24b6716-5dbe-4747-a647-8ac7b2807423)) - **Users Created Snips Faster Than Expected** - SNIPD launched with user-created one-to-two minute snips expecting low creation but high consumption; the opposite happened — many users created snips but didn't consistently consume them. - Short snips stacked back-to-back proved disorienting, showing that context and coherence (longer 5–10 minute moments) matter for audio-first consumption. Transcript: Kevin Smith Maybe there's a bit more context, I get the time, you know, TikTok was blowing up. So one of the concepts was actually we would have some users that create what we call snips. So these are bookmarks within an episode, like the moments where you, as earlier when I got off my bike. So basically, you save these moments, create these clips. And then there's a different set of users who are listening to snips for discovery, or maybe just because that's the new way of consuming some of this content. And we launched, we thought, it's going to be very difficult to get people to create snips, and it will be super easy for people to get people to listen, consume the snips. Then we launched and basically what happened was the exact opposite. Kevin Rose Yeah. I can see that. Everyone was creating these snips. Right. Because snips are personal though, right? Like for me, I'm like this point matters to me, but it might not matter to you. So in some sense, I don't want to listen to other people's snips necessarily. I find it interesting when I see heat around them and we can get into that as well. But yeah, sorry, continue. Kevin Smith Yeah, so exactly. People were creating all of these SNPs, but at the same time, it was difficult to get people to consume them. Yeah. The main difficulty that we've identified was actually less that the SNP is personal to you, because at the end of the day, we can actually see in the data that a lot of the moments they Resonate with similar people. With machine learning personalization, you can match the right users with the right snips. Kevin Rose Oh, interesting. Kevin Smith However, the problem is snips were like still today, they're like one and a half minutes long highlight clips. Now, if you're in an audio first setting where you're just listening, if you have one and a half minute clips that after one and a half minute, it jumps to a different highlight clip from A different episode, we're talking about a different topic with different context. It's extremely disorienting. ([Time 0:14:51](https://share.snipd.com/snip/c371c6b1-6d20-48d3-934d-ccabeacd33ea)) - **DJ Moderation Fixes Disorienting Clips** - The SNIPD DJ stitches coherent 5–10 minute moderated segments, introducing each clip so listeners keep context while jumping between episodes. - Combining user snip data + AI content understanding yields personalized, moderator-like playlists that feel like curated audio shows. Transcript: Kevin Smith So we call it Snipped DJ. A part of that is already live in the app. Part of it where we have currently as a beta feature activated. But basically what the DJ does, it finds the most insightful, most interesting moments of a podcast episode for you. It knows what you listen to. It knows what you snip. So it really knows your interest very well. Then finds those moments. And now we're talking about moments that are like five to ten minutes long not one and a half minutes so it's really a coherent section similar to a YouTube clip and it guides you through These moments like a moderator so the the DJ is not you know no no content is being copied no content is being summarized the DJ is like the moderator that creates this intro so cool hey Now we dive into the Huberman lab where they talk about this new finding. I've selected it for you because I know how much you're into the personal health stuff. You're going to hear, actually, Andy Galpin is the guest. And yeah, let's dive in. And we've seen some extremely positive feedback from our users so far. Yeah. Where it's very interesting how all of these ideas that we had in the beginning that didn't work are now coming back. And we're reapplying them and seeing like, no, now they work because the technology has advanced enough such that we can have a good AI voice guide you through. AI understands you much better. AI understands the content much better. And we have all of this data from our users who have been snipping these podcasts where we know these are actually the most interesting parts. So all of this is currently coming together. And we're almost going full circle, but with better technology now. ([Time 0:17:32](https://share.snipd.com/snip/9d1956fe-15c8-4e81-bdd4-6d31b9ea7e19)) - **Don't Build An Ad Skipper Build Revenue Share** - SNIPD refuses to build an explicit ad-skipper because podcast ads fund creators; instead they envision platform-level monetization like YouTube's revenue share. - As scale grows, SNIPD aims to be a net-positive by enabling payouts via subscriptions or platform ad-matching rather than removing creator revenue. Transcript: Kevin Smith No, I mean, this is a really good, really important topic. First of all, for years, we could have built a dedicated ad-skipping feature. It would be super easy. We know exactly where the ads are in all of the podcasts. We've specifically decided not to build a dedicated ad-skipping feature because we know this is the way that podcasters make money. We want to be a net positive contributor to the whole podcast ecosystem. Now, that being said, we do have some features where almost inadvertently you are skipping some ads. Like the DJ, the idea of the DJ was never to skip any ads. There's nothing in that feature, nothing in the prompt, nowhere that actually tells the DJ to skip the ads. But if it learns what you like, and if you don't like the ads, then it will not play those. Right now, the right way to look at it is that we're a small app and we're experimenting with this, and we're seeing what is the reaction of people. So right now, it doesn't really have any impact on any podcaster. But you mentioned the right thing, what happens if we get to scale, if this blows up? And I completely are of the opinion that the future of monetization in podcasts is actually to move to a model like how it works on YouTube, where YouTube as a platform, they pay out creators. Creators get a revenue share and now ads can be run out on the platform like YouTube does or you can have a subscription model like what we do. We don't add any ads to the app, but we have a premium tier. And with this, we can pay out creators. If you actually understand how all of this ads stuff works today in the podcast ecosystem, like it's so obvious that this has to be the future. ([Time 0:26:16](https://share.snipd.com/snip/82d1dc8a-ce62-4124-9125-93536e1bcfef)) - **Snip Heatmaps Give Creators Moment-Level Signals** - SNIPD captures which exact moments listeners snip and returns that granular heatmap to creators so they know which topics within episodes resonated. - Creators convert these snip-derived highlights into social clips and topic ideas, a data source unique to SNIPD versus Apple/Spotify/YouTube. Transcript: Kevin Smith Yeah, we're already circling that data back to creators, where once an episode has been out and our users have listened to it and snipped their most important interesting parts. We send that data back to the creators so that they know which moments resonated the most with their audience. And we are actually the only podcast app in the world that has this information. Like YouTube doesn't have it, Spotify doesn't have it, Apple doesn't have it. They can maybe tell you this episode was interesting to a lot of people but not like which topics actually within this conversation. Yeah. Kevin Rose The only thing I can think of is when you do a share on YouTube and you check the box that says share from this moment in time forward. That's probably an indicator of like this is a good start of something interesting. Kevin Smith Yeah. But that is something that like how often do you do that compared to actually when you consume a podcast. For us, our users, almost every episode that they listen to, they use the snipping functionality, especially also because it's so easy. You just have to tap your headphones so you don't have to get out your phone. And then the AI does the rest. So yeah, we're circling this back to creators. And what they're doing is they're using that as inspiration for social clips. ([Time 0:29:52](https://share.snipd.com/snip/4208944f-96c0-438b-942f-4553d00495c1)) - **Mix Models To Balance Cost And Accuracy** - Use a model mix: pick cheaper, fast LLMs for trivial classification and stronger models only for complex reasoning to control cost and latency. - SNIPD transcribes with Whisper, uses cheap LLM calls for book extraction, then stronger calls to validate and enrich database entries. Transcript: Kevin Rose You just jump to whatever suits the particular thing that you're trying to do? Yeah. Kevin Smith Yeah. So over the last couple of years, we've switched a lot of the models. So there have been phases where we used a lot of Gemini, OpenAI. We used some Anthropic. We also have actually still in production some perplexity stuff. At the moment, we actually think 90% on OpenAI. But it is truly changing depending on the models that come out, depending on the use cases that we have. Because, you know, not all use cases are, you don't actually want to use the best model for all use cases. So it would just be too expensive. Right, exactly. And you don't need that raw intelligence for every use case. Just to have one example, for podcast episodes, we identify all of the books that get recommended on an episode. This is actually a lot of different LLM calls underneath where there's one LLM call that actually looks at the entire transcript of an episode, identifies all of the books mentioned. For example, oh, this is Atomic Habits. Then there's a next LLM call that actually- That's a non-reasoning model. Kevin Rose It doesn't need to be reasoning. You can just find it. Yeah. So it's fast cheap it's fast it's cheap but that's not the transcribing model that was probably Gemini that transcribed it no there we're actually using whisper whisper okay yeah so Kevin Smith We have a whole transcription pipeline that we that we're running on our own GPU servers mm-hmm with transcription with speaker diarization So speaker diarization is identifying How many speakers there are and when does which speakers speak. Kevin Rose That's so cool. That's hard to do, I found. Yeah. Yeah. So is that getting easier? Because six months ago, that was really hard to do. Even I would drop in, you know, my audio into an editor that's out there. You know, there was like a bunch of these AI kind of editors out there. It would oftentimes say oh speaker number two is this person i'm like that you jumped the wrong speaker has that gotten pretty much like 99 accurate at this point or maybe 95 accurate Kevin Smith Okay actually just a couple of days ago there was a new model that dropped that i haven't tried out yet which one do you remember the name of it um think it was Microsoft. I have it. Okay. I'm happy to send it to you. Yeah, I'd love to check it out. Or if you want to check it out. Yeah. But I have it on my to-do list. I just saw it two days ago or something like that. Kevin Rose Isn't that the cool thing about this world? Yeah. It's like literally by the time this comes out, you're added to the show notes. It might be another model that leapfrogged that one. That's better, you know, every other week. It's crazy. Kevin Smith Yeah. But that is, yeah, that has been one of the challenges where, you know, that's also part of you as an app builder where you can go this extra mile of where are these models maybe not that Good yet. You can put in the work today to create a good experience by having the harnesses, by having the scaffolding around all of this. And maybe in two years, all of this will just be, you can just maybe one shot it. Maybe. But building all of these things around it today actually allows you to create a really great product and get that product and it just gets easier over time. Kevin Rose Yeah and so you're identifying the books you pull those out going back to that and then what what what where where from there do you go you've got that model. Kevin Smith We actually so there actually a lot of things are happening so basically you have this this mention of a book we now have another LLM call that just verifies whether this is actually a Book because sometimes these LLMs come up with weird things or like yeah this is is a book. Then we... Kevin Rose And that's a web tool-based model. Kevin Smith Exactly, yeah. So that one can research the internet and say, yeah, no, this is really a book. We go then to our database where we have an entire database of books that we bootstrapped. We find all of the books that are similar to this, basically like a nearest neighbor search. Now we give all of the results again to an LLM and ask, hey, is there a match between these? If there is, then we match these. If not, then we actually now create a new entry in the, or the AI creates a new entry in the database and fills all of the information and researches all of the information about the book, Like, what's the title? What's the author? What year was it published? What's like a synopsis of it? Finds good artwork. Yeah, exactly. Cover, yeah. And all of that, and then adds that to the database. And all of this is done and orchestrated by AI, basically an AI agent. And some of these calls can be very cheap calls, like using the cheapest GPT nano model basically. Right, totally. Kevin Rose Super fast, super cheap. Kevin Smith Yeah, and others, you want to have more intelligence to make sure this is actually correct, the database keeps clean. ([Time 0:34:12](https://share.snipd.com/snip/4361bec3-f3f7-4236-b8fb-458796c85777)) - **Keep Observability Simple At Small Scale** - SNIPD runs backend on Google Cloud and relies on simple GCP logging/monitoring with log-based alerts piped to Slack rather than heavy observability stacks. - This lightweight approach fits their small-team scale while capturing errors and weekly reports for processing health. Transcript: Kevin Smith So we just do, so everything runs on Google Cloud, like our backend at least. So you're all GCP on the backend. Yeah. And we're just using the plain old logging. If something fails, there's a log. And everything on top of that is just custom built, very simple. Like Graphana or anything? No, it's just basically the GCP internal monitoring tools and log-based alerts. And that then just sends messages into Slack, triggers something. So it's very straightforward and simple. ([Time 0:40:16](https://share.snipd.com/snip/a5ac6526-a99a-4fac-91db-107424837500)) - **AI Voices Still Fail Long Form Listening** - Current synthetic voice models (e.g., ElevenLabs) impress in short bursts but fatigue listeners over longer listening, making fully AI-hosted shows unpleasant. - Natural-sounding speech needs imperfect inflections and conversational quirks; SNIPD sees voice quality as underinvested by big labs focused on enterprise agents. Transcript: Kevin Smith So, you know, we started the conversation saying that our vision is to build the AI native spoken audio app. And maybe to take a step back here, like what, you know, we now spend most of the time talking about podcasts, which today we are a podcast app. We also mentioned audiobooks, but there's another category of spoken audio, which I think will be huge in the future, which is AI generated spoken audio. Today, you have to be a bit careful when you bring up this topic because right now we're in the era of AI slop. Most of the content that's created by AI. Kevin Rose I think we're going out of that, but yeah. Kevin Smith Exactly, exactly. So this is, I believe this is temporary and we're now slowly coming out of the other end. Kevin Rose By the way, I have a thesis that we will never even remember that there was ever AI slop. Because the window is such short in the grand scheme of things, we'll never even remember it. Do you know when the iPhone came out, the only way to get an app on the iPhone was to install a Safari shortcut on the actual iPhone main screen? Because there was no App Store yet. And so you had to do these mobile little, and they made all the UI disappear so it would act like an app. And this is before the app store. And they did this for probably six months before they announced the app store. No one remembers that. Like, you didn't remember that, did you at all? Like, you know, it was a small little brief blip in the grand scheme of what we now consider to be apps, right? And I just have a feeling that AI Slap is going to be the same way. We'll just like not even remember that it was bad. It'll just like be considered to be, because the vast majority of people haven't even used it yet. Kevin Smith It's a good point. It's a good point. I mean, the question is whether AI Slop will go away. You know, I can see a future where even in the future, even the models getting better and everything, that there will still be content out there that just because it's so easy to produce Um that some people will still do it and it will linger around uh in certain areas of tiktok and whatever um so i can also see that error but i think the the concept that people think anything Created by AI is bad, yeah, that concept will go away. Kevin Rose Yeah. Yeah. When you talk about some of these models coming together, one of the things that jumps out at me is, and I actually predicted this feature but didn't have the resources to build it, which Was with Notebook.lm, you can interrupt a podcast and ask questions of it. Yeah. And I was like, this has got to be the future. Because when Tim's talking about something, or whatever, Sam Harris, or you name it, and you're like, wait a second, what about this? And you interrupt the hosts, and so you're actually kind of in the room with them, and then they go off and grab the research, bring it back, answer your question, and then resume the timeline. Is that in your roadmap? Kevin Smith It is in our roadmap, but so we actually had an internal prototype for this a year ago. We tried it out and we got to the conclusion that right now it still sucks. The experience, like this is for me a typical feature that like in your head, it sounds amazing. But when you actually experience it, it's like most of the time, you're like, what am I even asking? Like, it doesn't happen very frequently. And then when you actually ask it, currently at least, the AI model's answer, it doesn't feel good. Like, basically, what you want. Kevin Rose It's not well-researched because it's a lightweight model is coming back to you with a snap answer. Is that what you. Kevin Smith That is one of the problems. But it's also a bit the, like, you're imagining is talking to Tim Ferriss. Right. That's sort of like when I hear users ask me for this feature, it's like, oh yeah, I get to talk to Tim Ferriss. Like, no, that's not what you're getting. Right. Talking to another person who basically listened to the episode. But the bigger problem right now is that these models, they are really not good at talking like a human. So you can actually really, really feel, and that's also still a problem today, you can really feel how all of these models have been trained to create text content. And it's very difficult to get them to create content that is like speech. I've also heard from some sources that there's not really a post-training step for these models that focuses on this. And it's just not a big money driver today. And I've spoken to some people who are doing stuff in this field, they have multiple calls that are doing nothing else but trying to reformulate the content that the LLM has created. Right, right, right. Okay, now make it sound more like a human would actually talk. Kevin Rose Right, but also you don't want to sound too human because then when you realize it's AI, it sounds fake. Like I'll give you an example. Like sometimes when Notebook Lbook LM comes back to you with a podcast and they're trying to banter. And one of the AI agents is like, I hate it when that happens. I tried that last week. And you're like, you never tried it. You're an AI. Don't act like you've tried this. You know, like there's stuff like that where it's like they're trying to be too human when it's just like, just get to the facts you know yeah that look that that i get but otherwise i don't Kevin Smith Agree with this because i think the more it's the more human it sounds like actually the like the i'm talking about the intonation the inflection point yes yes yes um the imperfectness Yes which is a perfect word because it's probably not a word right exactly lm would have not said that. This is what I'm referring to. Like, yeah, don't act like, yeah, oh yeah, I've done that last week. Kevin Rose That's probably an easy solve. Those are hard to snip out. Kevin Smith But these aspects is something that all of the voice models today, they are really bad at still today. I's like, I like saying it like, for example, 11 laps. You listen to an 11 laps voice for 10 seconds and you're completely blown away. Like, oh my God, I would have never thought this is not human. And then you listen to it for 30 seconds and you're like, huh, okay. ([Time 0:43:14](https://share.snipd.com/snip/e960fce1-e577-473e-b9a2-f2bc650e4e9d)) - **Open An API For Users' Personal AI Agents** - Build open connectors so users' personal AI agents can access listening history and clips; integrate via APIs/CLI for Claude, Codex, etc. - SNIPD plans an API so agents can query 'find moments I listened to about peptides' and return episode links and summaries. Transcript: Kevin Smith So I'm 100% sure that this is actually going to happen. And not just for people like the, it's called the nerds and geeks like us in the, you know, who want to have these PKM systems. But I think it will become a mainstream normal thing that everyone has this second brain. They might even not call it that. It's just your personal AI agent. So it will be very important that your personal AI agent has access to all of the data. And so we are thinking about how can we create this interface between SNP and your personal agent in a much more seamless way. So we're working on an API and command line interface that your AI agent can use to have access to all of the podcasts that you've listened to. Kevin Rose So you'll do connectors for Claude, and there'll be just an official way to tap into that. Kevin Smith Exactly. OpenClaude, Claude Code, Codex, all of these tools. So you'll just be able to go there and be like, hey, wait, I actually have already listened to various podcasts about peptides. Can you find those moments again for me? Can you summarize it? Can you link to the actual episode? ([Time 0:53:37](https://share.snipd.com/snip/57404f24-918c-40cf-869c-c95392eb0362))