Speaker 1: 0:00

What is AI? What is consciousness? One thing that I am worried about is and as soon as it's larger than zero, we should, If you're very early, optimize for learning.

Speaker 2: 0:22

Hello and welcome to another episode of the Accelerometer. I'm Dr Matthew Merman, ceo and founder of Anarchy. Today, we're going to be discussing AI and media consumption with Kevin Smith, the founder and CEO of Snip, an AI-powered podcast platform, which I guess is probably going to be one of the places that you're going to be able to consume this podcast. Kevin, can you tell us a little bit more about Snip?

Speaker 1: 0:49

You said it correctly, snip is an AI-powered podcast platform, but, very importantly, it's for people who listen to podcasts, predominantly to learn something interesting. We use AI to transcribe podcasts, create high-level overview summaries, break them down into chapters and identify highlights within an episode. By far the most important feature that we have, the one that really sets us apart from everything else that you get out there, is what we call Snipping. Basically, whenever you hear some great insight in a podcast, you can tap your headphones and our app summarizes what you just heard and saves it in the app as a so-called Snip. Now you can later come back to these, review your Snips, you can export them to your Notes app or you can share them with friends or on social media. Finally, you can actually discover great new podcasts filled with insights by browsing all of the Snips created by our community. Why just podcasting?

Speaker 2: 1:55

Why not use this in your daily lives?

Speaker 1: 1:59

So the answer to that is focus. So as a startup, it's very, very important to have a very narrow minded focus on one use case, and this was for us from the start podcasts. But in the future we do want to branch out and what we actually want to do is we want to become the number one spoken audio platform out there for people who listen to this content to learn something interesting. So in the future, there might be more and more use cases that we will incorporate.

Speaker 2: 2:32

What gave you the idea to build this?

Speaker 1: 2:35

That basically started at my last job. I started at an early stage tech startup here in Zurich and at the time I had no experience in the startup world. So one of the founders actually recommended this podcast called how to Start a Startup by Sam Altman, back then CEO of Y Combinator, and I started listening to that podcast in the morning while riding my bike to work. And basically you have to imagine, I was this kid, very inexperienced in the startup world, very inexperienced in the startup world, no connection to any of these great entrepreneurs out there in Silicon Valley. But there I was riding my bike listening to the best of the best give me their advice. So you know, people like, of course, sam Altman, reed Hoffman, peter Thiel, mark Andreessen, dustin Moskowitz all of these people talking to me directly in a podcast and I was sort of this fly in the room learning from them while being a nine hour flight away from Silicon Valley, and this experience just blew my mind. So that's basically how it started, how I got into podcasts, and then over time, I just started listening to more and more podcasts, but with that I noticed all of these issues that came up. So the most important thing was you know, I'm riding my bike here are great insight that I want to share with the team, but I have no way to write it down.

Speaker 1: 4:10

So the next day I remember, oh yeah, there was something. So I go back to my podcast player, try to find it in the episode. You know, use that skip 20 second button like 50 times and finally you found it. Then you want to share it with your team. There's no way to just share that one insight, so you share the whole episode and of course, and of course, no one listens to it.

Speaker 1: 4:35

Yeah, and then afterwards it's about discovering new content. It's still very painful to actually find a new podcast to deep dive into, invest these two hours, because it's just not skimmable. And all of these issues came together and the moment where I realized, hey, actually this is something I should tackle was when Facebook research AI group came out with their wave to work paper, where, basically, they brought the transformer architecture and self supervised learning into the speech AI space. And this was the moment where I thought, okay, nlp is going to continue to explode which it obviously has and speech AI is going to take the same path. And with these two things in place, you can actually now completely reimagine how you interact with spoken audio content.

Speaker 2: 5:37

What was the first thing that you snipped, oh?

Speaker 1: 5:40

wow, that is a great question.

Speaker 2: 5:43

I feel like that's like the dollar, that you hang, like at your counter when you open a restaurant, your first ever dollar made.

Speaker 1: 5:53

I have to be honest, I can't answer that question, but if I'd had to guess, it could be Sam Harris as a guest on the Lex Friedman podcast, where they talked a lot about AI and the future of AI, more from a philosophical perspective.

Speaker 2: 6:13

So do you remember what it was about?

Speaker 1: 6:18

Again, I'm just guessing here, but I remember that the conversation was a lot about what is AI, what is consciousness? Can AI be conscious?

Speaker 2: 6:31

There's like a deep insight into this question, of a deep enough insight for you to snip.

Speaker 1: 6:37

Definitely.

Speaker 2: 6:38

Okay, is AI conscious?

Speaker 1: 6:43

Wow. So I would say today's AI models, the public ones. We never know what open AI is cooking behind the, behind the curtains. I would not call them conscious.

Speaker 2: 6:58

Do you ever worry that, like the AIs might be conscious and you're actually like enslaving, like a little person to go and listen to these podcasts on repeat, over and over and over again and like be enslaved to snip things all day along?

Speaker 1: 7:17

I do sometimes wonder whether the current models that we have today are conscious just on a different time scale, and you're like micro consciousness. Yeah, sort of Um. But to answer your question, um, I would imagine that they enjoy listening to all of these amazing podcasts.

Speaker 2: 7:43

How many times you really enjoy listening to a podcast, though? Right, a lot, a lot. You can just listen to them on repeat.

Speaker 1: 7:50

Definitely yeah, the best ones, yes. But I mean usually usually I do listen to uh, two new podcasts that come out, but I have a couple of favorites that I do re listen. Either I re listen to the full episode or I just re listen to the moments that I've snipped. So what's your favorite podcast? So my all time favorite podcast is legs freedman.

Speaker 2: 8:18

I'm sad that it's not this one the one that hasn't come out yet.

Speaker 1: 8:21

Yeah, so I mean um, basically, over the short term it always changes a bit which one is my number one podcast. So the last two, three weeks, I think my number one is probably acquired Um of the last three months it might be the all in podcast, but all time it's definitely legs freedman. And then of course there's a future and, who knows, maybe this one will be the number one.

Speaker 2: 8:50

Hopefully, so we're working towards right. You ever think about starting your own podcast?

Speaker 1: 8:57

I have thought about that and, in a way, I have recorded a couple of podcast episodes, but these were just small announcements that we released inside of our app for our listeners, if we had a new great feature that we wanted to tell them about. Um, but one important thing was going back to your earlier question as a startup, it's very important to have focus and right now what we're building, we're very much focused on the listener, the podcast listener, and not the podcast creator. So this is why we haven't invested as much time getting to know the full process, the pain that you're going through setting all of this up and editing this conversation and creating this amazing piece in the end, and we focus much more on the listening side of things.

Speaker 2: 9:53

I guess it could also be a conflict of interest, right, you know you'd be competing with the people on your platform.

Speaker 1: 10:02

Um, yes and no. Yes and no, I mean, if you, if you look at Spotify, which is currently still the biggest podcast platform out there, a lot of their product managers actually create podcasts. So I think, if you're a platform that is focused on um, on creators, on podcast creators, I think it's actually very important that you create your own podcast to learn what it means, what your customers are going through.

Speaker 2: 10:35

I guess, like Netflix, isn't entirely neutral, is it right? Netflix has its own uh shows on Netflix.

Speaker 1: 10:43

Definitely, definitely. So I think there's actually a big difference there. So if you rewind, let's say two years ago, there was actually a moment in time where people thought it was moving into a Netflix style content war between the platforms. So Spotify very famously created a licensed deal with Joe Rogan. That's why Joe Rogan is only exclusively available on Spotify and they may deal with more and more creators. However, what our bet was what we saw happening is actually there's a huge difference in creating a podcast or creating the next game of thrones. It's actually much more difficult and there's much more red tape that you have to get through to be able to create a game of thrones than to create a really high quality podcast. So our prediction was always if you take a podcast away and make it exclusive to some podcast, to some podcast platform, there will be a new player coming up that will fill that void with exactly the same quality of content, If not even better.

Speaker 2: 11:58

You ever think about filling like creating AI podcast if you're focused on the listener, like one of the goals is to create or to have content for the listener to listen to. So why limit yourself to only human podcasts?

Speaker 1: 12:14

So Two avenues here. The first one is what is a podcast? And the second one is if you take the podcast that we have today, why do you even listen to them? So maybe let's start with the first one what is actually a podcast? So if you just look at AI generated spoken audio content, then we actually already have that on our platform. So, as I mentioned earlier, we're creating these high level overview summaries of podcasts to help you to get an understanding of what is an episode about before you dive into and invest three hours, and afterwards you can review key insights and these overview summaries you can also listen to. So this is already a generated spoken audio content.

Speaker 1: 13:03

However, this in no way actually competes with the actual podcast. And now when you say, okay, why do you not create AI generated content that actually competes with this podcast? The other question is why do you listen to podcasts? And for our users, we are very much focused on these interview style podcasts where there's a person who, in some shape or form, is interesting to the listener. If you replace that person with an AI, you have to answer the question why is that AI interesting to the listener? And that is something that in today's world, I do not see solved yet, and in my opinion, what's missing is that an AI is in some shape or form anchored in the reality that you care about as a listener.

Speaker 2: 14:04

And that's actually what we need right now is for chat you to come into the conversation and replace me.

Speaker 1: 14:11

Well, I would also disagree with that, because now you can ask, why do listeners keep coming back to the same show? So most often it's that they build up a relationship with the host, and there it might be the case that in the future you can create an AI. It doesn't mean that it will raise up that relationship, but it really has to be a full blown personality that you can get to know over time. And this doesn't seem to be solved yet, but quite some startups are working on that. So let's see what happens in the future.

Speaker 2: 14:52

I believe one of the startups that I interviewed one of the earliest startups that were on my podcast open souls specifically working on that, creating like ai's that look like consistent humans, like you know, you kind of have that at the moment, right now, right with like, you have different ai's, have different personalities, like if you were to have, like, a llama podcast, it would sound different than a podcast that says no, I don't want to ask you questions, I'm only an AI.

Speaker 1: 15:29

That. That is true, that is true, I mean. Now the question is why do I care? So, if you interview chat GBT about a certain topic, why do I care If it's purely informational, because you know chat GBT has a lot of information? Um then, why not just immediately create a lecture, which is a very different concept, a very different format than actually this interview style podcast, where a lot of the time it's really about getting to know the person.

Speaker 2: 16:05

Yeah. So, with your assistant, did you choose to focus on the interviews because they were what you most liked listening to, or was there something deeper about these interviews that spoke to you?

Speaker 1: 16:19

It's the. It all starts for us with listening to podcasts to learn something interesting, and then it's basically the decision of our users what podcasts do they want to listen to, and that's where they converge to.

Speaker 2: 16:33

So did you try putting like lectures on your platform?

Speaker 1: 16:37

Uh, we haven't yet no.

Speaker 2: 16:39

Okay, I look forward to like seeing like the Feynman lectures on the, because that would also be a useful thing to snap right 100%.

Speaker 1: 16:47

So I mean we get a lot of requests from our users to add additional audio content to the platform, which is also why we do want to branch out in the future and we see a lot of value in adding, for example, a Feynman Feynman lecture would be very, very interesting.

Speaker 2: 17:04

Um, what are you most looking forward to in AI in the next couple of years?

Speaker 1: 17:10

So, from a personal perspective, one element I'm quite excited about is the personal assistant, and really a truly personal assistant. This is something that you can already see a lot of glimpses off what basically chat GBT is doing, but there is still a lot missing. Basically, it doesn't have access to your data. It cannot really take actions for you, so it's not connected to all of the applications that you care about. So these two things together is also why my prediction is that actually one of the best players positioned in the market for this is Apple, because they already have your trust. You trust them to give them access to your data. All of the apps are on their platform. So if they offer something like that and say here's an API all of the apps can integrate into that, they actually have this market position to get them to do that. So that's yeah, that's one use case that I'm quite excited about.

Speaker 2: 18:29

And for you guys, what was the most difficult part about working with AI?

Speaker 1: 18:35

So this depends on which time so the pre chat GPT era or the post chat GPT era? Pre chat GPT, it was definitely the fine tuning, the LLMs and hosting them, and with the fine tuning there it's really getting it to mainly evaluate the output, whether it's consistently good across a large spectrum of, in our case, podcasts and topics. And then, of course, the hosting is much more complex and cumbersome than the calling an API and I would say post chat GPT, the most annoying thing so far were handling rate limit errors, handling downtime of open AI or any kind of API errors that you get, and creating the experience in the app that can take care of these errors.

Speaker 2: 19:53

What's something about the future of AI that you're worried about?

Speaker 1: 19:58

One thing that I am worried about is super intelligent AI. I'm also super excited about it, nick Bostrom style. So I'm on both sides. I'm super excited, but I also see the potential danger with it. Basically the whole concept of if there is an entity if you want to call it conscious or whatever, but it's just an entity that has some kind of goal, that came from somewhere and it's 100x more intelligent than us, then basically I see the potential of it creating a relationship to us similar to the relationship we have with animals. And then it's not about we humans, for the most of it do not purposefully kill animals, but over time, nonetheless, most of the species have been eradicated because we wanted to build some houses.

Speaker 2: 21:04

So you think that we are going to become building construction materials for the AI? I hope not. And a war of the world is a esch fear.

Speaker 1: 21:17

So it's not my prediction, but I do see that the probability of this happening is larger than zero, and as soon as it's larger than zero, we should pay attention to that.

Speaker 2: 21:32

So you guys were training models before GPT came out? Right when you were training models, were you ever thinking like we need to do something or create some company-wide ordinance, that if the AI has become too smart, we need to stop them?

Speaker 1: 21:51

No, no, not at all. I mean, we were not training models from scratch, so we were just fine-tuning existing open source LLMs and I wasn't afraid that the fine-tuning would make them suddenly conscious and make them try to take over the world now.

Speaker 2: 22:12

How do you want to approach AI safety in the future?

Speaker 1: 22:16

Do you mean us as a company or, more, us as humanity? Both? So for us as a company, the good thing about our use case is that it is relatively straightforward what the AI can do, and there I don't see the big risks that we were talking about earlier of an AI becoming the dominant species in our world. So we're quite independent of that.

Speaker 2: 22:52

I mean, if you guys are a platform that has summarization, you can imagine that a smart enough AI could summarize with subliminal messaging, right?

Speaker 1: 23:01

That is true. So this is basically the AI safety that we need to worry about. Are these summaries in line with what was communicated in the podcast? And one great thing here is that we actually have our users who listened to the actual content. They are the ones who are reviewing the SNPs, they are the ones who are then posting it in the community, and if they feel that this summary is not in line with what was said in the podcast, then they won't do that.

Speaker 2: 23:31

It can be in line and still like I mean we talk about translators having different interpretations of the underlying material, like different translations of the Iliad, like they're ranked, even Like that's a really good translation because I like his modern take on it right. So you could do the same thing with a summary, though, right, and you can have an AI like make us all more woke with its summaries of some content right.

Speaker 1: 24:05

So one beautiful thing in our case is that you have the original content always a click away, and that's actually what we encourage, right? So if you browse through these SNPs on our platform, it's actually to discover great podcasts to then deep dive into. So as soon as you do that, you're going to listen to the original vogue or non-vogue content.

Speaker 2: 24:31

So let's maybe talk a little bit about your startup journey. What's been the most challenging part of building a startup for you?

Speaker 1: 24:40

By far the most challenging part that I did not expect was the emotional roller coaster ride. I remember when I started off, one of the first founders that I knew that I spoke to. He actually warned me about this and I thought what are you talking about? But it's really this especially if you're just starting out, you have an MVP. You put it out there and you get your first user and he tells you this is amazing. You know you're on top of the world. You think, okay, we're going to conquer the entire world by next month. And then the second user comes in and he's just this is crap and your whole world falls apart.

Speaker 2: 25:26

Your first user thought it was amazing and your second user thought it was crap. That's the first time I've heard that one.

Speaker 1: 25:33

Well, I mean, that was just an example, a hypothetical example, but basically this up and down from for us it was mainly the user feedback and you know, then you get 10 positive ones, but then the 11th one is something negative and sometimes it's just, it's just the right amount of truth in there that you know deep down is true and you think like, oh, my God, yeah, we know this, this we need to improve or this is just bad, and then you start to question everything. So this I would say in the beginning that was the toughest and learning how to deal with that that it's okay that one in 10 people find your product crap and remind yourself wait, we just got nine out of 10 super positive reviews who love our app. Let's focus on those people.

Speaker 2: 26:31

With my last company, we're building a social mobile local app and our primary users were friends and like just the constant negative. Anytime there was the slightest bug. Suddenly all of our friends were just messaging us on every platform and like you've got a bug here. I was like I know.

Speaker 1: 26:48

I know.

Speaker 2: 26:50

Like how could you have a bug here? Like what are you doing to us? Just like, just made a mistake. You ever get that.

Speaker 1: 26:59

Yes, definitely, definitely from friends, from users. You know some one thing that we often hear from from our users new users who come in they are quite surprised when they actually hear that we're just a small team of five people. And then they start to understand ah, okay, I realize why you don't have all of the possible features that could ever be built, already implemented perfectly in this app. But until they do, it's like why don't you have this feature? Why don't you have playlists? This is the most obvious thing that you should have, so why don't you have playlists? It always goes back to focus. So we're a small team, we can't do everything, so we have to work on the one thing that we believe will have the biggest impact first, and then it's the second, and there's the third, and playlist is on the roadmap.

Speaker 2: 28:00

I'm not even using the app yet, and now I'm like I want playlists.

Speaker 1: 28:05

That adds your list of users. You want the playlist to me too.

Speaker 2: 28:08

Okay, and they're on the roadmap. Do you have an API open?

Speaker 1: 28:12

We don't have an API now.

Speaker 2: 28:14

I'm like if you have an API, then you can get other people to make your features for you and integrate them back into the app.

Speaker 1: 28:20

Yeah, so we do. We do have some integrations. We have some integrations with no taking apps, second brain applications, so, for example, notion or read wise, and we are thinking about adding a something like a Zapier API, such that you can build some stuff on top of that. But we have no plans for more integrated API.

Speaker 2: 28:48

Is this your first startup? It is, it is. Have you always wanted to build a startup? I would say.

Speaker 1: 28:55

I've always wanted to create. This is something that now, looking back, I realize that in everything that I've done, this was usually the driving force where my energy came from. This was also why I told you earlier, I joined this early stage startup here in Zurich. Before that, I was actually working for two years for one of the big banks here, and I just didn't have that feeling of creating something valuable, and that drove me away to a startup where we're building a product.

Speaker 2: 29:32

I bet. Working for a big bank, though, you know you created a lot of alpha and revenue for somebody.

Speaker 1: 29:40

For somebody somewhere maybe. I mean for me one of the shocking things after leaving the bank and working for a startup and starting to understand how to build a startup, that you actually have to at some point create value for someone, I actually realized how much of the work we did was wasted time.

Speaker 2: 30:09

I think that that's, for me, like one of the most satisfying parts of building a startup that it's not. It doesn't feel like wasted time. Even if somebody's not using every second that you're doing something is, it's for a reason.

Speaker 1: 30:21

Yeah.

Speaker 2: 30:22

What gave you the conviction to finally like jump ship, I guess, and starting to start up, or had you already left your previous job?

Speaker 1: 30:30

Let me say there's, in my opinion, there's never a perfect time to start a startup. I think it's very similar to there's never a perfect time to have kids. At some point, if you, if you feel that inside of you you have this, you want to create something, you have a problem that you see out there that you want to solve, then at some point you, just you just have to start. The beauty these days is, starting doesn't mean that you have to quit your job.

Speaker 1: 31:01

It doesn't mean you have to found like incorporate a company in what we are doing with code. It just means after work you sit down, maybe with a friend, and you start coding and you take it from there.

Speaker 2: 31:18

So with this company, at what point did you decide to incorporate?

Speaker 1: 31:23

So the way it started was a very good friend of mine, ferdinand. The two of us we at the time regularly were chatting about potential startup ideas in the AI space and when we were discussing applying AI to, to solving these issues that I had with podcasts, we actually decided to join the Haxeric hackathon, which is Europe's biggest hackathon, directly here in Zurich, and for us it was mainly to block an entire weekend to just hack a prototype together to see how good is the current tech out there. You know, can we actually already solve this, this problem, or is the AI not ready? And we participated in the hackathon and, as a big surprise to us, we actually won the hackathon, which was this big surprise. And you know, you're up on stage and they, they call your name and you win some money and you have this big adrenaline rush and this gave us a lot of energy Plus.

Speaker 1: 32:31

The other thing that happened was, after winning this, a lot of other participants of the hackathon actually came up to us telling us hey, I actually have the exact same problem, or I have a have a very similar problem with podcasts. Can I use your, your, your solution for this as well? And that's actually when we had the first validation like problem validation of other people, and these two things came together for us. Plus, both of us were not 100% happy in our current jobs at the time and these things coming together told us let's go for it.

Speaker 2: 33:12

So that that day at the hack Zurich was that the first version of your snipped product, essentially In a way, yes, in another way, no, because.

Speaker 1: 33:23

So let me tell you what we actually built. We built an search engine where you could search within an episode for a moment, based on LLM searches. So when we were on stage at the finals pitch, what we did was we used the now very famous Joe Rogan episode with Elon Musk, where Elon Musk smokes a joint, and it's a two hour episode. You know, the moment is like 30 seconds or 20 seconds. So what we did on stage, we just searched for smoking weed and the algorithm would find exactly that moment and we connected it with YouTube so it would pop up the video live on stage, and then here is Elon Musk smoking a joint, and so it was a great demo, I would say. But so this was, in a way, the first version of snipped, but in a way it also isn't, because, as of today, we actually don't have an LLM based search within episodes.

Speaker 2: 34:33

Did you go live with that product, with that demo after and get users?

Speaker 1: 34:39

We did go live with it, but without users.

Speaker 2: 34:45

What was the disconnect there? Why no users?

Speaker 1: 34:49

So I think it had a lot to do with the UI and the discoverability. I also still believe that this also the reason why we actually do not have this as a main feature in our app today is that it's not the biggest use case out there. So the user journey if you look at what happens on, if you open any of the big media apps, for example, tiktok, what actually happens is that you get content recommended to you. Instead of that, you start your journey saying, okay, I'm looking for exactly this particular thing, and then you go from there, and that is also why, yeah, it's not the biggest feature in our app, but we do have a transcript search where you can search through the transcript of an episode.

Speaker 2: 35:47

What advice would you have to any other founders out there?

Speaker 1: 35:53

If you're very early optimize for learning I think it's very easy to get caught up in oh, we need to improve this feature, that feature is not working correctly and we have a lot of feature requests coming in here. But in the beginning, a startup going back to the lean startup methodology a startup is an experiment. You don't know whether it's going to work out. So if you optimize for learning, you find out as quickly as possible whether what you're actually building will work out.

Speaker 2: 36:32

How did you guys optimize for learning?

Speaker 1: 36:34

This is really dependent on a day-to-day basis, just asking yourself whatever you do. So we operate on two-week sprints and once a month. So basically every two sprints we sit down for strategy session and those are the moments where we talk about what do we work on? So what is our current focus? And that's basically the moment where we ask ourselves why are we doing this? And of course, now, in the meantime, we know a lot more about what our users want. So there are now also much more things that we do to improve the product for our existing users. But if you're really early on and you're just starting out, the number one focus should be to learn.

Speaker 2: 37:25

Well, I think that's fantastic advice. Thank you so much for coming on, Kevin.