The Impact of AI on Podcasts: From Linear Digression to SuperDataScience

Introduction to AI and Podcasting

0:00Hello, and welcome to the Harvard Data Science Review podcast. I'm Liberty Vitter Capito, the feature editor of the Harvard Data Science Review. And this month, the HDSR editor-in-chief, Shali Meng, sat down with two podcast hosts to talk everything AI and podcasting.

0:21Katie Malone is the host of Linear Digressions, a podcast about data science, machine learning, and AI. She's a physicist by background and has worked as a data scientist in startups, high-growth tech, and enterprises, as well as teaching, speaking, and writing about AI. And John Crone is co-founder and CEO of the AI software company YCaret, author of the number one best-selling book, Deep Learning Illustrated, and host of Super Data Science, the data science

0:51industry's most-listened-to podcast. So, what can podcasting teach us about AI, and what can AI teach us about the future of podcasting? Let's get into it.

Goals of Linear Digressions and Super Data Science

1:03Katie and John, so nice to see both of you, even just virtually. And welcome to this episode of HDSR's podcast. And today, I want to ask each of you that what is your show trying to do, who's the full, and what do you hope someone takes away from listening? Maybe we'll start with Katie. Katie Well, thank you. It's a pleasure to be here. Appreciate the opportunity. I host a podcast called Linear Digressions. When we first started this in 2015, I would describe it as being about data science and machine learning. Like everything else, AI has eaten that world lately.

1:38So, now that's more of the focus, and that's kind of how I describe it to folks. Katie And it has a little bit of an interesting origin story. I was teaching a course, an online course, about machine learning at Udacity. And in the course of, no pun intended, putting that together, we had a lot of interesting content that just we didn't have space for in the regular course. And at the time, I was really interested in podcasts. And I thought, well, I wish there was a podcast where I could learn some stuff about data science, you know, not be too high level or too

2:09in the weeds. And it didn't really exist. And we had kind of this trove of content. So, a friend and I kind of pulled some microphones into a closet. And that's how it got started. Did it for five or six years and got, frankly, a little bit burned out on the process. But we came back about six months ago, and I've been doing it weekly since then. So, it's been a really interesting exercise to contrast what it was like getting started a while back to now. So, what I'm doing now is still back to those

2:43original roots, you know, really aimed at talking to folks who are interested in learning more about AI these days, interested in having a little bit of technical or scientific rigor to that conversation, learning something new, but not getting so in the weeds that you feel like you're reading an academic paper every week. And it's been a really fun experience. So, linear digressions, like I said, come check it out. Do I remember correctly, I was on your podcast? You were, yes. Yes. And I was also on Jones. Thank you, John. You invited me. I think, John, you probably have at least one of the most

3:17popular data science podcasts. So, what do you want the listeners to take away listening to your podcast? Yeah. I mean, exactly the same kinds of things as Katie was saying. We want people to learn something. We want them to hopefully smile a few times as they're listening to the show. We try to be a little bit lighthearted about what can be heavy technical concepts. And, you know, it should be entertaining a little bit at least, while also educating you a lot and keeping you up to date on all the latest

3:47things in data science, machine learning, and AI.

Using AI in Podcast Production

3:49So, we already mentioned now every conversation seems you cannot avoid AI and for mostly the right reasons. So, I want to start the part about how AI affects the podcast itself from the more kind of mechanical aspect, like AI, the use of AI in production. From research to guest preparation to editing, transcripts, promotion, translations, whatever it is, how much you're using AI now and how much you like to use AI if you haven't. And maybe this time start with John.

4:26Yeah. We use AI as much as we can without losing a feeling of authenticity. You know, I want it to be my voice in the program. Obviously, I'm conducting interviews or I'm doing deep dives into topics where it's important to me that those parts that my voice is really me. But we have AI everywhere else. I mean, so we have an operations manager, Sonia, who is incredible at stitching together lots of

4:57different AI workflows and automating episode production as much as possible. So, for example, the show notes, the episode summary, that's something that these days LLMs are so good at. And something that I actually feel really bad about is that we, for many years, we had a brilliant professional writer with a PhD who did such a good job of creating these. But, you know, she wasn't somebody from our niche. And so this year, we got to a point where Claude Fable 5 could create

5:36episode summaries as well as, you know, this professional human writer, but also had the in-depth knowledge of our niche that meant that there were, you know, fewer mistakes based on an understanding of what we're doing. So we obviously want to be augmenting as opposed to replacing people as much as possible. But every once in a while, AI comes in and is doing something actually better than we can even find a human to do. But one of the more interesting aspects of where we've included AI is in making animated shorts. We're on YouTube as well as on audio-only podcasting

6:11platforms. So we hire out to a contracting firm that uses AI. And this contractor takes our hour-long interview and converts it into half a dozen one-minute-long shorts, but they're animated. He uses AI for so many things, to figure out what clips to use, as well as to generate the video animations. And they are stunning. They're my favorite part of the show, and I have nothing to do with creating them. What you can do, text to video these days, is so compelling. Wow. Kitty?

6:42My setup is kind of similar, I think. So I've found that AI is a great sidecar for me in the production process. I mentioned a little bit that I got burned out at one point and came back. And a big part of the reason why I got burned out was I really enjoyed creating the episodes, recording. I could put up with editing. Producing just took something out of me. Writing those descriptions, even though it might not take that long, I just dreaded it every week. There's a lot of little tasks like that in making a podcast. And it's just me. I do this as a hobby. And so it

7:18takes care of a lot of that stuff for me, which is one of the things that makes it sustainable. One of the things that I have experimented with using AI for, and I've really had to find the right niche, John, I'm curious if this is similar for you, is around the research and preparation process. I find that having a good prep process where you're either learning about the topic that you're going to discuss, or you're doing some background research on someone that you're going to interview. It makes for good content. You can't skip that. But doing that on my own the first

7:51time around, there were a lot of blind alleys. I would wander down. You spend a long time reading something and it wasn't the best resource on a certain topic or trying to find the CliffsNotes, the summary, the high level stuff to drill into when you're preparing. So I find that having AI as a research partner and a little bit of a scout is super helpful. When I let it go too much farther than that and into anything that feels like scripting or even sometimes suggesting questions or lines of discussion is where it starts to get real jagged real quick. And I've ended up,

8:27I think, in a similar place where for reasons of authenticity, quality, and honestly, my own personal satisfaction from the project, I tend to stay hands-on. I don't have a lot of AI involved in actually creating the content or even necessarily like setting the editorial direction. Your answer just reminds me a lot more questions that I'm trying to answer as the editor for a journal, right? Because, you know, we face these issues when you have, for example, when you talk about research

8:58and, you know, I obviously use AI as well, not only just for research for doing podcasts or in the journal, but also doing my own research. There's really two big questions and I face this question really now daily, routinely because of editing the journals. One is how do you ensure that's trustworthy, especially in the scholarly research side? You know, trustworthy, this is obviously very important. The second part is how much you need to disclose, right? People say, make it transparent.

9:28You can push hard and say you disclose everything, but in reality, just like these days, you don't disclose that I'm using calculator to do my arithmetic, right? You know, there got to be some line somewhere. So I want to actually hear from you two, whatever you're doing in terms of podcasting, in terms of research, like in that aspect. One is the authenticity, which you already mentioned. How do we tell? And the other, the transparency in your, like when you have these podcast materials, how much you tell, so what this part is actually from AI or this part is from humans,

10:00if you do that at all. Yeah. I had an easy answer for that second one around transparency, which was, I just had a whole episode about how, cause I, it's an, it's a podcast about AI, right? Sure. And we had a series that I did about 10, 11 episodes that were just unpacking agents, kind of piece by piece. And as the icing on the cake at the end, such as it was, I walked through the agent that I use for the production process. I walked through some of the AI tools that I use

10:30for my research. It was a little bit of a gimmick, but I, in the episode, I recorded myself, did the speech to text. I put it in the transcript with the agents themselves. And then I had them like reply back, what do you do every time? You know, what's easy, what's hard, what goes well? And then used a speech generation software to have it be like an interview, but there was an AI on the other side. It was kind of fun. So I err on the side of extreme transparency on that because it's like good content. But I will say one of the other things that we do as part of the little ecosystem

11:03that I have here is a, a sub stack newsletter that's mostly summary and notes. And there's a little bit of original content that's not covered in the main episode, but that's all AI produced. And it's through sub stack, which has recently started an integration with Pangram. So there's a little button now that you can hit and it says, whatever, 70% AI generated or 90% AI generated. I always edit those newsletters, but it's been something that I've been thinking about a little bit more recently when all of the creation process that leads up to that point, everything that happens

11:35before the newsletter is like me and my kind of human curation and presentation. AI takes that transcript, you know, produces a newsletter. And then it says, this was like 90% AI generated. And I'm like, well, I don't know if that's really fair. Um, but, uh, it's an interesting question because I think there's different kinds of contributions that AI can make in some of these cases. And they might have different implications for how we think about the human authorship. And I don't know that we have all the machinery yet to tease some of that stuff apart.

12:05John, what's your take on this? As usual, I agree with everything that Katie says and, uh, going back to the authenticity bit, I want to reiterate how important it is that things like Katie mentioned editorial direction. That's something that I think, you know, as the host of a show, as the owner of a show, people are expecting our opinions, our guidance on where things are going. And they want it to be our unique take and be based on our unique experience. I think part of why people listen to a podcast is because there's personality in it as well. It's different than consuming something as

12:41an academic paper, for example, or, you know, consuming a blog, having that authenticity in the editorial process is important. And it's also fun. It's not something that I would want AI to replace me on. Although something that has been occurring to me recently is thinking, well, I wonder if it could make suggestions or ideas. Like what would be, I haven't done that yet, but it has, it has occurred to me that I'm like, I bet I could probably have some good brainstorming sessions on, you know, given recent topics we've been covering, what are some ideas for what we could

13:11cover next, uh, and use that in kind of a back and forth dialogue. And so this, this really gets into, you know, Katie's point there about this 90% AI generated for the newsletter. It's something I worry about that, you know, now any model that Anthropic releases is going to have these invisible watermarks based on the text where people can say, okay, uh, you know, probabilistically this was AI generated and that's kind of Claude approved. But what happens in circumstances where you come up with the ideas behind something, you are deeply involved in the research, you draft

13:47bullets of how you want the content to turn out. And just the prose is composed by AI. Then it's like 100% AI generated, but actually I did almost all the work. Right. Right. Yeah. No, it's a, it's a really interesting question. And, uh, but there's also the other side, right? There's the side of accountability, right? Whether something done by AI or something done by human, if there's a legal liabilities and how do you do the attribution there? So that's actually leads to my next question,

How AI Changes Content and Research

14:16which is, we already started talking about is not just about the use of AI in the production space, but the rather how the AI is changing the content. Yeah. I mean, I think, you know, if I want to be doing a deep dive into a particular topic or I want to be preparing for a guest, it would be insane to not be using AI, to be pulling resources and summarizing information for you. We have an interesting process at the podcast in terms of research where we have another contractor who does research and he's an amazing data scientist in his own right. He has huge amounts of

14:50agents working for him, helping him with the research, you know, curating things. There's still a human aspect to what he provides to me in the end. It was a pages of information on a guest that I'm interviewing with suggested topic areas, suggested questions. And so I think that there's this interesting dynamic evolving where we are more and more integrated with AI operations. And, you know, whether it's Sonia, who's doing our operations, whether it's Serge, who's doing research, whether it's Anthony, who's doing our shorts, all of those people are using AI in the

15:20flow of what they're doing. It's making their output better. And they're able to then provide me as the person who's getting the guests on the show, asking them questions with a much better context, much better ideas than I otherwise would have. And I personally would not want to go back to an era where I didn't have all of this magical power at my or our team's fingertips. Thank you. Katie? Yeah, I think on my side, I don't think that the advent of AI has moved us that far on linear

15:55digressions in terms of the content that we cover. Like we've always tended to be a little bit more on the technical side. Like I just my nerd brain likes to talk about how these algorithms work a little bit more than, you know, what's happening right now in the AI hype cycle or like, you know, there's there's plenty of other like, good people who are commentating on AI and different aspects of it in society. And I've been doing this long enough that I know where my sweet spot is. And it's not like algorithm of the week. It's about, you know, kind of thinking about more evergreen topics. So that hasn't changed much. And I think that's also reflecting something that's that's true for

16:30me, John, I bet this is true for you too. And Jowli, like, to a certain extent, you can only do this if you love it. And at least for me, like, that's the angle that I love. And it always has been and it's continuing to be there's just like, a lot of good stuff out there to add. I think what has changed with AI relative to data science and machine learning is it is a social topic of extraordinarily high salience. And I think that is a challenge, of course, like, it's a social

17:01challenge just for like, all of us as humans. It's a bit of a challenge sometimes for the the podcast and the content to, you know, find the the lane through all of that. That's, I think, true to the audience and the interest that they have and not trying to get sucked into any monster of the week type topic. But at the same time, I think there there are things that are technical topics that have real implications for how we live our life. Like, for example, we did an explainer episode

17:35very recently, we were just talking about this a second ago about how AI generated text is watermarked or detected or whatever. I think it's super interesting from a technical perspective. And my little nerd brain was just like, lighting up on, like, on all of the synapses when I heard about that. But like, there's going to be people that that, you know, has a false positive for somehow it misrepresents their their contribution to something. And, you know, somebody's probably gonna have like a really bad day or a bad month or a bad year because of that. I don't know who and I like I hate

18:07to say that. But I think the question of liability related to AI actions is similarly fraught. So anyway, I think it's, it's, it's an area that's very technically interesting. But what's, what's, of course, making it a little more complex and important, I think, is all of the implications that those technical pieces when they actually like meet the real world, they have an impact on all of us as like members of society.

AI in Educational Audio Formats

18:33One thing we find is very useful is to formulate things in the podcast version, having two people talk to each other. And this, as you know, LM notebook does extremely well, right? And they're so real. Do you guys see like really the use of AI and in the podcast format, particularly in the education space, any, any new directions or any things you would imagine that could be very useful, effective for education, but we haven't fully taken advantage of?

19:03So I learn a lot from podcasts. And my, my feed reflects that, I think, if you were to dig in there. So I can definitely see that. And I, and I think that there's a lot of possibility and notebook LM is probably the best example of this of, you know, not just that I can learn whatever somebody had the interest and inclination and expertise to make a podcast on, but that I can say, like, here's the thing that I want to learn about. And the podcast is the medium and AI can kind of

19:34fill in the blanks. I mean, I'll also say that some kinds of technical topics do require a certain amount of hands-on and struggling with the content. I'm a physics PhD. So I'm like, ah, no pain, no gain. So to some extent, I don't think it's a substitute for, uh, you know, the kinds of learning that you do through interaction where you're, you know, you're in the conversation and you're talking back and forth, of course. And I, I don't think you were suggesting that suggesting otherwise, but yeah, the idea that, um, you know, that kind of auditory explanatory, you know, dialing the, the level up

20:08and down, um, kind of format is, is now something that can also be customized and fit to an individual and their, their interest in what they're trying to learn. Like, I think that's an incredible opportunity. I think that's super cool. John, you have any thoughts on? So the first episode of every January, a brilliant data science and AI entrepreneur named Sadie St. Lawrence comes on my show and we talk about what happened in the past year and what we think is going to happen in the coming year. And whatever year Notebook LM came out, both of us had as our pick for the most jaw

20:40dropping AI moment of the past year was Notebook LM. It is astounding how well it can create a real sounding conversation. So I have a fellowship at a company called Lightning AI. And it's this sprawling business. And so recently the CEO of the business, Will Falcone, he started using Notebook LM to create podcasts about what we do at Lightning and his vision for the business. And then he has the whole company listen to these Notebook LM generated podcasts.

21:11One thing we have thought about, we haven't really done it, but it's not hard to do for journals like HDSR. We can creating a podcast for each article. We did some testing, which is great. It's just, few minutes of, you know, a great summary, you know, of that. So, so I have one more question

Organizational Impacts and Future Wishes

21:28before we do the magical one question. And so far we've been talking about the use of AI for podcast as a platform. I want to just have one question on the, how the AI is a fact organization and the individuals. And I want to start with John, because you started from saying like how the AI has changing, so to work for your, for your podcast and how does that affect you as an organization? Like, do you have to hire fewer people? I assume the revenue is better or, you know, what is the effect on organization and what would you like to see there?

22:01For sure. So, you know, I only have that one genuinely like regrettable example of headcount being reduced as a result of AI. I know that there are other kinds of businesses where that, you know, that kind of thing happens. But other than that one example I gave earlier in this, in this episode, I haven't personally been involved in or seen a kind of headcount reductions. What I do see is improvements in quality. The research quality is better. I can do a better interview because of AI.

22:31Sonia in our operations team, she can have AI workflows doing things automatically that previously she would have had to have hand typed. And so it saves her time and it allows her to do a better job because she can be focused more on ensuring quality, having a vision, being able to have more strategy as opposed to just doing some rote aspect of the role. So I think ultimately what it allows us to do is to create a better show, which downstream hopefully allows us to have a bigger audience,

23:04which hopefully allows us to be able to charge advertisers more and have more revenue. Katie, because I know you are kind of a one person show, right? Or at least two, whatever it is. So you may not be much organization, but how much, you know, affect you as a podcaster other than saving your time? It's like how much are you using AIs for you or broadly speaking as an individual? Yeah. So for the podcast, I'm very selective in where I use it, but it's basically the thing that makes it possible for me. I will say I find AI really fun. I tend to dabble widely. I am aggressive in what I

23:43try and I am aggressive in putting it down if it's not working for me. So I think maybe that speaks to one of the other things that's just been very interesting for me to watch. I think we're all as a society figuring this out right now, but, you know, there's a big difference between what the models are maybe capable of, you know, they're book smarts, if you like, and what process, like human processes, whether they're businesses or somebody in their podcasting process or, you know, me when

24:14I'm doing meal planning or, you know, some other routine that I have established, right? You know, figuring out where and how to make those modifications so that AI is actually like delivering value in a meaningful way. Like books are being written about this right now. And that's because it's a complicated topic. And I think this is something that gets lost when it's, you know, early adopters and enthusiastic practitioners talking amongst ourselves because we get excited and we're like, ah, the possibility. But most of the world is not like us. You know, most of the world has like

24:49many other things that they're thinking about and trying to figure out how AI fits into their workflow is maybe not at the top of that list. At the same time, AI is coming for everything right now. And so I think that's leading to some of the challenges and how to think about that as a society. So similarly to John, I don't have any, you know, case studies that I can point to that I've seen firsthand of large scale or even, you know, moderate to small scale job loss based on AI. I know there are labor economists who are watching this quite closely. And as far as I can tell, the signals are a

25:23little bit mixed on that one. But yeah, I think it speaks to the difference between a computer being really good at spitting out tokens and actual change in the real world. You know, there's a pretty big gap between those two things sometimes. Yes, very much so. And thank you to both of you again. I want to turn into the magical one question. If you had a magical one and what would be the one thing you'd like AI to do that is not there yet for your podcast? Oh, for my podcast. I was going to say like

25:56clean up my kids' toys. Okay. Well, that's a good question. Watch the dishes. Do my laundry. For my podcast. Okay, let's say it's for my podcast. I, okay, it is not there yet in coming up with really good interrogation lines for guests for me. It comes up with some okay stuff first, but I find that it settles into a rut really quickly. And I'm usually not impressed with what I get out once it's in that rut. So I think that the figuring out how to make it a little bit more creative and multifaceted in

26:32thinking about different ways to approach topics with like very interesting people that I want to be thinking about how to be, you know, interesting with them. Yeah. It's not there yet for me. Okay. Thank you, John. I would love for the entire hour long episode on YouTube to be animatable, like to have, you know, papers come up or for that to work in a way that, that it looks like an expert editor has been involved, an expert producer in terms of what content, maybe even that I had to have been

27:05involved in thinking of what should be popping up here all over the place in that full hour. I think that that would add a lot. And beyond that, I hope that AI doesn't do too much more because I mean, this is like, I guess I worry, I, I spend less time wishing for more things that AI can do on the podcast than I do time worrying about, Hmm, if people can just get whatever podcast they want for whatever thing they want to learn, if the barrier to entry is so much lower for people to be creating podcasts that are perfect professionally, because they're using some

27:40kind of notebook LN thing, maybe with their real voice and listeners don't even know that, you know, I guess I get worried that AI can maybe in the not too distant future compete me out of my job. Well, let's hope that won't happen. And I know that won't happen. Thank you both again. And I encourage all listeners of HDSR podcast to go listen to Linear Digression and Super Data Science podcast. Thank you so much, John and Katie. Thank you. Thank you.

28:10Thank you for listening to this month's episode of the Harvard Data Science Review podcast and make sure to check out Linear Digressions and Super Data Science. Thanks to Katie and John. To stay updated with all things HDSR, you can visit our website at hdsr.mitpress.nmit.edu or follow us on X and Instagram at the HDSR. A special thanks to our executive producer, Rebecca McLeod and producers, Tina Tobey-Mack and Erin Hieswetter. If you liked this episode, please leave us a review on Spotify,

28:44Apple or wherever you get your podcasts. This has been the Harvard Data Science Review, everything data science and data science for everyone. Thank you.