How to digest 36 weekly podcasts without spending 36 hours listening | Tomasz Tunguz (Theory Ventures)

0:00I have a list of 36 podcasts, but I don't have 36 hours every week to listen to 36 podcasts.

0:06Tomasz Tunguz:So what I did is I created a system that goes through each of those podcasts every day and downloads the podcast files and then transcribes them. Can you show us how it's actually built? Like where do you get this feed? It sounds like you run it locally. How does this all work? I wrote this thing called the Parakeet Podcast Processor. And this podcast processor basically takes in a file. And what it'll do is it will read the file, it'll download it, and then it'll convert it via FFmpeg. And that will take the audio and convert it to text. So here's the podcast summaries for today. There's Lenny's podcast, the host, the guests, a comprehensive summary.

0:43Tomasz Tunguz:So here's a conversation with Bob Baxley. Key topics and then key themes. The part that's most valuable for me are these quotes. And those quotes, I'll read them. It will suggest a bunch of actionable investment theses for a venture capital firm, which is put into the prompt like, OK, maybe we should be looking at AI-assisted design tools. You've gotten not only the content you want, but the user experience you want. You control it end to end and you can build this hyper-personalized software experience.

1:13Welcome back to How I AI. I'm Claire Vaux, product leader and AI obsessive here on a mission to help you build better with these new tools. Today, I have Tom Tungus, a legend in the enterprise software business and founder of Theory Ventures, which invests in early stage enterprise AI, data and blockchain companies. Tom is followed by over a half a million folks on his blog and LinkedIn, and he's going to show us today how he uses AI to keep up with all the podcasts, including this one, and draft blog posts that would be approved by your AP English teacher. Let's get to it. This episode is brought to you by Notion.

1:52Notion is now your do-everything AI tool for work. With new AI meeting notes, enterprise search, and research mode, everyone on your team gets a note taker, researcher, doc drafter, brainstormer. Your new AI team is here, right where your team already works. I've been a longtime Notion user and have been using the new Notion AI features for the last few weeks. I can't imagine working without them. AI meeting notes are a game changer. The summaries are accurate and extracting action items is super useful. For stand-ups, team meetings, one-on-ones, customer interviews, and yes, podcast prep, Notion's AI meeting notes are now an essential part of my team's workflow.

2:35The fastest growing companies like OpenAI, Ramp, Vercel, and Cursor all use Notion to get more done. Try all of Notion's new AI features for free by signing up with your work email at notion.com slash how I AI. To celebrate 25 ,000 YouTube followers on how I AI, we're doing a giveaway. You can win a free year to my favorite AI products, including V zero replet, lovable bolt cursor, and of course chat PRD by leaving a rating and review on your favorite podcast app and subscribing to YouTube. To enter, simply go to howiaipod.com slash giveaway, read the rules and leave us a review and subscribe.

3:27Enter by the end of August and we will announce our winners in September. Thanks for listening. Okay, Tom, I'm so happy to have you here because you're going to show us how you are solving a problem I'm creating for you. The problem I'm creating for you is I am creating yet another piece of interesting content that you have no time to consume. Certainly the format that we get it out. And I know to you, content is a really interesting source of ideas, of trends, of companies. So tell us what you built and why.

4:04Tomasz Tunguz:Absolutely. Well, thanks for having me on. So I prefer to read than to listen because I can skip ahead. And I think there's a lot of information inside of podcasts that people share that I would love to know. And so I built, I guess what I call a podcast ripper. And the idea is I have a list of 36 podcasts, this one included, that I really admire and I want to learn from. But I don't have 36 hours every week to listen to 36 podcasts, right? So what I did is I created a system that goes through each of those podcasts every day and downloads the podcast files and then transcribes them using, initially it was OpenAI's OpenSource Whisper, which takes audio and converts it to text.

4:51Tomasz Tunguz:And then there's a new version called Parakeet, which NVIDIA released that runs really well on a Mac. And so I'll take that text and then I'll run it through a prompt and it will spit out a whole bunch of different things. It'll spit out a high-level summary or whatever I ask it to. Okay, can you show us how it's actually built? Like, where do you get this theme? It sounds like you run it locally. How does this all work? So I initially downloaded the Whisper app, and what I did is I wrote this thing called the Parakeet Podcast Processor. And this podcast processor basically takes in a file, and what it'll do is it will read the file, it'll download it, and then it'll convert it via FFmpeg, which is a library that converts different kinds of files.

5:37Tomasz Tunguz:And that will take the audio and convert it to text. And then I use Gemma 3, which is really good at this, to actually clean up the transcript. So if we search for the Olama model, basically what I'm doing is I'm just cleaning up the file here. Your transcript editor, clean up this podcast while preserving all the content. keep the same length, remove the uns and the ahs, preserve all technical conversations, and that returns a clean transcript. And so on a given day, there might be five or six different transcripts that need to be transcribed. And then what I'll do is it runs through the Parakeet Podcast Orchestrator, actually it's just a podcast orchestrator, which is here.

6:25Tomasz Tunguz:And so So I'm storing each of the files that I'm transcribing in a local DuckDB, which has a little database that says I process this particular podcast on this particular day. And then I save the transcripts and I take all the transcripts on that particular day from the database, which is here. And then I send them through a prompt, which summarizes the daily summarizer. So it generates a daily summary document, which is here. It'll produce a file that looks like this. So here's the podcast summaries for today, June the 13th. So there's Lenny's podcast, the host, the guest, a comprehensive summary.

7:09Tomasz Tunguz:So here's a conversation with Bob Baxley. Key topics. So here he's talking about his philosophy, a company culture. And then key themes. And the part that's most valuable for me are these quotes. and those quotes are then, you know, I'll read them. It'll suggest a bunch of actionable investment theses for a venture capital firm, which is put into the prompt like, okay, maybe we should be looking at AI-assisted design tools. And then that might kick off a market map. We're really thesis-driven, so maybe that starts a conversation on a Monday and we decide to staff a market map. Then it'll produce these noteworthy observations which are actually put into tweets.

7:50Tomasz Tunguz:So here are the Twitter post suggestions. So I haven't done this yet. I'm still working on the prompt, but the idea is like, could we actually automate linking back to people who we really like? And then another part, this is a little out of order, but another part here is, are there startups that are mentioned within these podcasts that we should know, right? So here's Airbnb, Google, Amazon, Stripe. We know all these guys. I don't know what this company is. And so this might go into our CRM, right? To be enriched. And then the last is we'll actually generate prompts for blog posts in the style that I write.

8:26Tomasz Tunguz:And then this will go into a Python pipeline to actually machine generate blog posts. So before we get to the machine automated AI blog post pipeline, I have a couple of questions about this process because I think you did a couple interesting things. one i have a question is if you found higher quality by cleaning up the transcripts like how much did that incremental input quality piece actually help your your output so it helped so initially i was trying to get the answer was initially a lot and then over time less because initially what i was trying to do was to find these companies i was using named entity extraction algorithms from Stanford.

9:11Tomasz Tunguz:There's a Python library. And it was having a really hard time. And so I was cleaning up, cleaning up to try to get the performance to improve. And then I just pushed it to a really large, large language model. And it spits it out much better. And so the cleaning is not that useful anymore. Yeah, I was looking because I was looking at it, and you're focusing on like proper nouns, company names. And so I'm assuming if you want to extract something like Stripe, which has many, many meanings. Getting into a proper noun format, for example, would help with that extraction. But you're saying as you could just use as opposed to these kind of package libraries for specific machine learning use cases, instead just send it to an LLM, that ended up just meaning you could worry less about the input quality of your transcripts and more about that kind of prompting and structure here of the output.

10:02Tomasz Tunguz:Yeah, that's exactly right. So my goal initially was to do everything locally. And so I was using Olama, I was using that Stanford library, Parakeet is run locally. And then what I realized is, particularly for the named entity extraction, more powerful machines are much better. Yeah. And so, and then I have to ask another question, which is, everybody's going to look at this and they're going to go, what the hell is he typing in? Like, we have a couple of people that are like, why in the terminal? So I'm just curious, you know, did you ever think about putting a UI on top of this. You just, you seem very comfortable in the terminal.

10:36So it seems to work for you. I'm just curious about where you decided to focus your user experience efforts on this personal. Well, I love the term.

10:45Tomasz Tunguz:I read this blog post by Dan Liu with two U's where he was talking about latency and the latency between like the keyboard and the computer. And it turns out that the terminal is actually the application of the lowest latency. And the lower the latency, the less frustration you have using a computer. So during COVID, I decided to learn how to use terminal and since then i've sort of lived in it and so like my email client is a terminal based email client and i use i use that because it's really fast and i can also script different things so i can lead 10 messages at once or i can call an ai to actually automatically respond to an email or add a company to a crm so that was really important but at a high level like i think it's um i've just become really comfortable with it it's really fast and then the last thing i'll say is I think CloudCode is an amazing product.

11:34Tomasz Tunguz:And the great part about what CloudCode does is I have about 2 ,000 blog posts. I can just go into CloudCode and say, modify the files in this way or change the blog post theme. Or recently I launched a blog post generator, which takes all of the content that I have on the blog and you can ask it a question and it will write a blog post for you about your particular question. And I did that all using CloudCode. Yeah, I mean, I have two sort of thematic things that I think of while observing this workflow and your love for the terminal. I agree. Cloud Code is an amazing product and it's a really well designed terminal based product.

12:13I love it. I love that you have this constrained surface area in which to like communicate progress and latency and changes. And I think it's really thoughtfully designed. So for anybody out there building dev tools in particular, learn how to design in the terminal. And it's so, so important because you make really fabulous products for, I guess, people like you and me that say things like I picked up the terminal over COVID as my hobby. The second thing that I was thinking about is since generative AI has become mainstream, every single person has said somebody make a podcast digest application.

12:49Every single person I know is like it was one of the first projects I made. I made my kids a podcast digest, their favorite podcast. And it made little quizzes about the topics that they could answer. Super cute. So I think it was a very common use case. But what I was thinking is no startup is going to be like, you know, it's going to be a huge TAM company, a terminal based podcast, transcript, processor, and thematic extraction generation engine. And I think this is such a perfect example of like, yeah, there's probably something off the shelf that could do something like this. But you've gotten not only like the content you want, but the user experience you want.

13:31You control it end to end and you can build this like hyper personalized software experience, which I just it was not possible or wasn't efficient to do, I would say, until very recently.

13:45Tomasz Tunguz:Yeah, it fits the workflow, my workflow like a glove, right? And anytime something comes up and changes, like maybe there's a section that's out of order, like we found, I can just go into cloud code and update it, and it'll be done in 15 to 30 seconds, right? And, you know, I really wanted an email of this every day, and that was straightforward. So I agree with you. I think we're at a place where the marginal friction to achieving the glove-like fit with little utilities that maybe you wouldn't have paid for in the past is now, it's just so, it's so quick, right? Like you're just answering a couple of emails and it'll be done.

14:21Yep. You've seen the doom and gloom headlines. AI is coming for your job. But the reality is a little bit brighter. In Miro's latest survey, 76 % of people say AI can boost their work. It's just that 54 % still don't know when to use it. As a product leader and a solo founder, I live or die by how fast I can turn fuzzy ideas into crisp value propositions, roadmaps, and launch plans. That's why I love Miro's innovation workspace. It drops an AI co-pilot inside the canvas. So stickies, screenshots, and brainstorm bullets can become usable diagrams, product briefs, and even prototypes in minutes.

15:03Your team can dive in, riff, and iterate. And because the board feels like a digital playground, everyone has fun while you cut cycle time by a third. Miro lets humans and AI play to their strengths so that great ideas ship faster and happier. Help your teams get great done with Miro. Check out miro.com to find out how. That's M-I-R-O dot com. Okay, so you have taken all this content, including amazing content from the Lenny's Podcast Network, and you're processing it, you're extracting themes, you're extracting quotes, you're finding companies that may be interesting to reach out to, you're at least drafting Twitter posts.

15:52We will see if those actually get posted in production. And then let's talk about your second workflow, which is you extract insights that might be interesting for you to write about or add your perspective on. And then you actually turn those into drafts using AI.

16:08Tomasz Tunguz:There's a lot of stuff that's happening in the ecosystem. And every once in a while, I like to write about what somebody said in a podcast. And I think today, I was looking, the GitHub CEO is actually interviewed. and so Matt Turk interviews who's at another venture firm interviews Thomas and he talks about how AI encoding is the future and so what I really want to do here is let's suppose I really wanted to have a blog post that was tied to this so what I can do is I say like okay I have this podcast generator and I'll show it to you in a second and what I'll do is I'll take as context the transcription of that podcast which is here and then I'll define an output file and then I'll give it a little prompt which is like you know he said this quote which is actually within the podcast summary everything that I can easily replace with a single prompt is not going to have any value it will have the value of the prompt and the inference in the tokens but that's often a few dollars and I'll tell it okay go look for podcasts that are related to this and I've categorized them as AI and then here actually there's a bug so demo fail I was trying to fix it before I got on the video but the searching for the relevant blog post is failing and I need to figure that out it's it's run through a lancd vector embed in his database locally and then it'll generate a blog post and I'll share the prompt in a second and the best well one of the techniques that I found the most effective when generating blog posts is is to ask it to grade it like an AP English teacher.

17:42Tomasz Tunguz:And this goes back to my history. I remember not really loving to write until I took a class with an Army veteran. And he taught me to really love to write. And he was my AP English teacher. And so I really like receiving feedback in that way. I grade it on a letter grade and then tell me what I could improve. And then I'll iterate with the model until I get to an A minus. Got it. And so just before we go into the actual writing and I'd love to see a little bit of this AP English prompt. Are these two pieces connected? Your podcast summaries, do those go into this VectorDB that can then be searched through for relevant other podcasts if you're writing on a topic?

18:24Like, how does this all come together?

18:25Tomasz Tunguz:Yeah, so right now it's just the blog posts that I've written in the past, the 2000 blog posts or so, that go in, and the major reason I add those as context is I'm trying to capture my style. and I have to tell you like that's really hard like I have fine-tuned open AI I have fine-tuned Gemma models and getting the voice and you'll see it in the output it sounds like a computer when it writes even without additional context and it doesn't the other thing that I have not been able to figure out is I think it's really important in one blog post to link to other blog posts that I've written just because the knowledge builds on itself and obviously outside as well but I haven't been able to figure out how to get it to link effectively?

19:08I think this is a common feeling with AI-generated writing. No one is satisfied with style. Even when style is exceptional, I think I've seen examples, especially with some of the newer commercial models, actually writing really lovely prose and really lovely language. It's just, it's so personal what your style is and how you would write something, the rhythm in which you would write it, How would you punctuate and break line? All that kind of stuff is so personal that I have, like you, had a very, very hard time getting it to write like me. And I think even harder, which is why I appreciate that you're not yet posting this.

19:48It cannot, it can't tweet like me. I can't, I cannot.

19:51Tomasz Tunguz:No. The short ones are the hardest, you know? I guess they say that about writing generally. Have you felt like any of the models have done better or worse at writing like you? Or is it just like they only get 70, 80 % there and I just accept the fact that I'm going to have to rewrite things? Well, they have different voices. I don't think any of them are close. Like, I think Gemini is more clinical, is the way that I put it. I agree. Claude is more warm and verbose. Yeah, he's very, very garrulous, like just wants to keep talking and wants really long sentences and really long paragraphs. and OpenAI, I think the models each have slightly different personalities.

20:39Tomasz Tunguz:So there, I don't think there's like a single characterization. So I've been, I think I've been iterating to, I used to use Claude 3.5 a ton and I uploaded all of my blog posts in a project and I, and then I'd have it iterate there. Now I can kind of do it with Claude code or using this prompt. So that's a little less useful. But what I found is you really need to add your own voice and then you need to tell the AI to keep the things that are wrong, right? Like this, it's kind of funny thing to say, but as you were saying, Claire, before the way that you punctuate, I really like ampersands, right?

21:16Tomasz Tunguz:And I like adding spaces before colons. And I like starting certain sentences with or having little incomplete clauses because I think they keep the reader moving. But an AI won't do that. An AI will only deliver you a grammatically perfect specimen. Yeah, we're going to have one very nerdy English language moment, which is I like to start paragraphs with a conjunction. I love a and or a but. Yeah, it pulls you in. so okay you and i are gonna work well we'll build like a a micro sass on good good writing models um and prompts that that people can use so okay so we accept that it's not going to write exactly like you but you've created this grading process to say well is at least good and so i'm curious, can you walk us through how it gets to an A minus 91?

22:16Tomasz Tunguz:Yeah. As an A plus student, I don't know, a 91 would really stress me. Tell me how you kind of wrote the prompt and then why you picked like A minus as your bar. Yeah, for sure. Okay. So the way I broke the prompt, I told it what I wanted. And I asked an AI to critique, I think I asked Gemini to critique Claude's output. So it's kind of using a student teacher or critique model and then what it does is we'll walk through the prompt in a second but it goes through three grading attempts. So it reads the file gives it a grade and a score and then it the things that are the most important that I found particularly for readers are the hook which is the first few sentences or the lead you might call it and then the last is the conclusion and making sure it ties back because then you have a complete um you have a complete post and so it goes through this three times So you can actually see like here, it gave itself a 90 and then a 91.

23:13Tomasz Tunguz:And then at that point, it basically was good enough. It was satisfied with the hook. So if we, let's see, if we read the blog post generator, you can see what it does at a high level, right? So it finds the blog post. It generates an initial blog post, grades it like an AP English teacher, improves, and then auto-generates a URL-friendly slug. So it actually writes it in the right format. and then it can use OpenAI or Olama. And then the prompt is here. You are an expert blog writer specializing in technology and business content. And then here I add in the blog posts and it kind of shows the patterns.

23:52Tomasz Tunguz:What it also does is it dynamically calculates the number of paragraphs from relevant posts and uses Olama to summarize the stylistic patterns of those related posts. So I might write a little bit differently when I'm targeting a Web3 or a crypto audience than say I might when I'm analyzing the public disclosures of a company. A snowflake just announced earnings, let's say. And so it's dynamically injecting that here. It shows a bunch of different examples. And then, you know, here's what I think makes my blog post tick, right? 500 words or less. I have like 49 seconds with a reader. No section headers.

24:31Tomasz Tunguz:I ran an analysis of dwell time. as a function of how many headers there were. And it turns out headers were terrible for dwell time. People just bailed. Flowing paragraphs, each paragraph transitions smoothly to the next. Actually, the AI consistently critiques my transitions and says they're too harsh. And going back to the A minus point that you made before, I think I lose five or six points because of my transitions because they're abrupt. And then limit each paragraph to at most two long sentences. And then the structure of the blog post. I think this is a really interesting story at the top, and I want to make sure people don't miss it.

25:09I've seen this before, which is like take this example and describe it back to me and use it. And so you're saying, I'm writing on this topic. Go find the blog posts like this topic. Analyze them for format. Like what is the structure? How am I writing things? And match, stylistically match this subset of my blog posts because I do vary style by topic.