Screensharing Kevin Rose's AI Workflow/New App

0:00Today's episode is a show and tell episode with none other than Kevin Rose. He takes us through his entire AI workflow and lets us in and a new product that he's developed that he hasn't shared anywhere. And we learn about how he thinks about building new products, Claude Code, Vercel, all the tools he uses. So it's super, super fascinating. Kevin is just one of those iconic entrepreneurs and to have an inside look into his AI workflow and how he's building products in this AI age is absolutely fascinating. It got my creative juices flowing. I think it will get yours too. And if you stick around to the end of the episode, you will understand how to build products in the AI age, how to think about it, and what tools to be using.

0:46Enjoy the episode.

0:55We got the one and only Kevin Rose on the podcast. I'm super excited to have him. He has been going down the AI rabbit hole for a while now. And I wanted Kevin to come on just to share, you know, what are the tools he's using, some of the AI workflows that he has. And he's just going to screen share. Kevin, by the end of this episode, what do you think people are going to get out of it? like if they stick around yeah well i certainly first thanks for having me on it's i love the content and everything you're producing it's so important right now especially with things moving so fast but i think at the end of it you will see that things that are technically out of bounds for you like things that you just think that you cannot do are very much possible as a solo engineer solo designer and now we're finally at the point where i'm not even calling it slop anymore like they call the ai slop or whatever it may be or however you want to look at the code it's damn good and it's getting better by the week and um i think that you'll just i hope that you'll be inspired to go build something amazing because i'm going to show you something that is a little mad sciencey weird and all over the place and you'll get to see kind of a raw version of my brain and how deep I go on some of this stuff.

2:15But I think that's the beauty of it is this idea that we can go build anything. And oftentimes when we do, it ends up being a little bit of a, I don't know, a sandbox that can be a little too big and messy. And then you have to refine it back to something that's actually usable. So I think that the future engineer and the future developer, on the future product builder here, it's not going to be what you build as much as what you don't build, if that kind of makes sense. Because it's going to be so easy to build anything and everything to pair it back to something that's really usable, I think is going to be a real skill.

2:52Yeah, the hard part is the clarity. How do you get the clarity to know what to build? At least that's what I've been struggling with. Yeah, and that actually rolls right into the project that I've been kind of working on just for fun. And yeah, so I'm excited. All right, let's get into it. Okay, so I'll show you kind of this, you know, quote unquote vibe coded, or let's just call it coded project. And I'll tell you the inspiration, and then we'll get into some of the kind of dirty details and how it functions and what's possible as someone that kind of walked into this thinking, actually just asking myself, can I pull this off?

3:32so let me go ahead and do a little screen share here all right you should be able to see Tech Meme up there right? Yep. Okay awesome so first thing and I hope we don't in the edit this has to make it in I have nothing but a huge amount of respect for Gabe and what he's created in Tech Meme what I'm building today is not meant to be a competitor I don't plan on launching it in its like form it was mainly a personal curiosity of can I build something that's on par or better than TechMeme by myself and call it like a week and see what that would look like. And one of the nice things about TechMeme is, for those that don't know or for those that are lightly familiar with it, it's been around for a long time.

4:20Gabe's been building software since I was back in 2004, which is crazy. We were both kind of experimenting in early social news and what that meant. And what you see here is an aggregator that pulls from RSS and other sources. I don't know all the sources that he pulls from, but then also pulls from social media as well. And when you gather momentum, whether it be multiple news stories coming in or multiple people tweeting about something or talking about it socially, that is considered signal. And then that is considered used as part of ranking and showing you what's prominent here. And so what you're seeing is actually visually, the higher an object is here, the more weight it carries with a user.

5:05Because if you come in here and you scroll down to the bottom and you see something with very little tweets on it, it doesn't nearly carry as much impact as something with 15 or 20 different X posts underneath it. And so I like what he's doing here. I think it's a really cool way to say, maybe you might recognize some names, especially in tech, given how small the ecosystem is, you can probably look through this list and say, oh, I know these two or three people. If they're talking about it, what are they saying? And might I just hover over and see what their comment was about a particular story?

5:35So a lot to love here. I mean, Gabe's done a great job at creating this. But I have a slightly different area of interest. A lot of this is big tech news. And I kind of wanted to dive into more of what's going on in AI because AI is moving just so fast. You know, how can I slice this in really interesting ways to find my version of this? And this is stuff that we're playing around with at Digg with the reboot as well. So it kind of goes hand in hand with some of the kind of exploration that we're doing in a lot of these different areas. And so my job these days, you know, with Alexis Ohanian at Digg is we have Justin as a CEO who runs kind of the day-to-day and builds out, you know, the Reddit competitor version of Digg.

6:15and then we have me that I'm kind of more in the labs area where I'm pushing on the edges of ideas and saying what might we do and does any of this make sense to roll back into the main product and then Alexis is kind of just an overarching great idea so much depth of knowledge around both micro and macro communities and how they work and how they scale and what to do and not do and some of the missing tools and that's kind of how it all comes together So long story short, this is what I built. So first thing I want to show you is that there's going to be some errors here because it's been a minute since I have actually done the whole, actually touched this code base.

6:58But what you see here is I have 63 sources of information coming in. And so these are from just your classic RSS feeds. Yes, RSS still does exist. A lot of sites do break RSS, and so there are actually ways to go and add in kind of scrapers that will create dynamic RSS for you, even though RSS is no longer around. But you can see here, it's a bunch of different RSS sources that are coming in here. Some of them are Reddit, some of them is TechCrunch, but all largely tech related. so because this hasn't been run in a while we've got a lot of jobs that have to run in the background here to go and recrawl all of this so I haven't run this project in a few days and so it's essentially going out and hitting all these different RSS feeds, pulling in and ingesting all that information bringing it into the engine and then expanding upon and there's orchestrators and there's ways to actually go in and resolve certain chunks of information and find out more about individual stories.

8:05I'll get into that in a second, but we're playing a little bit of catch-up here. If the news stories, for some reason, don't look exactly fresh-fresh, you'll know why, but they'll be pretty damn close here in a couple minutes. It just needs to go and do this catch-up. Going back to the RSS feeds here, okay, great, we've got sources. Now, what do we do? Well, sources lead to articles. And when we go over to articles here, you can see now it's starting to pull in some of these articles in real time. So we've got MacRumors 13 minutes ago. And then we've got a little status check here. And the status check is really what happens once the article is pulled in to the system.

8:46Like, how do we process that article? How do we figure out what to do with that article? So, you know, 21, 25 minutes ago, 29 minutes ago, the Verge 9 to 5. And so we're pulling in and saving all of this in a Postgres database. and then we're also going in and doing things like figuring out who the authors are and this comes down to author reputation and so we can start taking a look at different authors that are here and I can click on any of these authors and actually see what articles they've actually contributed as well. So this individual author here and what are all of the articles that they've published.

9:23But what's more interesting here is actually what we're doing with these articles. So what happens is, as I start to pull in these articles, and this is where it gets pretty crazy because I didn't have this, I didn't really know, I understood technically how to pull this off in terms of the tools that would be required, but I couldn't write the code to do it. And so once we go in and we take an individual article, let me just get into an article itself here. Let's see if that will take me into the actual,

10:02Kevin Rose:that might pop me out. Okay, here we go. So here's an article right now that just came in and talks about the new AirTags that just got announced. And this came in from MacRumors. It's got an ID associated with it. So I was setting all this up in the database. It has a cluster membership associated with it, which we can get into in a little bit. But also it has a pipeline status. And so what happens is it comes in via RSS. I hit iFramely, and iFramely gives me back additional metadata about it, so like title, description. Sometimes it can pull in even additional kind of deeper context about it, article data, things like that.

10:39And I want to store all of that. That's really important for me to have that rich data. And then I can see that Gemini's been running against it as well, which we can get into in a minute as to the reason why. And so the winner has been resolved, the TLDR has been generated, and the embedding has been done. So the winner is basically saying, okay, there are a bunch of different factors here that might come in. Like, for example, if I'm pulling this in with RSS, I might not get a description there. Because RSS sometimes is kind of truncated, I might not get a really good description. But if I go and I hit iframely, then I might actually get a longer and better description.

11:16Or if I go and hit fire crawl, I actually might get the full actual paragraph or the full body of the article. So what we do is we have a judge that looks at it and says, okay, which is the best of these three that came in and pick a winner. And the winner is the one that actually gets stored to the database. So that's what winner means. And we do that on a variety of different things. We, I don't know why we say we, I do that. I'm the one that actually ever coded this thing. So basically we can see here that for this particular story, it did not have, I think we got a failure. We're coming from Mac rumors.

11:51The winner was actually Gemini. So Gemini is used as a last resort. So you can see here, fire crawl, a return day 403. We had a hard air there. iFramely, we had some kind of low quality signals. We got a successful crawl. We had low quality signals and then RSS. Um, good. It's used as a fallback, a good amount of characters here. We can see what the actual, uh, summary of that content was. And we got ours out of RSS. But as you can see here, it's actually pretty crappy RSS pull. There's not a whole lot of data there. You know, it just doesn't look that well or it doesn't look that good. So what we did is we hit Gemini and said, Hey, Gemini, turn on your ground truth, turn on your search, go out and figure out what is this actually about.

12:36And oftentimes, they won't come out and say this, but they kind of crawl the article and give you back what the article actually says. And they might reword it a little bit, but it's largely what the article says. And I can prove that because you can do that with Reddit articles and things like that where most everything will fail because Reddit's so hardcore about blocking it. Greg, does that make sense so far? That does make sense. I just have one question around, Why did you choose Firecrawl and iFrame-ly? Can you just explain what they do for people who aren't familiar with those services?

13:11Yeah, for sure. Let's go ahead and take the actual article here. I'll share this tab instead. You can see this is the MacRumors article, right? Now, if I take this MacRumors article and I copy and paste it, and then I go into iframely here. I can paste in a URL, any URL. And obviously I'm doing this via API, but I just want to show you how it works. And I hit check URL. And what it's going to do is it's going to give me this card back. So it's almost like those cards that you see on X, where you get a piece of rich media, you get the best possible image, you get a beautiful title, nice little description here.

13:54Sometimes it's a little bit longer. And then I can just jump in here and look at the JSON and see like, okay, what did we get here? and kind of expand out and say, did we get a high-quality logo? I can look for and parse through all of this and try and find high-quality data. And then also, I mean, it's called iFrame, largely because if there is an embedded piece of media, you also get all of those embed codes as well. So you get all of it. So I can actually take, and the reason I like Gemini is because they're tied in with YouTube, obviously, and I can get transcripts back from full YouTube videos, which is great for vector embedding.

14:29when I want to figure out what it's actually about. Okay, does that make sense? Crystal clear. Okay, and so you can consider Firecrawl to be a very similar thing. It is a little bit more on the crawling side, so a little bit more fine-tuned around crawling. They have some AI aspects to it as well, where AI will actually try and go out and figure out how the best way to kind of scrape the content. And then, yeah, lastly with Firecrawl, I think it's just, they have some stealth modes that you can turn on. So some of these news sources, they get really picky about what they're allowed to be crawled.

15:04And so you can kind of like hide and turn on stealth mode and then get to the actual data. And it's not my intention is not to like, you know, get around their ads or do something like evil here. It's really just to figure where's the signal and to start to cluster these things together. So back to, and by the way, the working title for this is my little thing that I code on nights and weekends. It's called Nylon. It's my little incubator that I work on. So that's why you see Nylon up here in the corner. All right, so scrolling back down, we can see that, okay, so here's the resolve content. iFramely, we got the picture, we got the author, we got the summary.

15:40And here's the main content from Gemini. So we use that, Gemini won here. iFramely won the summary. Gemini won the actual main content. And then what I hit is I hit GPT-5 mini, largely because it's fast as hell and cheap. and it's quite smart. Anthropic has a handful of models that also fall in this camp. I mean, they all do. It's really, at the end of the day, because I'm using vector embeddings from OpenAI, it's just when I have one model provider, I just stick with it. Unless I really need to bounce around, which I did for the actual Gemini crawl and understanding of other things. So I try to keep it as simple as possible, but sometimes you need multiple models.

16:23And I will say Vercel's AI gateway is a great way to kind of code once and just flip a model on the go. So I highly recommend checking out Vercel AI Gateway as a way to quickly swap models rather than having to recode it. All right, so I want a TLDR. That's important. I just want that human readable so you can see that. So I want a TLDR that is a vector TLDR. And so for people that don't know, vector embeddings are really interesting. I had never worked with them before. I'd heard about them. I understood technically how they function. But the point is that if you take a keyword-rich and deep understanding piece of content, you can create mathematical representations of that content and embed these with OpenAI and store them in Postgres by using a vector extension.

17:10So they're actually stored in your database as those pieces of math. And when you apply some of these clustering algorithms on top of them, they get really good at nuanced information where keyword search would completely fall down so the old school ways back in the day in 2004 when i launched you know dig as a social news site if you search for you know apple um releases whatever it's it's fine it's looking the word apple's looking for release is looking for whatever it was back then that they were doing like an iPod or something, right? And it just does it based on just, can I find that text in the database?

17:49And if so, show me back the article. The beautiful thing about what we have today with our understanding of linguistics and around using vector embeds and algorithms on top of that is that you can say, there is a difference, even though they both have the same type of keywords, but there is a huge difference between Apple sues Google and Google sues Apple. And that is impossible to do with keyword search because you're not understanding at a deep level what's going on here, right? So anyway, this is a very rich, purposely rich, longer form version of the TLDR used just for vector embeddings. And then I also wanted to create some key points here that we can use to feed into other models later when we're comparing the difference between articles, when we see multiple articles starting to get clustered together.

18:44And then I don't use this, but I asked AI, hey, write me a spicier title, a title that people might click on more, find more interesting, and just to kind of rewrite the title. And TechMeme does this too. When you go to the front page of TechMeme, it's not the title of the article, it's actually what their editors chose to write. So I just wanted to see how this looked. And then I wanted to put it in one of three different categories, tech core and a couple other categories, largely because there's a lot of stuff that come through these tech feeds, especially when you add in like Forbes and some of these others where it's like, you know, does not relate to actual core tech or AI or the things I care about.

19:17And I just want to put them up in. And so here's the embedding. I'm using the large model from OpenAI. You can see when it was generated, when it was done. And then just some information about making sure that we don't recrawl the article and have all that. So that's done. Okay, any questions so far? We got one article into the system. uh no keep going keep okay so now let's talk about clusters so clusters uh well actually let me show off show off one other thing so how does this actually work so how does that all that work on the back end right and there's a couple different ways that you could do this you could say i want to kind of um write this in next js and i want to have this just be a function and if you want to get a little bit more fancy, you can say, okay, I'm using Supabase, so I'm going to throw on some cron jobs there and fire off these things, and I don't know if they fail.

20:16It gets a little dicey here because oftentimes bad things can happen. You can have an RSS feed that gets blocked temporarily, you can have timeouts, you can have a model that actually doesn't complete. and so I need durability around a lot of this stuff, right? And so what I do for that, actually let me just do a screen share again, is for the durability side, okay, I use a service called trigger.dev. And so I like trigger.dev because what it does is it allows me to create these functions and they're all TypeScript that live in the cloud and that they are fired off either when I call them from my app or at a certain cadence.

21:02And so there are these orchestrators that will go in here. You can see the expansion orchestrator. That's when it wants to go in and expand a story and actually see more information about it. You can see clustering different things here. Here's my fire crawl. Here's my Gemini. Here's my iFramely. So anytime I fire something off to be enriched, I actually create this new little micro instance that goes out and runs on its own. You can see here are the ones that are executing. See the little spinners here? So these are all kind of running in real time. And then you can see the compute charge to go off and execute these and how many milliseconds they took.

21:40Now, the thing that's interesting here is you actually get to see the whole chain in which everything went down here. So you can see, okay, I was looking to kind of resolve a story ID, locked in candidates. I got one picture, one summary. Did it resolve the winners? Yes, Gemini was the winner here. It was published at, and then it finished the whole process. The nice thing about this is that if something fails, I get retries for free. And so I will automatically, if an AI TLDR for a vector embedding, which is very important as we're building clusters, fails once, fails twice, it will continue to retry and automatically spin up these instances and try again and then will report back to me via Sentry or any other type of monitoring software that I have on the back end that I had a failure, why did this actually happen?

22:37I like that because a couple things, if I'm developing this locally and I don't have this on production, my data continues to be enriched in the database and so I can continue to develop before I actually deploy to production so when I'm running this locally like I'm showing you here it's functioning and continuing to build things out there's a few things that are still tied locally that I need to catch up on to fire this back up because I just don't want to be burning through cache for no reason if I'm not using it does that make sense? Yeah, I also think if I remember correctly trigger.dev is open source so we like that I also think it's relatively cheap Oh yeah, it's really inexpensive.

23:21It's something like$10,$15 a month for 50 tasks or something, last I checked. I think when we looked on there, we were actually seeing the per... I'm running thousands and thousands of these. It's under$100 a month. When I say thousands, I mean per day. Obviously, you can choose your instance type within anything else. These are really lightweight, non-computational tasks to do. Some people use Trigger for things like using FFMPEG to do encoding. Things of that nature that are going to require a larger instance are obviously going to be more pricey. It really depends on everything in cloud on the workload.

24:09Totally, yeah. I just don't want people to see this and be like, oh my God, this is probably thousands of dollars a month. It's shockingly inexpensive for what it is. Yeah, and the other thing I was going to tell you is that Vercel now has something called Workflows that they've launched in beta that are basically, it's for free, and it is trigger.dev for free. But it's part of the Vercel ecosystem if you're into that ecosystem. But it's a way to monitor, retry, and have these long working tasks. As you know, with edge functions, that's been the challenge, right? You don't really get a whole hell of a lot of time and things get stale.

24:48Anyway, this is nice. I like it. You're right, though. It is an extra expense, but I think it is relatively inexpensive for this type of task. All right, so back to the clusters. Now what we've done is we've run an algorithm on top of that. We've got all the vector embeds and we're starting to build clusters. As you can see here, this is today. I don't know that everything's caught up. It looks like we've enriched, oh yeah, we're close, 2 ,284 of 2 ,288 stories have been, so 99.8 % has been enriched in the last 24 hours, meaning that they've gone through that entire pipeline. Of that, ones that were more or less failures that we actually had to reach out to Gemini and get some extended data from them, 353 stories were done there.

25:35And you can see here we're starting to get some actual weight. So 105 sources reported on this EU investigation into X over Grok-generated sexual images. NVIDIA invests$2 billion into debt-ridden core weave, 47 stories. And you're starting to get something that looks kind of like a tech meme-ish type feed. And then obviously you can slice and dice this in a variety of different ways. But this is where it gets even crazier. So I promised you crazy stuff. This is where we go down the rabbit hole. So let's just take this idea of CoreWeave and NVIDIA investing$2 billion. So we'll click on that cluster.

26:17And now we have 47 stories. Nine were done via RSS and 38 discovered. So what does that mean? That means that when I see enough signal, and I consider that three or more RSS stories talking about the same thing, I then hit search APIs. You can hit Brave's search API. I use Tavly's search API. and you can say, go out and find me other stories outside of my RSS ecosystem that may match this and bring them in. So I can expand the scope and see, is this a broader story that has more context that I'm just not seeing because of my finite set of RSS articles? Does that make sense? Absolutely. Okay, so, and then I'm looking at, there's average distance between articles and similarities.

27:04This is all part of this algorithm right here that I'm using for the clustering. So first story started four hours ago. Last minute was 13 minutes ago. So then I created something called the Gravity Engine. And so this is kind of like a editorial-type score, and then it has actually, it votes on it on how important this story is, generally speaking. And I rate these in a few different buckets here, like the impact, another one called it's gravity, which we can get into, the confidence ratio that this pertains to technology and the things that I'm interested in. And then this is all the stories that it's evaluated.

27:38And then here's a little matrix here that I've had built out where you can see you have impact and then you have gravity and it's in the high, high area here. Viral potential and how bubble size, like how big this will eventually get and how many people will be talking about it within tech is 60%. Early trend, like a growing trend is 75%. Intellectual gravity is 86 % and impact is 88%. And so a total editorial vote of 85%. And then we can get into the rubric here. So these are the impact dimensions. So X-axis drivers are industry impact. So it's 90%. So the massive$2 billion investment solidifies CoreWeave's position as a key new cloud provider, intensifying competition with hyperscalers.