Episode Transcript
[00:00:00] Speaker A: I want to start by talking about notes. So N o a T S AI So kind of granola replacement local recording for calls.
Local meaning like local models too. Right? You plug your own model, bring your own keys kind of thing.
[00:00:19] Speaker B: Yeah. I mean we had initially thought of it as fully local and then and that was about a year ago and the product just wasn't there. Like it couldn't do it. And then maybe a few months ago basically could do it. And do it means like hit close enough parity where it's 90 plus percent as good as something like granola and fully locally. And it was. And then the world moved in a year. We should have open sourced it a year ago because we would have been first or close to first and built a business off of that. But at the same time like the tech wasn't there so it would have been janky for real commercial use.
Now it's there, but the market's very crowded and it's consolidating and the play is actually very different than a meeting recorder a year later. Right. It's more about what Granola's gotten to, which is essentially the power of what happens when you can record everything in a company. That's what they're going for. Right. So the play is very different than individual use of I'm going to record a meeting for my use. It's actually we're trying to capture the knowledge across the entire company from these like live meetings that we have. How do we do that? Right. And then there's all these cultural and all these implications of all of that that thankfully Granola is dealing with and will probably open up people's minds to these things being good. Because when I was at Dropbox there was only a handful of us using Granola. And that's when Granola first came out. It was a couple years ago and I was one of the first earliest users and there were just some things that they did were, that were brilliant back then that now like they've lost touch with on the UI ux.
So anyway, that's the, that's a reaction I have to this.
[00:02:07] Speaker A: No.
[00:02:07] Speaker B: Yeah.
[00:02:08] Speaker A: And, and it's like that's why I wanted to start here because that is like three or four things I want to talk about. Right. Like local models, product motion. Right. Like product, product marketing. Marketing in general and like how you can go bottom up right. In, in an organization. So I want to start with local models because there's just so much to it. You said something I think some people won't really Pick up on intuitively, which is like, the tech isn't there. You don't mean like the JavaScript app you built? Like, that's been there. It's like, hey, Quinn, 3.8, 27 billion parameter model now can run on a reasonable piece of hardware and is good enough to do everything you need it to do for. For this application.
[00:02:54] Speaker B: Yeah, more or less. It's actually Quinn. And then for another application, type ahead. Gamma 4 is what really changed the game on that product. And then now we have a third product that has no AI, but it's meant for AI workflows where we help bring a window from another Mac. And I think we were the first to do it, but whatever, it doesn't matter now. Bring a window from another Mac into your current Mac instead of bringing the whole screen. People want the whole screen, but the window is just a different way to think about it. And it actually came off of our own use cases where we needed to manage computers remotely and not remotely and only wanted the window. But now you could have 10 computers and 10 windows in from those, one from each computer on one computer. And that's very different than trying to do screen sharing through tailscale and all this other stuff or any other form where you're only able to. Where you're only able to do the whole thing. So it's like you can bring a tab in. Right. Like with Chrome and stuff like that. So that's the third product out of that business that's a fun one and trying to do a few different things there, but there's no AI in it. Yeah, it's just meant to help with AI.
[00:04:08] Speaker A: Yeah, yeah. So you and I chatted on Twitter a little bit after, like, one of your live streams. You do weekly live streams, which are amazing. Very. I think you strike a really nice balance of, like, speaking about complex things at a level everyone can understand all the way up to, like, power users, which is hard. As someone who does a bit of this, it's easy to, like, get lost in the weeds or the clouds maybe.
Either way.
We were talking about, like, hey, I want to get a local setup that can run good models.
And you kind of said, like, hey, you have a bunch, but you kind of need. You need like 100 gigs of ram kind of as the short version. Right. To run, like, really good models. And so this is like a Mac Studio or a couple DGS sparks, or
[00:04:55] Speaker B: a MacBook with 128 gigs of RAM.
Right. Like the MacBooks can have that. Mine don't because I have 64, but I have five 500 and 12 gig RAM machines. So I don't have this problem, so to speak, of not enough compute running around or not enough RAM running around at the same time. Like those large models are still super slow on the equipment I have. Yes. When I put two GJX Sparks and a few Mac studios together, things get a little better, but only a little better. Not like enough tokens per second to run enough jobs all at the same time. And everyone's got used to doing that. If you're doing coding, of course.
So I think there's all these weird things about local models in terms of what people keep saying is possible versus the actual reality and practicality of making that possible. Is it true that if you have enough ram you can run close to frontier level models? Yes. And I say close because it's all still debatable because we're dealing with Kimi and all these things. And we're also dealing with models trained with a Chinese flavor. That's the way I would say it. We haven't seen model, we have from Nvidia a little bit, but models trained that are open weight, that are US flavor. And what's not about geopolitical or anything, it's just every model, every time you train a model, like this is the short of it. Every time you train a model, you're hoping that the training you do optimizes for the thing that you did the training for.
So if you're training for coding, you're hoping that that training run turns into a better model for coding because you put the training on that emphasis. The problem with models and training is it's a black box.
So even the OpenAI and anthropic, they don't know what the model will be great at. They do know what they intend the model to be great at based on training. So they only know what they trained the model on, but they don't know once it's done what the other end is going to look like. Is it going to be good at the things we trained it on? Probably. Is it going to be great?
Don't know.
Not even a maybe. Just don't know. So when you really dig into all this stuff, you're like, we're training models. It's a black box. People are trying to figure out the black box, but it will probably always be a black box. That's what it feels like because of the way it does computation and things like that that human brains don't. We can't process like that.
[00:07:33] Speaker A: Yeah.
[00:07:34] Speaker B: Right. So then we're trying to use AI to tell us what's going on while it's training and in order to help optimize that training. So then the output is something that we can expect. So the problem is every three to four months there's model cycles and depending on what the models are trained on, they're going to get a different. You're going to get a different reaction from the model or a different observation of how it works and all that. This was a lot.
This was a lot more dramatic a while ago.
That's why you had people stuck on GPT4O and didn't like the new one. Or, you know, Sonnet was really good and Opus, but Opus is more powerful. But Sonnet still does this thing better. I like the way it talks to me and all that. It's because they don't know what's going to come out the other end until it comes out.
Yeah, that's the truth.
[00:08:32] Speaker A: It's a challenging marketing. People don't understand.
[00:08:35] Speaker B: Yeah, yeah.
[00:08:35] Speaker A: It's a challenging marketing problem. Right. Like we're going to spend all this time and money, if you're a frontier lab, developing this thing and then we have to figure out how to position it.
[00:08:43] Speaker B: Right.
[00:08:44] Speaker A: And that's kind of what you see
[00:08:45] Speaker B: from, you know, and these companies aren't good at positioning. That's the other problem.
[00:08:49] Speaker A: Yeah, right.
[00:08:50] Speaker B: Because the positioning problem is what can it do? And then how do we frame it for other people so they understand what more it can do than the previous ones.
[00:09:01] Speaker A: Right.
[00:09:01] Speaker B: The more part is where things get really tricky. That's why you have evals and benchmarks and the whole scale, AI, data training, all this stuff. It's because we today don't know how. When we train a model, what's going to happen on the other side? Like one to one.
Of course, if we try to train it on code, it's going to be better code.
How much better? No clue. Right. That's kind of the current consensus. And that's why I say, like, this is the. This is. This is the thing. If you're using Kimmy or Quinn, you will get Chinese character output when the model feels like it cannot describe something in English to you. And there is a better foreign word. Foreign as in foreign to English speakers for the thing.
[00:09:47] Speaker A: Interesting.
[00:09:48] Speaker B: That's what I mean by flavor.
It will still spit out some foreign language that it was trained on, usually Chinese, that me as not a native non Chinese, you know, I don't speak any Chinese or can read It. So what do I have to do? Why is the AI doing that? What was it trained on? That it was doing that. And again, none of this is geopolitical for me. Like, forget all that. Like, that has its own, all kinds of issues and implications and someone else's problem. My, my problem is simply I get the model, I try to use the model. This is what the model does.
[00:10:22] Speaker A: Why don't you think there's been a US based open weight model?
[00:10:27] Speaker B: There have been recently. There's one called Poolside, or not one called Poolside, but a company called Poolside that's throwing it out. Nvidia did a massive deal with them and kind of bought the IP in some funky stuff, not bad stuff, seems legit. And Nvidia is really the one that seems to be training open weight, open source models. They bought Hugging Face. So there are moves, there are significant moves happening from Nvidia in particular, who says they're building tooling for AI factories.
And if that's what they're doing, they want all of us to be running AI factories off of their software and hardware and all that. And now it doesn't even matter because if we're using Hugging Face, Nvidia is happy because they bought it. So to me, we are there. We are definitely there. We're just maybe six months away from seeing some significant competition from US Labs, or not US Labs, but I would say Nvidia in particular, maybe others.
But we also have new stuff coming out, right, all the time and new random things that some are interesting, some are not. To your point about the hype, and I think cutting through the hype is near impossible for most people. That's kind of what I would say. And so I'm just trying to do that as we talk, which is like, hey, let's unpack this. Let's make sure we all understand what the heck is probably going on. Because again, no one knows. Even the folks at the labs don't know because they don't know what the other labs are doing.
[00:11:56] Speaker A: I run a little bit of local models and my mix of local versus frontier is like 95 plus percent frontier models, cloud based there currently. And this will probably change in the future. Cheaper, better, easier.
What's your current mix of like local versus cloud and frontier models?
[00:12:19] Speaker B: I use everything and I use everything aggressively.
So if you ask me about Instinct, I have an opinion. You ask me about Muse, I have an opinion. You ask me about some latest grow model. If I don't have an opinion, my opinion's five minutes away.
[00:12:33] Speaker A: Yeah.
[00:12:33] Speaker B: Because what I would do, and it's probably not five, it's probably more like 30. Because what I would do is I go to a bot. Like, someone just did this to me. They're like, hey, have you tried this model? Answer is no. So I'm like, no, I'm about to try it right now. Throw it to a bot and say, hey, this model came out. Buddy asked me about it. Can you do your thing? And then it goes and evals and does its thing against my use cases and whatever. And then it spits out, is it better for what I'm trying to do or not? Or is it just better against some benchmarks that are out there? That's fine, too.
[00:13:00] Speaker A: Yeah.
[00:13:00] Speaker B: And I'm done. So. So appreciate the question. But to me, it's more like, do I have all the things set up to be able to run at the level I want to when new things come out?
So when instinct first came out, like, so. So I'll explain how I think about this. Twitter is where everything breaks. Like, and I mean that in both ways. News breaks, models break, life breaks. Right. It's where everything breaks. If someone asked me, what is that thing? X. Sorry, not Twitter.
It's the place where everything breaks. Got it? Cool. Our narratives break, our news break, stories break, models break, and break is the word in both definitions. Yeah, Right. Because someone launched some trained model months ago. I forgot the person and all that stuff, and it was bogus.
His model broke on Twitter both ways. He broke the news and said, I'm doing this crazy thing. Turned out it used to be full distillation, which is using another model or something like that. And then his old model broke. Right. So I'd say one big piece of the puzzle for me is I'm hyperactive on that platform because that's where things break. So that gives me that kind of signal. Right. Then on top of that, I have machines set up so that they can easily test this. So if I want to test instinct, I can go to a machine and be like, sign up for this thing and just test it out. And it's usually open call or Hermes agent, and I'll pick the one that's isolated enough, doesn't have access to my emails or anything weird on the computer. It's pretty clean. And it goes and tries the thing out, and then it tells me what it thinks of it, and then I'll still use it. Don't get me wrong, I'm not like a nut, but, like, that's the first line of defense. So my understanding of instinct comes from both the AI analyzing the crap out of it in ways I can't, and then myself using it as a user and just seeing what it can do.
So when you ask this, I'm like, oh, are you set up?
The question I would reflect back to someone is, are you set up to experiment the way you want to? So for me, I want to experiment with the cutting edge. I'm addicted to Twitter for that reason and that reason only, because everything breaks there. And number two, I have my machine set up so I can screw with the local model. So when someone says something smart or points me to something, I can just point a bot to it and the bot will handle it. Right. And to be frank, like, Muse and Grokbot make that even easier now because they are commercial, productized things that let AI use a computer on your behalf, and it doesn't have to be your computer.
So now it's like, go to Grokbot and be like, tell me about this local model. It could do that.
[00:15:40] Speaker A: Yeah, yeah, yeah.
[00:15:41] Speaker B: And test it. It could do that.
More or less, give or take.
[00:15:46] Speaker A: Yeah. I think in practice, for so many folks, all the way from maybe folks who are kind of a couple steps rungs lower than you on the kind of, like, amount they're all in on the bleeding edge of AI, like, folks who are pretty deep enthusiasts, all the way to, like.
[00:16:05] Speaker B: I work at a corporate job and
[00:16:06] Speaker A: I want to stay on top of things, but I'm not subscribed to eight different things.
The challenge is almost a practical one. I think of, I have my workflow, I have my tooling, I have my subscriptions, I can do the evals, but then there's another step of, oh, I'm going to change my actual workflow to use this other thing. Now, I think there's a difference between I'm going to play around with this and when I want to get a piece of work done, I'm going to go over here.
Is that me because I'm not sophisticated enough, or are you actually switching the tools you're using on a. On a pretty regular basis based on, like, what. What is working best?
[00:16:45] Speaker B: Like, this week, I think AI is very fungible, and people don't really, like, grasp what that means.
So to me, it goes back to what I said earlier. Am I set up to experiment the way I want to? The way I want to experiment is something new comes out, I can test it for value, but that value is up to me to determine what is valuable.
So if I Know, I have these workflows on these machines and the machines are set up in a way where I can just throw another model and be like, tested against what we're already doing. Is it better? And it can give me a rating or a ranking and I'm good to go. But this is like treating it like a workflow, a process, a system, something like that. And then if you take the factory analogy all the way, I think it's a good analogy because it takes it out of the humans. An AI factory means you're not trying to get humans to run any piece of that factory.
So then you're just trying to figure out how much of that can an AI run and where do I need to insert myself. So if you think about that, my machines and everything I'm doing is a factory and there's a bunch of inputs and widgets and stuff that come out of it and a bunch of workflows. And so when something new comes in, I should be able to easily take test it and figure out what it can do for me. So I don't think there's a wrong or right way. But if you're happy with your level of experimentation and your level of ability to pay attention to that, then you'll have something closer to what I have. But then you have other folks, like my wife Amy, who I didn't want to throw any AI at. We threw a bunch of bots in Slack. She uses them all. Good.
She's been running Crazy Egg for the last 18 plus years, since the last time I touched it until recently.
And so she's in tech, has to deal with all the stuff most of us deal with to build tech products. But she doesn't care for agents unless they're easy. So we put some in Slack. The engineers manage it at Crazy Egg. I manage one and we use them and they're connected to all the systems. So now a support ticket can turn into a pull request. Okay, fine. This is fairly standard if you're trying to AI pill a company or whatever you want to call it.
But I threw her muse because I was like, it's not instinct. So it's not some stupid startup trying to do this. No offense to them, I think it's great that they're doing this, but it's a stupid startup trying to do something that is probably going to require a lot of money, capital and people.
[00:19:12] Speaker A: Big boy company, you know that.
[00:19:14] Speaker B: Yeah. And like, and I can say that right now because they didn't hire a CISO yet. From what I know, and that's a company that needs to hire a CSO asap, especially with what Meta is doing. That's why I'm. I'm saying that very deliberately, which is like, they need to know a game they're playing. I think they know a game they're playing on capital, but I'm not sure if they know that what game they're playing on infrastructure and tech, because Meta is playing deep game there. They basically got a lot of crazy security controls inside of Muse, even though Muse is a consumer app. Why?
Trying to make those IT and security people happy. Why?
Because that's a transfer. That's an opportunity to transfer people's opinion from one way to another off of a product and how seriously they're taking the product. Because when you went to real engineers that used to write code without AI and you told them what Open Call was doing and you could get an agent to write code and run the code, they wanted to shoot me in the head.
So Meta's got. Is on point with everything they're saying and doing.
So when I think about it like that, I'm like, why would I let anyone in my family use a startup for this kind of personal productivity that AI can do?
As long as everyone's okay with the privacy implications of that. Giving it to Facebook. But the thing is, you're going to give it to a startup or are you going to give it to Facebook? You've already been giving it to Facebook. Why would you give it to a startup too? That's stupid.
Personal safety and privacy. That's just stupid. Sorry, like,
[00:20:39] Speaker A: I want to, I want to talk about Notes as, as a example of go to market, because I think this is, this is one of the discussions that I have with folks a lot, which is like,
[00:20:53] Speaker B: even if
[00:20:53] Speaker A: you're an AI company, like, go to market is hard.
There's so much noise. I think people's attention span is consumed with just keeping up with AI. Much more so than like, hey, how can I actually do things?
Like, in my business, how are you approaching, like, growth with Notes as a, as an AI company? Right. Because, like, I know, I know your last, like, product launch before this was like a few years ago. Like, how is the reality now, like, different for you guys in, like, trying to grow this? And it's a free tool right now, right?
[00:21:34] Speaker B: Yeah, I would say that I wouldn't ship a product publicly without a good solid idea of distribution. And so a majority of the distribution on our products are not on Twitter.
So if I thought that Twitter could Get me distribution, then I'd be doing very different things with notes and these products than if I thought there was a better or different way to get distribution. So we're trying every channel, not Twitter, for distribution, even though it might look like we're trying Twitter, because of course we're going to try Twitter and let people know what I'm working on. But I don't think that's how I'm going to get users for that product. It's just other distribution channels that are more interesting or lucrative than X or Twitter for distribution.
Most mostly for monetization, more so than distribution. It's just like a lot of people, again, because Twitter breaks things.
A lot of people are trying to sit on the edge if they're on Twitter. And so providing them with products that are more either prosumer or work focused doesn't really accomplish the goal of getting them to convert to become paying customers.
[00:22:45] Speaker A: Okay, what does that mean then?
So I get like. Not Twitter, I'm sure, like there's whatever channels out there. There's other social platforms, there's YouTube, there's paid, there's cold email, there's events and networking and referrals and things like that. You guys kind of trying everything. Do you have a hunch of like, what works? And you don't have to like, spill the beans. But like, I think that's probably the big question that folks have is like, how do I think about growth right now?
[00:23:09] Speaker B: I think that we.
Yeah, in that business, we focus very much on building products that we think have a large enough appeal that more channels are opened up than less.
So we could be running Google Ads or TikTok, Meta, you know, all that stuff, even a bunch of stuff on Reddit, and we can see it working more so than other channels. And the simple reason is like that those products we built have millions of users, Tam, if not tens of millions. Right. So if anyone can use these things, that's one piece of comfort for me in being able to open up the channels that require that kind of mass coverage. So that's the way we're thinking about it for that business. If we launch anything niche, like we launched this product called Beam, or not launched it, but are playing with it that, you know, lets you do the window thing. I was saying that's probably not mass appeal.
So that one's actually interesting because we actually have a Twitter group for that product and a group chat and have just, you know, 30, 40 early users on it. It is a paid product today. We're probably Gonna bump the price up and do a bunch of other things and experiments on marketing. But the way I think about this today is like, and I've been saying this to a lot of people to explain what I think's going on is that in the past your go to market was a wrapper around the product.
Today the product is a wrapper around
[00:24:37] Speaker A: the go to market.
[00:24:38] Speaker B: So Type Ahead is very focused on paid ads and we have a price point. We've been experimenting and testing it there Notes is very focused on freemium but we don't know how that's going to play out fully yet. There's some paid opportunity there too, but that one is a little more TBD on perfect distribution because we waited too long to open source it and we're not going to open source it. Most likely if we would have open sourced it, that would have been the channel, which is GitHub. But we don't have a play there and I do want one there on on GitHub with that business, but not yet. Then we have Beam, which is much more specific and much more of a niche that could potentially be either used or monetized by folks on Twitter. Because those are the like that crowd is the main crowd that has multiple machines hanging around and trying to do stuff with compared to finding those people elsewhere besides subreddits and things like that. Right. And there's a bunch of like product stuff we can do like connect like make it so that it's easy to bring a Mac window into omark or bring a marquee window into a Mac. Right. And things like that. So those are things on our mind. They're on our roadmap. But without nailing this thing that we're doing now which is Mac to Mac. Doesn't make sense to go do that yet. Right. In our minds because we're really testing the core use case we have today and if other people have that problem and there's just a lot to build on reliability for the type of products we build mainly because it's either local AI or comms between computers and stuff like that and requires a bunch of work and performance issues and stuff we run into. So I like working on little like right on the edge or little tweaky problems that someone could ship a wide coded version. Please do. But it's not going to do the job well, you know that's kind of the other thing. So yeah, those are the three different ones and how I think about it today. But there's some ideas with notes that I'm very excited about. I just don't know how it's going to play out fully and should have, should have, have released it differently.
[00:26:29] Speaker A: Yeah, I mean I see, I see a lot of the AI tools and AI products that are free. Like you mentioned the Mac tool, like probably won't be free. That makes sense, I think. But like, do you think most software products need to have a free option these days?
[00:26:48] Speaker B: No, I think products need a distribution model that matches with their business model.
So that's what I mean by like your go to market is that the core, not your product, if you go to market is at the core. You need to address that first. I know people say, oh, we always worry about distribution first, all that. Yeah, yeah, like you said it. But you didn't do it. Now you have to do it. And you shouldn't be saying it because it's reality. Like you don't need to say distribution first. You just need to realize that it's go to market or nothing because you, because people can put out more stuff than ever. So then all you got is go to market and figuring out how do you actually make that the core. And then wrap the products. You build around some go to market motion or set of motions for the products. Right. So we will keep tweaking the products until the go to market works. That's what I mean. While in the past you would tweak the go to market until it worked. Right now that's what I'm saying. The product is a wrapper on the go to market. Right. Because we would tweak the product in the go to market in parallel. But it was product first. Now it's go to market first and you're tweaking the product to make that match happen. Which is a flip from how we used to do it. We'd keep tweaking the product and the go to market would kind of come in place. Now you tweak the go to market and then a product sort of comes in place. We've learned a lot from even doing paid ads on where type ahead should go. If we didn't do the paid ads, we wouldn't understand where it should go. And we weren't using feedback in a traditional customer feedback sense. Although that was great. We were actually using signals that we got from the market on the paid side. Cause that was our distribution channel. And now in the next like 3.0 of type ahead, we're going to tweak. We're going to tweak a lot to align with what we learned about go to market.
[00:28:26] Speaker A: Okay.
Go to market kind of being the core of, like, the business motion. The product needs to kind of fit within the go to market strategy.
Are you kind of. It sounds like you are across, like, the several different business units of like, kind of bottom up or selling, like enterprise or like direct selling to the consumer. Like, those models, like, are. You're. You're kind of indifferent as to, like, which approach you're taking on, on your products these days.
[00:28:55] Speaker B: Yeah, very much so. I, Yeah, I.
[00:28:57] Speaker A: We.
[00:28:58] Speaker B: We want to be ready to tweak our go to market at any time.
[00:29:01] Speaker A: Yep.
[00:29:01] Speaker B: Just like we used to be ready to tweak our products at any time. But tweaking products now is much easier than ever. Tweaking go to market is not much easier than ever. It's only one piece of. It is probably easier, which is the creation of the assets, but everything else is much harder, actually.
[00:29:18] Speaker A: Yeah.
I want to ask about Crazy Egg you mentioned. Like, you haven't kind of touched it in a while.
What's that? Like, I would feel a lot of, like, guilt and remorse, I think, if I didn't touch Kastos for a while, even though, like, I'm not in it full time every day. But, like, how has that kind of journey been emotionally for you?
[00:29:42] Speaker B: It's not very emotional. I've been involved in the business for a very long time, and I did have, like, calls with people usually once a week, a couple calls throughout the last however long.
So there's no emotional part about it.
I had other things I wanted to do and I did them, and I learned a lot doing those things. I think coming back to this business 20 years later in a different reality of where the world is, is still the same reality if I were building something else. So I think the.
There is no difference except I have resources that I wouldn't have if I started from scratch. And so then. And because it's a relatively small company from a headcount standpoint, we have always been very already sort of efficiency pilled, for lack of a better word. And now with AI, we can do that even more. And we're just starting to do that more and more and more and learning how to do that.
You know, if we tried to do some kind of, you know, synthetic heat map and make, you know, because that business was originally about website heat maps and try to like, make AI make a heat map and make it accurate.
Took a lot of money and time not too long ago.
Today, 20 bucks, right. I can train a model on Stuff, you know, and it's not a big deal. So there's some advantages from having lu the luxury of time to some extent, especially right now, as long as you know where you're trying to get to and with that business. This is a very obvious place that everyone's trying to get to if they have any website business, which is make the websites change automatically. There is no other kind of path for anyone that's trying to do anything related to websites except make them change automatically.
So that's just a very obvious future.
And we're not there yet. Not even us. No one is there from what I can see, because this is a place I experiment with a lot and look into.
That's what we want to build for. So to me, it looks like if you're in a somewhat legacy business and trying to transform it, you still need the same things as before, except you might need to reevaluate the how more often than not. So even today, with some new model that came out a couple days ago and one that came out a couple weeks ago, were like, oh, this stuff's easier just because what's possible now?
[00:32:07] Speaker A: Yeah, yeah, yeah.
[00:32:09] Speaker B: So. But the North Star is still the North Star, is my point.
[00:32:16] Speaker A: As you're on Twitter where everything is breaking and you're looking at new products, evaluating them, what are you, I guess, both, like, excited about that, like, isn't true today, but that you think is probably coming and like, what's keeping you up at night?
[00:32:34] Speaker B: I think change is keeping me up at night, even though I'm pretty used to it compared to most people, I would say.
So that keeps me up at night. What gets me excited is the same thing, because in change there's opportunity. It's just in this pace of change, opportunity is somewhat fleeting, and it's never really been fleeting in this way in my experience.
And a lot of stuff is up for grabs that wasn't. And determining what's durable is super tough. Right. It didn't, it didn't dawn on me that, you know, right when I got to crazy egg that, oh, the future is websites changing themselves and AI doing it, and even AI visiting websites and what the implications are, which people think it's AEO and geo, and it's absolutely not that. But websites changing themselves, it's very obvious that if you're an analytics company focused on websites, we're not focused on products or apps or mobile or anything, we're focused on websites, the websites are going to change themselves. So as long as that's the core Like Objective in North Star, we're just going to scramble to get there. And we want to be a big piece of how websites transform in the future. And we'll die trying. Right. And that's very obvious. So I think there are obvious futures for most products today. And trying to find that obvious future and then being very flexible about how we get there compared to the past is really the key.
So when you talk about Instinct and Muse, I'm just looking at the future possibilities for instinct and I don't see them because Muse will be a better product and they have too many advantages, in my opinion. Does that mean instinct shouldn't exist? No, they should try. I respect it. I respect the people building it and the investors. But the problem is there's a very obvious future to that business. And if that future, someone else beats them to it and they can't get distribution fast enough for it and the data loops working, should they even try? And I don't know the answer to that and I don't know what their vision is, but it looks exactly the same as what Muse from Facebook or Meta is building. So then what? Even Anthropic and OpenAI don't look exactly the same anymore. They have different versions of the future they want to go after.
[00:34:49] Speaker A: It's a daunting thing that, like, you saying that. And I think about my business, think about like all my friends businesses. To say, like, we need to think about what the future version of our business will be or what our business will look like in a future reality like that, That's a heady kind of thing.
I'm just not that smart. So how do you think about not my business, but, like, if you're to abstract it to like, questions you should ask yourself or like a thought process you would go through, maybe that you went through with crazy egg, like, how can folks be like, okay, here's my business today. Say we're kind of like relatively legacy.
[00:35:22] Speaker B: No, no, let's do your business.
Cool.
[00:35:25] Speaker A: Let's do my business.
[00:35:25] Speaker B: Yeah, yeah. What's your business? Break it down.
[00:35:28] Speaker A: Okay, so we're Castos. We're a podcast hosting platform. So we're a content management system for podcasts. So if you want a podcast, it has to be hosted somewhere. We do analytics, we do monetization.
[00:35:41] Speaker B: Yeah.
So you obviously have choices. Right?
At the bottom line of the choice in my mind is everything around hosting a podcast is commoditized.
That's where I'm coming from.
So if just like website analytics is commoditized. Right. So if it's commoditized, then what is the obvious future off of that commoditization that can exist and that you can be a part of? To me, when I, and I'm dangerous on this topic because I do content stuff. So to me it's like full on automation of what happens after so I can actually grow my podcast.
So if so to me it's like this table stakes to record and host and do all of that.
It's much harder today, even today, although there's hype on Twitter around this of cutting up the stuff and making it consumable in other ways, which is what provides distribution for a podcast, then you have all kinds of complications like video, audio is AI good at cutting it up, rejiggering stuff or making social media stuff. For me it is good enough. But all the data and context, you have the historical data and context on my podcast as well as everyone else's out there and all that stuff is all great context and data to be able to feed the engine for doing the other parts of this product or business that you might not be doing too much of today, but it feels very obvious right now. The problem there is like everyone's going to go after these potential realities that are very likely to happen.
So you need to figure out what your differentiation or point of view is because differentiation is still super important, right? Like at Crazy Egg, we've found an ICP that we're very excited about that we're just going to focus the entire business on and ignore every other ICP we can find because this one fits what we think we can do and what we think we can do with go to market. If it doesn't work, I'll know in 12 months. That's the problem because it's making a bet on something. But that's okay for us. So I think you have to determine a similar type of thing and with an obvious future. The obvious future is a recording happens.
Everything else is taken care of. What is everything else? That's up to you to decide. But when I play out your business, that's where I go. There might be other things and you might have other inputs. Happy to talk about now or some other time. But like off the inputs you gave and what you said you're doing, to me, that feels obvious. But even the platform we're sitting on right now wants to do the same thing.
[00:38:17] Speaker A: Yeah.
[00:38:18] Speaker B: So then what do you do?
[00:38:20] Speaker A: Right to, to, to your point earlier around like the, the product is a wrapper around distribution.
So that means like in our case, let's say repurposing content and getting it ready to. To publish is. Is the thing that distribution for us would be like social based. Right. It's like we show up everywhere and show off ourselves and our customers on social because that's what we do. And it's kind of the most like, aligned with what our product does. Is that how you would think about it?
[00:38:52] Speaker B: Yeah. Like, why aren't you doing social media scheduling?
Why aren't you automating that? Why aren't you getting really good at that? There's so much data out there that you can use to get good at that. That's not even your data. That's data you can go scrape and get. So to me, it's like it's changing the job that people hire your product to do today. If they're hiring it to manage my podcast, cool.
I think in the future they should be hiring it to grow my podcast. Because if you can't grow my podcast, why am I using you? And that is going to be a reality at some point. I don't know if it's six months, a year, or two years. Right. But within six months to two years, most of the platforms that do what you do are going to need to help a podcast grow. That's my opinion. Yeah. There's other caveats. Like, if you go enterprise, maybe that's not as big of a goal. Although I would argue it's same goal. They want a podcast and all I'm going to is what is the ultimate value prop for a product in the podcasting category? The ultimate value prop is I sign up for it. It helps me get more readership viewers, whatever. Viewership, Right.
So then you can even start thinking, be like, oh, what metric do we optimize for? I just gave you the one.
Right. It's essentially attributed viewership to your business.
If. If someone comes on and they can attribute new viewership or viewership increases or whatever attributed to your business, you win. That's my take today. That's not the metric. That's not the bar. The bar is. It just works for hosting. Right.
Cool.
[00:40:27] Speaker A: Oh, it's shifting, but yeah.
[00:40:28] Speaker B: Right. So now the bar has moved. That's table stakes. You want to be great at that. You don't want to let go of that. That probably informs the rest of it.
[00:40:35] Speaker A: But.
[00:40:35] Speaker B: But ultimately, if you're not increasing my viewership, why should I use you? Because someone like you is helping me increase my viewership in the future. Right. Like, that's the kind of logic tree. I would go down and try to break it down like that.
[00:40:50] Speaker A: Cool. Helpful. Helpful.
I'm not going to. I'm not going to. Just because I'm genuinely very curious. I want to hear more specifically around things you're excited about, like, I don't know, local models or, I don't. I don't even know. There have to be things where you're like, gosh, this is a pain. And when this is solved, it's going to be awesome.
Around, like, AI and AI tooling.
[00:41:16] Speaker B: I think I'm sitting in a moment where I've already seen the realization of things I wanted to exist, and my opinions are playing out. So I'll give an example. So there is.
There was openclaw, there was Hermes. I know you're familiar with all this and use them and all that. And you were using them fairly early, like me, and those really shed light on what I said earlier. Computer or AI can write code and run the code. That was really the big switch that freaked out all the engineers. If Facebook is launching a product that does that today after they tried to buy Open Clause founder, acquire him or whatever, think about how far we've already come in like six months.
Fine, nine months. Let's give it nine. Not six.
You want to say a year? Cool. Open Cocky out in November. Fine, right. But like this year, 2026, right, we went from AI not being able to use a computer or a browser to AI fully, reliably being able to use a computer in a browser, right?
So when you ask me that, I'm like, dude, yeah, my priors, as the nerds call it, are already solved for. They've already changed with the reduction of friction. I'm telling my lovely wife to use Muse because it did all the things I was trying to do with openclaw, or not even me, I wasn't trying to do productivity things, but the things that people were trying to do with that stuff is now one click. And don't even get me started about Grokbot and what that does, right? If you have the concept of these things being teammates, which I think is reasonable for now, I don't think they're teammates in the long run, but it's totally the right paradigm right now.
They came out, Muse came out. So I can talk shit on instinct as a startup because Muse is probably better and will probably be better and will probably move fast enough knowing that the A team is working on it. That's my metric. If Muse was not the A plus team working on it, yeah, Instinct has a big shot at going after something. But now they're compared constantly, right? And here's the funny thing. Meta knows what to do with competition. Okay? Now, Grokba, Hermes will keep trying. And they have some angles. Open claw, I don't know. Although it's gotten a little better. So we're starting to use it a little bit more than we were recently.
But I think it might be too little, too late. But I don't know. This is armchair quarterbacking, right? But I'm trying to prove my point. So then you have Grokba. So if someone wants Hermes or openclaw, what's my answer today? Hey, Grokpa, you want the easier version? Try that first. Just try it first. I'm not saying openclaw, Hermes suck. No, what I'm saying is there is at least 50 to 80% of use cases that most people were trying to do with those things that get solved with this thing. And between Muse and Grokbot, dude, 90%, 95%. Like, how much coverage do you have? 85. If I want to be nice, I probably would peg it at 85% of what we were trying to do with Openclaw and Hermes agent can be done between Muse and Crocba.
If that's true, you're asking me for something where I've already. I'm already at a place where like, yeah, I thought this was going to become popular, right? It already has. The next thing is unseen, but there is some stuff coming out right now where it's like, oh, like there's some cool stuff coming out, but I can't. It's not in the prediction area yet because it's like two weeks old. And I don't like making predictions. But there's Town, there's Victor, there's like one or two others that are somewhat popular. But why would I use those?
Hyper Agent from Airtable. These all. At one point when these all came out, I looked, took screenshots of all of them, sent them to a friend, and said, hey, can you tell the difference?
And the answer was no.
No. And they all had complicated, heavy interfaces that didn't even look as good as ChatGPT. As in, they were at least usable. ChatGPT is very usable for all the crap it has. Claude is not in the same way ChatGPT is, but they care about consumers. You could tell they do. You could feel it when you use their products. You can't feel that from Claude, and you can't feel it from most of the products out there. You could Feel it from Grokbot. You want a teammate. You want a whole set of teammates that are fungible. Grockbot does it for you, right? You could feel it with Muse. You want a personal assistant that does badass shit. You got it right there, right? Sure. Instinct with the text messaging is cool, but I use poke before that. What is the difference between poke and instinct? One difference. It came out later and the models are better. That's it. There's no other difference between poke and instinct in my mind. And yes, some of the paradigms with computer use, but what is that? The models got better, otherwise they couldn't use the computers before. Yeah. So I guess for me, the things I get excited about is we're starting to get into a world where the input for AI might not just be text, and we're starting to see that now. As of a few days ago, I was trying to get everyone off the wait list for that types fast or whatever type. I always forget the name. But that thing, the Jet or Jeb or whatever.
[00:46:20] Speaker A: Yeah.
[00:46:22] Speaker B: I was trying to get all my friends on it before they were letting people off the wait list. Because I was able to do that, right? Why I don't get hyped. I don't care about that. Like, I've never, like, I only give people an instinct invite. If they ask me, I don't give them one, but they ask me because they don't have one or whatever. So I have a bunch. I can give them out.
But this one, I was giving to everyone I know. I was like, hey, do you want access to this? I can talk to them and get you access off the waitlist now. Now they've let everyone off the waitlist, which is awesome and all that, and are kind of grooving and moving, but holy shit. Like, what they invented is pretty baller, but the reality of what that means, we don't know yet. So it's hard for me to say blah. But the second I saw Open Call Hermes within about a week, I was like, oh, crap. This is every. This needs to be everywhere. I even tried a bunch of stuff around it and realized that I don't want to be an agent company launching generic agents.
You know, I have a whole different point of view on. On that.
That hopefully I can share in the near future. But, like, yeah, so when you ask me, I'm like, yeah, there's stuff I'm excited about, but I'm very easily excitable but not hypeable, if that makes sense.
[00:47:27] Speaker A: Yeah, you can.
[00:47:28] Speaker B: I'll cut through the hype. But I get excited about real things and I try to point out what the things are that are real. And usually I'm pretty accurate on that.
That's why I keep talking about these comparisons, because people see something and it's really fast and easy to build something. For example, Muse added phone calling. Instinct added phone calling. Okay. Now the question is, was Instinct working on it before Muse added, I don't know, but I wonder.
So now you're just in this weird race between these companies. And that's why I actually love Grockbot. I don't think anyone's racing against Crockbot today, and I don't think they're racing against anyone. But that DNA at Cursor, it's Cursor's DNA coming in there is inventive. And I'm not saying SpaceX AI doesn't have that. They don't have it for software the way Cursor does. Tesla has it for software, by the way. But SpaceX AI wasn't focused on all that. Now they are, though. And Grokbot is a manifestation of both teams coming together, from what I can tell. Right. And that's pretty awesome. But that's differentiation. They found a differentiated point of view. That's going to work on X, in my opinion.
[00:48:33] Speaker A: Right.
[00:48:33] Speaker B: Because X is their distribution. Back to that whole point. Right, Right. While what Muse is doing with Facebook, that's their distribution. They learned a ton from Threads and they're using. They're just pushing it the right way because they know how to do that from how far they got with Threads. So thus, what is instinct?
And I don't. I don't have the answer to that.
[00:48:52] Speaker A: I don't think they don't have an answer.
[00:48:53] Speaker B: I don't even know if they do. Yeah. And if they do, I'd love to hear it, but I haven't heard it, and I haven't heard anything differentiated. And maybe that's okay. With so much money and finding, like, getting 1% of Facebook, I don't know. Maybe.
I don't even know if they can do that, frankly. But I'm hopeful. I just. Yeah, I just don't see it. I want to see the logic.
[00:49:14] Speaker A: Yeah. Yeah, man. This is awesome. It's been a long time since we caught up, and I'm glad to be able to chat.
A lot for me to think about with my business and hopefully a lot for other folks to think about with their business as we talk about, like, product and go to market and how to think about the future. Reality of our businesses.
Because I think it's. You mentioned, like, what's, what's defensible and what's durable.
I think that's the big question right now. It's not, can we build product? Can we do Go to market? It's like, how do we align the strategy of those things to be able to be successful if we do it right.
[00:49:48] Speaker B: That's right. That's right. And I. I think the biggest tip I would give everyone is they always say you have to pick who you serve. I think you have to pick who you serve.
Like, that's why Go to Market is not the rapper anymore. You have to pick who you serve. If you don't pick who you serve up front, you're.
Now, if you want to go not pick, then be deliberate about not picking, knowing that you're going to have to pick. So then create an experiment so you pick. So one move for instinct, which they're probably not going to do, is pick a market and stick to it. And the only reason they're not going to do it is they can't do that if they're competing with Facebook because they're not going to pick a market.
[00:50:22] Speaker A: They need the tandem.
[00:50:23] Speaker B: They're going to do everything for consumers. Yeah, exactly.
[00:50:25] Speaker A: Yeah, yeah. Cool, man.
We'll drop notes into your products down below and you on Twitter. Anything else you'd like to share. Share with folks?
[00:50:36] Speaker B: No, all good.
[00:50:38] Speaker A: Join the Friday calls. They're awesome.
[00:50:40] Speaker B: Thank you.
[00:50:41] Speaker A: Yeah.
Heaton, thanks so much, man. This is great.
[00:50:45] Speaker B: Yeah, of course. Happy to chat. Always happy to chat.
