Serial Founder: Stop Using AI Like Google! Get It to Do Real Work

0:00 My bots's made 1,800 commissions. Then I asked them, "Hey, this is a little excessive. What's going on?" We were doing this for each one of these things. Maybe you don't want to do that. Max 200. But that bot found the other bot and said, "Hey, go do this thing for me while I go do this other thing." And another bot got the message and I could see it in the chat. It says, "Oh, the product marketer told me to do this, so I'm doing it." What is this starting to sound like? This is your team. And this feels like power. Actual automated work that I now don't have to even think about too much.

0:30 So, we talk about agentic stuff all day, every day. He and I feel like we're in threads, multiple group chats, coffee chats in person about this stuff. So, we figured we'd just get on camera and talk about it, share it with the rest of the world. I feel like the most pressing thing that I wanted to talk about right now, cuz we haven't really caught up on this, is Grockbot. So, right now, everybody is spinning up bots all day.

0:58 You know, you and I have been doing this for months now, but now it's being distributed to everyone. And I think even today, Curser announced that their $20 plan has access to Grubbot. So really, they're pushing really, really hard for everyone to have it. And everyone's sort of creating these personas. Everyone's doing a bunch of stuff. So if someone's getting started, what's the simplest thing you tell them on Grockbot to do? >> Yeah. You know what I really like about it is the fact that it spins up a computer for you and these are all the reasons we were putting openclaw and Hermes and all this stuff on computers so that the AI could use the computer.

1:38 So I think what they've done is really packaged up AI being able to use a computer. So you can go all the way from something really tiny that you couldn't do with chatbt easily or anything. And I'll give one example which is like LinkedIn doesn't let you scrape. There's no good MCP or anything to let you go do anything. There are things you pay for, but they're very expensive. With Grockbot, I can literally have it check out my LinkedIn. I can have it save every single page that it goes through and save it somewhere, right? That it can go access it again. I can even, if I want to get fancier, go put all that on GitHub. Like just yesterday, because of a bot, I made 1,800 commits >> on GitHub because of a bot. because I was just and because one of the things I'm doing with Grockbot that I don't see a lot of people do but I was doing this with open cloud and all that which is have the bots do something and have it all stored in GitHub because then I can have other bots whether in Grockbot or somewhere else just go look at the GitHub repo and know what the heck happened and what's going on right and now because Grockbot is outside of my other agents or even chatgpt and stuff like that the GitHub is the one that can connect all those things together so that's the magic trick that I I figured out with even Hermes and OpenClaw put everything in GitHub. So yesterday my bots made 1,800 commits. Then I asked them, hey, this is a little excessive.

2:59 What's going on? And they're like, well, we were doing this for each one of these things. I'm like, okay, maybe you don't want to do that. Max 200, you know, or something like that. And it was doing, you know, it's figured it out and it's doing that. But there was a moment with Open Claw. And that moment wasn't just like when I came to your conference, the first one in SF and like there were so many people there and Ashton Kutcher was there and this and that and the other.

3:23 It was actually well before that when we were actually implementing it and using that stuff cuz there was so much energy that people had because we were trying to implement this and it was like it was like we had a gift handed to us and we needed to open it up and figure out what to do with it and we didn't want to just play with the cardboard. We actually wanted to play with the toys, right? And so I think like that that that time was a moment and that moment helped me realize when I helped me realize that when I see something like Grockbot and realize how much friction it reduces for people, I'm all over it. I'm like, yeah, I totally get it. Like you're totally right, but this is really freaking good.

4:03 And they're going to keep iterating it. This team seems to know what to do because the first thing they shipped was essentially open claw for anyone and no one had done that before. And you could say, I see this, I see that over here. But like the simplicity of this thing and how they did it, I'm still blown away. Yeah, I think that's totally the people don't realize how just removing the steps and the friction kind of opens it up to way more adoption cuz you know you and I had to sort of deal with all the different times that our agents would just die and have no idea why or what and so now and we were like I remember you were saying like you're now doing DevOps >> and so like this completely removes the need for that for everybody.

4:50 And so giving it a computer, you know, giving it access to the internet and people are like, "Oh, but you can do this with chatbt work or you can do it." You can't, right? You can't see the computer, log in to your LinkedIn account, you know, and do all those things and and right now they're trying to get there, right? It's interesting to see how >> they ship some stuff today. >> Yeah, exactly. today for people to deal with login in an encrypted way and and doing interesting things, you know, but people are pushing to the edge. One of the craziest things that I saw today is is like that Elon is actually feeding people like pushing people to the edge where someone said that they wanted to connect their bank account to Grockbot but were hesitant and he just responded, "Hey, you should totally do it. If you lose your money, we'll reimburse you like your bank account." which is so wild to see of like how far we've come but also the capabilities of these things. I think like one of the things that people are saying is that oh how can you trust it? How do can you know that if you're giving access to these things like how do you handle that? Like how do you approach those types of questions or fears that people have around you know giving it access to your LinkedIn or giving it access to your GitHub? I give it access to LinkedIn and I say don't do any don't post anything or even with email don't do anything unless you've had me approve it. I think we we chase automation with AI a little bit too much instead of just saying hey can we just check it and verify what it's doing and make sure that the access it has is constraint enough. You know, one of my favorite things about actually Claude is they have a lot of granularity on permissioning when it comes to the connectors. They have tons of granularity. You can make some things only read only and not write and stuff like that. I haven't seen Grockbot go that far with the connectors quite yet, but they're probably going to have to.

6:46 There's all those kinds of things that I would suggest. But in some other ways, I almost want to shake people and say, "Hey, if you're worried about this, don't do it. Do something else." Like, >> right? There are plenty of things you can do. like you don't need to give it access to your bank account, you don't even get access to your email and still have a lot of benefit from just trying, right? And like right now, like I said, this is not an ad, but like for 20 bucks, they they let you try it and it's kind of like a no-brainer to try it. And so you go ahead and you connect us to your LinkedIn. And so one of the neat things they also did is for my opinion is they made it really easy to make those connections, right? The MCPs. And they don't even talk about that. They're just like, "Hey, it's just plugins.

7:24 you're connecting to these things and then you can give it access to it. And then so besides for LinkedIn or are there any plugins that you use? >> There's a bunch of dope plugins. My favorite is still the Matt Van Horn who I think you had on not too long ago last 30 days. I have a bot for last 30 days. I even posted about another bot using that bot >> to do some things. So in some ways it's almost like with the way Grockbot is set up today and this is not how COD and CHPT are set up whatsoever. So whoever thinks you could do that over there even with the stuff they've added right now is incorrect. Not yet. I'm sure they're trying to figure out how to do it. But there's also some value in the simplicity of I download this thing, it just does this one thing that's really expansive. So to me it's like if you when we were building these inside of Slack each agent was not like I just type and now all of a sudden I have an agent with Matt Van Horn's thing one of the first things I did is the last 30 days that's the thing I just gave it to GitHub and I said make a bot you bot that I'm making are just going to run these this type of research let me know what you need connected and I'll connect it and give you the API keys or whatever it did that within five minutes I gave it all the things it needed and now it's there and it was really awesome that it's there that when I go there I can say hey go look up all the last 30 days about Grockbot or what have you. Then I started talking to other bots doing other things like product work. Like I have a product marketing bot and that bot needs to do research. But that bot found the other bot and basically said, "Hey, go do this thing for me while I go do this other thing." Basically, right?

9:03 And another bot got the message and I could see it in the chat. It says, "Oh, the product marketer told me to do do this, so I'm doing it." Like, what is this starting to sound like to me? This is your team and you just build out agents really quickly that are all your team. Obviously, you need to know who you want on your team. You can debate how much scope should they have. Just try something. So, if I the way I thought about it is I have product marketer. I don't necessarily want the product marketer going and doing the research. So, I had the last 30 days already in there. Then I built the product marketer. But the product marketer should have context on my products and my categories and how we think about things or the positioning documents or internal documents we have.

9:44 I just fed it all that and now I have a product marketer to work across all of my things in Grockbot that also can call the other agents or other bots or whatever in there and do its job. So now when I want content and things like that, in the past I would either use an agent in Slack, which I still do because it still makes sense for me, or I would try to do it with Chad GBT or something and it works as well. But this this feels like power and it feels like actual automated work that I now don't have to even think about too much as long as I have the agents that are able to do these things. So, I need a new marketing page, a new campaign, or some positioning update, or even want to incorporate what I learned about a competitor or alternative. I can do that very easily because I have a few of these bots set up and they work kind of magically work together. I think one of the biggest things is they magically work together. And I was trying to do that in Slack with channels and things like that and it was definitely more of a pain by by far. So, one of the things that you know around that which is super fascinating is that it's kind of reminds me of sort of skill use, right? So, you have like a bunch of skills and so like knowing which skills to call against, >> but it's kind of a little bit more than skills, >> right? It's it's skills that also can act on its own and sort of instead of like spinning up sub agents on the fly for a skill. It sounds like the way you have this with like a last 30 days which started off as a skill but now you've upgraded it to like a permanent bot because you if you want to you can talk to it directly and just say hey like what's what's happening for this specific keyword or industry on your own or your other bots can run it as a sub agent to like help them in their thing.

11:43 So it kind of changes the shape of I guess getting to what the output that you want. >> Yeah. It's just a different way to get to almost the same goal. And we're all just trying to get to that goal better, faster, cheaper, whatever way you want to think about it. This is the fastest way I can get to those goals. And I tried doing it other ways and it's fast, but it's not this fast and it's not this easy. and the history occurs and if you do what I do and just tell it put everything in a private repo in GitHub, it'll go save everything there. So then that's saving all the context and data it's creating somewhere. You know the funny thing is like until recently I don't think we had a similar platform or or tooling the way that let's say Epic has tooling internally because they have unlimited tokens with Claude or whatever you want to call it. We now have that power and capabilities and actually probably more flexibility and more logical things with the computer that you get as part of your entire sort of Grockbot account that like we can test these things and do these things so easily. So like back in the day I'd have to be like oh you want to use last 30 days oh like connect it here do this thing put it in your cloud co-work and do this and do that. It's easy. It's easy if you are willing to do all that.

13:01 But now with a Grockbot, just put the link and say, "Hey, figure this out. I want you to I want you to be a bot that runs these for me." And that composes it all the way down. So what do I need to know? I just need to know that that thing exists that wants to do that can do the thing I want. I mean, you could even get into, oh, you could just ask the AI and it'll go find it. Yeah, you want to do research using AI. You will probably find last 30 days or maybe there's something else out there I don't know about that'll help you or a bunch of APIs out there and then the agent will just take care of it. Again, I feel silly saying this to you because we were already trying to do all this stuff, but with Grockbot, it's almost like what all us crazy people were trying to do, jump through all the hoops, DevOps, etc., deal with the updates of these tools and how that breaks everything and all that, everyone can do. And and even how fast that happened, it's like what five six months since OpenCL, all that got popular. It's pretty impressive that we now have the tooling. There's been a lot of experiments out there as well. You know, one thing I'll say like because I get hit up a lot with startups and stuff telling me, "Oh, you can do this with my startup tool or something." I'm like, "Yeah, but like this is XAI or SpaceX or whatever they're calling themselves now.

14:10 I think it's SpaceX, right? Like you you're not going to trust them. And look what Elon's saying. Connect your bank account. You lose money, I got you." Like holy Like really? So to me, like I know there's like Victor and Town and and Hyper Agent and all these things. My trust with a bigger company doing this correctly is way higher than any of these startups. Maybe it's me jaded from being in startup land and knowing how things change and maybe they sell and stuff like that. But I feel like I can depend on this tool and the team behind it in a way that I wouldn't want to maybe I'm weird, but I wouldn't want a startup powering all this for me.

14:45 Anthropic, cool. Cool as in on this vector. Open AAI. Okay, awesome. Google, of course. Meta, maybe. But I'm bullish when they do something. but it's probably not going to be for work. The other thing I wanted to say about Grockpot is it straddles a really good line between personal and work. I can use it for both and I don't feel like I'm compromising. A lot of tools make it feel like a compromise. >> Interesting. And what do you think that like what do they do in the product to like straddle that line really? Well, >> I think today they they did simple things that are a little messy but not their fault until they figure out what they want to do about companies and organizations and all that because they haven't figured that out or or you know all that. I'm sure they're going to figure it out is basically just the fact that I can connect multiple accounts from the same service >> into it. And that was like >> I learned this 10 years ago when I was building enterprise search tool. You were probably one of the users that wanted me to do that if I recall correctly. Yeah. Right? It's like 101 product development. Hey, you have connectors, especially if it's a proumer. They're probably going to want you to connect multiple accounts from each connector. Figure it out.

15:54 >> Right. >> Yep. >> That's what I mean by I have love for the team simply because they're doing the right things and listening >> and and building well compared to other things I've seen. Like Claude breaks all the time. Every update or every like day there's something weird that I see that I'm like, "Oh, they vibe coded it right." I can't say the same about chat GBT. They seem pretty solid. Every time I see stuff, I don't feel like it's vibe coded. Grock doesn't feel vibe coded either, but it has the advantage of all the patterns that we saw from all these harnesses and agents that were open source and it's kind of applied a lot of them. So to me, the magic is how simple they made something very complicated and some of us know how complicated it used to be.

16:31 >> Yep. Yeah. So there specifically you can connect let's say your personal Gmail account as well as your work Gmail account and that goes a long way like for you know setting up a Gmail account on open claw is like laughable you know comparatively so they made it really really simple. Have you messed around with channels at all? >> Yeah for sure channels are awesome. Yeah. >> And so that allows you to to add multiple bots into like a it it differs from group chats, right?

17:04 >> I So take my product marketing example. Put the product marketer in there. The agent I'll put the last 30 days agent. I'll put myself in there. Maybe there's a couple other agents that I've made that maybe have like an agent that has context on something else or even a copywriter if I want to go there, right? and shove them all together and literally make them jam on what I want them to jam on together like explicitly compared to implicitly where they might choose which one they want to talk to.

17:32 And so we go in there. I've also done things where I'll ask the product marketer, hey, who should we have in this channel with us? And it'll tell me, oh, this one, that one, the other, because I've made a lot of these agents now or whatever bots. >> And it'll tell me which ones it wants in there. I even have one bot that's a librarian. >> What does that do? It takes books and things, content, books, PDFs, whatever, and turns them into stuff in a GitHub repo.

17:57 >> Oh, interesting. >> Because I found myself feeding it. Yeah, I found myself feeding it these things. I don't want to always keep feeding the AI even. Now I have Chad GBT look at the GitHub with the library because I had those bots or that bot librarian bot handle it. And what I mean is take a book, find all the frameworks, build the markdown. I mean, the N you name it. And I'm trying to get the librarian to essentially extract everything I can from books, PDFs, and even transcripts, anything like that, and turn it into AI compatible stuff in a repo that I can then if I'm using chat GPT for something, just say, "Hey, go look in the repo, find this thing." Done.

18:35 Because they're all connected to GitHub, right? And these are all private repos and stuff like that. These are things companies can do, too. I'm doing it as an individual with my personal account and stuff like that, but you could easily do that as a company and get really organized around it. And I just I I was already trying to do this stuff, but the friction was too high. Like you try to do this with open claw and build a librarian like that, which I've tried multiple times, something just doesn't work right somewhere.

19:00 >> This just straddles that whole situation really well and makes it so that I can build things like that. Yeah, it's I mean it's rather than you having to deal with the DevOps, they're doing it for you. They're making sure that you know if it breaks, it's trying again, it's reloading, etc., etc. I really like the idea of having kind of GitHub as your your sort of backend system and I would imagine they'll eventually build it into the product like Cursor is building their own.

19:30 >> Yeah, they have origin I think it's called, right? I I actually am very bullish on what they can do if they plug it into Grockbot properly and differentiate otherwise there's very little differentiation compared to Git and GitHub. >> Yeah. Yeah. I think it's just a matter of scale and you know there are other companies that are focused on that stuff too like storage and so I think we can talk more about this GR bot stuff because I think it's super fascinating.

19:57 Do you do you subscribe to the chief of staff meme and kind of have one bot that is, you know, overseeing all the other bots or is it on like a task specific? So if you have a product marketer that you talk to directly versus, you know, an engineer type of bot. >> Yeah, with this GitHub thing, the chief of staff is more a storage knowledge thing and context and less so a bot itself. So I I don't prescribe. Short answer, I don't prescribe to it. Longer answer, I'm not saying it's the wrong paradigm. I just use the paradigm differently where it's like, well, I also never I don't really believe in bottlenecks. So to me, the chief of staff concept is introducing a bottleneck. I'm not sure if I like that bottleneck. Now, that being said, do I have bots doing things for other bots and those being the ones I trust with finding out stuff or doing stuff? Yeah, absolutely. But that's more bot management type stuff and I have some bots that help with that. But I think the overall concept is for a chief of staff I'm trying to cover 50 to 80% of it by having the context and the data sitting there accessible to any of them.

21:08 But even today like when it did the 1,800 commits the first thing I said was like hey it looks like you made the repo really messy. Like one why 1800 commits two looks like you made the repo messy. Can you go clean it up? Do I want a chief of staff to go do that for some other bot? No. That being said, if I am the type of person that wants an admin or chief of staff that's my assistant personally, I think it makes perfect sense to use that concept and go after it. But personally, I'm maybe a little more fluid and I probably hop between tools a lot more than most people.

21:43 >> And so what other tools are you using besides Grockbot? So I my current order where I probably use them equally as much right now is ChatGBT and Grockbot. And the reason for that is I actually think ChatGBT is a very underrated product today. >> I refer to the desktop app or the web app. What are you referring to? >> I just mean the product Chat GBT. >> Okay. And the reason I say that is like the way they've done memories and projects and all this stuff like it's gotten so much better over time beyond just the models getting better. And when I use Claude, it feels kind of clunky and it doesn't remember things or remember me and my stuff in the same way. I think they've improved some things with co-work having access. I don't remember what they did, but they said they did something recently. But that's been my biggest chatbt. I still use it a lot and the the specifically five I know people like hearing about this stuff but like specifically five six salt high is my just go to sometimes medium but usually on high if my usage is fine and and it it it has been a flattening out for me of like oh I don't really reach for other models much when I have >> not even fable >> you know the short answer is I save Fable for the engineers >> okay >> they really need it like and I'm not trying to engineer those things in the way an engineer is like a real engineer like someone who not like me and you try to engineer but a real engineer needs Fable I'm there so we have accounts for them just for that purpose >> but for the my use cases I found fabled to be heavy-handed and not necessary in most cases even like ultra or extra thinking or whatever on solid gets me pretty darn far >> but I also am a little allergic to interface for CL. I just think it's terrible.

23:38 >> They've done a bad job building that product in my opinion. >> Terrible job. >> Yeah. >> And I'm comparing it to Chat GBT. Chat GBT is not a terrible job, >> you know? And what I mean by that is just vibe coded, broken, breaks, and I am I will be the first to admit I have a lot of baggage on that because I was around when it kept breaking every day. >> Mhm. >> With every released. But I still feel that.

23:58 >> I also would say I I'm in agreement with you. I think that both companies are vibe coding their stuff. The difference is is that nobody at Enthro is using the app anymore. Like they're all using Claude Tag. If you talk to anyone, they're like, "Oh yeah, yeah. I just have like my own channels in Slack and I talk to the Claude Tag all day." And so when I heard that, I was like, "Oh yeah, no wonder the product's because nobody uses it." Well, and their excuse before that, not theirs, but ours was, "Oh, they just care about code and the API."

24:36 >> Yeah. Yeah. >> And cloud code, right? Or whatever. >> Cool. Got it. Terminal, etc. Cool. So, that was their excuse then. This is their excuse now. I'm not saying they're giving it. We are for them cuz we love them obviously. But that product is when you try to use it straight up. And there is no excuse, vibe coded or not. That means some to your point, they're not putting the judgment on what they're shipping. Do you will trust a company like that? That's all I I would ask like does that embody trust? And the answer is no. Forget even what Dario says and what opinions people are. The product needs to make me trust your business and your company. That product does not.

25:12 Other products of theirs maybe do. That one does not. While with open AI and chatbt honestly I trust them because the product works and they haven't lost my trust. Claude loses my trust like all the time with the product. >> Like >> yeah like don't automatically like shut you down. say, "Hey, you can't use table because you're trying to do a security audit on your own code." >> And then it's like, "Nope, sorry. You got to use our dumber model to do it."

25:38 >> And so, they don't I'm writing an essay on this about companies and how they build products because they have inspired me. >> And the reason they've inspired me Yeah. Yeah. I'm like, "Wait, like >> y'all should be building the most vast products for us on the planet. if you have that many tokens and you're number one on revenue as a lap, why are they not doing that? And I think your points are all correct if like they don't dog food it or whatever. But at the same time, a lot of products I built, I didn't dog food them every day cuz I built for someone else. So I don't think that's necessary. I think you just have to know that like you don't ship You actually have judgment. You care, you know? And maybe they say, "Oh, things are changing so fast. It's okay."

26:19 But the trust is a problem. If I can't trust the product to do what I want, the way I want it, when I want it, is it really a product for me? And the answer is no. And and that's what Grockbot so far has totally crushed and destroyed because the flexibility is there and the simplicity is there. Thus, the friction so low. I feel really good giving it to people. And I'm coming from a place of having tried giving all these things to my friends. I know you do this too and people I know, right? Grubbot stickiest by like a million miles, not even a mile, right? Like it's just stickiest. I I send something to it, they're like, "What do I do?" I'm like, "Figure it out. You have questions, ask me." They don't ask me questions. Few minutes later, couple hours later, next day, they're like, "Hey, it did this. You were right." I'm like, "Oh, cool. What else are you going to do?" And then they have ideas, but they they never sat there and said it broke or it did this or it didn't do that. They were just like, "Yep, I tried to get it to do X and it did X for me as well as anything could and I don't know how I could have got X done without it." Obviously, these are people that might not be aware that you could do X with Claude or >> you know, Chat JBT, but like that's the whole point of the world we're in right now. So many people don't have that knowledge and shouldn't necessarily.

27:33 They just want things to work. That's my mini rant. I'll get off my high. >> Yeah. No, I think it's right frame because you know the people who say oh you could just do this that and the other thing like you can just contort this way and all of a sudden you can get chatbt to do the same thing Grockpot does you can do you know this other thing and he's like >> yeah yeah but no one's going to do that >> right and like the simplicity of >> putting cute eyes on it or a cute animation to just you know indicate that this is you know working for you is all that it takes to get to remove the fear and the friction of just getting started.

28:10 >> When you use a product, you want to feel like the people that built it care about you, the user. >> There's no situation where you don't want to feel like that. None. So, if that's the case, then Grockpots's really nailed that feeling. And you're right, it's those little things, too. It's like you can hover over that little colored thing anytime and it will tell you what it's doing. sounds really small and we have little pieces of it in other places like Enthropic has some stuff Chad GBT does but the way they do it because there's no other crap involved is so simple right they don't have the tabs or any of that I don't think they're going to add all those things because they seem to understand elegance of the simplicity that they've created with the interface so I think that's really it and like honestly it's like the kind of interfaces I was mocking up for this stuff when I was like thinking about what do we do with this how do we how do we get out of Slack >> my biggest thing is how do we get out of Slack now that we're in Slack because Slack is not doing us any favors. We can't we don't have access to all our data when we use Slack, >> right? And so they they gate it or they put it behind like a big pay wall in order to to get it totally. I think the interesting thing that people have tried like Buzz, you know, my co-founder like likes to say that Buzz is is like too, you know, cippher punk for its own good, but they have sort of similar ideas around the simplicity and playfulness around it, but I still think like it's not the right shape.

29:33 >> Buzz is really clunky. Like it's just really, really, really, really, really clunky. And you can tell that it's very optimized for engineering. >> Yeah. And I'm not sure if that's doing them any favors. And also I ask people who tried Buzz and they're like, "Yeah, it tried to install it. I couldn't install it." >> Like that's just a basic I don't know what the issue is, but that's a basic problem. But love that they're doing it.

29:53 It's good that we have options, but I don't think Buzz is the option that I would go in any reality, frankly. >> So you're basically in this stuff all day, every day. Are there things that you don't delegate to AI? my opinion on an interface we're about to ship or like my my opinion on a on a on a website. Like I'm now flipped the other way where like I've never been into the flashy websites in the first place. Like a lot of people like the like the fact that you can do all this crazy stuff with CSS, HTML, this that and the other. I'm not I've never been into that cuz that gives me like Flash vibes, you know, back in the day when sites were all flash and I'm not into that. Like just my own thing, all that. Super dope. Not my thing. I I like seeing it. But like I want to build sites that don't look bipod. So when the AI does it, great. It made a wireframe. Let's, you know, judge it. So I I to be frank though with the answer to your question is judgment. If I'm going to publish something, post it, ship it, let it be in front of other people that don't care how it was built or created, then I need to really be the last smile. that I know a lot of people have different takes and try stuff, but like show me their stuff. I'll probably have a negative opinion about some part of it because they didn't pay attention.

31:11 Some of this is just AI slop words and sentences and all those patterns. Other ones are eyelashes on designs or a specific aesthetic palette. It's not even a color palette. It's like a aesthetic palette that includes colors and the way it matches the colors. You could tell that ChachiBT or Claude designed it. And like >> I don't want to see websites like that. I don't want website and again I know I will. I don't want websites to be so muted that they all have the same patterns and we all get used to them.

31:39 >> But isn't this something that you had I remember you built when you were at Dropbox you built like something to kind of like help people like provide like their taste or how they want things to look for. >> Yeah. So, so when I was there, one of my favorite things that we built was an ability to take what you already have and turn it into an AI powered template and recreate it. So, for example, if you had five roadmap docs that were great, maybe they pass reviews or whatever it is your job is and they're great, we would take those five and our AI, and this was about a year ago. Yeah. Our AI would create an AI powered template.

32:21 Then you could take a transcript from a call, shove it in there, and it'll make the exact type of road map you want document. Or you could take, you know, notes or anything. Or you could even just talk into it and you'd be able to get it to build the scaffolding and probably get really far and draft you something. In fact, I pushed really hard to call them drafts because at that time, I think that was a better way to think about it. So, but the big thing out of that that's the lesson or a thing that I keep reusing and taking back is you have to show AI what good looks like. And so, anything I try to do, I'm always trying to find a good example to give it and then have it deconstruct it and almost create these AI powered template type things and then I'll go have it do the thing I want. So, for example, if I want to build a new homepage and I like a bunch of role models that are out there as well as some alternatives or competitors to what I'm doing, I would give AI all of them and give them the URLs and all that. And I would actually say, go get all the info. Oh, they're using copy. Great.

33:22 take that copy and write me a positioning doc for each and every one of those websites and businesses from the websites from the source material of the actual websites which I like the words they use and all that then it would create that then if I want to create my positioning doc now I have these five positioning doc from other derived from other websites because I like their words and what they're saying and then I can take those and say okay I got these five great now make me money and so I try to just systematically do that workflow in almost everything I do there's not anything I can think of right now where that workflow isn't a part of the sort of or those steps aren't a part of what I'm trying to what what the workflow is to get like good outputs it's garbage in garbage out >> right >> that's all that's all I'm going on >> and so do like do you have a do you put all that stuff in GitHub like how do you keep track of what goods look like for yourself like a lot of people you know back in the day would have like second brain systems >> swipe files is like a marketing one yeah all GitHub GitHub all day or it's sitting in folders in my in my local computer but I can sync those with GitHub and all that. So GitHub all the way. I don't see another like people have pitched memory and concepts storage stuff. I've tried to implement stuff even QMD and all these things and embeddings. I like that. I do use that quite a bit, but at the end of the day, if it's sitting in GitHub, I know any bot, any agent can go get to it or I can download the zip and just upload it somewhere where I need to.

34:53 >> Yeah. Yeah. And regarding that, like you had mentioned that you save stuff locally. I know that you're a big proponent and believer of local AI. So, we talked about kind of grabbot and that's, you know, cloud AI. How are you navigating or what are you delegating towards your local AI inference? Yeah. So I think local AI for me started as like a a a really big curiosity and then it's ended up as it has ended up as like basically maybe an obsession of some kind because when we had the open source agents and those were the only game in town that and and Claude cuz they were the ones that essentially Peter derived openclaw from we only could like we could put them in the cloud or we can put them on our computers and he's got a bunch of other stuff that's really cool that was essentially pieces of open claw open clause harness essentially and when I saw all that stuff I'm like oh I have a lot of use cases where I need the computer I I need the AI to browse browse websites or even use a computer I'm building desktop apps so a lot of the stuff I'm doing with local models I think one of the biggest use cases that's always running is essentially QA so I have a box that sits there and is trying to run QA on our desktop apps because our one of our desktop apps works across other apps like it's meant to integrate with every app you use that requires a lot of testing these days when someone does a support ticket like the other day not working in Microsoft teams properly and it's just an autocomplete on a form field right that's the tech >> this is for the product type ahead >> yeah it's for type ahead and we were able to spin up micro we were able to have the bot just by saying hey this don't work in Microsoft teams and with everything else we set up. Again, this is not like out of the box, but it's all set up. You just hit the bot with that in Slack and say, "Hey, this ain't working like go fix it." And then it would go figure it out, test it, make screenshots, videos, figure out what's wrong, debug it, and go all the way to a PR and saying, "Hey, it's fixed. Here's a PR someone, some human should probably review." I can't do that with Grockbot either yet. And I don't think anyone should try to do that with a lot of these things because there's a lot of scaffolding and stuff to do. But the machines are very capable when you do that.

37:12 You do that right with so just I want to just explain for people who are not familiar with type ahead. Type ahead is kind of like an amazing product where if you're familiar with Gmail like autosuggesting you know what to type next. This kind of runs locally on your machine and it works in any app and like I've tweeted out this before but you know I'm a big voice to text person but when I'm not using voiceto text I'm type ahead all day and it's been like really amazing product so thanks for building it >> for sure and you were saying like you can do this in Grockbot but it's a Linux box it's not a Mac that they've spun up for you right to your point so if you want to do anything in Linux or the web yeah Grockbot will let you do it but the power of a Mac and then you throw in the local model. So for example, when we're summarizing anything on a website or any of that, we're going to use a local model that's running on the machine instead of going to the cloud because the tooling is already set up to do that. But also, we will take things we've done with the cloud and outputs we have there and then essentially use evals and whatever fancy fine-tuning, whatever words you want to use, whatever you want to call it, to get the local models to get as close as possible to what we get from the cloud. So then we're just bringing work essentially workloads off of the cloud into our local sort of environment and testing all the local models. So this is why it's a little bit of obsession and all this stuff can get super insane and useless. But any workload that's happening with the frontier models, we're trying to figure out how can we make them happen without those frontier models on our local machines.

38:46 >> Yeah. Which is always fun. like I'll experiment with local machines running different harnesses or te testing stuff but I haven't gotten to the same extent that you've had with the amount of machines and testing because I guess you're testing more products and so those are kind of we talked about you know how to use AI like what both right now simplest removing friction with grogbot and on local side but you know as a product builder and CEO of crazy egg How are you thinking it from a a product building perspective or a user perspective? Like are people now thinking from like an agent first perspective? Like how are you designing products differently now that agents are starting to become more distributed?

39:35 They're not necessarily here for everybody. You know, there's still there's still years ahead of that. But how are you thinking about building a product, you know, now in 2026? The moment right now in my mind is basically this statement, which is if we were to build it today, what would we build? I don't think I ask myself that as much as I'm asking myself that these days. So, probably a couple times a month, I'll just go, "Oh, what would we do?" And if it's any different than what we're doing, I really need to think really hard about what we're doing. And so, >> found meaning if you were to start Crazy Egg today, >> what would you do? Got it.

40:16 >> Yeah. And that's been the most clarifying answer. Even outside of Typad, we have a new product and it's like a meeting recorder. I know you've used it when it was really janky when local models weren't there. They're there now. >> But >> the way open source works and the way all that works is it's usually first or early and then a lot of promotion on something. So, we were going to open source it a year ago. But a year ago, it wasn't good enough. But then 6 months after that, so about 6 months ago, people started open sourcing their versions of these media records, but they were Forget even the UX. Like the UX is terrible in most of them, which is a pet peeve. But okay, you can get away with that if the tech's good.

40:52 The tech wasn't good. And then recently with some models and some UX stuff we've done and tweaking we've done, we made it so you can do a lot of the stuff that locally that Granola does native. And but now competing with Granola isn't what our objective is. Well, a year ago it would have been, right? And so constantly just asking us asking that as things evolve. And you could say, "Oh, why'd you spend a year blah blah blah?

41:14 Well, I didn't have a choice cuz the local models weren't there and that was the whole stick. But now, even though it has local models and does all that for you, we hook into the codeex harness so you could use codecs with it, no problem. Or we hook into bring your own key, so you can get all that because it's a meeting recorder and all that also a note-taking thing and you can chat with all your all the knowledge that's in there. So, you know, it's like I think one one of the problems today is where does your product start and where does it end? which means somewhere another product starts and somewhere another product ended right I'm not sure if that paradigm which was a very classic paradigm of software is as relevant but then I don't know because if you don't do that then everything can do everything and if everything can do everything do we need anything again not getting philosophical but like this is what I struggle with dayto-day in trying to figure out how to build this stuff which is like one things are changing so fast if I were to start today what would it look like and two again this is separate But like building so easy, we're still we're we're back to like, you know, again what everyone's starting to say a lot more, which is what about distribution or how you go to market and all that. So if I were to really take a harsher look at it, it isn't what would we build today? It's how would we go to market today and build the thing that works with the go to market. And I think if you don't do that, it ain't going to work. Another reason I'm very bullish on Grockbot, they have an entire top social media platform as marketing and even if they don't do anything extra with that, which you know they probably won't, but even if they don't, which is where we're kind of at right now, it's Grock. That's X all day.

42:57 >> So I think that advantage would make me if I were trying to compete with Grockbot kind of annoyed and being like, "Oh crap." and I'd be studying the crap out of their numbers to see if it's worth competing because it might not be if you can't if your goal is to win and be number one. If your goal isn't to be number one, that's a very different story, you know, or maybe you find some differentiation like do it just for hospitals. I don't know, random idea.

43:23 >> No, I I think that that doing it for hospitals is the answer is how you compete, right? Like spe specializing, you know, you had shared a lot right now and I was just trying to process it all. I think that when you could build everything where anyone can build everything like where do you focus you know and so the meditation around hey what would we build today and then you know it sounds like the adage of like you know serious serial entrepreneurs and seasoned entrepreneurs focus on distribution first and figure out yeah how do you it's not the product it's the distribution that sort of gets you to get a chance to keep continue building that product.

44:03 >> Yep. >> And so the the question around Yeah. How would you compete is I think by specialization, right? And so just because everyone's going to have every single feature on your phone, right? Like in the same way, you know, there was my my friend like Andrew Parker, this is maybe 20 years ago now, took a screenshot of Craigslist and and was like, "Oh, there there's a startup for every single Craigslist section that became like a billion-doll company in of itself, right?" So like housing became like Airbnb and Zillow and all this other stuff, right? And I think that and because Craigslist to some extent was like the everything website for the dot era and I think we're going to have the same thing where like Chad GBT Claude Grock is going to try to be the everything app for everybody and they're going to try and that'll be the era of this time and whether that happens you know over 2 years or or 10 years we'll time will tell but then I think you'll have the specialization of the people that are focused on the flows for a hospital or a specific type of vertical that they just can't compete right and so maybe the market won't be as large because of distribution or maybe they will I don't know but I think that there will be a chance I still believe that specialization around this stuff like I'm a believer that we're going to have you know thousands upon thousands of agents that you can interact with, you know, disposable agents, like just a agents on agents, right? And so, we just don't have the right surface areas for them yet. And people are starting to build them out. And so, yeah, it's a definitely definitely an interesting question in 2026 of like, yeah, what do you build? How do you build it? How do you focus on distributing it in order to be be number one? sometimes like the distribution like my favorite distribution strategy right now has been the instinct launch, right? Which is cool. Like someone, you know, got an invite to instinct, played with it, got I was like, "Yeah, this is great.

46:15 Someone built out a seamless way to talk to have an agent do stuff for you. Great." So I who spent, you know, 400 hours building out my agent, no one's going to do that. They did it for everyday users, >> but it's Poke. It's Poke. I don't think it's extra special. And I I think there's a whole thing to be said about that. Yeah. Yeah. >> I was bring I wasn't mentioning it for its product chops. Sorry, I was for >> their distribution were was BC's and on X and that allowed them to and this is a proven distribution hype for fundraising, right? And I think they're like in the last three weeks they've had three different funding routes it sounds like like so >> and that's not true obviously it's just the optics and what's being pushed out right I mean the other thing is it was really funny that nobody who was talking about it that was a VC talked about who the founders were there was a game being played on that one just saying >> yeah yeah >> and it was very transparent and I'm like well great like I don't like products that we got gamed regardless of how good they are I hope they do well whatever startups love startups, but no, that ain't it. That's not like I get it.

47:29 >> That that's not Yeah. Yeah. Yeah. Yeah. But that ain't not even that. It's like it's distasteful. I'll just say it. It's completely distasteful in my opinion on how they did it. >> And I think the VCs also showed lack of taste. >> Yeah. >> Yeah. Yeah. Yeah. It it was it was a game. But to be honest, I think that going back to removing friction, the fact that you could just text a number, you didn't have to, you know, do anything. People didn't play with Poke because Poke actually had friction. You know, when you started, when it first started, it came out. You had to like haggle with the bot to to get >> But that was cute at the time. And that got them the what they needed or wanted, >> which was back to your point, distribution. I mean, really, you're talking about hacking X.

48:17 >> That's all this is. It's all just hacking X. And that's why I think it was distasteful. It's like >> you wouldn't tell us who that you backed it, and you wouldn't tell us who you're backing, yet you're talking about it as a VC, >> right? >> There's there is some fiduciary going on there, >> no matter how you look at it, right? And I'm not saying it's not a gray area and all those people are lovely, but it was distasteful to someone like me that can see through it and be like, "Oh, this is not how I want startups to blow up." You know, I want them to blow up because people use them and want to share them genuinely, right? And also, there's other ones that have done that. Town did the same thing. And like, I get it. You should do that and leverage your investors. No issue there, but it's just showing lack of taste across the board.

49:00 >> So, here's the pop quiz. What startup has done a tasteful job of whether launch or distribution? >> I think Poke was super tasteful because they made the product the distribution by making it so interesting on that you had to haggle with it like like look distribution hacking is psychology hacking. >> Yes. >> Like and and so that's what I'm saying is distasteful. It's distasteful when you do it and it's super transparent and you can people can look back and be like, "Oh," and you can't do it for if you're a VC, not to harp on VCs, they're all great, but if you're a VC, you can't do that over well, you know what I mean?

49:39 You can't do that over and over and over again. And that's what I mean by the tastefulness. It's like there's a lot of nuance to this very specifically, but I think about this a lot because you're just hacking psychology. If you do that the second time, the same VCs for something else, you're just eroding the trust that people have with you. A lot of distribution is also trust, >> right? >> Yeah. >> That's why the hacking part is funny.

50:02 You're actually trying to hack >> trust. >> It's like trust me. So, what was hacked with poke? Not trust. It was simply human psychology of this weird experience that they probably tried and were like, "Holy crap, this is a fun experience. Really weird. people might talk about it, right? Again, there's no guarantee there. The difference is if your VCs start talking about your company because they gave you money in your seed round and stuff like that and it's on X and people trust these VCs, great, right? You're going to get distribution. How many times can startups do that? The same startup, I don't know. Different startups, again, if it's tasteful, it all will be fine.

50:39 But if it's not, you're just going to erode. Everyone's brands are just going to get eroded over time and we're done, right? Everyone's cooked. That's the problem I see. So, I think if you're thinking about distribution, I would be thinking about channels that don't require all this other funny business. Yep. >> And ones that actually have a you pay, you get a user, they sign up, they pay you, you know, your CAC, you know, but I'm I'm an old school marketer like that. It's like we should know our funnels and things like that. And yes, brand and all this stuff is part of it, but high level it's like I'd be trying to figure out how to get distribution.

51:14 In some ways, I'm almost of a different mindset than I have been historically, which is like I'd rather just figure out what is worth distributing through the distribution channel more so than I have this thing I want to go distribute through that distribution. >> Yes. I think that's kind of the the frame is start with distribution first and figure it out that way, >> right? >> Yeah. That means you kind of need to reserve this idea of what your product is like and put that in the background because you have to be open to what you learn.

51:44 >> I feel like we can keep on chatting because I feel like there's something else I think. What what do you think right now, you know, leave it on on a note around just overall the day-to-day stuff. What do you find yourself wishing more people are doing that you don't see them doing with the tools that we do have access to these days? >> You know, I think we have the Google problem. >> Okay. Same. And the Google problem to me is like you and I and a bunch of people let's say at this pace is probably 50,000 100,000 people are using this stuff very aggressively trying to tinker with it have the wherewithal to deal with all the friction and are tech curious not even early adopters like before early adopters like tech curious like like if it's new we're like going towards it and running at it right and trying to figure it out together like I could see crowds of us just running at things right That's not everyone. So, what I wish everyone was doing is just learn one thing you could do that's extra that you're not doing with AI today. I mean, literally like the thing I was saying earlier, which is like feed it context, show it what good looks like and then use it. Because that's what I mean by the Google problem. The Google problem is people got so lazy and aren't generally lazy about this, they just type in really short stuff. They don't like re keyword it, you know, they don't redo it. They don't try again. So, a lot of people are still there with AI where they're not they might in replies go back and forth, but they're not really trying to use AI in a more advanced way and it's a hell of a lot more powerful than Google.

53:18 >> Yeah, I think that that's the interesting thing is like just trying and I think those people are like afraid that it's not going to kind of work. So they're not really sure, you know, when like to do it or they're going to break something. But you're saying like, "Hey, there's no extra charge basically. Just just try it. Try a little bit more every day." >> Like did it did it cost anyone anything to do site some site and search the site using Google or file type colon PDF?

53:52 >> Right. Yeah. Like advanced searches like nobody did. Right. >> Like Right. And and we did we enjoyed it but we were tech curious. Right. So, this is why Grockbot is so awesome. I could throw anyone at it and they're not thinking about Open Claw or Hermes or, you know, tokens or anything. They're not even thinking about tokens with the thing. They don't even let you choose the damn model, which is all where we're going anyway.

54:14 >> We were going there. And they happen to be now part of a lab because this is the cursor team, right? part of a lab and are thinking from first principles, so to speak, about where we're at right now and what is a product that being created, not the parody product that matches Anthropics or OpenAI's products. One of the reasons Grockbot works is because it doesn't hit those paradigms at all. Actually, BAT, sure, but that's everywhere. It just thought of it from a different perspective or they thought of it from a different perspective or at least I feel that when I use it, which is this is a place where you make these bots and talk to me. That's not chat GPT and that's definitely not cloud either, right? Like that's I think the big unlock and and I see every lab wanting their piece of this pie. The problem is they had separate apps, they combined apps. Cloud is still a mess as we talked about and so what are they going to do?

55:06 So when I think about it from chess, you know, and now this is Elon's chess game. He's the new player that has legitimately entered in now. He's playing chess. It's a chess game, right? And one of the moves is do something they they are not going to do anytime soon or cannot do. >> Yeah. >> Can they build yet another separate app? That's what I'm trying to get at. And the answer is no because Claude would have to pivot pretty hard. Maybe they'll do it to let me make bots in there >> today. As in the model and all that definitely a version right now on each team's you know PM has like a codeex version of bots.

55:43 >> But check it out. Check it out real quick. Then we're back to the Google problem because they are just adding more crap into their already bloated products. Then what? Like are they going to go highlight bots as the thing thing? I don't know, right? Like that's what I mean by they'd have to hard pivot to touch Grockbot today. And yes, they should do it. They will do it. I mean, you know, this is how PMing works. Got it. But we're back to the same problem because these products create sprawl and and all this baggage with all these toolings and things. And you you we both know people don't want to watch videos to learn how to use interfaces.

56:19 >> Yep. Exactly. Okay. I feel like I mean I can't keep going, dude. But let's let's call it let's let's wrap it up. Appreciate you. Thanks for taking the time to chat. I know you're hosting some webinars and teaching folks. Every week I try try to host a live AI session and teach people stuff. Try not to get as ranty as I do with >> that. >> Yeah, it's on heaton.com. So my firstname.com. >> Great. >> That's where people can find me.

56:45 >> But I think they they got a good preview. >> People got a good preview of like what we do when we hang out. So just saying that's what this felt like. This is great. Totally. Yeah. Okay. Catch you later. Cool.