Manus AI replaces your AI tech stack? (Full Demo)

Everyone's talking about Manis AI, the Chinese AI app that basically can take your thoughts, turn your ideas into fully automated businesses and products. Now I wanted to try this, but I didn't have access, so I called my friend Min Choy, who came on the podcast, and we had a crazy idea: can we build a business? Actually, the idea was a docy killer multi-billion Dollar business in some, a few prompts. In this episode, we go through it and we show you the ups, we show you the downs, we show you what's good about Manis, we show you what's bad about Manis. But if you're here, if you care about AI, if you care about building businesses, understanding how Manis works is a must.

I was so impressed with some of the stuff it was doing, and I was so not impressed with some of the other stuff it was doing. This video shows you all about that: the good, the bad, the ugly. I recommend you watch the entire episode, and you make the judgment call for if you can go and create a business fully automated with a few prompts on Manis AI that generates $1,000 per day. Enjoy the [Music] episode time, baby.

All right, I brought Min Choy today. Min has been going viral recently because he's been playing around with Manis AI, and he's going to give us a first look into how to use Manis to actually build a business that works for you. (Role-playing as Greg:) "Men, thanks for taking the time on coming on the show. What is the viewer—the listener—what are they going to get out of this episode by listening all the way through?"
(Min replies:) "Thanks for having me, Greg. This is a very exciting podcast to join—it's my first time, so thank you. So, this Manis has been going viral and very crazy in China, and it just, like, came over just a few days ago and I got an invite so that I can go play with it. I realized that this thing is just more than the typical AI that we've been seeing, and the thing that is different with this is that we've had in the past something where we wanted to make this agent AI where the AI can actually go and run tasks and run tools, as you've seen on cloud computer use, browser use, and similar to that. This is sort of like a more advanced, the next stage of that. I think that we're getting closer and closer to sort of the glimpse into what the AGI is going to look like."

What I want to show you today is that I want to take some of your startup ideas, correct, and I want to see how far Manis AI can take it. We're going to try to do this step by step and in a very process-oriented way, and we want to see if this thing can really help entrepreneurs or startup owners to really kickstart their ideas into action. I think that this will show what this type of technology can do for us in terms of growing a business.

Let's see—I mean, I'm curious. This is my first time getting a glimpse of Manis. I hear big things, so yeah, let's just get right into it. Yeah, let's start with some of your great ideas. I think that one of the things that I read recently from your post was something about DocuSign. I think DocuSign might be a little bit complex, but let's give it a shot and see what Manis can do for us. While you pull that up, my thought on Docy Sign, Men, was basically that what's happening is, you know, DocuSign makes, I think, like billions of dollars or something, thousands of employees for an electronic signature; and with AI, it's going to just be a lot easier to create software. We'll see by the end of this episode if we can compete with, you know, the DocuSigns of the world, and instead of charging $10 a seat or whatever they charge, you'll be able to make it free or charge just a very small amount. So I think that there's an opportunity to build— I mean, I think multiple people could do this, can make about, you know, at least $1,000 a day in revenue, and so I'm excited to see if this could work.

All right, so what do we— it says, "Hello, Men, what can I do for you?"
"Yeah, so this is a very familiar interface that we see all the time, you know, ChatGPT kind of had this type of interface, and everyone else kind of started copying that interface design. So, Manis also has a similar—you got a little chat window where you speak with the agent; so you just kind of simply say what you want. I like to start with a very simple request. Like, you know, let’s say, create a DocuSign clone."
(Interruption in role-playing:) "Quick break in the pod to tell you a little bit about Startup Empire. Startup Empire is my private membership where it's a bunch of people, like me, like you, who want to build out their startup ideas. Now they're looking for content to help accelerate that; they're looking for potential co-founders; they're looking for tutorials from people like me to come in and tell them 'How do you do email marketing? How do you build an audience? How do you go viral on Twitter?' All these different things—that's exactly what Startup Empire is and it's for people who want to start a startup but are looking for ideas, or it's for people who have a startup but just they're not seeing the traction that they need. So you can check out the link to startupempire.co in the description."

Simple, like a super simple request.
"Y, very simple." I start by saying there, "The cool thing about this is that Manis is sort of a multi-agent under the hood, and it's got a planner, executor—that type of thing. And if you remember, like, very early in the days when we used to have this thing called Baby AGI by, Yoi—I don't know if you remember—that kind of used to have that sort of, that setup where it had like task creator and like an executor, etc. So this thing does a similar thing, and then what it's going to do is, GNA, it's going to take your request and then it's going to try to research into what that is and what it needs to do to create it. Even giving a simple request is doing many things like on its own, like it's planning, it's designing, it's executing, and it's going to try to run things, and then analyze the results and then try to kind of replan or update as needed if it's going the right way."

You will see a lot of things; there's a lot going on. It's doing a lot of planning that I'm researching right now, so it's doing it, you know, sort of the Googling underneath. It also checks out their website—I think it takes AI; it also takes screenshots, and it tries to do stuff on it. I did run into one interesting thing before, which was that it got stuck on like captchas and stuff, and I thought that was kind of funny because it wanted me to kind of intervene and do something with it. So it has some sort of the "human in the loop" feature in there as well. It will ask you for confirmation or, you know, additional context as needed.

So, this is so interesting. If I'm understanding this correctly, you know, before Manis you kind of had to go to a bunch of different places—I call it the AI cocktail approach. You would go to maybe a ChatGPT or a Grock 3 to kind of come up with the idea and create a PRD, and then you'd go to, you know, V0 to actually design what it would look like, and then you'd go to, you know, a bolt, a lovable, a cursor, depending on where you are, and, you know, in terms of technical prowess, actually go and build the idea and deploy it. What you're saying here is that what's happening on the screen is it's kind of doing all those things at the same time.
"Exactly, yeah. It's actually what it's doing is it creates a sandbox; like a virtual sandbox of its own, and it runs, like, you know, command line executions. It runs a browser, runs tools—which is on the right—it’s right now browsing and just capturing stuff, capturing information. I think what it's doing is it's crawling; it's scraping all the data and information it can get. Since I said DocuSign, it goes to DocuSign and checks out what, you know, type of information and the features the product has, and then it will probably test it out and see what features it can capture as part of its planning. And remember, we just started with this very simple request: just say, 'Go create me a DocuSign clone' and it's now just going there and doing its own thing. It's just pretty wild to see where we are with this agentic AI that is able to kind of do most of these tasks on its own, understands your request, and if it doesn't have all the information, it goes and researches on its own. I think that's like the key to a lot of this."

And like you said, we've had many AI tools that we had to use separately, and we've done a lot of copy-pasting in between. Now a lot of that has improved, a lot with Undercursor, Lovable, Repet, and all that, but this is a little bit different because it does more than just coding—it actually does the research. It actually goes to the browser and actually captures the information from there directly, and then it takes all that context into its planning and then it formulates a plan of execution and then it goes and executes them. So this is a very interesting way.
Now, you see on the right side of the markup the way it has that markup—it this reminds me a lot of the Baby AGI days. Baby AGI used to kind of do this in the command line where it shows its planning phase and all that stuff. I thought this was interesting so it's doing, like, step by step on what it needs to do to create this DocuSign clone. And that's all I said: just go get create me a DocuSign clone, and it's getting all that information and it's planning, replanning, and executing. It's very interesting to see, and it's going to take a couple minutes to get to the end. Also, sometimes, as tasks are needed, it may have missed something in the process of doing this.

Have you ever had an instance where you're like, "Create this app" and you're looking at maybe the markdown, like, for example, if you go back to the markdown—can you go back to the markdown while it's, or it's...
"Yeah, I mean, you could like look at the twoos—oh wow, that markdown was just like the raw source code, but this is the preview. So cool."
So, I'm actually going to read a little bit of this. We do have a lot of listeners on Spotify and Apple for the Startup Ads podcast, so it says here: "Research DocuSign core features; identify key functionality to implement; docu—document user flows and technical requirements."
Now, there could be a case where you're like, "Whoa, whoa, you know, I don't want you to do XYZ. I don't want you to identify the key functionalities; I just want you to maybe only focus on this one feature."
"Is there—have you ever asked Manis to do something and then kind of said, 'Oh, I made a mistake or you did something wrong?'"
"Yeah, you can actually talk to it in the middle of its progress."
"So, which key function do you want to limit this to?"
"Well, that's a great question. I'm going to pull up the DocuSign website right now, because it looks like when I'm on the DocuSign website it looks like there's—they've created a lot of different products at this point. For example, they have the eSignature product, the payments product. I identify all I care about is the eSignature product, so I just want to confirm that we're only talking about the eSignature product. So let's only focus on the eSignature key functionalities."
"Good, yes, that's cool."
"Well, that's actually a really helpful feature, because, like, if you're using ChatGPT, for example, it's in thinking mode, and when it's in thinking mode you're kind of twirling your thumbs unless you stop it and then, yeah, ask it to do something."
"This is probably one of the coolest features I've noticed on this particular product: that I can interrupt it in between without it, like, you know, stopping its work. It’s going to try to..."
"As you see here, it's going to say, 'I understand you want to focus specifically on the eSignature key functionalities,' and it will adjust the approach to prioritize the eSignature feature of DocuSign. And then it says, 'Let me continue in setting up the project structure with this focus in mind.'"
"So that's just one of the coolest features I've seen on this—that you can actually have, sort of, a human in the loop to interrupt the process to readjust the planning or the setup."

I also just noticed that it's very clean. I feel like I understand what's going on; it's very pleasant to the eyes. A lot of the demos I've seen, most people just kind of kick this off and walk away, and then have a coffee, a snack, or lunch, and just come back and check where it's at. It's kind of funny, and I know you've seen those videos with the guy or the person in a Tesla driving, and you know, the person is on his way to some coffee meeting, like, "Oh, you know, create me some talking points for this coffee meeting," and it's just kind of in the driving zone while the Tesla's FSD is driving him and also Manis AI is preparing his meeting notes.
"It's kind of funny—we'll include that video so people could see that."
"What's that—we'll show that video on screen."
"Yeah, yeah, yeah, yeah, yeah. I don't like that video that much because it just feels like I don't know what's going on other than I see a car, and then the screen's kind of small. So I don't know if it's doing anything, but I thought it was kind of cool just the fact that he was doing it."

I thought it was cool.
"How long do you anticipate this taking? Is it like going to be 10, 15 minutes, 20 minutes?"
"Depends. You know, if it's something very simple like some of the research stuff, it actually doesn't take that long. I’ve done a research prompt to say, 'Hey, buy me all the research on this particular topic,' and it'll go and just, you know, do this googling and searching, and then it'll find it, and then just kind of bring the report, very similar to, you know, the deep research and deep search etc."
"So while this is running, what we can do is start a new session and say, 'Let's get, like, top—Greg's startup ideas—and see which is the most viable to run on Manis,' while it's doing the DocuSign."
"How's that?"
"Beautiful, yeah. I think for me it's most viable to run on Manis, but also the one with, 'What does Manis think has the highest probability of success?' So I'm interested to know. I'm interested to see: How intelligent is Manis, right? There’s been on this show, on the Startup Ideas podcast, I've given away, I don't know, a thousand-plus ideas, and the reality is, some ideas aren’t that good, some ideas are good, and some ideas are really good."
"Yep, yep, and it's really—the listener today, it's the humans, the listeners, who have to figure out: 'Okay, is this a really good idea, or did Greg have too much wine and I don't know what he was thinking when he came and shared that idea?' You know, you've got some really great ones. I'm surprised that people haven't taken some of those ideas and put them to work, you know, all these AI tools do; but a lot of them just DM, like, they keep it to themselves, and they just DM me and they're like, 'I built this thing and it's doing like—yeah, it's crushing.' I guess if this was, you know, so easy then everyone will be all like rich."

All right, so I asked Manis, while it's doing his DocuSign creation, to research into your startup ideas, and the top ideas that Manis thinks have the highest potential. That's what it comes up with, and let's see if we can have some actionable, you know, startup ideas that we can also ask you to create while it's doing the DocuSign creation.
"I have an edit to this prompt—sorry to do this to you, Men."
"Okay, but I would say, I think we should ask for ideas specifically that I've shared on X and ideas that I've shared on the Startup Ideas podcast and the Greg Eisenberg YouTube channel. I think we should just qualify that that's where we want it."
"What do you think? Should we qualify it or should we just let Manis figure it out?"
"Let's let Manis figure it out."

All right, yeah, the thing with this type of agentic AI that I like is that it's easier to start with. If you notice on the other AI tools, people are running into issues because they don't know where to start—they don't know how to ask, and you know, they're maybe not as good at prompting or have templates and etc. to get their ideas across.
"But you clicked the browser, by the way—sorry—could you click the browser? Like, can it go more full screen?"
"Okay, wow, so this is it, actually. So this is—it's on our, it's on my holding company's website right now—which is, by the way, insane. I'm surprised; checking you out, it's checking you out."
"Oh my God, it's going to kind of crawl through your site and just check out what you've got, and just grab as much of the information as possible. I'm just surprised, honestly, that it knows how to navigate the website, because it's, like, we created a world, so it's kind of like confusing."
"Yeah, it's using a tool called, I believe, Puppeteer, which is like an automation tool. So it's a tool that's been around for a while in test automation world, and it's using that to do a lot of these things. It's similar to what BrowserUse and Cloud Computer use."

"Interesting."
"See, all I, you know, I just have your name, and it's just a very broad ask, and it just goes into this research mode to kind of find out about you—that now it's going to your Twitter. That's crazy, but it doesn't have, say, Twitter X, right? It doesn't have an X account though, right? So it can only see a couple of my tweets."
"Oh wow, so it knew to click because in my bio it says, 'Click for free startup ideas.' It knew to click that. Wow, that's crazy."
"Yeah, it's a different type of agent—agentic AI, I would say. And I know there has been a lot of things going on just the other day, that people found out that, you know, Manis AI is really under the hood running, like, clawed."
"Yeah, yeah, there was a lot of memes going around talking about how, yeah, this is just a rapper, etc., that may be, and then maybe very oversimplifying it. I think what they've done is create a whole framework under the hood. I believe that they got an army of agents under the hood with many different fine-tuned AI models that it's using for specific tasks, and that's why it's doing what it's doing."
"And I think they've done a really good job, in my opinion, in terms of setting up this type of AI that can do all of these tasks and use all these tools, and you're, you know, less of a hassle for anyone to kind of start using this. Because sometimes you don't know what to ask, and you just kind of have this question and you're thinking, 'You throw a simple question and you're not going to get the information that you want.' But this goes a step beyond that—it will just crawl to your site, check out your bio, check out your website, and try to understand you as much as it can, and try to grab all the startup ideas that you have. And it says it visited your personal website, checked out your social media profiles—it's got a little to-do checklist that it goes through—and it actually goes and does it."
"Which is kind of remarkable on where we are with these kinds of tools, and I think this will help people, who are non-technical. And I think you know, AI tools like Manis AI will help close that gap. I think a lot of people are just kind of, I don't know, lost. I've noticed they're always asking questions like, 'How? Where do I start? What can I do? How do I ask? How do I prompt? What template do I use?' It's sometimes just a bit too much for a lot of non-technical folks. I think this closes that gap."

"Yeah, I mean, it's non-technical and it's also, like, this is onerous and a lot of work to go and do all this stuff. So if you can have Manis running during the day while you're working—or even if it works at night too—I mean, anytime, I guess—it just seems like it could help you, even if, I mean, it could help you come up with the ideas that you want to do and then put it out in the world."
"Allegedly, let's see if it actually could do it. By the way, I just heard my doorbell ring and I was like, 'My first reaction—is Manis? Is Manis at my door? Is Manis just, like, coming? You know, Manis, is that you?'"
"They'll be pretty wild, like, now a robot shows up at your door and, you know, 'Hey, I want to ask you a couple of questions about your startup ideas.' They'll be, like, crazy. I feel like this guy's stuck here, collecting Greg Eisenberg startup ideas from his podcast, YouTube channel, and website. Maybe it's just, like, a lot of—that's a lot. Let, let—what happens?"
"I feel like I'm stuck, but because typically when I see it using the browser, I will see, like, you know, navigation in the browser. Yeah, it was doing that before, and now it looks like it's just stopped. So could we ask Manis: 'Are you stuck?'"
"Yeah, let's ask that. Checking out your videos too, maybe Manis is like, 'I don't want to give these guys my startup ideas; I'm gonna go and build these myself. I don't want to—are you stuck? Are you stuck, bro?'"
"Wow, there we go. Okay, so it says, 'I'm not stuck. You point the finger to YouTube page.'
"Yeah, okay, blaming YouTube."
"Okay, man, my dog ate my homework too."
"Let me continue researching Greg Eisenberg startup ideas from multiple sources and compile a list of his most promising concepts. I'll focus on gathering specific startup ideas."
"Okay, let's check back on DocuSign. See where is—see where he's at."
"So he's creating the app now. Another cool thing about this is that you don't have to have a project structure like you see on other tools, where people say, 'Go clone this template, this boilerplate code,' and you can use that to build your app with Undercursor, Lovable, and other tools to build that for you. Because their template is more optimized—it's already like skipping a lot of steps. So you don't have to ask Undercursor to do it for you. And with this one, you don't have to do anything at all. You just kind of say, 'Go ahead, create me this app,' and it'll just create its own project structure and everything."

Now, I know it's similar to what, you know, Undercursor and WindSurve do as well, but it's just another, easier step for people who want to get started with this kind of tool.
"How worried are you that this is going to [be on] Chinese servers, and how do you think people should think about that? How is it going to turn out?"
"No, no, like I know that Manis is a Chinese company, right?"
"Yes, so there are some concerns with people around, you know, having a Chinese company handling your data. Should people be worried about that? How do they think about the fact that this is a Chinese company, not an American company?"
"Yeah, I mean, I think that there is a valid concern around that. I think one of the biggest things is that if you are a Chinese company, I think you're obligated to give access to the government. Because of that reason, I know people are very reluctant to use this. And I think, you know, again, it's a valid concern. I'm a huge security, data safety, and privacy person, so I don't want people to go and just start throwing any personal information and give any kind of stuff on these kinds of tools. But at the same time, I want them to see where other, you know, competitions are at. You know, if we don't know our competition, then we're going to be behind. And I think that, again, this is a lot. Some people are calling this the second deep-seek moment for China, and, I mean, I'm starting to see a lot of that lately. I think that we need to get ahead of ourselves with these kinds of tools, and we do have the technologies—we just didn't have someone to put the technologies all together and package it like what Manis has done. I think it's good that they have done it. And now that we know what can and cannot be done, I'm sure there will be many US companies going out there and trying to build something very similar to this tool, which is great. I think we need options, we need different alternatives. I know there are already several open-source projects that popped up that build around what Manis has done, so I think that's great. And I'm a huge open-source supporter—I use open source for almost everything. So I think there's nothing wrong with using the tool that is located in China—just be very careful and mindful of what type of data you give to it or give access to, because some of these may ask you for certain things, and you don't have to give that information. Just keep it big, just you know, do simple stuff like try it out. Like, go clone this website or something and just kind of see how the tool works and what you can get out of it. I think that's the important thing—the learning. You know, I'm a lifelong learner, so I'm always trying out new tools and new technologies, so people should not be scared to try them out, but always be mindful of the data that they share."
"Cool, that's fair."
"Yeah, I mean..."

Let's see where we're at with the Greg Eisenberg startup ideas.
"Yeah, he's still stuck trying to collect the data about you."
"And this happens. It's not like perfect—it's still in beta. There are a lot of errors still, but it's still cool to see that, how it was browsing your website and all that stuff, which is neat. And you can go back and see what it was doing and all the search that it got. It found your website, your X page, your YouTube channel, your LinkedIn, your Instagram—it’s scraping all your social media profiles and checking you out, basically, because they want to understand you and want—me, because I asked specifically by your name so it wants to know who you are and all your startup ideas. So it's going to go and try to find everything."

"Okay, wow, Manis computer has encountered a critical issue. You can choose to reset Manis's computer or start a new session. So what should we do?"
"And this does happen. When something like this happens, you could either opt to reset the computer or start a new session. But let's try to reset the computer. So it's going to try to reboot this sandbox and try to restart the task. The cool thing is, it remembers where it left off. Now, it's not a guarantee that it will be able to fix whatever issue it ran into, so we'll see. So it's going to reset. Let's go back to DocuSign and see where it's at. It's still doing its thing; it's still building the project. I like that it's kind of, you know, very worry-free. It's doing all the stuff for you. I just sit back and relax."

"Would it be annoying to ask Manis, like, 'Hey Manis, what's your status'?"
"Yeah, let's try that. See what he says, just like every annoying boss."
"But remember, the AI is, what is the word, ever—like, not an annoyed intelligent assistant, right?"
"Yeah, okay, so—"
"Oh, it says, 'I don't know if you're like, it was trying to troll us or something. As soon as I asked that, it was like, 'I'm done, like, so were we waiting this whole time? Until I ask, I'm like, 'Okay, all right! Would you like me to deploy?' So now I ask you a question: 'Do you want me to deploy this to the public network?' I say, 'Yes.'"
"And what is that going to do? It’s got its own public, like, similar to what Bel Repet has. It’s got its own little deployable website, and it's going to give you a subdomain with their primary domain attached to it, so you can actually see the website in action, and it doesn't make you sign up to any service when you do this, right? Does that work?"
"Not yet. Yeah, not. I think that when you start asking questions like— and I haven't gone to that level yet where I want to say, 'Hey, set up a Stripe and stuff,' only because, like we discussed earlier, they would force me to give up some of my information. So I haven't done that. I didn't want to go that route yet. Maybe I'll create a dummy account and just try it out. But yeah, I mean, because the glass is half full, the downside of that is if you ask it to set you up with a Stripe account, all of a sudden they hand it over to the Chinese government, or whoever, and then all your money is gone."
"Yeah, yeah, for your audience."
"Yeah, don't do that. Like I said, always when you're on the Internet, always be careful of what you share. So that's why I always tell all my family and friends and colleagues. I mean, when you're in the real world too, right? If you're hiring someone, or you're getting to know someone, and you're just giving them the keys to, you know, would you give them the keys to your apartment? Like, maybe not initially."
"Exactly, yeah. It's compared to the physical world—it's so much easier to get those stolen from you in the digital world, so you've got to be even more careful."

All right, so now it's working to deploy and demonstrate the eSignature. I'm kind of excited.
"I don't know, I'm kind of optimistic that GNA will work, but I'm also like, 'How would this actually work?' This is such a complex scenario: we asked it to do such a complex thing—a multi-billion Dollar company in one prompt. Like, come on, you know what I mean? Can it actually..."
"Exactly, exactly."
"All right, so Manis has encountered an issue with the deployment process. So it's going to try to, I guess, try to see if you can—"
"Oh, so, so he's not going to run it. So he doesn't want to create it in his own production environment. He wants me to run it locally, and that's fair. Yeah, because, I mean, I'm not a developer or I'm a bad developer, let's say. Actually, I'm not a developer. I'm not a developer, and I would do that, right? I would want it to try locally first before deploying it."
"Yeah, so it's still doing stuff. Oh, it's just zipping it up. Okay, I'm GNA to ask Manis to try again to deploy."
"Wow, you're getting greedy, M. You're getting greedy. It ran into a terminal error while I was trying to zip. I don't know if it was because I just kind of interrupted it. It says, 'Manis encountered an issue on the right-hand side. Manis encountered some issues while performing this action. No need to worry, it will handle the errors on its own.'"
"Yeah, that's another part of the cool feature: it's able to reflect on itself, on the errors that it ran into, and it'll try to figure it out on its own instead of me copy-pasting or clicking stuff to say, 'Hey, there's an error.'"
"So these are the types of seamless operations, or user experiences, for this type of agentic AI, and I think that's the next level. This is where we're getting one more step closer to the AGI, where we don't have to know more than what we need to know and just let AI handle the stuff that probably knows it better than us anyway on some of the things."
"Well, yeah, it's abstracted."
"Right, it's abstracted, very abstracted, very black-box. Some may like that, and some may not, but I think for most users, they rather have it black-boxed. They don't want to worry and think about it. This audience—this is a black-box audience. The people listening to this are like, 'Yeah, I don't care how it's done. I just want to push my ideas live and I want to push a lot of ideas live.' Just give me the results, man."

"So yeah, so I convinced it to try to deploy, so it's, again, trying again—let's see what happens."
"What what's interesting about your style, Men, is that you're a man of few words. You just tell Manis a few words, not long prompts, and that's your strategy."
"Yeah, and I do both: I do simple prompts and I also do very, very, very complex prompts because I want to give specific directions. But for your audience and my audience too, on X, I try to keep it very simple; otherwise, most people are just going to get lost on what to do. I like to demonstrate to the audience that it's a very simple prompt that I use and I got these results, and that just kind of always makes people go, 'Oh, wow. That was very easy.' Like, yeah, this is the future I want to get into. I don't want to have to think so much to come up with a very elaborate prompt, you know, when I can just say, 'Hey, I want this—just get it done.'"
"And I'm sure along the way you're gonna have a lot of questions and so sort of like an iterative approach to getting to your results."
"I think that's sort of the step I'm taking with this, and that's the sort of the step I take usually for my audience as well. I don't try to overcomplicate it, and I think that's sort of how these agentic AIs are going. I think they're trying to take a lot of the load off your thinking, and trying to understand what you want. Maybe in the future, they'll just know exactly what you want after a while, you know, like after working with you for a long time, it will probably already know your history of what you wanted and you don't even have to tell it again. I think that's the AI future that we're going to start to see."

You got me thinking of something I think could be big, and I don't think we'll have time to cover it today—maybe we will, maybe we won't—but this Manis, it seems to me that it's not just about how you can build a product and launch it. It's also probably very useful for growth tactics and actually marketing and growing whatever it is you built. I'd imagine that you probably could do things like SEO. I'd imagine that you could do things like help you create a content calendar—writing, researching based on different styles—getting basically organic content and that sort of thing.
"Am I right or wrong on that, Man?"
"Yeah, yeah. I think absolutely. You're right on point on that one."
"And I just got me thinking: maybe we should have Manis go to one of your sites and just see how is my SEO on this website. Great, let's pull up latecheckout.agency, because I know the SEO on that is probably garbage—it's literally a landing, you know, a quick framer landing page. If you go to latecheckout.agency—it's our design agency."
"Well, I guess you don't even need to go to it; we can just ask Manis to go to it, right?"
"Okay, yeah, this website—okay, all right, let's ask him. Let's take this URL, and say, 'Critique for me. I want to—here's how I would say: I have a feeling that my design agency could be getting a lot more SEO traffic. I'm looking for help from you.'"
"Oh, I'll wait till you type."
"Okay, help me improve SEO for my—let's not say my website; it's my design agency, this one."
"Yeah, okay."
"Yeah, it's—yeah, okay, that's fine. Let's remove your name maybe."
"Yeah, exactly. Let me improve SEO, but I want to be clear that, you know, basically, this firm—latecheckout.agency—it works with companies like Dropbox and, you know, Grammar, Character AI, all these software companies and sort of big companies that are looking to be AI first and transition to be AI first and build AI products for them. I want it to not just have better SEO; I want the executives from these tech companies to come to the website, right?"
"So you want those tech companies to come to this website."
"You see, all the companies we work with—Character AI, Bolt, Salesforce, Slack—and look at all the executives we work with at those companies. I want more of those executives to come to this website via SEO. I think it's important, because when you're doing SEO, not just to do it for a million visits a day—if I got a million visits, I don't really care; I'd rather have a thousand visits, but they're highly targeted."
"Yeah, there we go. You said it better than I could."
"So for the listeners on Spotify and Apple, I want a more executive—of AI-first companies—to come to this website."
"Okay, so it says, 'I understand you want to attract more executives to the latecheckout.agency website. I'll continue my SEO analysis with that specific goal in mind.'"
"Very cool. I love how it's running, like, we're software testing Manis AI in real time."
"Totally, again. They're going to remind us that, oh, it's in beta."
"Yeah, yeah, our bad. We're just testing. I believe a lot of the software is still in beta. There are a lot of issues behind the scenes that we just don't talk about."

Oh no, it has run into an issue—this guy is not going to deploy.
"Oh, so it's—this is a new message for me because I guess I haven't gone into very long context, but Manis's performance may decline with an extremely long context."
"That's what it's saying. So I think it's running into a context limitation. I think this is probably the biggest issue that the Manis AI team needs to address before it goes live, because I think a lot of people are going to run into these similar issues, especially power users like me. I'm going to keep pushing it to its limits, and we're going to run into this type of error all the time. And this is not the first time I've seen the error; it's happened before, other than this context error. It even asks you like, 'Oh, I ran into an issue; I need to reset the computer; I need to do this,' and then it just asks if you want to go ahead and restart or start a new session. So basically, starting a new session means like it may have run into some kind of long context issue that it's not able to process, and you have to start, like, a whole new session. And that's kind of the downside about this tool—the context size.
"So far, so good to know."
"Yep."
"So we were not able to complete its stock design website, so that's a bummer. I think that we probably could do a local deploy and try."
"Yeah, let's try it. I mean, even the fact that we can do a local deploy is absolute insanity; actually, it was trying to zip this thing for me to download, and it failed earlier. So let me check if it was able to finish creating the zips. I mean, it's created all the document structure and the code files, but it doesn't have the zip. Yeah, it's so interesting to see how different people prompt. I, right now, I'd probably be prompting, 'I'm just trying to get this to a local deploy; I want to create something that's beautiful that I can show on the podcast.'
"Do you think that you can make that happen? Like that's how I—because I talk as if I would be talking to a coworker."
"Yeah, I think that, you know, it's probably quicker to run it locally to see the results right away. Because deployment is not as straightforward with these tools. I know Repet has done that really well, and I think that's got some good future in where they're going. Their AI agent is getting better and better, but Manis has to work on their deployment process as well. I think they're just trying to do too much. I think that's part of the problem; they're trying to have all of these services and products packaged into one thing, including deployment, and obviously it shows that deployment is not an easy one they've solved yet, in my opinion."
"I know there have been other, very simple ones that they've done that show, like, if I do a game, like 'Make an airplane game,' like PeterFlyPeter.com, right?"
"Yeah, fly.peter.com, you're talking about the Peter levels."
"Yeah, yeah, yeah, so Peter Level is basically made like a flight simulator using AI, 3js. Oh, maximum daily usage limit."
"Yeah, there's a limit; I believe it's 10. I actually used a few before this episode, so we can't do this one today, unfortunately. You know what's also cool? It feels like we have, like, four employees working at the same time. What's cool about Manis is it really does feel like you have multiple employees working at the same time. Like, 'Oh, we've got this SEO, this marketing person; we've got this researcher; we've got this cracked PMD developer.'"
"Yeah, yeah, you got one guy working on DocuSign, one agent working on your startup ideas, and one working on improving the SEO for your LCA. So it's crazy."
"Another thing is that the problem I see is scale. Obviously, AI scale is a challenge for a lot of companies, and the reason why they're very, right now, they're invitation-only, with an invitation code that you can only have access to Manis. And they're limiting those codes too. I've been privileged to get like 10 invitation codes from them. I actually gave that out to users to my ex-account. I actually asked for 10 more. I was going to offer it to your audience, saying, 'Hey, Greg's audience, you get 10 codes if you want to try it out,' but that didn't happen."
"So, I think the thing I noticed is that very early on when people were starting to get access to Manis, their web has been hammered. And the reason we're seeing a lot more errors today is, actually, I was not seeing this many errors in the beginning because there were as many people who had access. But now that more access has been given out and more people are on the system, I start to notice a lot more frequent errors. And I think that's part of the scale issues they're going to have to address again before they go live. And do we know when they're going live?"
"I don't know yet; I have not heard any."
"I think they're trying to tease early. It looks like they launched Premier, kind of like, it's been like six days now—they actually launched it like six days ago, but it didn't get go viral until like two days ago, basically. So when it came over to us and people in the US started to get access to it, that's when it started to get real traction. I was one of the early ones who got access, and it was working pretty good. Obviously, I've done stuff like pretty basic, you know, comparisons like doing the deep research or deep search type of thing, making a— and as you can see on the left side, I had more, but I deleted them because they kind of crashed. But, you know, building a simple SaaS website, like a clone of Apple.com—obviously, Apple.com is one of the well-designed websites. So, I think it's a commonly asked thing—like people, you know, go to Apple.com and then, you know, build a casual multiplayer game. The research one works really well, obviously; simple games work well; landing pages work well. I think those are the ones that have been very successful with Manis AI so far."
"I think the other one that it does well is something like this, like you're analyzing your website for optimizations, scraping data, because it is able to control the browser and go to the website and actually crawl through it—like clicking stuff—it can actually scrape data using it. So I think that's one of the very easy-to-use and probably frequently used tools to get information for people. And as you get more complex, like DocuSign and stuff, it does have its challenges. But I like to push things to its limits."

Just to wrap up, maybe let's just go through what we've done today.
"Let's go through the DocuSign, let's go through the startup ideas, and let's see what we were able to do in, you know, 60 minutes using Manis."
"So we started with this: create a DocuSign clone. And in the middle of it, we say, 'Hold on, let's just focus on the eSignature.' So we say, 'Okay, just focus on that,' and it replanned its project and set up all the code—and I didn't actually get to see the code yet—but it looked like it set up its typical Next.js app, set up its routes, and then it created an interface with a login page, a register page, routes, password middleware routes, etc. So it came, it looked like it went through a lot of the important things that it needed, at least the basic ones. Once it got through creating this code, it got stuck on getting the deployment successful—so we got document upload, document management, and the signature field placement. So I'm actually excited to see if it actually runs. But it wasn't able to deploy, so what I would do is I would actually like to run this locally and then I can send it over to you to see what that looks like."
"Again, one thing that I do when it asks me to run it locally is I have to scan through the code, just in case. I don't just download it—that's why I didn't want to do it on this recording, because I didn't want to look paranoid. Keeping it real here, trying to play it safe, you know, because I don't know what's gonna happen. And I do like to create this into my own sandbox. I don't download this stuff to my local computer; I create a virtual sandbox in the cloud and then I run it there, just in case as well. So I don't want to do that on this session, but I will do that and then I'll show how that runs locally."
"Then let's go back and check your top startup ideas about highest potential. Again, it got stuck. So the thing with this one is interesting is because research has been very straightforward for me and it worked. And like I mentioned earlier, since people got access to this application, I've seen a lot of these issues where it hits a wall. And this should have been an easy thing that should have completed, but it wasn't able to. And again, like probably the time of the day is not the best. I think a lot of people are probably on it and asking stuff, so you may be running into more of these issues. I may try to run this again later, like, late at night when the traffic is a little bit lower."
"Then the other one is that we tried to ask it to create an airplane game, like PeterFlyPeter.com, in 3js, deploy in public. And you reached the maximum daily usage limit. Again, this is a beta stage. I believe there's a 10-request or 10-session limit in a day. The cool thing is that even though it says this, I can actually go into another session and run it. I believe it's the sessions that matter, so I'll use the one that is broken, which I think right now is still working. So this one is broken, but I'll use this one to say—I don't know if this is a loophole that I found, but when you reach the maximum limit, just start a different session and it will let you do something else. That's awesome."

Now, see, it's like, "I'll help you create a flying game." So it's just like a new session basically, but the only difference is that it’s got all this other stuff in the context. So hopefully that doesn't interfere with what I just asked.
"Oh, might Victor from Hugging Face actually have done something similar? But he didn't specify fly.peter.com; he just said, 'Create a flying airplane game where I can control the airplane and whatnot,' and it just kind of looked like he did it in one shot. And I know people are asking, 'Was this one shot, or was there follow-ups?' But I'm sure there were follow-ups. It doesn't actually produce everything in one go sometimes, and you may have to tweak some of the features—especially aesthetics. I noticed that, so you may have to just throw a very simple aesthetic prompt, and then you may have to say, 'Oh, can you make this a little bit prettier, or make it more—, I don't know, more objects in the scene,' because if you just ask it to create a flying game, it may just create like a sky and a plane field, and you may have to give very specific instructions on 'Add a building, add a tree near a lake,' or something to make it more interesting."
"Okay, so yeah, we might even show a clip of that one-shot prompt, just showing what that could look like."
"Men, just a minute—literally, this is just right after the pod was recorded—saying, 'Our flight game has been created.' So we basically one-prompted a flight game. It looks like a more basic version of Peter Level's game, but it's pretty cool. Here you can see the plane flying, and I believe you can shoot—well, maybe you can shoot, but that's the idea of the game: you shoot down these balloons, which is pretty cool, and you can play it. It's deployed; you can play it on your web browser. I think the next step for me would be, how do I make it look a little more visually cool, and also add multiplayer. I think one of the reasons why Peter Level's game has done, I mean, there's a few reasons why Peter Level's game has generated 67,000 in revenue in the last 30 days. Part of it is because of his audience, you know—he has a big audience—so he can distribute it. But the other part is that it is fun. I know I tweeted about his game and a lot of people were like, 'It's such a stupid game,' and I'm like, 'I don't know—I was playing it for 30 minutes. I was enjoying it. I could have watched The White Lotus, but I was playing it and enjoying it.'
"So, Manis did it. It took one little prompt and we got a game working, and it was deployed. So we've gone over the DocuSign, we've gone over the game, we've got the SEO, and we've got the research assistant. We did all this in like an hour—which is kind of crazy."
"We haven't been able to successfully deploy, you know, we barely deployed the app, but it got us almost there. We just have to review the code, run it locally, make some edits, and then finally deploy it. Overall, my take on this, Min, honestly, is that it's—yeah—it is the closest thing I've seen to AGI yet. And it seems like it's close. It's not there yet—I need it to deploy, I need it to deploy—but the fact that it feels like an operator mixed with, like, a Repet mixed with a ChatGPT, I think having this is super, super, super helpful for people who want to automate building their businesses, which is why I had you on the pod."
"Min, I want to thank you for coming on the pod. And I want to ask you, you know, where can people find you? We'll include those links in the show notes."
"Yeah, hey, thanks for having me. This was fun, and people can find me on X.com/minoyU—that's currently my main page."
"Cool, all right, we'll include that, Min. Thanks for… I mean, you had the beta access; I didn't. So thank you for hooking it up for everyone. I think a lot of people are appreciative of that, and, you know, let's go build some businesses. And I got an extra key for you so that you can have access if you want."
"Oh my God, yes, please do, and I'll play around with it. If people want more Manis content, please let me know, and I'll see you next time, Min."
"All right, also, Min, I know this is your first podcast that you've ever done, so thank you for sharing it with us."
"Hey, thanks for inviting me. I've always meant to do one; I just didn't get to that step yet, but maybe this will be my kickstart into the podcast world."
"Absolutely. All right, catch you later."

All right, man. [Music] baby.