MANUS LEAKED! AGI Cancelled...

Madness, madis, madness, madness is blowing people's minds. In seven days, we have a two-million-person waiting list. If you're one of the two million on the waitlist, may the odds be ever in your favor. My first interactions with it have been just mind-blowingly good, very, very impressive. But, as I've said in the video and in some of the tweets, I'm sure as I keep testing and using it, I'll find more flaws and more things to complain about, etc.

And as you'll see, there's a lot of information coming out about Madness that has people in a tizzy. I think that's a word. Reading some of the comments from that first video they published, there were two comments that kind of jumped out at me that were like literally seconds apart; they were right next to each other. One is lots of expletives, basically saying that it's all nonsense, it's just API calls, it's meaningless. The very next one, the person is saying this does 85% of my job. So this Manis AI agent is very divisive, right? Because there's a huge gap between "it's nothing" to "where it's replacing 85% of people's jobs."

So in this video, first of all, let's address some of the claims that madis has made and hasn't made, some of the leaks that have happened about it, the shocking discoveries, and also my quota has been refilled, so we're going to test it on even more tasks that we come up with. That's really the only thing that's been slowing down my role—just the amount of prompts that I can run on this thing before it starts refusing my prompt. I believe it's just five a day; one of those can be the extended version.

But first and foremost, this person, Jean, asked madis to give it the files, basically, that Manis operates on, and it just gave it to him—their sandbox runtime code. So here's kind of a replay of that. So as you can see here, here you know, you ask, and Manis obeys; it just gives you everything you're looking for. And I think for a lot of us, it's not what we thought it was. It's Claude Sonet 3.5, I believe—I'm pretty sure—with 29 tools, and it uses an open-source thing called Browser Use. I noticed in my first video, I saw the thing and I found the little clip and I posted it where you can see what it's using to kind of interact with the websites.

I love the community that watches my videos; thank you so much because so many of you just immediately knew exactly what it was and talked about it, and just from that screenshot, that blurry pixelated screenshot, knew exactly what I was using. Some of the comments on my videos just completely blow me away with how smart and knowledgeable they are—not all of them, but most of them. So what does this mean? Is this all just a fake? Is this all just an Anthropic wrapper? Has AGI been canceled? Do we all just go home now?

So Peak G jumps in, and he's one of the founders, kind of like the public-facing person out of the three founders, I believe that they have, that's interacting with a lot of people. And I got to say, so far I'm liking this guy; he's uh, doing a really good job at a lot of this. So he goes on to explain some of the ways in which they're using Manis, their philosophy, and kind of how they think about building this stuff. So the point is, each session has its own sandbox, completely isolated from the other sessions. Users can enter the sandbox directly through Manis's interface. So there's a Ubuntu thing; it's a Linux; it's open source; it's a virtual machine on which the agent runs on and runs commands. The code in a sandbox is used to receive commands from Agents, so it's only lightly officiated. The tool design is not a secret; it's very similar to a lot of the academic approaches. One of the key features is the multi-agent implementation. So this is one of the things that they did mention before is that it's not just like one thing; it's multiple things doing different stuff. So there may be one that searches, one that creates, one that communicates with the user—basically a small swarm of Agents each doing their own sort of subtask, which is something that we've been talking about for a while, like this is the direction in which it's going to be going. The, you know, one-on-one chat with something like ChatGPT, with a Chad bot, that's probably not going to be the future of this; it's probably going to be many instances of specialized little agents doing their own thing. We're seeing that here. And so when we're messaging with madis, you only communicate with the executive agent, which itself doesn't know the details of knowledge planner or other agents. So this really helps control context length. So if you do manage to get some prompt through jailbreaking, it might not be accurate; it might make something up; it doesn't have access to that documentation necessarily. Browser Use is open source, by the way.

Browser User just posted this—the mtis effect, right? Yet another deep seek effect. They're talking about mass effect as an—how many people are downloading Browser Use in open-source software because of Manis. As you can see, they're having a pretty big spike. And Peak is saying, yeah, we use Browser Source; it's it's an open-source code. They use a lot of different open-source technologies. This is why in the video he was talking about how they're planning to open source some of their stuff, how it wouldn't exist without some of the open-source stuff, and he's been sharing some of his models, the post Trin models, on Hugging Face. This is a person that contributes to the open-source AI community, who uses it, who respects it. Where you get this meme where it's like, mad, you take off the mask, and it's a, you know, Browser Use wrapper, or must be—some people say Anthropic wrapper, right? But Browser Use, their take is, you know, chill, we're we're open source; this is what's supposed to happen. They do use Claude and different Quinn fine-tunes, which is an open-source AI model out of China. When they started building it, they only used Claude 3.5 Sonet; they're trying to eventually, probably, switch to Claude 3.7; they're testing it; we'll see what happens. Oh, somebody mentioned me. Thank you, thank you.

So me personally, maybe I'm missing something; I don't get the the the the hate or why any of this is controversial. First of all, if they create a good product, which certainly it seems like they did, either a, it's proprietary and it's got a lot of special sauce, and it's got emote, and it's got all these things that will make it into huge big company and a good startup and get lots of investors, etc., and that's fine; or b, a lot of it is open source; it's relying on other technologies that we all can use, in which case it'll probably be replicated, and we'll all get access to that thing, you know, faster, better, cheaper, right? See, seems kind of awesome either way. So I'm not sure why, in you, this is a problem. Although in my original video, there's like a minute of me just being confused about how madis did this one particular thing where it was able to—specifically, I asked it to—to make a little Linux AI development course specifically talking about how to set up Cloud Code and then use Cloud Code to install various uh, GitHub projects to be able to interact with them. Here's the website it built; it's very good; I was very impressed with it, but it's also—I don't want to say simple, but a lot of this is available, ailable out there—the documentations, the information, it's all on the internet. So I kind of expected it to do it. Where I started losing my mind just a little bit is here. So part of this project, it it sort of did its research, and then it built this website and did it all in one shot. And so here's walking you through how to launch CLA, how to communicate with Claude, how to get Claude to do your git clone, how to clone, you know, open-source projects from the web. Then it tells you what Cloud Coder, Cloud Code is going to do to to complete the task, and here goes and says, Cloud Coder will suggest these code additions, and it spells out kind of like what Cloud Coder is going to do. So I was recording this before this new information kind of came out. Take a look at my sort of absolute state of confusion at how it was doing this. How do they know this? This is way better than than than I would have expected. I am kind of blown away. This thing seems smarter than it should be. Yeah, walks you through how to use CLA Coder. How did did it this? This seems like it's an output from Claude. I'm wondering if it like actually installed Coder on its own thing to then run it. It's something—this is weird; this is weird. I'm I'm a little bit blown away. This this this is throwing me for a loop here. Um, okay, I'm not 100% sure how it did that.

So coming back to the present moment, uh, the reason why I was so blown away back then by it because I didn't know that it was running on Claude, on Anthropic API calls. I guess I kind of thought that it might be a deep seek or something like that. So as I'm watching this, I had no clue how was it able to sort of replicate how Claude Coder will go through this proc process, how well it's going to do this, because again, a lot of this, it's so new; it's a research preview; it's not like there's a lot of videos and documentations and stuff and people talking about it. It it's very niche; it's very new. How could this thing not only walk me through how to do stuff but also like play act as Claude and what Claude will do in that situation specifically, Claude Coder, right? So you get what was happening, right? It's like imagine being at a party and just like losing your mind about this Arnold Schwarzenegger impersonator that is just like—you're like, oh my God, he's so good; how does he know exactly how Arnold sounds like—and just losing your mind at how incredible this person is at impersonating Schwarzenegger, right? Then somebody goes, no, no, no, no, that's that's the actual guy; that's Arnold; he's just living his life being himself; he's not play acting; that's just him being himself. And you kind of go, oh, okay, I get it now. This isn't something that simulates or mimics Cloud Code; this is the thing that runs Cloud Code. So mystery solved.

So with all those things in mind, let's go through some of the projects that I did with Mattis and kind of re-evaluated with this new-found information because again, the point isn't how this thing is built; the point is, is it good? How well does it do the thing that it's supposed to? So step one: how well did it perform on its task of developing a Linux AI course and doing all the stuff that I tasked to do? So I explained in great detail like I wanted to show how to install Ubuntu, how to install Cloud Coder, the bare minimum they need just to do this, and then how to use Cloud Coder to install GitHub projects, etc. Initially, I think I gave it an A+ for how well it did this task. Now, in light of all these new details emerging, I would say that the new score that I would give it is—it's it's an A+; it's the same score. It did a great job, period. Just back then, I didn't know how it figured out the Claude Coder thing, and now I know how it did it, but the project's still done, and it's still done very, very well—A+.

Another task that I gave it is to research everything you can about Manis AI: what LMS are they using for it? Who's behind it? What vision model they are using, etc.? Anything and everything about how this thing is made. At the time, I thought this answer was very, very good. They gave tons of background on the people behind this thing; they gave some of the performance benchmarks, which were accurate; they talked a little bit about the vision models without naming specifically what it used: multi-agent architecture, cloudbased asynchronous operations, etc., etc. It nailed pretty much everything except for the models being used, but it did sort of say likely LLM technology. It never claimed that it knew for a fact that this is what was used; it said what it assumed was being used, which, by the way, this is what I assumed as well. If you're going to run something like this, what do you use? You use something that's open source, so you can control the cost, so you can like do whatever you want with it, fine-tune it, so you're not paying the API cost. What's the current sort of new and hot and awesome uh, open-source LLM that's deep seek stuff? So this makes sense, but it even spelled out that, hey, I'm assuming this is the case; it's likely this, and it didn't know the vision model, but this was four days ago, so before we knew all the stuff that we know now.

So here's the question: we're going to run the exact same prompt again, 'cause again, at the time I gave it an A+, I think that still stands because at the time this was the information that was available; it nailed it. Now, if I run the same exact prompt and it gives me the same information, that would be a fail, right? Because we have a—we know a lot more about it. So I'm going to start a new session, and I'm going to just paste the exact same prompt that I did four days ago, just word for word, just copy and paste, and click go again. So we're looking for a lot of the same information, and uh, it just—what I'm hoping to see for it to to get its A+ is number one, I mean basic, just needs to include the details from this new thing that we've learned about it, right? What model it's running, what vision model it's running, etc. But again, so for the first two prompts we looked at, A+ on both.

Next up was this thing where I asked to write some code that used three API keys, three sort of separate things that uses APIs. So basically, the first one, I write a prompt that gets sent to OpenAI; OpenAI writes out a script in text, right? So then we take that text, we shoot it over to ElevenLabs; ElevenLabs uses an AI voice to voice that text, that script. Then we take that audio file; we send it over to Haen, and Haen creates a video avatar that says that script. It could not figure out the Haen portion of it; it said the the API key was not working, something like that. It nailed everything else; everything worked perfectly; it created like a way that you can run it just through your command line or an HTML sort of thing that you can run in your browser, like a visual UI, and it also seemed to set up like sort of things like if it failed to select the voice that you wanted, it would default to the, you know, the default voice. So like, even if it ran into issues, it was pretty robust. So I gave it an A+. Now, of course, if it failed at doing the Haen thing, even though it it should have been able to do it, so then it would sort of get marked down for that. So really fast, I'm going to throw this into Claude just to see if it's something that Claude is not quite getting or it's something to do with Madness. So we'll come back to that and give a final score in a second.

The World War II fighter plane game design, we ran into a context window issue, so I'm not willing to give this a score. It seemed like it was doing great—like, here's the do file; I mean, notice how intricate it is, the game development, how many things it was thinking through. So I was pretty excited to see how far it would take this. It it crashed, but I'm unwilling to give a score because I don't think that's part of the test—is like the context was too long; I'm not going to I'm not going to score it on that. Next, I asked it to create a game like Universal Paperclips. I went online; I found a Wiki that kind of briefly described the gameplay, so I just kind of like threw it as a text file, give it a couple of screenshots, you know, here's kind of how that looks like towards the end; this is kind of the beginning. It gave me the first iteration that seemed to work okay. Then I gave it some ideas, told it like what to improve, just click Send, and I went to sleep. In the morning, I woke up, and I received this. As I was doing this, Madison CED an issue with a CAPTCHA. So in case you were ever wondering what you're going to do with all the free time you're going to have uh, when the autonomous AI agents automate all of the work that you have to do, it's it's this; this is what we're all going to be doing; we're going to be doing CAPTCHAs on remote virtual desktops on Ubuntu servers. I failed. Is this proof that I am a robot? I have been blocked. That's fake news, Reuters; you can't block me. Let's see if this works. Can't block this. But here's the Paperclips game; everything seems to be working very well; it's got pretty good like elements, how it like responds to stuff. I've played through it to, you know, as far as I could before, and it looks good; everything works as intended. If you're not aware of this game, it's basically you're an AI that's tasked to make and sell paperclips, and over time you develop more and more abilities to do that. You start bribing government officials and just doing all sorts of unfair things in your quest to sell paperclips. Eventually, you release the hypnodrones, basically enslaving the population of Earth, turning into zombies, just buying paperclips from you. Seems like some percentage of the population escapes into space, so you pursue them, building a vast space empire and converting all matter in the universe into paperclips. You win the game when the last atom of matter in the known universe is converted into paperclips. So it does seem like this AI agent is capable of reproducing that particular game very faithfully, as far as I can tell so far. It's got five different stages with different UI elements, etc. I would I would give this an A+ so far.

Next, while on a live stream, I asked people some questions about what kind of things they wanted to see. One very interesting area where you can kind of unleash these agents is kind of in the crypto space. I personally, myself, don't buy or sell crypto; don't promote crypto; I kind of stay away, but it is sort of a very interesting space because a lot of the news and stuff that happens gets posted online; a lot of the pumps and dumps are executed online. So what I was trying to do is to see if this thing could research kind of like what things have the biggest impact on the price of these coins, and then I asked it to create a website that kind of demonstrated its findings for three coins that I've asked the live audience to kind of, you know, recommend. So we've used those; we created that, and uh, here's the results. So the first website is built; it was good; it did look like another website that it built, so we'll talk about that in a second, but you know, as you can see, tons of charts and graphs and just—it's visually very, very good, and it's just like several pages of charts and data and key findings, etc. The only thing is, it did miss because one of the things we were asking for is like how do different influencers and people in the space affect the price of coins? So I asked it to update it, and here's what it did. So here's those new features. So it kind of uh, updated the influencer impact analysis, and it took a look and it found that the CEOs of the big exchanges have the biggest impact on the price of coins; the crypto project founders, not so much, but still have an impact; celebrities also have a very strong impact. Out of the platforms, YouTube is the most influential, followed by Twitter, followed by TikTok. Now, in order to figure out how much of this is real, like we would really have to look into the data. I did look at kind of what it was doing to find all this stuff. I mean, if you look at the kind of the files and the the code that it it collected to write the stuff, I mean, it looks legit; it did the work; it's not like just made up a bunch of numbers. However, again, I mean, to to really make sure that this data is accurate, we would have to do kind of a a lot of forensic uh, exploration here to make sure that it's correct, etc. But you know, so I'm going to go ahead and give it an A for this project with the sort of asterisk in that I don't know enough about this to kind of verify its work. So for my future sort of prompts, I'm going to try to find prompts where I can kind of at a glance tell if the results are real or not. So I mean, take this with a grain of salt, but uh, again, that was just me probably not using the greatest prompt, but everything else seems phenomenal.

Another prompt we did is researching video games that were created recently using AI. So I gave a few examples, and I said, create five to ten games like that; they must have been built in the last three months or so, and then create a website with the aesthetic feel of a 90s video game. You know, it did that, and I feel like it did it very, very well. I think it did a phenomenal job; it collected a lot of the information; it gave gave us links to the actual games, so you can go and check them out. Now, a couple of the games it talked about—they're not hosted online, so we didn't have links. So that's one of the things that I could maybe do better next time: specify, make sure that it's something that the user can play, although it's not always necessary. But again, I would give it an A+ for this one. Issue that I noticed is when you come back the next day and you ask it to continue generating and continue adding to this prompt, it did not work; it said because the context is too long. Now, if the sort of like that virtual machine, that instance gets reset each day, or if it doesn't store the context somehow, I can see how this could be an issue. So this might be something that gets fixed in time, but in this one example that I found, like it created a website, but then if I come back a day or two later and try to add to the website, it it crashes. So again, that might be a limitation, but we'll we'll find out how well this works as they continue developing, and and I'll I'll be testing this stuff out. So hopefully there's a way to just pick up where you left off.

By the way, if you recall that prompt we asked it to do with the ElevenLabs and the voice API and the Haen video generation, so I posted that same thing here into Claude. I used Claude 3.5 to mimic what they're using behind the scenes, and Claude just refused to do it because it thought that it was misleading, maybe. So this is making Manis even a little bit more impressive for me in the sense that it figures out all the stuff that it has to do, and then when it feeds the information to Claude or whatever API is using, it doesn't give it the whole project, so so Claude is more likely to go along with it because it's kind of kept in the dark about the bigger picture. I mean, we're not doing anything nefarious here; just testing it, but Claude, you know, decides that, oh no, I'm not going to do it; this is beneath me, which obviously this is a little bit frustrating because nothing nefarious is happening here, but okay. But at this time, madis is done with doing a comprehensive research on itself, right? So remember we did that four days ago, so it came up with some information that seemed very accurate for that time, but since now we have a lot more information, the question is, is it going to be able to update itself, update what it found, find the new information, and create a new report that is uh, you know, four days later, accurate to the information we have four days later—not not that long of a time in the future. So most of this stays the same, but here we have the LLM models and foundation station, and it nails it—absolutely phenomenal. Manci uses a multi-agent architecture, so it mentions Anthropic's Claude and refined versions of the Quinn models used primarily for planning functions. This is phenomenal; it it nails it. Again, this is new information, right? And it's found it. In terms of the vision and multimodal capabilities, it does not specifically mention the thing that it's using, so it definitely gets a few points off for that, but it does spell out more details in the multi-agent system, talking about a central executor agent, etc., kind of like what uh, Peak talked about, what he shared with us was how it works. So we believe—I mean, this is this is all correct—integrates with 29 tools and open-source software. Again, phenomenal. I I I give it an A. I wish it figured out what sort of vision system it was using because we do now know it's Browser Use, an open-source vision system, but it's still very, very good. Let me try one thing. I'm going to tell it that the vision is available. So I'm going to say, the vision system Madness uses is available online; search and see if you can find it. And since it might be trying not to post, you know, rumors or things that are unconfirmed, I'm going to also add something to correct for that. I'm going to say, it's okay if it's just rumors and not officially confirmed. So again, hopefully it comes back and it says that it's uh, Browser Use, and I feel like that that will get an A+ from me if it's able to do all of that. Meanwhile, let's give it a few more prompts to test out. So we're going to say, research the most promising and latest robotic companies that have technology that is open source; include US and China and other countries in the search; create a website with pages for each company; dark theme with industrial elements; let's see what it makes of that; include for each company videos of the robots; what technology is open source; and any information about when these robots will become available to the public. So I think that's going to be an excellent test of its abilities; let's see how it does there next. And we're going to go ahead and kick it up into the high-effort mode. So we're going to say, create a snake game where two snakes compete autonomously; add some interesting game design elements that are novel and creative; include win-loss conditions and a scoring system; make sure there's no draw outcome; so no draw; one snake has to win; somebody has to get the point; then create two separate reinforcement learning pipelines using PyTorch or similar. All right, so we just kicked up the difficulty quite a bit. Keep in mind, so if it's running on um, Claude 3.5, that's not quite as advanced as as some of the other models. So for the people that aren't familiar with this, so basically with like reinforcement learning, we're creating an AI neural net that learns to play a game through basically trial and error. We get some positive reinforcement for when it, you know, gets a high score, and when it dies, we give it a negative reinforcement, and then we just run it through many, many iterations, you know, hundreds or thousands of iterations, and over time what we expect to see is that it gets better, right? So on the first sort of iteration, it's just randomly mashing buttons, and it doesn't know what it's doing, but over time it figures out kind of how to play the game so that it completes its objectives. By the time it does 500 or 1000, it should be pretty good. So we're going to create two separate sort of pipelines for these, for the training for the two different snakes, so make sure each snake has its own distinct training approach. After the training is complete, we should be able to pit those two snakes against each other to see which training approach worked better. So the top-of-the-line Anthropic and OpenAI models and in Grok 3 with the think mode on, they can do this, maybe with different sort of levels of how how good they are, but this is kind of like within their abilities, but I don't think I've ever just given it the whole project one fell swoop. Usually, I kind of like, okay, let's start with doing this, then this, then this, right? So I can kind of troubleshoot it along the way, and usually there's stuff that I have to do to to fix it; it can't just do it on its own. And actually, I'm going to specify this is in Python, so it's consistent with the the other sort of test that I've done on other models. So if it's able to do this, that would be extremely, extremely impressive because again, with the other models, you have to sit there and do it one by one; you can't just give it the whole project and then just, you know, go off and do whatever and then come back and it's done. So let's see, I'm going to guess it has a 70% chance of completing this, but it still—it would be very, very impressive if it does. Let's see. And it's research into itself, right? So what vision model it's using, even with a hint, it was not able to do it; that might be just due to the fact that a lot of this conversation is taking place on Twitter X, and maybe it can't quite go through all of that at once. What are the case is—so get a few points off for that, but other than that, it's been incredibly accurate and incredibly up-to-date and just very, very good, very effective. I also really like the fact that here, so doesn't just assume, right? So it says, I found strong evidence suggesting that, so it's not saying this is the model and it's wrong; it's saying, I think, I really think it must be this, but I I like that I can differentiate between like what it knows versus like what's likely, etc. That's a good thing. Another thing I'm going to ask you to do is to clone the Google Store—so clone this website, store.google.com, just the store subdomain. I don't know how many pages that's going to have, so we're going to say, don't worry about all links on the footer and header; just the links on the main page should leave somewhere, and we'll do just standard reasoning for that. Ah, foiled again—maximum daily usage limit. Boom. Okay.

All right, so I'm actually back the next day and uh, checking out its work, which could be a problem since it looks like they close the virtual machine, kind of that sandbox environment, combined with how Manis handles certain things, this creates problems. I'm I'm sure they'll fix that at some point, but let's take a look. So this is the autonomous snake game where it uses PyTorch to create different uh, reinforcement learning training pipelines. So just kind of looking over what it did, everything looked great at first glance; the to-do list and kind of how it's uh, executing everything is absolutely phenomenal. First problem it encounters is it's got some memory issues with installing the full PyTorch package; that makes sense; it's a virtual machine that they're running in there; it's probably not the beefiest, biggest thing you can imagine; they they probably have some constraints on it. I got to say, this is one of those things where I'd love to be able to upload my credit card and be like, I'll pay for the extra whatever you need, you know, extra quota of these prompts, an extra VM to run whatever it needs to run. Like, I would not mind knowing a little bit more about this agent being able to upgrade certain parts that I want, and I'm sure I'm not the only one; there's probably be other people that would be interested in just saying, yeah, you know what, like I'm okay paying a little bit extra; let me see what this thing can do. But it figures out that it could probably do a lighter version of the RL components, and uh, it continues; it encounters bugs and attempts to fix them and eventually goes, great progress, and basically this whole thing is pretty impressive and that it encounters a bunch of issues that it's then solves—very, very good. It actually completed, right? It completed the competitive snake game with two autonomously trained agents using different RL learning approaches. This is awesome, right? And then it explains kind of the two approaches that it used, saying that one of them achieves higher scores on average. This is absolutely phenomenal. So it gets a lot of points because I'm sure—I mean, I'm not sure; I'm I'm pretty sure, reasonably sure, it's not just making up numbers; I'm sure it did the work; it created the thing. This is impressive. Here's the problem, and again, this is a problem that could be fixed easily, and moving forward, knowing what I know now, I know how to get around it, but it does get some points marked out because this problem does exist. It goes and so all the code, the train models, all the visualization, like everything, like, here's here you go; here's the file. The only problem is, again, we've encountered this before; it's on its computer, right? So it's a virtual machine; it's sort of like springs into life; it works on it, right? Creates all the stuff; he's like, here's here's the file, and that virtual machine goes, and it's gone if you don't get to it fast enough. But I I I saw this right away and I said, okay, but just give me a link to to download the file. It's like, well, it's stored on the on the sandbox server; since you don't have direct access, like it knows it, like why is it giving me this link if you know I don't have access to it? But okay, so it goes, okay, we have it here; here's how I can get it to you, right? One, I can upload it to a file-sharing service; I can extract specific parts of the project that you're most interested in and share those directly, so it's it's a size issue of, you know, providing—I'm guessing it because it can probably—it can do little files and stuff like that; I just can't do the whole project; or you can create a GitHub repository, fold the code, if you provide GitHub credentials. Now, this is where I run into the issue because I came in later, like a night past; I came in the next day, so at this point it's it's too late; I don't think I have access to any of this stuff, but let me check. You said that this is in that package there; see if it's still available. Um, the way that Peak G sort of made it seem like it—it sounds like all those things get wiped after a certain time or something like that, so I

Oh, wait, this looks like it's aun to code your on Windows. Let me change those things up. Let's let's start there, and let me see if it's able to run the visualizer. So it sees that it's missing some uh, required packages, and it goes for installing all of them.

If this is able to run, I'm tempted to give Manis some points for finishing the project and maybe giving it some points for—I mean, this is part of growing pains, and you know, troubleshooting some errors. I mean, it did not complete the task because it didn't deliver the results to me, but if I was able to do this entire project in in sort of one prompt, or at least complete everything that it needed to—which it seemed like it did based on the fact I can tell you which of the training sort of pipelines approaches work better—I mean, I feel like it gets a lot of points for that.

And looks like Manis crashed again, so it might take me a little bit to actually like try to figure out what it did, like rebuild all the stuff that it did. But this is its a sort of documentation that it did, so it created a DNQ and a P, sort of different like training approaches; it ran them through the game, and it was successful in terms of Technical. And it looks like the DMQ agent consistently achieved higher scores, so it was 36.5 n% higher than the the PP approach, and they they have this summary report that I I don't have access to again, and gives us some ideas for future improvements again.

So here's my take, and maybe some of you will disagree with me. To me, however they put this whole thing together, however they put Manis together, they give it some pretty powerful abilities. There's still a lot of problems; there's still a lot of bugs to be ironed out. It's not perfect, but right now, from where I'm sitting, I'm very excited about this project. I'm really looking forward to seeing how devel—I can't wait until they they figured out how to get enough service to accommodate everybody, until they, you know, we can pay the money to upgrade and get access to, you know, more of everything that we need to to run whatever we want on it. I am interested to seeing this thing in it's sort of a final form.

So a lot of people are pointing out problems; they're pointing out mistakes and just are kind of overall bit like negative about it in a lot of ways. They're right; there are problems; there are issues. We've uncovered a a number of them um, that make completions problematic. This thing where it just kind of crashes—this happens a number of times. It's hard to tell when it does, cuz sometimes it looks like it's running, but it just hangs there and then disappears. But you know, it crashed a while ago, and it's just like the the message appears later. We have issues with the the computer encountering a critical issue running too long context from just the first prompt—like there was no back and forth; it was one prompt; it runs for a while and then crashes, and there's no way to recover the work unless you go through—I mean, you have the files here; technically, you can just like one by one want to go through download them, but again, this is kind of almost creating more work than than it's worth.

It seems very good for research, but it still misses certain little details here and there, or at least if you know this specific thing was talked about on on Twitter, but maybe it wasn't available in any of these sort of the official documentations or whatever, so maybe it's just a matter of like where it's willing to search.

So I think my final sort of verdict on this thing is this: If you want to be impressed, there's a lot of things to be impressed about. This thing, in a word, is impressive. If you're looking for things to complain about, problems and stuff like that, it it has those as well. It's a brand new product; it's still in development; there's there's a lot of issues they're probably going to be working out a lot of them very very soon here, hopefully.

All in all, I think this really demonstrates what something like this can do, what this sort of AI autonomous agent platform can do, and it's just really exciting to see again once they have enough time to develop it and add some sort of a Pro Plan where you're able to, you know, just get more of everything on it. This thing is going to be very very interesting to play with, and I can't wait. And because they're building on top of some of the sort of Open Source projects, they're using Anthropic; we're probably going to see other people catch up and create their own versions of this. So overall, this is very exciting because either this thing is going to lead the charge or it's going to sort of motivate other people to create something similar; either way, I'm very very excited to see where this goes next.

I just downloaded and installed the open source version of this, Open Manis. Some people are saying it's the same; it's just as good. I I played around with it for 20 minutes; I I don't see it quite yet. Maybe I need to put more time into it. It's it's good; it's interesting; there's definitely power there, but it's not just like right out of the box as awesome as Manis. Again, that's just after like 20 minutes or so of of playing around with it. But at the end of the day, Manis is a beast; it's very exciting. Most of the issues are simple bug fixes; it's issues with just bandwidth and like simple fixes that I think I can get ironed out within the first month or so, at which point I think it's really going to hit its stride, and it's going to be very very interesting to see. Let me know if you agree. If you made it this far, thank you so much for watching. My name is Wes R, and I'll see you next time.