AI just BROKE all of the software industry...

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: AI just BROKE all of the software industry… · Channel: Wes Roth · 26:09 · Watch on YouTube

Topics: AI agents, Enterprise AI applications, Cybersecurity and software security, Software engineering with AI, AI safety

Summary

The video argues that AI agents have reached a point where they can reverse engineer software from its observable behavior, enabling people to recreate useful tools and combine elements from different games. Wes Roth demonstrates a video-editing tool he built to replace a frustrating subscription product and discusses a community-made Adobe-like suite, while noting that legal questions remain. He also describes how modifying local game assets could enable unfair advantages in multiplayer games, choosing not to reveal specific details. The video introduces tools and workflows for reverse engineering and modding, shares examples of game projects, and ends with advice to use coding agents cautiously, back up data, and take security seriously.

Key points

The speaker presents AI-driven reverse engineering as a way to reproduce software functionality without using its original source code, and describes this as clean-room engineering.
A reverse-engineering benchmark is cited in the video as evidence of AI progress, with the speaker reporting high success rates for GPT-6 Astra.
The speaker says he used Astra to build a usable video editor based on the features and layout he needed, then canceled his subscription to the original product.
The video discusses an Adobe-like creative software suite, presented as a user-created alternative, and frames such projects as possible pressure on subscription software businesses.
The speaker warns that changing local game assets may enable unfair advantages in multiplayer games, and withholds the specific game and exploit details.
Recreating software and games can provide verifiable tasks for AI agents: their outputs can be compared with the original behavior or code.
The video highlights Snowboard Kids decomp and Reverse Engineer Anything as tools or approaches for reconstructing and understanding software.
Examples of game crossovers and mods include Minecraft elements in GTA V, Spider-Man in a Batman game, and Mirror’s Edge-style movement in Skyrim.
The speaker recommends starting with one game and one change, backing up saves, and using research or reverse-engineering tools as projects become more complex.
Agents with broad computer access can cause security and data-loss risks, so the speaker advises backups, caution, and potentially using a separate computer.

Chapters

Start	Chapter	Summary
00:00	AI-driven software and game mashups	The speaker introduces examples of games and software being reverse engineered or combined, and previews projects and ways viewers can try similar work.
03:16	Reverse engineering and clean-room development	He discusses benchmark results, the distinction between copying source code and recreating functionality, and the clean-room engineering approach.
05:52	A personal video-editor replacement and ‘spite SaaS’	The speaker describes building a simpler editor with an AI agent and considers how user-made alternatives could challenge frustrating subscription software.
11:09	Risks of modifying multiplayer games	He explains that altering local game assets could change what players see and potentially create unfair advantages, while withholding actionable details.
14:56	Reverse engineering as a source of training tasks	The speaker argues that software recreation can generate verifiable tasks and introduces Snowboard Kids decomp and Reverse Engineer Anything.
18:07	Tools and examples for game modding	He describes reverse-engineering tools and demonstrates crossover mods that combine mechanics or assets from different games.
21:55	Building and iterating on game projects	The speaker shares experiments with RimWorld-inspired projects and recommends beginning with a small change, then expanding the work carefully.
24:36	Using coding agents safely	He shows his agent setup and warns about broad computer access, context loss, and data deletion, recommending backups and security precautions.

Quotes

Every software as a service is now open source, whether it likes it or not. — 01:38

It took about 2 hours. — 06:25

So, spite stores are literally a thing. — 10:08

There’s nothing foreign. There’s nothing extra. — 13:52

Being able to just kind of like conjure worlds into existence just by uh speaking words into this codex, you do feel a little bit like a wizard of sorts. — 25:39

Concepts and entities

GPT-6 Astra, SREBench, clean room engineering, Snowboard Kids decomp, Reverse Engineer Anything, Auto Rim MCP, Adobe, Wes Roth

Full transcript (original language)

00:00 So, I’m not sure if you’ve heard, but we’ve crossed a certain threshold with AI a few weeks ago. And now pretty much any software on Earth can be reverse engineered. Someone reverse engineered the entire Adobe suite of products, Photoshop, Premiere, etc. Now they’re working on reverse engineering the Microsoft suite. They’ve also put Mario 64 in Elden Ring, Spider-Man in Batman: Arkham Knight. There’s a mashup of Modern Warfare, Skate 3, and Minecraft. Or for example, World of Warcraft and Skate 3. And you got World of

00:31 Skatecraft. There’s a Mario 64 in Garry’s Mod, Mario 64 in Sonic Generation. I’ll skip like the next 50 of them because it’s either Minecraft plus something or Mario 64 plus something. You can do Mirror’s Edge style parkour in a Skyrim, Snowboard Kids, a Nintendo 64 game, reconstructed and ported to PC. You can take any game on any console or PC and port it to anything else. You can make it VR. Pretty much anything you can think of it exists. Or if it doesn’t yet exist, well then you can make it right now. So, in

01:04 this video I’ll show you what people have built, how you can start building something similar, and some of the projects that I’ve been working on. But first, what would it look like if you put a Minecraft Nether Portal in the middle of GTA? Just take a look.

01:38 So, Fireship did a video about this a few days ago, and I think they put it best. Every software as a service is now open source, whether it likes it or not. So, if you put Mario into Elden Ring, then Mario has all the familiar movements that you’re used to in Mario 64. When you bring Minecraft into GTA, all the same mechanics of Minecraft exist, but they exist in that GTA world. Suddenly, you can start experimenting with some of the combinations that the original developers never shipped, could never maybe even imagined. So, the

02:11 bigger possibility here is this. You take a software that’s already doing something useful, something that you’re using for work or for entertainment on a daily basis. You take some part of it, you understand how that part works, and then you reuse it or something new around it. Now, the games make that possibility very, very easy to see, especially if you’ve played those games, you can see how Mario moves, and yet it’s in the world of Elden Ring. You can see how it’s a mashup. It very much captures the essence of one and it puts it into the essence of the other. Now, imagine applying that to all sorts of

02:45 software. And this is a fairly recent development. Really, where that inflection point began is with GPT-6 Astra. Specifically, on this one little benchmark called SREBench. It’s a reverse engineering benchmark. The test is simple. It asks if an AI agent can figure out how some software works without being given the source code. So, it’s shown the output, like, here’s what the software looks like, here’s what it is, can you recreate it, you know, kind of with your own source code, but recreate all of the functionality. And

03:16 as you can see here, GPT-6 Astra is getting a phenomenal success rate at very inexpensive prices. So, Astra solves 88 percent of the tasks in a single attempt and 99.2 within four attempts. So, the point is that now pretty much anything can be reproduced, reverse engineered. And while there’s some discussion about the kind of a legality of this, as as far as we know, as far as we understand, this seems to be very legal. This holds up. The key here is not to use the source code so because that is copyrighted by the

03:47 original creator. As long as you’re not looking at that and just copying it over, but rather you’re creating everything from a scratch, that is perfectly legitimate. Again, we’re probably going to see some lawsuits that kind of, you know, do battle over the finer legal points of this, but this is known as a clean room engineering. This is a concept that’s been around. One team takes a look at some product or software and kind of just breaks down the entire architecture, how it was built, and they try to kind of write a high-level functional specifications describing what the system does. Now,

04:18 they’re only allowed to capture the functional ideas, which can’t be copyrighted. Then, a second team just builds those ideas into real life. This way they’re able to kind of replicate it but without anyone actually just copying the existing design. By the way, here’s the paper on which that benchmark is kind of a formed. The next challenge for Agnik Cybersecurity, a realistic contamination-free reverse engineering a benchmark. And so, we have a people from Google DeepMind, OpenAI, Anthropic, UCLA, UC Berkeley, Columbia, etc., which

04:49 is kind of interesting to see because the life expectancy of these benchmarks that are coming out for AI, they’re just getting shorter and shorter because GPT-6 Astro, of course, it gets a a near-perfect score. This paper was first submitted August 11th of this year. Now, for me personally, for last a few years, I’ve been using a particular video editing software to help me edit my videos. More recently, I have a video editor who is absolutely awesome, but whenever I have to get something out quick or it’s a super simple, I use this particular company for my video editing. I’m not going to say the name just to

05:22 avoid any issues, but recently the quality of that software has been kind of going downhill. They try to get you to buy more and more stuff. They have these annoying pop-ups even though you’re paying a monthly fee for the software. And then, more recently, there was actually a bug, some glitch that prevented me from completing to be able to, you know, export the video that I was working on. Their support was useless. They give you a little chatbot that can’t do anything. If you manage to actually email the actual desk, you know, they email you back like 3 days later and don’t really help out at all.

05:52 Now, I’m sure you’ve encountered in your own life a software like this or companies that kind of do this stuff. And in the past, you either had to kind of just deal with it or find some alternative, some other company, maybe learn a new software, etc. But, no longer because I just took a GPT-6 Astra on ultra mode and I told it to I didn’t say like replicate it exactly, but I described the layout to the functionality that I needed. And again, for most of the software that you use, if you think about it, you don’t need 100% of all the tools and bells and whistles that they have. You probably

06:25 just need a 5% of the whole and you need that to work flawlessly. And that’s exactly what Astra did. It took about 2 hours. I set a goal, it ran for a long time, it built everything out. There was a little bit of a back-and-forth and I’ll probably continue iterating on it, but it’s like 80-90% of the way there and it’s very very usable. I’ve canceled my subscription with the original company and I’m never going back. Now, I’m pretty sure some of you are Larry David fans. So, you might remember that episode where he opens up a spite store.

06:56 Basically, there’s some coffee shop, if I recall correctly, there’s some coffee shop that just doesn’t do something right to his liking. Like the muffins are too soft or the biscotti is too hard, whatever. So, he, using, you know, his money and his capital, just opens up a store right next to it, selling basically the same thing. The goal isn’t to make a profit. The goal is just do a little damage to a business that made you mad. Anyways, basically, this really catches on with a celebrities and, you know, people with fame and a money because a lot of them decide to open up their own spite stores to put out of

07:29 business various other businesses that made them mad. Here’s the thing. We will see that play out in real life right now. Mark my words. If some software as a service business makes the wrong customer mad and could be any customer with you know a max chat GPT plan. And if that customer is petty enough and has enough time on their hands, well, then that business can expect to see a clone of itself. That clone might be completely free. It might have better features. And you can really get some really like viral marketing campaigns

08:02 for these stores. They’re not stores they’re not spite stores they’re spite SaaS sassy spite whatever you want to call it. The point being is if some business made you mad and then somebody creates this spite SaaS software project, not only are you going to use it, you’re also probably going to promote it. And really no one has made more people mad I think than Adobe. First of all, they tend to bundle all their stuff together so it doesn’t matter that you need just one thing, you’re going to pay for all of it. Also they created some pretty evil cancellation policies where basically

08:32 you paid more if you canceled. They apparently allegedly allegedly is the word allegedly they hid their cancellation fees and charged the 50% of the remaining subscription when you cancel. So literally sometime in the last few weeks I believe somebody took the entire Adobe suite the creative cloud as they call it and basically just replicated it in its entirety. By the way, you get that this is not sponsored right? I’m not getting paid no one’s getting paid to talk about this project. This is a spite store and a lot of

09:04 people have spite for this particular business that this is a cloning. Now I’ve downloaded the suite here it is. I haven’t had a chance to check out all of it quite yet. I haven’t used Adobe’s video editor all of that much although they have their own clone. Some people there’s mixed reviews of it. Some people don’t love it. I’m probably not going to try it out cuz you know, I’m making my own. But this photocraft which is you know, allegedly something similar to Photoshop. For those of you that have used Photoshop in the past, does it this look familiar? It might. Some might even say it’s a pixel perfect copy, which of

09:38 course it’s not because Photoshop doesn’t say Photocraft on it, right? This little icon is a different, of course. But, I feel like for pretty much everything else, it’s more or less identical. Like I I know where all the buttons is. This is like this first or second time that I’ve opened this software, and I know where all the tools are, which is kind of amazing. That’s that’s a quite a intuitive interface, if you will. So, just from the few brief moments that I had to use it, you know, it feels very very similar. It’s lightning fast. It loads really really quickly. And again, I know some people

10:08 have said that there were issues with this, that it was slower. I have not experienced any of that. So, spite stores are literally a thing. I think Larry David was ahead of his time. And honestly, I I could not be happier because I I think before a lot of these companies, they did kind of abuse the fact that they had this moat around their software. Now, if your only moat was the ability to build this software, then well, there is a no moat. So, I in general, I’m pretty happy about this because I think there’s going to be this this is going to put some pressure on these software as a service businesses,

10:38 you know, like Adobe, to not, you know, kind of abuse their customers because there’s always going to be an alternative. I’m not saying Adobe was abusing their customers, just you know, allegedly abusing their customers. I got caught up in their cancellation racket one year, and I was not happy. Now, that’s kind of the good side of it. There’s also the bad side. And I went back and forth quite a bit with this because I did not know whether I should talk about this or not. Because I was able to do quite a bit of work with

11:09 this, but if I publish kind of all of my findings, it could do quite a bit of damage. Basically, when it comes to two certain competitive online video games, these have massive massive player bases. There’s entire e-sports around them. There’s many communities and millions of people playing them. And one of the kind of issues is when you’re playing an online game, the server only really controls a small portion of what’s happening. So, most of the stuff that happens happens on your local machine and only the needed information is uh

11:40 transferred back and forth to the server. So, I was pretty blown away when I was able to modify a lot of stuff in one of these games. And I don’t want to reveal the specific game because it’s a vibrant game, huge community, many players. But, when you’re playing it online, a lot of the kind of the the assets, the 3D models, all of them are kind of the they’re on your computer. Now, that these AI models are getting really good at reverse engineering these games. Not only that, but they’re really good at creating 3D assets. They can take, for example, an asset in the game that looks like this. It’s a I’m going

12:12 to call it a tank, just to not give it away. And they can, for example, remove the top portion of it so that it’s still there, but it’s uh see-through. So, before there might have been some real visibility issues. So, that’s there’s a pros and cons to using this thing, whereas after it’s kind of unfair, especially when the other players in the game don’t have that same advantage. Lots of characters in the game, they have kind of that internal rigging system, kind of like the {quote} skeleton that shows how they move. You’re able to add what’s called X-ray dots or X-ray pixels to those rigs.

12:45 X-ray pixels are basically pixels that you can see through anything, right? They’re always rendered on top. So, if there’s a person behind the wall, this rigging will be drawn kind of on top of the wall. You you see this through the wall. You’re able to add mini maps in a game where you’re not supposed to have a mini maps. You’re able to add light sources to objects that normally wouldn’t have a light source. You’re able to change textures. You’re able to do a lot more than I thought you were supposed to be able to do. It goes a lot further than I expected. Now, some of you may be saying, well, yeah, these are

13:18 cheats and the cheats have been around for a while. People sell cheats online, etc. Here’s the issue. Most cheats, if I understand correctly, they use kind of a different approach. They use these memory injection techniques. Oftentimes they require something running alongside the game. This is very, very different. It’s different because there’s no script. You’re not running anything alongside the game. There’s no second program. You’re going into the game itself into like the assets of the game and you’re changing those. Then when the game starts, it just pulls from those files and it just displays them in the

13:52 game as it normally would. There’s nothing foreign. There’s nothing extra. Now, to be clear, I did not use this in a competitive environment, so I didn’t use this to to cheat or to get an extra advantage. This was an experiment. I wanted to see if it was possible. And the answer is yeah, it’s possible and the amount of stuff that you’re able to do with this is kind of scary. So, it’s going to be kind of interesting to see how well game companies are going to be able to adjust to this. Because now if anyone can sort of rebuild, redesign, reverse engineer, and change the game in any way that they see fit and do so kind

14:26 of like on their local machine and then participate in in multiplayer, well, things get weird. And again, I apologize that I’m not showing you all the details, but again, if I did, I feel like it would ruin that game. It would ruin the game for a lot of people. And I’m not even going to walk through exactly how I did this specific thing. This is more of a just for you to know and be aware of this because it’s very likely that a lot more of this is coming. So, I’m still not a 100% sure that I should be talking about that. Those things are technically considered exploits. And for me personally, once

14:56 I’ve tested them out, I’ve since then deleted everything. But of course, there will be tons of people that will abuse it and continuously use it to give themselves an edge in the game. But coming back to more AI centered topics, if we kind of take this idea one step forward, the idea that these models can reverse engineer, for example, games and various complicated software. This is huge as far as training data goes. If you’re able to create software, reverse engineer software, each time you’re doing it, kind of going from a functions to source code and back. This is going to be extremely good for training future

15:29 models. Because at the end of the day, the results, they’re checkable. In fact, one of the tools that I’ll show you in just a second, it’s called Nigel. Basically, imagine giving this model some specific movement function, right? Mario doing a triple backflip or some parkour move. Then you let the coding agent try to create that specific movement. Then you run the verify fire and you kind of recheck what happened. So you rebuild the game. In this case, they were rebuilding an old N64 game. But the point is that creating that sort of specific movement, it becomes a problem with a verifiable solution. Can

16:02 you make the character move in this specific way? Can you create this a function or this a building, whatever game mechanic or visual or or whatever you’re trying to recreate. This is great training data for these models. Since the models are now able to take any software, any game, uh just the amount of training data that now they have access to is massive. The fact that you don’t have access to the game’s source code is almost an advantage because you really don’t need it. You have some observable behavior. The AI agent can attempt to multiple different ways of recreating that. And you’re able to

16:34 check the answer against how the original mechanic works. Now, if you wanted to kind of mess around with this for yourself, one interesting project is Snowboard Kids decomp. So a decomp is a decomposition. And the basically the goal was in this particular case to decomposition this game from N64 called Snowboard Kids. Here on the on the package it says only for Nintendo 64, interestingly enough. But with AI, it’s for anything. And again, the loop is very simple, right? We choose some next unfinished action. We let the coding

17:04 agent try. It writes C code, compiles it, and compares it with the original instructions. Then it runs the verifier and either it works or it doesn’t. This example checks the exact rebuilt code, whereas a movement test might check behavior such as like the jump height. So, you may give a character the same inputs and see if let’s say the jump is as high or it lands in the exact place on the track. If it hits the wall at the next spot, etc. So, we’re able to like recreate the behind the scenes physics or whatever you want to call it by just making sure that our character, the observable way that they behave is

17:36 similar or exactly the same as the game. So, you can imagine how it would be very difficult to describe in English to the agent exactly how to, you know, make the the same game feel right. So, like you got to make it like smooth and fast, but it doesn’t really translate exactly into kind of like the feel that you’re going for. Whereas with this loop where the agent can just iterate and try again and again, eventually you’ll find some a very close approximation of the original code. And then maybe it’ll be the code itself will be different, but the actual output will be very very similar. And then kind of the other big tool that

18:07 people are using for this is this reverse engineer anything or REA. So, basically running this on whatever software you want to kind of reverse engineer, it allows the AI to better understand how it works even if you don’t have the source code. So, it’s able to help explain a behavior, recreate some of the features. And I found that while Astra can often dig pretty deeply by itself, but something like this allows it to dig even deeper in in my experience. I haven’t put too many hours into this yet. Probably ran it for about an hour or so and I definitely saw some pretty good early

18:39 results. So, I do have some ideas to test with this next, but this is the big one that a lot of people are using and talking about. So, I’ll put those links in the description if you want to check them out. If you’re not sure how to use it, again just use it your favorite AI agent, take the link, give it to your agent, and that will be able to walk you through all the steps needed. Now, obviously got to be a little bit careful with these links, you know, there’s still fears of prompt injections. Certain websites can attempt to hijack your agent for something bad. With something like this where it’s extremely

19:10 popular, so many people looking at it, the code being a visible. I feel like most people would say this is fairly safe, but again, use your own judgment. And then for specific games, there’s these pass-through mods. So, for example, that Minecraft by GTA V, this is kind of what it looks like. The idea there is it’s taking a certain portions of one game and putting it into another. And here it spells out what portions come from which game. So, for example, if you’re putting down a Minecraft block, those placed blocks become visible GTA props, so people and cars will stop at your wall. So, as you can

19:43 see, you’re kind of using both games. So, there’s the Minecraft block, but there’s also something analogous in GTA that’s being kind of placed there at the same time, so that the pedestrians and cars know how to react to it. TNT and creepers blow up in both games. Each Minecraft explosion is also a GTA explosion. Minecraft weapons with GTA effects, right? So, here we’re using what’s already been coded up and made in GTA as the final sort of result. And you’re able to have the mobs from Minecraft fight the police from GTA, etc. Here’s Ark web, so this is where you’re putting a Spider-Man into the

20:15 Dark Knight game. So, this is what the gameplay looks like. Now, this is probably one of the more limited things you can do. This probably looks cool, but it probably looks better than what it actually is. And the setup that they have is just no technical knowledge required, just to set up with Claude. Here are the instructions. Again, all links will be in the description. And I’ll also link the Mirror’s Edge movement mechanics in Skyrim. But, here’s the thing. Now that you have these sort of demonstrations for your AI agents, as well as the tools that I showed you, basically, you can take all those links, copy it into ChatGPT or Claude or Anti-Gravity or Grok, whatever

20:49 you want to use. At this point, it should have enough examples and code and everything it needs to kind of get to the point. Start by maybe just telling it, “Hey, kind of understand how all this is fitting together.” And then, if you have your own special idea for a crossover or something that you want to mod, you know, just ask it and and how far it takes you. Right off the bat, I’m going to tell you that Astra is a lot easier to work with in terms of what it’s willing to do. Some of these are things that I was testing out that could be used as exploits. Claude refused to do it. I think it said something along the lines of, “I’m going to stop you right there. I’m not going to do any of

21:22 that.” Blah, blah, blah. Like it just went off about why this is a bad and certainly the fact that you’re able to mod multiplayer games, but there are ethical considerations there for sure. For a lot of these single player games, Claude will do perfectly, but putting Astra on ultra mode, especially if you’re using that kind of the extra fast mode and just watching it crank through building assets and doing a lot of this work is is very exciting, I got to say. If you wanted to try this for yourself, it’s pretty simple. I would try one of two things. Either take an existing a game and then try to mod it in some way

21:55 or if you want, try to have that game recreated so you can kind of build more stuff around it. One of the first things I did and I showed it way back when Astra got released was with the game RimWorld. It kind of looks like this. You have a number of colonists, they try to survive. It’s a spiritual successor of Dwarf Fortress if you know about that game, but this one’s a lot easier. So they mine and a farm and do all sorts of stuff and also survive various attacks. So the first thing that I did was I asked Astra at that point and I replicated some of this with Claude as well, but I asked them to just add a

22:25 hook to the game to where the large language models are able to play the game. So basically I would hook up ChatGPT and it’s would play the game. It would control the flow of time, what the different characters did and it would to the best of its ability just try to progress through the game. It added a little dialogue box in the top right so I would be able to see kind of what it’s doing. That was kind of fun, but it was more for a benchmark. And then the second thing that I did is I had it recreate this with our own graphics, all our my own mechanics, etc. And it did. It basically created a separate standalone game. It called it Harfline. I had to look up what it what it called

22:59 it. By the way, it used the open source stuff for that. So, there’s this a thing on GitHub called Auto Rim MCP. So, you’re able to have an LLM control a RimWorld colony through an MCP. So, it just hooks right into the game. So, you’re able to use something like this. Or what I did was actually told Astra to recreate something similar. So, I think it used some of the ideas here, but it’s kind of built its own version of it from scratch. So, initially just start with one game and one change. If it’s a game that you care about the the progress in the game, make sure you have a separate copy of it back up your save games, etc.

23:31 For more complicated stuff, it’s often a good idea to tell the coding agent to do some research online, see if there’s already assets out there they can use something like this for example. And then just slowly over time try to add different things that you want to see in the game. As you keep adding things, as you’re getting deeper into this modding or reconstructing, reverse engineering, whatever you want to call it. If you’re seeing that the agent has trouble really understanding all of the game’s mechanics, how everything works, then bring in something like reverse engineer anything. That would be a very good next step. Or if you’re trying to recreate something very faithfully, very kind of

24:05 similar to the original, then this one, the Snowboard Kids decomp, it probably going to work pretty well. So, here they call it the Snowboard Kids, but there’s actually no games here. They just created this code and all the code that you can use and they’ve created the code the Snowboard Kids. That was the the test environment for it. But you can use something like this to for example recreate Pokémon Red and add a whole bunch of new Pokémon and adventures and whatever you want. Then add the the ability for your favorite large language model to play through it. Here by the way is Hearthline. So, this is kind of

24:36 what it looks like, at least the first UI that we’ve built. And for a lot of this, I personally like to use ChatGPT Codex for desktop. I run it on my computer, so I give it full access. I like running it on ultra. And of course, it will give you a warning message like, “Are you sure you want to give it full access?” And then put it on the extra crazy scary ultra mode. Because at this point it does have access to everything on your computer and also it’s able to browse the web. So, definitely kind of understand the security implications of this. We’ve all heard stories of things going wrong and a lot of valuable data

25:08 being deleted permanently. A lot of this happens when the context gets compacted, right? So, it when it runs for a long time, it reaches the context window and it says, “Wait, I got to compact to the context.” And sometimes on the other side of that, it loses some important piece that was supposed to remember some strict detailed instructions that you gave it. So, if you’re running it like this, I would definitely just understand what you’re getting into. Make sure you have some of backups. Make sure you take your security seriously. Or just run it on a separate computer. But literally, this is the setup that I’ve built most

25:39 of my stuff in. I’m just incredibly impressed at how far this will take you even without introducing some of the tools that we talked about. So, let me know what you thought about all of this. If you’re thinking about building something, then definitely let me know. If you want specific deep dives and tutorials on any aspect of this, uh just leave a comment down below. But more than anything, just just try it out. Being able to just kind of like conjure worlds into existence just by uh speaking words into this codex, you do feel a little bit like a wizard of sorts. Anyways, with that said, if you

26:09 made it this far, thank you so much for watching. My name is Wes Roth. I will see you in the next one.