7:43
I see some spicy videos and articles saying that software is over. We'll see if that is true. They say that because modern AI tools are so powerful that they can just look at a video game and reproduce it, give you a playable version just by looking. And yes, hundreds of such AI assisted reverse engineering projects have appeared, some of which are playable in your browser. Even Photoshop and other [clears throat] Adobe tools are being reproduced. Okay, how does that happen? Well, an AI system can look at the program itself, not the source code, compiled code, figure out what it does and write a new program that does the same. But it gets crazier.
It can even just look at the behavior of the program or the game and reproduce it from that too. Then it finds differences, mistakes, and keeps improving until they are nearly the same. Now, is that legal? I am not a lawyer, but courts have ruled that the functionality of software isn't necessarily protected by copyright. So if you are an ambitious engineer, independent reimplementations can be legal. Now, models, textures, game characters, now those can be copyrighted. Now, as always, we are here to talk about the scientific aspect of this. So let's cut through the noise and look at what the actual research works say. One piece of research tested the ability of AI systems to reconstruct entire software projects without source
code. And on one particular program, it didn't start out too well. It only passed 15% of the tests. But hold on to your papers, fellow scholars, because in just nine months it jumped to nearly 100 percent. And you see that even these are older models. The newer ones are way better. And it gets razier. Researchers took a tiny 12 million parameter AI model that could almost run on a toaster these days. And [clears throat] it could recover mathematical equations with around 90% accuracy on specifically constructed test with software we don't have the source code for. But wait. So it can reconstruct the mathematical rules based on observations. That's kind of like doing science. Now that was
reproducing functionality. What about appearance? Well, researchers also used AI to reconstruct web pages by just looking at them and got incredibly high visual similarity scores, 0.938. That basically means nearly identical layouts and visual elements. And get this. It did that while using 95% less code than the original website, probably by pulling the ads and tracking out of it. I wish I was kidding. Okay. So with these research results in mind, here is why this is working so well. One, we have near infinite training data. Clone an open source repository and run the app. You now have a training pair of
source code and a working app. And you can learn to translate between the two. Second, now you have a new piece of software or game you wish to recreate. Then you have a ground truth oracle right there. What does that mean? Well, the AI writes a first version and we start playing hot or cold. Are we close or are we far? You can run and test the original software to your liking for as long as you wish. Then the AI adjusts, tries something new and tests until everything matches. But there is a problem. What is the problem? Well, Frontier Labs are pulling the brakes. For instance, Anthropic updated their Claude 5.5 Opus AI to discourage recreating existing games. And some developers reported that it also refuses
legitimate reverse engineering tasks too. That is another argument for us to own our compute and have free and local models that nobody can take from us. Something that we talk about here all the time. Super important. So I think flat out saying that software is over is a bit dramatic. But there definitely is a drastic increase in AI capabilities in this area. And that means yes, it seems likely that you will be able to run your own Photoshop and Final Cut Pro variants that are written for you and tailored specifically to your taste. Now there is a wider industry question on how useful that is. After all, if you wish to work
at a graphic design company, you need to have skills using standard industry tools like Adobe's apps. We don't quite know what the value is in skills with home brew software that no one else has heard about. But on the other side, we could say that as long as the work is done and you can collaborate with others well, just use what you like. So home brew software with interchangeable formats could get really big in the future for free or for nearly free. That's amazing. Now we mainly talk about the scientific aspect of things here. We do the papers. But I just wanted to mention this because there is some external context to this question as well that I think we should know and think about. So I think
there is so much to be excited about. Software tailored specifically for you. That sounds amazing. What a time to be alive. I use Lambda to reproduce AI research papers often in minutes. It's great to train your own models, fine tune an existing one, run inference, text to image or video easy peasy. Running a DeepSeek chatbot or agent super fast, super reliable. Lambda gives you powerful NVIDIA GPUs to run your own experiments. I test ideas from papers I cover and moments later results. Love it. Seriously, try it out now at lambda.com/papers.
