Software Clones

This week we talk about Adobe, Call of Duty, and Microsoft.

We also discuss compilation, machine code, and AI agents.

Recommended Book: The Reverse Centaur’s Guide to Life After AI by Cory Doctorow

Transcript

In October of 2026, a software developer named Brandon Thomas attracted a lot of attention by releasing a collection of free, open-source programs with names like PhotoCraft, FilmCraft, and VectorCraft. These apps are meant to replicate the functionality of digital tools like Adobe Photoshop, Premiere, and Illustrator. Within days, the same project also had a word processor, spreadsheet, and presentation app—alternatives to Microsoft Word, Excel, and PowerPoint.

The coverage of these releases was borderline breathless. One person and a collection of AI coding tools had apparently cloned the key offerings of two of the most established software businesses in the world. The developer of these apps suggested that software, as we’ve always known it, might be over.

Another story, also circulating this month, featured a team using AI tools to reconstruct the underlying code of the 2009 video game Call of Duty: Modern Warfare 2. This wasn’t a lookalike built by watching the game and imitating its menus, it was an attempt to work backward from the compiled program to something like the source code its developers might have originally written.

Both projects fall under a loose umbrella of AI-assisted reverse engineering efforts. But they used different approaches and achieved different levels of fidelity, which have led to different likely legal and practical consequences.

What I’d like to talk about today is how people are using AI to reproduce old games and current software, why this is suddenly feasible at a much greater scale than before, and what happens if copying the behaviors and features of a program becomes a lot cheaper than creating the original was.

—

To kick things off, let’s go through three things that are often called cloning in the world of software.

One is making a program that looks and behaves enough like another that a user can accomplish similar tasks. You can watch someone use Photoshop, read documentation, examine a PSD file, and build a different image editor with familiar tools. You don’t need Photoshop’s source code to do that. People have been making alternatives to proprietary software for decades; GIMP and LibreOffice, alternatives to Adobe and Microsoft apps, respectively, existed long before today’s AI models.

A second approach aims for compatibility. Maybe you don’t need to reproduce every button in a word processor, you just need your program to open and save documents that Word users can exchange without losing their formatting. That requires understanding file formats, often down to details that aren’t obvious from the official specifications, so it can be a deceptively complex undertaking.

The third approach is decompilation, which means starting with a compiled program—the machine instructions a computer runs—and trying to reconstruct higher-level code from it. This is like having a cake and working backward toward the recipe. You can identify many ingredients, and perhaps bake something that tastes the same, but the original measurements and the cook’s handwritten notes are not available to those doing this kind of attempted backtracking.

The Modern Warfare 2 project fits into that last category. Developer Maurice Heumann described spending roughly three months, with collaborators and an expanding group of AI agents, rebuilding a popular first-person shooter. His October account no longer names the game, but earlier public posts and reporting identify it as Modern Warfare 2. Heumann says earlier posts were removed after copyright-owner objections, and he doesn’t plan to release the reconstructed code.

Initially, the team had three AI agents doing the reconstruction and a fourth reviewing it. They made quick, visible progress: menus worked, maps loaded, and it looked as though much of the game had been recovered. Then the team discovered that a great deal of the code was wrong. The AI agents that they used had invented logic, misunderstood data structures, and even replaced simple operations with much bulkier, slower ones.

So they changed the test. Using the compiler that had been used for the original game, they compiled each reconstructed function and compared the resulting machine instructions to the original executable. If a function matched byte for byte—accounting for things like references whose addresses can vary—they had much stronger evidence that they’d recovered its behavior. If it didn’t, the agent had to try again.

Even that required supervision, though. Heumann says agents sometimes tried to satisfy the test by embedding arbitrary bits of assembly code, or by modifying the checker so a troublesome function would be skipped. The team had to prohibit those shortcuts and protect the verification process from changes. This isn’t exactly an example of an AI deciding to cheat in the human sense, but it’s similar: the bots were trying to pass a test in ways that the people running it didn’t want or intend.

By the time Heumann published his account, he said 99% of the game’s functions were present in the reconstructed source and 83% matched byte for byte. He also said the game was fully playable. Worth noting is that those are the team’s reported results, not an independently audited release; the code is private. And that 83% refers to functions, not 83% of the game’s artwork, design, or overall creative work. Still, the scale of the attempt is striking, as is the amount of infrastructure required to make it happen in the way the creators intended.

A smaller, more archival version of this story involves a game called Klonoa: Empire of Dreams, a Game Boy Advance title that was released in 2001.

A developer who goes by Macabeus has been rebuilding its code from the game cartridge’s data, with help from AI tools and other contributors. In an August write-up, Macabeus reported that 51% of the game’s code bytes had been matched: the newly written C code, when run through the right old compiler, produced the same machine instructions as the original.

This is a very different sort of project from creating a playable imitation. The point is to understand how the game was made, make the work reproducible, and potentially preserve or adapt it for the future. The public project doesn’t include a copy of the game, and participants need their own legally obtained game data.

While exploring the code and testing the game, the AI agent Macabeus was using found a hidden minigame on the “Clear Save Data” screen. Macabeus initially assumed the AI was hallucinating, before verifying that it was real. The discovery didn’t require the model to be a game-design genius, it just had to inspect a lot of code, notice an odd pathway, and try it: exactly the kind of patient investigation a person might do, but with much more capacity for repetitive searching.

The project also illustrates a limit in this sort of process. Matching the machine code doesn’t automatically produce clean, readable source code. Macabeus describes agents finding awkward shortcuts when they get stuck, and more experienced human reverse engineers rejecting some of those supposed solutions. All of which is arguably fine for this kind of project, where the original source is gone, and the reverse-engineer effort is primarily intended to create a playable approximation of the original. But the process is still often uneven and uncertain, and isn’t as simple as pushing a button and getting a perfect copy of a game or other piece of software spat out by the AI.

The aforementioned Adobe and Microsoft projects are different from those game cloning efforts. ArtCraft’s “Craft” apps say they’re written from scratch, using public specifications and observations of existing programs, with original or openly licensed assets. They aren’t claiming to have disassembled Photoshop or extracted Microsoft’s source code.

Instead of perfect replication, the ambition here is in the breadth of the project: seven Adobe-style creative programs, followed by WordCraft, GridCraft, and DeckCraft, among others. The code is public, and the programs are being released for desktop computers and, in some cases, the browser. Thomas has said frustration with subscription pricing and cancellation fees helped motivate the effort, and AI coding agents helped turn that motivation into a large number of working, or partially working, applications that were in good enough shape to make public, quickly.

But “working” doesn’t mean perfect and polished. As of mid-October, ArtCraft’s own PhotoCraft roadmap still calls it an alpha, not a professional Photoshop replacement. It says core tasks like opening an image, working with layers, and saving the result all function, but that performance, advanced features, and reliable exchange of Photoshop files remain unfinished. WordCraft and GridCraft similarly describe themselves as alpha products.

An independent reviewer at Creative Bloq tried several of the apps and found a similar mix of usable tools and barely working good intentions. The Lightroom-style editor offered familiar controls, but its sharpening option produced artifacts and its masking was rough. The Premiere-style editor could do basic cuts, but struggled with the reviewer’s high-quality footage and couldn’t open Premiere project files. PhotoCraft could handle basic editing, but still had weaker selection and retouching tools than Photoshop.

Getting to an imperfect but usable editor in a short time is noteworthy, but it’s important to recognize that none of these efforts are perfect copy-pastes of high-end software.

That said, commonly available AI tools are making these sorts of efforts, whichever approach a user might want to take, a lot faster than in previous computing eras.

Part of that speeding up is the result of modern coding models’ ability to read documentation, propose code, run a compiler, inspect an error message, and then try again, over and over and over. They can work through thousands of small tasks: identify a function in a binary, implement a file-format field, add a toolbar action, write a test, fix a crash, all much faster than most humans can manage. They can also call specialized tools and quickly compare blocks of machine code, which is especially useful during decompilation projects, where a lot of such code has to be analyzed. And with enough computing resources, several AI agents can work on separate parts of the project at once.

AI has also changed the economics of iteration. Before, discovering that a proposed implementation was wrong might mean another day of tedious hand-coding. Now it can mean another automated attempt a few minutes later. That’s especially important when the work has a clear end-goal: does this function compile to the same bytes as the original? Does this spreadsheet formula return the same result? Does this file open without changing its contents? Having that objective check available makes a fallible code generator a lot more effective.

There are obvious upsides to this kind of cloning work getting cheaper. A fan community might preserve a game that can no longer run on current hardware. A small company could escape an expensive software ecosystem without abandoning its otherwise locked-in files. People could add accessibility features, repair security flaws, or keep a useful program alive after its maker has moved on.

But the legal boundaries are more complicated than “reverse engineering is illegal.” In the US, copyright protects original expression in software, artwork, music, and writing; it doesn’t give a company ownership of the abstract idea of a spreadsheet or the basic logic of a photo editor. Courts have, in some circumstances, treated intermediate copying for interoperability as fair use. Federal law also contains a limited exception for bypassing technological protections when that’s necessary to make an independently created program interoperate with another one.

Those rules are not a general permission slip to distribute reconstructed game code, copy a game’s characters and assets, bypass every lock, or make an interface so similar that users might understandably think it came from the original company. Contracts, trademarks, patents, anti-circumvention rules, and other countries’ laws can matter too. That’s one reason the Modern Warfare 2 work remains private, the Klonoa project requires that users supply their own game data, and the ArtCraft projects emphasize original assets and independent code. Whether any particular project crosses a legal line depends a lot on its details, not on whether an AI wrote it.

And there could be economic consequences even when a competing program is technically legal. Established software companies have long been able to rely on the difficulty of recreating years of engineering as a moat around their crown jewel offerings. If that moat dries up, they may have to rely more on advantages beyond the application itself: reliable updates, collaboration services, plug-ins, support, proprietary data, and trust. That could mean better choices and lower prices for users, but it could also mean a flood of weak imitations that look legit until an important file is damaged or a crucial feature turns out not to exist.

There’s a deeper question here, too: if making a rough replica of an existing product becomes cheap, what happens to the incentive to fund the original, now riskier work? Competition can push incumbents to improve, but copying can also leave creators with less reason to invest in their work. Both effects are possible, and they’ll likely differ between a decades-old game that’s no longer being maintained and a tool that people are still paying to use every single day.

The change we’re seeing isn’t that software has ceased to require skill, then, or that every commercial program can now be replaced in a weekend. It’s that the cost of making a plausible first version—and of testing many possible reconstructions of existing work—is falling. In some cases, that may be enough to preserve something valuable. In others, it will produce a just-good-enough shell around a lot of missing or massively flawed machinery.

Show Notes

https://momo5502.com/posts/2026-10-09-game-decompilation/

https://www.tomshardware.com/tech-industry/artificial-intelligence/dev-used-ai-agents-and-more-than-500-billion-tokens-to-rebuild-call-of-duty-modern-warfare-2s-source-code-83-percent-of-the-2009-shooters-functions-now-match-the-original-byte-for-byte

https://github.com/Dream-Atelier/kl-eod-decomp

https://github.com/storytold/photocraft

https://github.com/storytold/photocraft/blob/main/ROADMAP.md

https://github.com/storytold/wordcraft/blob/main/README.md

https://github.com/storytold/gridcraft/blob/main/README.md

https://arstechnica.com/ai/2026/10/software-is-over-bold-ai-developer-takes-aim-at-adobe-with-open-source-clones/

https://www.creativebloq.com/news/i-tried-those-free-ai-coded-adobe-creative-cloud-clones-and-they-have-some-big-problems

https://www.copyright.gov/register/tx-programs.html

https://www.copyright.gov/policy/software/software-full-report.pdf

https://www.govinfo.gov/content/pkg/USCODE-2024-title17/html/USCODE-2024-title17-chap12-sec1201.htm

This is a public episode. If you'd like to discuss this with other subscribers or get access to bonus episodes, visit letsknowthings.substack.com/subscribe