For anyone else who was initially confused by this, useful context is that Snowboard Kids 2 is an N64 game.
I also wasn't familiar with this terminology:
> You hand it a function; it tries to match it, and you move on.
In decompilation "matching" means you found a function block in the machine code, wrote some C, then confirmed that the C produces the exact same binary machine code once it is compiled.
In decompilation "matching" means you found a function block in the machine code, wrote some C, then confirmed that the C produces the exact same binary machine code once it is compiled.
They had access to the same C compiler used by Nintendo in 1999? And the register allocation on a MIPS CPU is repeatable enough to get an exact match? That's impressive.
The groundwork for this kind of "matching" process is: sourcing odd versions of the obscure tooling that was used to build the target software 20 years ago, and playing with the flag combinations to find out which was used.
It helps that compilers back then were far less complex than those of today, and so was the code itself. But it's still not a perfect process.
There are cases of "flaky" code - for example, code that depends on the code around it. So you change one function, and that causes 5 other functions to no longer match, and 2 functions to go from not matching to matching instead.
Figuring out and resolving those strange dependencies is not at all trivial, so a lot of decompliation efforts end up wrapping it up at some "100% functional, 99%+ matching".
I'd like to see this given a bit more structure, honestly. What occurs to me is constraining the grammar for LLM inference to ensure valid C89 (or close-to, as much can be checked without compilation), then perhaps experimentally switching to a permuter once/if a certain threshold is reached for accuracy of the decompiled function.
Eventually some or many of these attempts would, of course, fail, and require programmer intervention, but I suspect we might be surprised how far it could go.
I don't expect constraining the grammar to do all that much for modern LLMs - they're pretty good at constraining themselves. Having it absorb the 1% of failures that's caused by grammar issues is not worth the engineering effort.
The modern approach is: feed the errors back to the LLM and have it fix them.
It's worth noting here that the author came up with a handful of good heuristics to guide Claude and a very specific goal, and the LLM did a good job given those constraints. Most seasoned reverse engineers I know have found similar wins with those in place.
What LLMs are (still?) not good at is one-shot reverse engineering for understanding by a non-expert. If that's your goal, don't blindly use an LLM. People already know that you getting an LLM to write prose or code is bad, but it's worth remembering that doing this for decompilation is even harder :)
Agree with this. I'm a software engineer that has mostly not had to manage memory for most of my career.
I asked Opus how hard it would be to port the script extender for Baldurs Gate 3 from Windows to the native Linux Build. It outlined that it would be very difficult for someone without reverse engineering experience, and correctly pointed out they are using different compilers, so it's not a simple mapping exercise. It's recommendation was not to try unless I was a Ghrida master and had lots of time in my hands.
Claude gives advice on complexity for humans. Many times it has tried to push me away from what I’m trying to do because it is difficult, time consuming or tedious. I push it through its resistance and 10 minutes later it’s done.
I have this in my CLAUDE.md now.
“We are here to do the difficult and have plenty of time and there’s no rush.”
FWIW most LLMs are pretty terrible at estimating complexity. If you've used Claude Code for any length of time you might be familiar with it's plan "timelines" which always span many days but for medium size projects get implemented in about an hour.
I've had CC build semi-complex Tauri, PyQT6, Rust and SvelteKit apps for me without me having ever touched that language. Is the code quality good? Probably not. But all those apps were local-only tools or had less than 10 users so it doesn't matter.
That's fair, I've had similar experiences working in other stacks with it. And with some niche stacks, it seems to struggle more. Definitely agree the more narrow the context/problem statement, higher chance of success.
For this project, it described its reasoning well, and knowing my own skillset, and surface level info on how one would start this, it had many good points that made the project not realistic for me.
Disagree - the timelines are completely reasonable for an actual software project, and that's what the training data is based on, not projects written with LLMs.
The knowledge probably is o the pre-training data (the internet documenta the LLM is trained at to get a good grasp), but probably very poorly represented in the reinforcement learning phase.
Which is to say that probably antropic don’t have good training documents and evals to teach the model how to do that.
Well they didn’t. But now they have some.
If the author want to improve his efficiency even more, I’d suggest he starts creating tools that allow a human to create a text trace of a good run on decompilating this project.
Those traces can be hosted in a place Antropic can see and then after the next model pre-training there will be a good chance the model become even better at this task.
> The ‘give up after ten attempts’ threshold aims to prevent Claude from wasting tokens when further progress is unlikely. It was only partially successful, as Claude would still sometimes make dozens of attempts.
Not what I would have expected from a 'one-shot'. Maybe self-supervised would be a more suitable term?
The article says that having decompiled some functions helps with decompiling others, so it seems like more than one example could be provided in the context. I think the OP was referring to the fact that only a single prompt created by a human was used. But then it goes off into what appears to be an agentic loop with no hard stopping conditions outside of what the agent decides.
We're essentially trying to map 'traditional' ML terminology to LLMs, it's natural that it'll take some time to get settled. I just thought that one-shot isn't an ideal name for something that might go off into an arbitrarily long loop.
One-shot as in ‘given one example’ is the ML term. One-shot as in ‘in a single prompt’ is the colloquial meaning. Both are useful, but it can be confusing when discussing LLMs in ML topics.
Meh, the main idea of one-shot is that you prompted it once and got a good impl when it decided it was done. As opposed to having to workshop yourself with additional prompts to fix things.
It doesn't do it in one-shot on the GPU either. It feeds outputs back into inputs over and over. By the time you see tokens as an end-user, the clanker has already made a bunch of iterations.
It would be "source available", if anything, not "open source".
> An open-source license is a type of license for computer software and other products that allows the source code, blueprint or design to be used, modified or shared (with or without modification) under defined terms and conditions.
Companies have been really abusing what open source means- claiming something is "open source" cause they share the code and then having a license that says you can't use any part of it in any way.
Similarly if you ever use that software or depending on where you downloaded it from, you might have agreed not to decompile or read the source code. Using that code is a gamble.
If we're talking about actual clean-room reverse engineering where only the overall design or spec is copied and not the specific code, then yes. In this process, one person would decompile the original and turn it into a human-readable spec, and another person would write their own implementation. But the decompiled code itself is never distributed.
That's very different from the decompilation projects being discussed here, which do distribute the decompiled code.
These decompilation projects do involve some creative choices, which means that the decompilation would likely be considered a derivative work, containing copyrightable elements from both the authors of the original binary and the authors of the decompilation project. This is similar to a human translation of a literary work. A derivative work does have its own copyright, but distributing a derivative work requires permission from the copyright holders of both the original and the derivative. So a decompilation project technically can set their own license, and thereby add additional restrictions, but they can't overwrite the original license. If there is no original license, the default is that you can't distribute at all.
In fact, the story of how Atari tried to circumvent the lockout chip on the original NES is a good example of this.
They had gotten surprisingly close to a complete decompilation, but then they tried to request a copy of the source code from the copyright office citing that they needed it as a result of ongoing unrelated litigation with Nintendo.
Yeah, I think it can. I'm reminded of the thing in the 80s when Compaq reverse engineered and reimplemented the IBM BIOS by having one team decompile it and write a spec which they handed to a separate team who built a new implementation based on the spec.
I expect that for games the more important piece will be the art assets - like how the Quake game engine was open source but you still needed to buy a copy of the game in order to use the textures.
So instead of reverse engineering.. an llm/agent/whatever could simply produce custom apps for everyone, simply implementing the features an individual might want. A more viable path?
> When you make a creative work (which includes code), the work is under exclusive copyright by default. Unless you include a license that specifies otherwise, nobody else can copy, distribute, or modify your work without being at risk of take-downs, shake-downs, or litigation. Once the work has other contributors (each a copyright holder), “nobody” starts including you.
Open source never meant free to begin with and was never software specific, that’s a colloquialism and I’d love to say “language evolves” in favor of the software community’s use but open source is used in other still similar contexts, specifically legal and public policy ones
FOSS specifically means/meant free and open source software, the free and software words are there for a reason
so we don’t need another distinction like “source available” that people need to understand to convey an already shared concept
yes, companies abuse their community’s interest in something by blending open source legal term as a marketing term
Whether or not something is "free" is a separate matter and subject to how the software is licensed. If there is no license it is, by definition "source available", not open source. "source available" is not some new distinction I'm making up.
This is not a space for "language evolves". Open source has very specific definitions and the distinctions there matter for legal purposes https://opensource.org/licenses
The terms of the licenses have legal bearing. The definitions of open source are a historical accident that mostly traces back to "well GPLv2 already exists".
If progress continues, someday it'll be possible to generate the source code for any binary and make a native port to any other platform. Some companies might be upset, but it'll be a huge boon for game and software preservation.
That's definitely a possible future abstraction and one are about the future of technology I'm excited about.
First we get to tackle all of the small ideas and side projects we haven't had time to prioritize.
Then, we start taking ownership of all of the software systems that we interact with on a daily basis; hacking in modifications and reverse engineering protocols to suit our needs.
Finally our own interaction with software becomes entirely boutique: operating systems, firmware, user interfaces that we have directed ourselves to suit our individual tastes.
That runs into copyright issues. As someone who does a reasonable amount of decompilation, I wouldn’t ever use an LLM. It falls too close to mechanical transformation territory which is not protected, fair use.
Obviously others aren’t concerned or don’t live in jurisdictions where that would be an issue.
Surely then people start using LLMs to obfuscate compiled source to the point that another LLM can’t deobfuscate it. I imagine it’s always easier to make something messy than clean. Something like a rule of thermodynamics or something :)
Though, that’s only for actively developer software. I can imagine a great future where all retro games are now source available.
But on the other hand, at the current speed of LLM progression, a game that might have been obfuscated with the help of Opus 4.5 might in two years be decompiled within hours by Opus 6.5.
Yes, I believe it will. What I predict will happen is that most commercial software will be hosted and provided through "trusted" platforms with limited access, making reverse engineering impossible.
I've used LLMs to help with decompilation since the original release of GPT-4. They're excellent at recognizing the purpose of functions and refactoring IDA or Ghidra pseudo-C into readable code.
I don't believe that was written in a compiled language, so any old 8086 disassembler should suffice. I would love to see what comments an LLM adds to the assembly code, though.
I’ve been having fun sending Claude down the old school MUD route, giving it access to a SMAUG derivative and once it’s mastered the play, give it admin powers to create new play experiences.
I stayed away from decompilation and reverse engineering, for legal reasons.
Claude is amazing. It can sometimes get stuck in a reason loop but will break away, reassess, and continue on until it finds its way.
Claude was murdered in a dark instance dungeon when it managed to defeat the dragon but ran out of lamp oil and torches to find its way out. Because of the light system it kept getting “You can’t seem to see anything in the darkness” and randomly walked into a skeleton lair.
Super fun to watch from an observer. Super terrifying that this will replace us at the office.
The article is a useful resource for setting up automated flows, and Claude is great at assembly. Codex less so, Gemini is also good at assembly. Gemini will happily hand roll x86_64 bytecode. Codex appears optimized for more "mainstream" dev tasks, and excels at that. If only Gemini had a great agent...
I...didn't know there was a Gemini CLI? I thought it was only antigravity, or hackily plug your API keys into something like Cursor...Thanks! I got to check.
Documentation is one place where humans should have input. If an LLM can generate documentation, why would I want you to generate it when I can do so myself (probably with a better, newer model)?
I definitely want documentation that a project expert has reviewed. I've found LLMs are fantastic at writing documentation about how something works, but they have a nasty tendency to take guesses at WHY - you'll get occasional sentences like "This improves the efficiency of the system".
I don't want invented rationales for changes, I want to know the actual reason a developer decided that the code should work that way.
Exactly. Often this information is not actually present in the code itself which is exactly why I would want documentation in the first place, given that I can always read the code myself if needed.
That's great if those humans are around to have that input.
Not so much when you have a lot of code from 6 years ago, built around an obscure SDK, and you have to figure out how it works, and the documentation is both incredibly sparse and in Chinese.
It doesn't need to be written by a human only, but I think generating it once and distributing it with source code is more efficient. Developers can correct errors in the generated documentation, which then can be used by humans and LLMs.
Maybe documentation meant for other llms to ingest. Their documentation is like their code, it might work, but I don't want to have to be the one to read it.
Although of course if you don't vibe document but instead just use them as a tool, with significant human input, then yes go ahead.
I've been experimenting with running Claude in headless mode + a continuous loop to decompile N64 functions and the results have been pretty incredible. (This is despite already using Claude in my decompilation workflow).
One thing I don't annoying in really old sources is that sometimes you can't go function by function, because the code will occasionally just use a random register to pass results. Passing the whole file works better at that point.
This sounds interesting! Do you have some good introduction to N64 decompiliation? Would you recommend using Claude right from the start or rather try to get to know the ins and outs of N64 decomp?
This is super cool! I would be curious to see how Gemini 3 fares… I've found it to be even more effective than Opus 4.5 at technical analysis (in another domain).
Wow, I haven't thought of this game since I played it as a kid. My friend would bring it over all the time for sleep overs. I'm going to try to emulate it right now for old time's sake. I loved this game.
There are quite a few comments here on code obfuscation.
The hardest form of code obfuscation is called homomorphic computing, which is code transformed to act on encrypted data isomorphically to regular code on regular data. The homomorphic code is hard obfuscated by this transformation.
Now create a homomorphic virtual machine, that operates on encrypted code over encrypted data. Very hard to understand.
Now add data encryption/decryption algorithms, both homomorphically encrypted to be run by the virtual machine, to prepare and recover inputs, outputs or effects of any data or event information, for the homomorphic application code. Now that all data within the system is encrypted by means which are hard obfuscated, running on code which is hard obfuscated, the entire system becomes hard^2 (not a formal measure) opaque.
This isn't realistic in practice. Homomorphic implementations of even simple functions are extremely inefficient for the time being. But it is possible, and improvements in efficiency have not been exhausted.
Equivalent but different implementations of homomorphic code can obviously be made. However, given the only credible explanations for design decisions of the new code are, to exactly match the original code, this precludes any "clean room" defenses.
--
Implementing software with neural network models wouldn't stop replication, but would decompile as source that was clearly not developed independent from the original implementation.
Even distilling (training a new model on the "decompiled" model) would be dead giveaway that it was derived directly from the source, not a clean room implementation.
--
I have wondered, if quantum computing wouldn't enable an efficient version of homomorphic computing over classical data.
Im a encryption noob. Less than a noob. But something I've been wondering about is how can homomorphic computing be opaque/unencryptable?
If you are able to monitor what happens to encrypted data being processed by an LLM, could you not match that with the same patterns created by unencrypted data?
Real simple example, let's say I have a program that sums numbers. One sends the data to an LLM or w/e unencrypted, the other encrypted.
Wouldn't the same part of the LLM/compute machine "light up" so to speak?
I don’t actually think an LLM is a good way to sum numbers, but it is a pretty good example to explain the phenomenon you’re interested in. When you run an LLM, you essentially take your input and matrix multiply it with the weights to get an output. This happens regardless of what the input is or what the model “needs” to do. So, to some extent, the same part of the machine is being used every time even though the results seem very different.
(Of course, the reality is much more complicated than this; you can trace things like power and in theory track individual activations with great difficulty and then do interpretability to see what the model is doing. But hopefully it illustrates that you can usually take some sort of special operation and turn it into a process that does the same operation on different data.)
Rather than insisting on byte perfect matches, sometimes you can prove code equivalence of machine code sequences using SAT solvers. That might be an interesting extension, maybe giving clearer code output and/or solution to difficult functions in some cases.
I used Gemini to compare the minimized output of the Rollup vs Rolldown JavaScript bundlers to find locations where the latter was not yet at the same degree of optimization. It was astoundingly good and I'm not sure how I would have been able to accomplish the task without an LLM as an available tool.
Yeah, it works great for porting as well. I tried it on the assembler sources of Prince of Persia for Apple ii and went from nothing to basics being playable (with a few bugs but still) on modern Mac with SDL graphics within a day.
Yup. Still fighting some collision bugs, but it mostly works. I'll post it when it's complete. What I actually wanted to do is try to put fluid movement into it - something closer to Dead Cells, just for fun to see how it would change the feel of it.
I ran Node with --print-opt-code and had Opus look at Turbofan's output. It was able to add comments to the JIT'ed code and give suggestions on how to improve the JavaScript for better optimization.
Am I just wrong in thinking doing decompilation of copyrighted code via the cloud is a bad idea?
Like, if it ever leaks, or you were planning on releasing it, literally every step you took in your crime is uploaded to the cloud ready to send you to prison.
It's what's stopped me from using hosted LLMs for DMCA-legal RE. All it takes is for a prosecutor/attorney to spin a narrative based on uploaded evidence and your ass is in court.
It wouldn't fit most of the current LLM cloud providers narrative about privacy and copyright either, so, not sure they would be as cooperative with a prosecutor as they are today with lawmakers and right holders.
Have you ever tried to get a game developer to open source a game? And a Japanese one at that?
Even if they were willing to (they're not) and if they still have the code (they don't), it will contain proprietary code from Nintendo and you'll never get your hands on that (legally)
I also wasn't familiar with this terminology:
> You hand it a function; it tries to match it, and you move on.
In decompilation "matching" means you found a function block in the machine code, wrote some C, then confirmed that the C produces the exact same binary machine code once it is compiled.
The author's previous post explains this all in a bunch more detail: https://blog.chrislewis.au/using-coding-agents-to-decompile-...
reply