I'd be a bit suspect of an LLM getting an emulator right, when all it has to go on is docs and no ability to test (since pass criteria is "behaves same as something you don't have access to")... Did you check to see the degree to which it may have been copying other NES emulators ?
Highly complex, fairly novel.
Emulators themselves, for any chipset or system, have a very learnable structure: there are some modules, each having their own registers and ways of moving data between those registers, and perhaps ways to send interrupts between those modules. That's oversimplifying a bit, but if you've built an emulator once, you generally won't be blindsided when it comes to building another one. The bulk of the work lies in dissecting the hardware, which has already been done for the NES, and more open architectures typically have their entire pinouts and processes available online. All that to say - I don't think Claude would have difficulty implementing most emulators - it's good enough at programming and parsing assembly that as long as the underlying microprocessor architecture is known, it can implement it.
As far as other NES emulators goes, this project does many things in non-standard ways, for instance I use per-pixel rendering whereas many emulators use scanline rendering. I use an AudioWorklet with various mixing effects for audio, whereas other emulators use something much simpler or don't even bother fully implementing the APU. I can comfortably say there's no NES emulator out there written the way this one is written.
> I'd be a bit suspect of an LLM getting an emulator right, when all it has to go on is docs and no ability to test (since pass criteria is "behaves same as something you don't have access to")... Did you check to see the degree to which it may have been copying other NES emulators ?
Purely javascript-based NES emulators are few in number, and those that implement all aspects of the system even fewer, so I can comfortably say it doesn't copy any of the ones I've seen. I would be surprised if it did, since I came up with most of the abstractions myself and guided Claude heavily. While Claude can't get docs on it's own, I can. I put all the relevant documentation in the context window myself, along with the test rom output and source code. I'm still commanding the LLM myself, it's not like I told Claude to build an emulator and left it alone for 3 days.
Even with your own expert guidance, it does seem impressive that Claude was able complete a project like this without getting bogged down in the complexity.
rust + wasm simulation of organisms in an ecosystem, with evolving neural networks and genes. super fun to build and watch.
>which AI you are using?
using chatgpt/claude/gemini with a custom tool i built similar to aider / claude code, except it's very interactive, like chatting with the AI as it suggests changes that I approve/decline.
>No sign so far of AI's usefulness slowing down as the complexity increases?
The AI is not perfect, there are some cases where it is unable so solve a challenging issue and i must help it solve the issue. this usually happens for big sweeping changes that touch all over the codebase. It introduces bugs, but it can also debug them easily, especially with the increased compile-time checking in rust. runtime bugs are harder, because i have to tell the ai the behavior i observe. iterating on UI design is clumsy and it's often faster for me to just iterate by making changes myself instead.
Given that you've built your own coding tool, I assume this is as much about testing what AI can do as it is about the project itself? Is it a clear win as far as productivity goes?
As far as productivity, it's hard for me to quantify, but most of these projects would not be feasible for me to pursue with my limited free time without the force multiplier of AI.
Tech stack is nothing fancy/rare but not the usual ReactJS slop either - it's C# with OpenGL.
I can't comment about the best practices though because my codebase follows none of them.
Yes, the user has to know enough to guide the AI when it's failing. So it can't exactly replace the programmer as it is now.
It really can't do niche stuff however - like SIMD. Maybe it would be better if I compiled a cheatsheet of .NET SIMD snippets and howtos because this stuff isn't really on the internet in a coherent form at all. So it's highly unlikely that it was trained on that.
I vibe code the vast majority features nowadays. I generally don't need to write a single line of code. It often makes some mistakes but the agent figures out that the tests fail, or it doesn't build, fixes it, and basically "one shots" it after it doing its thing.
Only occasionally I need to write a few lines of code or give it a hint when it gets stuck. But 99% of the code is written by cursor.
Specifically for the front end I mostly vibe code, and for the backend I review a lot of the code.
I will often follow up with prompts asking it to extract something to a function, or to not hardcode something.
It seems to me that where we are today, AI is only useful for coding for very localized tasks, and even there mostly where it's something commonplace and where the user knows enough to guide the AI when it's failing. I'm not at all convinced it's going to get much better until we have models that can actually learn (vs pre-trained) and are motivated to do so.