On Monday, 6 October 2025 at 11:48:57 UTC, Dennis wrote:
> [snip]
Sometimes people ask Walter for tasks to work on. After hearing the tasks he gives, the enthusiasm quickly dies down. For example, apparently no one wants to write an 80 bit floating point emulator.
To be fair, not many people could write an 80bit floating point emulator. When people are complaining about stuff over and over, it's usually not low-hanging fruit that someone could figure out how to do in an hour. It comes back to the number of people in the community with both the skills to do these things and the motivation to do them. That's a hard problem when it's a voluntary project.
I know there was a big back and forth a few months ago about using AI in the D code base. Rather than re-litigate that, I'm curious if you are aware of Retrieval-augmented generation (RAG) [1], or know if any of the other major D developers are aware of it. I have heard of some people working on one for a less well-known probabilistic programming language that I follow (Stan).
This is basically a technique where you can take an off-the-shelf AI and provide it with some information that provides some additional context. The idea would be that off-the-shelf AIs may hallucinate more for D because it is a less well-known language. But if you feed in additional context, then you may get better results.
OpenAI has a pretty easy way to make Custom GPTs and add knowledge to it. So for instance, we could load in the D language specification and docs. We could load in the style guide. Maybe load in Ali's book. It might take a little coding, but there would probably be a way to load in the DMD code base itself. The limitation here is that you only load in 20 files up to 512MB and I don't think zip files work, so you would need to convert multiple files into one large text file. The end goal would be a custom GPT that is more specific to the D language.
So for instance, suppose we create a custom D GPT, then feed in the standard for 80bit floating point (maybe also an example of another type of emulator), then ask the custom GPT to write an emulator for it. See what happens. At least with a project like this we can create unit tests to compare the behavior of the hardware vs the emulator. That makes it a good test case. And if the concern is something being pure AI, then a human with more experience could take this as a starting point to clean up into a final version. I think it would be less work than starting from scratch.
And there might be other problems where this approach might provide a good starting point. Like how no one wants to be getting debuggers to work better.
I'm not saying that AI is a panacea, but it keeps getting better. I've had better success with ChatGPT5 than prior versions. It still hallucinates and gets stuff wrong, but the rate is lower than before. A custom GPT for D might help reduce hallucinations and incorrect responses further for D code. it could help kickstart some of these more difficult projects or otherwise reduce the burden on experienced D developers. Or otherwise, it could help people new to D come to the language, like how run.dlang.org can make it easy to get started with D (heck I bet you could combine the two, talk to the GPT in natural language to get it to write the code, then put it into run.dlang.org).
[1] https://en.wikipedia.org/wiki/Retrieval-augmented_generation
Permalink
Reply
