GPT-RAY
An experiment to learn how to best interact with this amazing new tool.
Introduction
Although I have a solid 10 years programming experience, I have no prior knowledge of light transport and rendering software.
So I thought why not let ChatGPT help me write a raytracer.
As I am in no position to judge the correctness of the code (or at least I will pretend not to be, for the sake of the experiment), I will have to try to keep improving iteratively through further chat interaction.
01. Initial render
The basis of the code, up until the point it could output an image for the first time, was generated within one and a half hours! This is the initial render, personally I think this is an amazing result in such a small timeframe, although there are probably errors in the code.
02. Clamping
Adding clamping clearly improves the output image.
03. Adding a Plane
Adding a plane to the scene seems to create a proper mess. Time for some chat guided debugging...
04. Positioning
After correctly positioning the plane and sphere, and adding field of view, the output looks better, but it looks like a new kind of clipping has been introduced.
05. Multisampling
Adding multisampling to improve the output quality.
06. Aspect Ratio Fix
Fixed a bug with the aspect ratio calculation and lowered the ground plane a bit.
07. Improved multisampling
Improved the multisampling method by asking for a different method.
08. Improving the Shade function
Improving the shade function by trying to incorporate transparency, refraction, reflection, diffusion and specular light.
Also adding a recursion depth limit to prevent stackoverflow. It currently looks far from perfect, but it is a decent start.
09. Improving the shade functions
After having split up the Shade function into separate parts, I could optimize and tweak them individually.
10. Further improvements
I have rewritten the shader functions, but this required a lot of debugging. And some aspects still feel incorrect/off.
At this point I'm not sure whether it is any good to continue down this path; I feel like I need to learn more about raytracing myself before I can redirect the AI to find potential bugs.
But then I could have just written the code myself...
Interlude
Now that the overall structure of the code is clear to me, I decided that the next best thing to do would be to learn some more about the subject, and then try to work with GPT again to improve certain parts of the code.
So... I threw out all the code and implemented Ray Tracing in one weekend in Golang.
This functions in a somewhat similar way, but I quickly understood that some parts use completely different paradigms and that there is quite a lot to choose from when it comes down to different ways of rendering. This was something that was completely unclear from just interacting with GPT, and would have required more understanding on my part to be able to pose correct follow-up questions.
I seem to get an idea about the type and depth of knowledge currently contained by ChatGPT, and while it is certainly impressive, it requires quite some thinking to write decent prompts and debug the returned code and statements.
Onwards
To be continued...