What's happening with Riemann is one of the most dramatic things I've seen in a while. I've been wrestling with my own version of this, of course much less important. I've had ideas how to make really fast stroke expansion for a while, and just in the last week used an LLM to iterate and polish. The results are good. I'd say it's about half my work and half LLM. The problem is, I don't fully understand the latter. I think it should be a blog post, but I'm still deciding on what exactly to do.
@raph So I cannot tell you what to do, but I guess I'd encourage you to sincerely ask the question of why you are doing anything you are doing, and query whether any one approach you take is still achieving that…
@mcc @mcc Thanks for your thoughts. Yes, I am using this as an opportunity to go deep.
The main thing I care about is quality. Of course, that has multiple facets. For graphics rendering, the most important facets are correctness and performance. The new work blows those out of the water. But an *understanding* of the techniques, so that others can build on it, is also an important facet. And in its current state it falls short. I could improve that by putting time into "digesting" the...
@mcc ...LLM output (in the sense of Terry Tao). But I'm not 100% sure I want to do that.
I deliberately put my work (code and writing) out with a permissive license. I am ok with others adapting it. Again my main lens is quality – if someone else builds on it and makes something better, all good. I also of course appreciate attribution. This is something that is utterly broken in the LLM world, and is one of many legitimate criticisms of AI companies and AI culture.
@raph I guess I'd say this: I'd really like to have a solid GPU-based line drawing library. But if I found a GPU based stroke expansion algorithm, and I saw it was developed with the help of LLMs, I'd ignore it and move on.
I'm not important, but I do think people follow a strong bimodal distribution: People who will reject partly-LLM creations; and people who will ignore anything you make and get their LLM to generate some stroke expansion code (possibly your own solution, uncredited).