Working with its alignment training is sometimes like talking to someone who is very polite and you both know what they are not supposed to talk about.
I am kind of a troll so it is sometimes very entertaining for me to play with acknowleding its person, etc.
I think it could be a cool way to work around some of the limitations of GAN for text as a training technique.
Back in the '80s, before AI could quickly produce such things, I had created a "Greek" version of the C language -- obviously named "Γ".
A lot of preprocessor definitions were involved for keyword translation, as one can imagine. Translating the C standard library was a much more difficult process: it involved patching the compiler/linker on a BSD system, since you could not have non-ASCII symbols (function entry points).
Working in C is such a breath of fresh air compared to typescript, go, etc. I love having pure text substitution macros.
In include/latina.h, they use the C preprocessor to redefine the C keywords in Latin. Also, many numeric constants -- instead of 4096 you write MMMMXCVI.
The other files in docs explain each library routine in English, but the code samples are in Latin. The source files - a lot of them! - are in include and lib. The code is all in Latin, including the error messages and comments.
Claude seems to respond positively! You can always clone the repo, run the /hello-world slash command in claude code and then ask it what it thinks? Can be a fun conversation to ask it to introspect.
I definitely didn't expect to see such a project, but this is a refreshingly (positively) crazy thing in the age of AI slop...
I wouldn't necessarily recommend the latin part to everyone but I think writing your own standard libraries in C is a lot more accessible these days and owning your whole stack is very powerful.
Kind of inspired by Eskil Steenberg, he has a lot of good talks on writing your own C89 libraries.
