An idea to defeat
#GenerativeAI in
#FreeSoftware:
Just use a #ProgrammingLanguage that isn’t popular (e.g. #Haskell or some #Lisp dialect) to write your code, but publish human-readable intermediate form of that code in the public code repositories (e.g. the C programming language). Share the actual source code privately with trusted contributors in non-public branches, and require GPG signatures on actual contributions.
You could argue that not sharing source code is against the GPL, but the GPL does allow you to share the code as a hard copy printed on paper and sent over snail mail. Or you can just wait until the person asking is an actual human that you can trust not to use the source code for LLM training.
LLMs are unable to learn unpopular programming languages because they don’t have a sufficient corpus of training data to learn how to code it, so if your receive a contribution in C, thank the contributor but inform them that they will have to rewrite the contribution in your Lisp dialect before you can accept it.
#Scheme dialects like #Gambit , #Chicken , and #Bigloo would work well for this. So would a #CommonLisp implementation that translates to C such as #ECL . Although keep in mind that the idea is to use a less popular language, so you may have to further obscure these languages a little bit, but not in a way that would be difficult for humans. For example, using a macro system, you could use df instead of define, rename types of things like string? to utf8str?, use generic functions with mulitple dispatch so append will work on strings, lists, vectors, and bytevectors. Small tweaks like this might throw-off an LLM asked to write source code in Lisp.
#tech #software #LLMs #LLM #FOSS #FLOSS #OpenSource #SchemeLang #R7RS