While I expect that a language specifically for LLMs would not be a terrible idea, I think what is actually needed is a different interface for LLMs to use.
Currently Claude Code and Codex are primarily interfacing with code directly mostly using shell tools, while humans typically use IDEs.
Yes, I realize that there are lots of IDE-integrated LLM solutions, but this isn't exactly what I'm thinking about...
I'm thinking more of an API/MCP that the agent would need to use to work with the code which provides all the necessary interfaces, and designed specifically for AI agents.
I'm actually working on one... well, it's a lot more than a C++ to C transpiler, but it's just one of the features it contains. It's actually a C/C++ derived language which currently aims to fully support C23 and C++17 in addition to its own extensions which bring in features from other languages like PHP, Rust, Python, Perl, Ruby, JS, etc, as well as auto-including headers, and auto-resolving namespaces. It can JIT execute and also generate executables, as well as emit standard C code you can compile with GCC or CLANG.
So this makes me think that a "better" idea is let C be the actual underlying foundation for all of these languages because most languages are written in C (and C++) in the first place, so instead of forcing this strict divide, why not allow for every language to be transpilable into C so that the end result is that you could code in any language you like, because in the end, it can all be compiled with a standard c compiler? This could just be the new rule going forward... want to make a new programming language? Sounds good... so long as it can be transpiled into C code. Doing some weird shady stuff that generates machine code directly? No problem -- put it into a library that can be linked in.
Oh wow, interesting project... I've also been working on a "better c" for a few years, and while it was originally a JIT language (using jitasm), I recently switch to MIR, but since I already had my own lexer/parser/etc stack, I ended up changing it to basically lower everything into a node_t compatible AST tree and handing that off directly to c2mir, bypassing its parser.
Currently Claude Code and Codex are primarily interfacing with code directly mostly using shell tools, while humans typically use IDEs.
Yes, I realize that there are lots of IDE-integrated LLM solutions, but this isn't exactly what I'm thinking about...
I'm thinking more of an API/MCP that the agent would need to use to work with the code which provides all the necessary interfaces, and designed specifically for AI agents.
reply