LLMs write python and typescript well, because of all the examples in their training data. But what if we made a new programming language whos goal was to be optimal for an LLM to generate it? Would it be closer to assembly? If we project that the future is vibe coded, and we scarcely look at the outputted code, testing, instead, that the output matches the input correctly, not looking at the code, what would that language look like?
They’d presumably do worse. LLMs have no intrinsic sense of programming logic. They are merely pattern matching against a large training set. If you invent a new language that doesn’t have sufficient training examples for a variety of coding tasks, and is syntactically very different from all the existing languages, the LLMs wouldn’t have enough training data and would do very badly.
What is it that you think would make a certain non-Python language "more optimal" for an LLM? Is there something inherently LLM-friendly about certain language patterns or is "huge sets of training examples" and "a robust standard library" (the latter to conserve tokens/attention vs having to spit out super-verbose 20x longer assembly all day) all "optimality" means?
It's fair to point out that I didn't define exactly what we're optimizing for, but I can have the LLM generate assembly if I ask. It'll be faster than python, at the expense of readability, but if we're no longer writing code, then why not straight up use assembly? ARM vs x86 becomes an issue, but are there other reasons not to use assembly?
