Ask HN: Are High Level Languages Necessary in AI Future?
The natural language interface with my AI tools is spreading quickly, and I'm starting to wonder about a future where I dont have to read the code.
If humans aren't read/writing the code, why does Claude need to use python? Code in managed, dynamically typed languages spends more tokens on doing the actual task and less on managing resources and satisfying the compiler. (Explicit type information might be useful for "reasoning" AI systems, but the job seems overall easier when it isn't actually forced to do the type calculus.) Plus Python reads more like a natural language overall, especially since it isn't using as many non-verbal tokens like closing braces to indicate the end of a scope. Plus there's a ton of example Python code out there specifically in contexts where people are talking about programming and explaining what the code is supposed to do. I don't think it's at all surprising that LLMs have learned (whatever that means) Python (and JavaScript) better than other programming languages. Great question, been asking myself that for a while as well (no sarcasm, I swear). After all, why not take our natural language instructions and output perfect machine code? Why mess with the HUMAN interface to programming a computer? Well, the only right answer can be that LLM's don't know the FIRST THING about computers. We humans don't program computers with natural language. We program them in PLs, mostly with Python and JavaScript. So those are what an LLM can mimic the best. This implies that LLMs will always be limited in what they can do when it comes to programming. Biggest argument in favour of certain languages is the wealth of documentation and guides and tutorials and stuff that train AI. I think there's something to be gained for constraining AI to produce human-readable code too as a security and quality control mechanism, even if it's only for other AI to monitor lol.