Hacker News new | ask | show | jobs
by rafaelmn 42 days ago
If anything LLMs should use something higher level because it compresses the context and makes programming closer to natural language they are trained on.

Forcing LLMs to do a shitty job of what a compiler can do deterministically is not a good approach IMO.
1 comments
_heimdall 42 days ago
Low level was the wrong term for me to pick there. I was meaning more along the lines of "purpose built". I.e. I could see a language, potentially still an abstraction requiring a compiler, that isn't meant to be particularly meaningful or inspectable by humans. For LLMs your right, conciseness would be important and that would likely mean it would be compiled.
link
not_paid_by_yt 36 days ago
you're both under and over estimating this technology I think, we are further away from SWE becoming a niche job than you think, but also LLMs really are less impacted by programming language than you think, this is a very pre-2010s approach to NLP, it really matters quite little to a language model what language the code is in, you could have it write it in Python using exclusively Chinese logographs for names, and exclusively French transliterated into Mayan script for comments and it might barely perform worse than if you asked it to do something normal, but that also goes in the other direction, trying to design a language that's "friendly" to LLMs will likely not do much good either. The reality is there's a massive amount of data on existing languages and they are good enough for LLM usage.

I think if changes will happen in this direction it will be around formal verification, it's more difficult to trust LLM generated code if you aren't completely brainroted on press releases, and traditionally formal verification was seen as more effort than it's worth, but a tool that's great that generating insane amounts of text is quite well suited to formally verifying code, and assuming you can figure out how to make the specification readable to people, you just don't (in theory) need to ever look at the code or the proof, even if it might be helpful to anyways in practice. I'm quite excited the future of software might be fully unhackable software being the standard, with hacking becoming something that's talked about as a brief criminal fad in the beginning of the 21st century, like how you might hear about coin clipping from before fiat currency.

link