Ask HN: What does the ideal programming language look like in the age of LLMs? | Hacker News Reader

Ask HN: What does the ideal programming language look like in the age of LLMs?

2 pointsxlbuttplug25 comments

What language - current or hypothetical - is optimal for LLM assisted/produced code?

I have a couple preliminary thoughts:

- Prioritize human readability: With LLMs progressively taking over more of the coding responsibility, and developers being relegated to more code reviewing, we might as well make life easier. It’s ok for syntax to be verbose since it comes at no cost (to LLMs, anyway). Information dense code is fun to write but not so much to read. Perhaps something close to pseudocode.

- Ability to self verify: Reduce the burden on the developer of validating that the produced code is correct and not riddled with hallucinations. Code that compiles is a good baseline. Then we have static analysis like what Rust does. Then comes formal verification and being able to prove that the logic meets the spec.

I’m also curious how well LLMs will be able to write code in such an ideal language, given that they haven’t been trained on such.

Have you asked an LLM? Copilot made the following recommendations:

Python - AI/ML development

Typescript - Web-based LLM apps

Rust - High-performance inference

Clojure - Functional orchestration

Julia - Scientific AI

It's a bit surprising to see Clojure in this list, but I would say everything else is the usual suspects.

Most popular languages were designed with human writability in mind. I’m wondering how languages will evolve if that becomes less of a factor going forward. Of course, for the time being, we need a middle ground to satisfy both humans and LLMs writing code.

Edit: tweaked the title slightly

I see human readability as being a key criterion moving forward since humans will be reviewing the code for the foreseeable future.

I also see functional and pure functional languages gaining in popularity because they're easier for LLMs to reason about, and they're easier to apply automated verification rules.

This is in the application space. It'll be interesting to see if any significant changes happen in the systems development space. Will LLMs drive the further adoption of Rust into the systems space?

Interesting Ruby isn't in the list, since I think it would be the closest to natural language. Echoing OP's thoughts about Human readbility being the most important.

But I might think the other direction, the most data-driven computer-readable code with embedded meta-data would be the best interface.

English > Mandarin > French

The LLMs will decide on python obviously.

Reply on news.ycombinator.com