Syntax is a UI problem, but programming languages are more than syntax sugar for assembly. Features like Rust's borrow-checker, type systems, even small things like name resolution and async/await provide "semi-formal"* guarantees. And LLMs are no substitute for even semi-formal guarantees; people have spent lots of money and effort trying to integrate (non-NN, rule-based) algorithms and formal methods into LLMs, some even writing LLM programming languages, because they may be the key to fixing LLMs' accuracy problem.
* Not formally defined or verified, but thoroughly defined and implemented via reliable algorithms. A true formal specification is completely unambiguous, and a true formally-proven implementation is completely accurate to its specification. A thorough language specification or reference (like https://tc39.es/ecma262/ or https://doc.rust-lang.org/nightly/reference/) is almost unambiguous, and a well-tested compiler or interpreter is almost always accurate, especially compared to LLMs.
Cases where true formal guarantees matter exist but are rare. There are many, many cases where LLMs are too inaccurate but well-tested algorithms are good enough.
You can certainly think of them that way if you want. And, you'll be wrong, at least IMO, and in what seems to be the consensus of people who study and work on PLs.
"UI" issues, while not at all unimportant, are surface level. Semantics are fundamental. Get the semantics of your language wrong, and you are stuck. Indeed, this is what happens with most popular languages that aren't well-considered:
- lang author creates some random new lang, throws whatever feels like a good idea to them into the lang
- lang community discover that these features interact in rather surprising ways, ways that keep tripping up newcomers to the language, and would like to change it.
- lang author, having had X more years of experience now, identifies that changing the behavior could break user programs who rely on that behavior (wittingly or un-)
- lang author decides to write a new, backwards-incompatible version, community has 20 years of turmoil before recovering (python3), if it ever does (perl6).
Speaking as an academic PL researcher: you are right that this math-centered view is common in the PL research community, but that doesn't make the opposing view "wrong". There is also plenty of work in the intersection of PL and HCI. Both perspectives are crucial, and neither field is intrinsically superior to the other, despite what some individual researchers feel.
A human view of programming languages - Amy J. Ko
In computer science, we usually take a technical view of programming languages, defining them as precise, formal ways of specifying a computer behavior. This view shapes much of the research and development that we do on programming languages, determining the questions we ask about them, the improvements we make to them, and how we teach people to use them. But to many people (even software engineers) programming languages are not purely technical things, but socio-technical things. In this talk, I discuss several alternative views of programming languages, and how these views can reshape how we design, evolve, and use programming languages in research and practice.
The upshot is while the "PL as math" and "PL as tool" views have been well researched, if you're a researcher getting started in this area there is plenty of fertile ground out there. We might just have this view that "PL is math" because we've done the most work there, but there are many more applications of programming languages we have not thoroughly explored enough to say they are not just as important.
Yes, “wrong” was a bit of a strong word, but do note that I intentionally framed it as an opinion and not as, like, objectively wrong.
UI is still important, but I do not think it really defines a language.
Let’s say I created a new Haskell compiler, and it had the best error messages in the world, a hiccup-free LSP experience, maybe an alternative syntax format that’s more familiar to mainstream devs. But it still has all the same semantics. Is that still Haskell? Why or why not?
Now let’s say I took whatever has the best tooling currently, idk what it would be, maybe TypeScript, but gave it say non-strict semantics. Is that still typescript?
If you consider the programming experience/UI to be the language along with the tools that support that language, then you don't need to concern yourself if using Haskell in vim is the same thing as using Haskell in an IDE that supports it really well, they aren't the same thing.
If you take it to an extreme, like with Smalltalk, you wind up with languages whose environment is considered essential, they are basically part of the language at that point.
To most people it is about IDE and development tool chains.
Or we can think of it from the perspective of mapping programmer’s mental model to the problems they want to solve. Interface to the Mind, so to speak.
I don’t disagree, but if you set it up as “ui vs math”, you signify certain things. And based upon your comment it did not say “they mean how the programmers mental model maps to the problem”. In fact I’d say that’s more math like than not
I have been intentionally coding with Claude in the past few months and I started to think about programming from problem solving standpoint other than generating artifacts that conform to whatever languages, libraries and frameworks I happen to use to solve the problem at the time.
We can call this mental model or not. But it seems to be a different abstraction than we are talking about at PL level.
To me, LLM is part of the interface between me and the solution, that is what I mean by UI.
The real work is still in my head. PL language is just a medium.
> What might a programming language designed specifically as a UI for coding agents look like?
A bad idea, probably. LLM output needs to be reviewed very carefully; optimizing the language away from human review would probably make the process more expensive. Also, where would the training data in such a language come from?
So then a programming language designed explicitly for coding languages would need to take human reviews into account, what is the most efficient and concise ways to express programming concepts then?
In the end, we circle back to lisps, once you're used to it, it's as easy for humans to parse as it is for machines to parse it. Shame LLMs struggle with special characters.
Surely lisps don't have drastically more special characters as other languages? A few more parens, sure, but less curly braces, commas, semicolons, etc
Also feels like making sure the tokeniser has distinct tokens for left/right parens would be all that is required to make LLMs work with them
Don't get me wrong, they do work with lisps already, had plenty of success having various LLMs creating and managing Clojure code, so we aren't that far off.
But I'm having way more "unbalanced parenthesis" errors than with Algol-like languages. Not sure if it's because of lack of training data, post-training or just needing special tokens in the tokenizer, but there is a notable difference today.
Yeah, makes sense it sounds like that. But the crux is probably that most of us learned programming via Algol-like languages, like C or PHP, and only after decades of programming did we start looking into lisps.
But don't take my word for it, ask the programmers around you for the ones who've been looking into lisps and ask them how they feel about it.
I don’t think that would be a real issue in practice. Coding LLMs need to be able to cope with complicated expressions in many languages. If they can produce legitimate code for other languages, they can be taught to cope with s-expressions.
Until such a point where have agents not trained on human language or programming languages, I think something that’s also really good for people as well.
- one obvious way to do things
- locality of reasoning, no spooky action at a distance
- tests as a first class feature of the language
- a quality standard library / reliable dependency ecosystem
- can compile, type check, lint, run tests, in a single command
- can reliably mock everything, either with effects or something else, such that again we maintain locality of reasoning
The old saying that a complex system that works is made up of simple systems that work applies.
A language where you can get the small systems working, tested, and then built upon.
Because all of these things with towards minimising the context needed to iterate, and reducing the feedback loop of iteration.
You're contradicting yourself. Raw, fully optimized executables for production would mean machine code for the target platform, not an intermediate bytecode that still requires a VM.
Not really, it would be one of the steps in the chain between design and implementation for a specific hardware platform. That is, unless that code is only ever to run on a single hardware platform, a rare occurrence outside of embedded applications.
The same reason an actual AI wouldn't play chess by brute forcing every possible position. Intelligent systems are about reasoning, not simply computing, and that requires operating at the level of abstraction where your intelligence is most effective.
