Comment by zoogeny | Hacker News Reader

Comment by zoogeny | Hacker News Reader

Parent Full thread

I'm not here to predict the future, rather to reconsider old assumptions based on new evidence.

Of course, LLMs may stay as "autocomplete" forever. Or for decades. But my intuition is telling me that in the next 2-3 years they are going to increase in capability, especially for coding, at a pace greater than the last 2 years. The evidence that I have (by actually using them) seems to point in that direction.

I'm perfectly capable of writing programs in Perl, Python, JavaScript, C++, PHP, Java. Each of those languages (and more actually) I have used professionally in the past. I am confident I could write a perfectly good app in Go, Rust, Elixir, C, Ruby, Swift, Scala, etc.

If you asked me 6 months ago "what would you choose to write a basic CRUD web app" I probably would have said TypeScript. What I am questioning now is: why? What would lead me to choose TypeScript? Do the reasons I would have chosen TypeScript continue to make sense today?

There are no genies here, only questioning of assumptions. And my new assumptions include the assumption that any coding I would do will involve a code assisting LLM. That opens up new possibilities for me. Given LLM assistance, why wouldn't I write my web app layer in Rust or Zig?

Your assumptions about the present and near future will guide your own decisions. If you don't share the same intuitions you will come to different conclusions.

> Given LLM assistance, why wouldn't I write my web app layer in Rust or Zig?

Same reasons as with no LLM assistance. You would be choosing higher maintenance burden and slower development speed compared to your competitors, though. They will get it out faster, they will have fewer issues, and will be able to find people to support it more easily. Your product may run faster, but theirs will work and be out faster.

Lets imagine we are assembly programmers. You have a particular style of assembly that you believe gives you some advantage over your competitors. The way you structure your assembly gives you a lower maintenance burden and faster development speed compared to your competitors.

I show up and say "I have a C compiler". Does it matter at that point how good your assembly is? All of a sudden I can generate 10x the amount of assembly that you generate. And you are probably aghast, what crappy assembly my C compiler generates.

Now ask yourself: how often do you look at generated assembly?

Compilers don't care about writing maintainable assembly. They are a tool to generate assembly in high volumes. History has shown that people who use C compilers were able to get products to market faster compared to people who wrote using assembly.

So lets assume, for the sake of understanding my position, that LLMs will be like the compiler. I give it some high-level English description of the code I want it to run and it generates a high volume of [programming language] as its output. My argument is, the programming language that it outputs is important and it would be better for it to output into a language that low level native binaries. In the same way I don't care about "maintainable assembly" coming out of a C compiler, I don't care about maintainable Python coming out of my LLM.

> In the same way I don't care about "maintainable assembly" coming out of a C compiler, I don't care about maintainable Python coming out of my LLM.

A well tested compiler is far more deterministic than an LLM, and can be largely treated as a black box because it won't randomly hallucinate output.

Humans aren't deterministic. I've trusted junior engineers to ship code. I fail to see a significant difference here in the long term.

We have engineering practices that guard against humans making mistakes that break builds or production environments. It isn't like we are going to discard those practices. In fact, we'll double down on them. I would subject an LLM to the level of strict validation that any human engineer would fine suffocating.

The reason we trust compilers as a black box is because we have created systems that allow us to do so. There is no reason I can see currently that we will be unable to do so for LLM output.

I might be wrong, time will tell. We're going to find out because some will try. And if it turns out to be as effective as C was compared to assembly then I want to be on that side of history as early as possible.

> Humans aren't deterministic.

Exactly, which is why I would want humans and LLMs to write maintainable code, so that I can review and maintain it, which brings us back to the original question of which programming languages are the easiest to maintain...

Well, we're in a loop then because my response was "you don't care about maintainable assembly".

I want maintainable systems you want maintainable code. We can just accept that difference. I believe maintainable systems can be achieved without focusing on code that humans find maintainable. In the future, I believe we will build systems on top of code primarily written by LLMs and the rubric of what constitutes good code will change accordingly.

edit: I would also add that your position is exactly the position of assembly programmers when C came around. They lamented the assembly the C compiler generated. "I want assembly I can read, understand and maintain" they demanded. They didn't get it.

We're stuck in a loop because you're flip flopping between two positions.

You started off by comparing LLM output to compiler output, which I pointed out is a false equivalence because LLMs aren't as deterministic as compilers.

Then you switched to comparing LLMs to humans, which I'm fine with, but then LLMs must be expected to produce maintainable code just like humans.

Now you're going back to the original premise that LLM output is comparable to compiler output, thus completing the loop.

There are more elements to a compiler than determinism. That is, determinism isn't their sole defining property. I can compare other properties of compilers to LLMs. No "flip flop" there IMO, but your judgment may vary.

Perhaps it is impossible for you to imagine that LLMs can share some properties with compilers and other properties with humans? And that this specific blend of properties makes them unique? And that uniqueness means we have to take a nuanced approach to understanding their impact on designing and building systems?

So lets lay it out. LLMs are like compilers in that they take high level instructions (in the form of English) and translate it into programming languages. Maybe "transpiler" would be a word you prefer? LLMs are like humans in that this translation of high level instructions to programming languages is non-deterministic and so it requires system level controls to handle this imprecision.

I do not detect any conflict in these two ideas but perhaps you see things differently.

> There are more elements to a compiler than determinism.

Yes, but determinism is the factor that allows me to treat compilers as a black box without verifying their output. LLMs do not share this specific property, which is why I have to verify their output, and easily verifiable software is what I call "maintainable".

An interesting question you might want to ask yourself, related to this idea: what would you do if your compiler wasn't deterministic?

Would you go back to writing assembly? Would you diligently work to make the compiler "more" deterministic. Would you engineer your systems around potential failures?

How do industries like the medical or aviation deal with imperfect humans? Are there lessons we can learn from those domains that may apply to writing code with non-deterministic LLMs?

I also just want to point out an irony here. I'm arguing in favor of languages like Go, Rust and Zig over the more traditional dynamic scripting languages like Python, PHP, Ruby and JavaScript. I almost can't believe I'm fighting the "unmaintainable" angle here. Do people really think a web server written in Go or Rust is unmaintainable? I'm defending my position as if they are, but come on. This is all a bit ridiculous.

> How do industries like the medical or aviation deal with imperfect humans?

We have a system in science for verifying shoddy human output, it's called peer review. And it's easier for your peers to review your code when it's maintainable. We're back in the loop.

That is one system. Are there zero others?

Funny thing about this thread and black and white thinking. I feel a different kind of loop.

> Do people really think a web server written in Go or Rust is unmaintainable?

Things are not black and white. It will be less maintainable relatively speaking, proper tool for the job and all that. That’s why you will be left in the dust.

Again, your competitor will get there faster and with fewer bugs. LLMs are trained on human input and humans do not do great at low level languages. They churn out better Python than C and especially when it comes to refactoring it (have observed that personally).

Reply on news.ycombinator.com