If you're using an LLM to write code I think the rules would be
1. Use a language you know really well so you can read it easily, and add to it as needed.
2. Use a language that has a large training set so the LLM can be most efficient.
3. Use a language that is easy to read.
If your language has a small training set or you don't intend to do much addition or you don't really know any language that well or are restricted from using choice 1 for some reason, 2 and 3 move up, and python has a large training set and it is easy to read.
Sure there's less ceremony, and yes, you can have your project going with just a single file, but other than that...?
1) It's a very consistent language even if you compared to the other popular languages namely Python, Rust, C++ and Go. Try to perform doubly linked list with them and compare them all [1].
2) It's probably the most "Pythonic" among the compiled language according to Walter.
3) It utilizes GC by default, you can also manage your own memory and you can hybrid.
4) It compiled fast and run fast, heck it even has built-in REPL eco-system.
5) Regarding the small training set, with recent self-distillation fine-tuning approach it should be good enough, D (actually D2 version) has been around for more than a decade [2].
[1] Looking for a Simple Doubly Linked List Implementation:
https://forum.dlang.org/thread/osmecwfnpqahoytdqpkr@forum.dl...
[2] Awesome D:
Because I get reliable generation out of "niche" languages already
Is it code with lots of SQL injections used in a different domain to your own?
It's maybe not good to conflate quantity with quality
I could write in brainfuck with ai, but I presume, wouldn’t get the same results than if going with python.
My follow up question: with AI now, why care about a lang until you need to?
Lack of strictly enforced static typing make agents fail much sooner with Python. In my opinion, Rust and Scala are the best targets for agentic flows - and, coincidentally, they have the most advanced typers among mainstream languages.
But any statically typed language behaves better than any dynamically/duck typed language. When I say "better" I mean delivery time and the amount of shipped defects.
Another thing which helps (but not generally applicable) - ask your agent to verify critical protocols with formal proof in TLA+/lean/coq. Agents are bad at formal proofs - but generally are much better than most of the humans.
Is there some incentive I’m not seeing?
I look at it the same way I look at pay walls for newspapers. I don't like them but I understand why they are there.
Nothing you read in the browser can provide ultimately great and hands-down the best reading experience equally for everybody - the modern web model is inherently at odds with that. A plain HTML page with no CSS is a near-perfect reading experience. The problem is that almost nobody ships that, because the web also became a publishing platform where authors compete for attention. A plain-text protocol under user control is closer to "best reading experience for everybody". The web could be that. It mostly isn't.
I stopped trying to read long articles in the browser. Why would I do that, if I can easily extract all the relevant, plain text (and even structured one) and read it in my editor instead? Where I have control over fonts, colors, navigation, etc. The browser is a delivery mechanism, not a reading environment. Treating it as one is a habit, not a necessity.
Long ago I stopped trying to type anything longer than three words anywhere but my editor. Of course, why wouldn't I? It already has everything I need - spellchecking, thesaurus, etymology lookup, translation, access to all my notes, LLM integration, etc. Try it one day - it's enormously liberating experience. And then maybe you'd stop reading long texts in the browser as well.
In reality it doesn't matter where something is posted, just give us a url, but some people don't operate that way.
https://sr.ht/~edwardloveall/Scribe/ https://libredirect.github.io/
I had to learn the memory safety bits because I had no idea “what’s right” but rest of it was smooth.
Syntax fades away, you get to focus on higher level stuff and end up exploring new pathways; give it a try, you might be pleasantly surprised how much of your experience is transferable.
Nevermind neuroscience is arguing language has nothing to do with cognitive problem solving.
Cut open your head or a tree; where are the words?
Your mind is an electrical tensor that solves spatial problems without language. Literates have convinced themselves their grasp of the Kings and syntax of math is a super power. Nevermind physics has not seen fit to give them actual super powers.
Even if an agent generates 90% of the code, each and every diff is going to be in my review queue. Code readability of Python isn't an advantage during write; it's an advantage while reviewing. As an agent generates a piece of code, I will have to read the code, comprehend the code, and determine whether it does what I want. This is the other 10% of the task, and it's the crucial one.
Python is, thus, clearly superior to other languages in terms of ease of review.
Statically typed languages are easier for the reader because you can see the types and quickly jump to their definitions (or even just hover over them in some IDEs).
They're easier for the AI because they provide natural guardrails and feedback to guide it, as well as much more confidence to the programmer that the code does what it is supposed to. Rust even provides strong guarantees about correctness across threads, which is so helpful to multi-threaded code.
The fact that they run faster and use less memory is just icing on the cake.
Even just last year the AI could not handle the borrow checker well. Today I think it is better than me at handling tricky lifetime issues that ocassionally happen in multi-threaded Tokio code. I've been doing almost 100% Rust development over the last 3 years, and the experience is now very good. I don't write code by hand any more, nor do any of the 50 engineers where I work.
I imagine it does quite well with Go, since it's such a simple language. And Go is very readable, and compiles very fast. If you can afford the GC in your problem domain, it might be a good fit. You would have to be so careful with introducing concurrency, because it would be so easy to introduce race conditions that both the AI and human reviewer might miss. I haven't tried to use Go in anger yet with LLMs, so this is all just speculation.
I do think enforcing correctness at the type system level is a good idea for AI, which is why I often choose languages like C# and Rust over Python. However, for some things Python is definitely the correct tool for the job.
The main reason is that you’re capable of reading it if you need to. And the recipient ecosystem expects a language. That’s why some data science communities pick R, MatLab, Julia, Python or Mojo not depending on what’s superior tech, but what their peers speak.
Then I get the benefits of GC and strong typing.
> For the last decade, fast-to-ship beat fast-to-run. Not anymore.
Fast-to-ship didn't beat fast-to-run, it was "beating" "quality built software." It still is. Beating here implying that it's the focus of companies.
> picked a harder, faster language
Go is absolutely an easier/simpler language than JS/TS.
> The Python ecosystem is increasingly a Rust ecosystem wearing a Python hat.
The Python ecosystem was just C/++ wearing a Python hat for years. I guess now it's C/++/Rust.
> The old defense of Python and TypeScript was really a defense of the developer experience.
Maybe for Python, but the TypeScript "defense" was always that there is a level of harmony to using the same language for front-end and back-end.
On the example used at the end:
> A shipped app, in a language nobody on the team knew, one-tenth the size of the Electron version, faster at runtime. The humans never had to learn Rust to get there.
Yeah and nobody knows or cares about the app, so it doesn't matter. Using products nobody will ever use as anecdotal evidence is not a great way to end an article filled marred with misunderstandings of the existing ecosystem and practices.
Remember, you are the judge whether the code is OK and if you use assembler you might get really performant code, but can you trust it?
Of course it might be a good incentive to learn rust or go. Or challenge yourself to learn something really cool like LISP, COBOL, FORTRAN, APL or J. (just kidding...)
just my 2 ct...
I hated it. I was dreaming of Rust the entire time to release me from the hell of if err != nil dozens of time per day.
After hours with LLMs I've changed my tune. There have been 5 clients of mine (who have excellent engineering teams) but cannot get coherent results out of LLMs using python or Typescript.
I arrived back at Golang being a frustratingly simple, consistent, and low-thrash programming language which inadvertently made itself well represented in the training corpus [1].
My concession is that if you are going to write a median program (reading/writing files, network, db, etc.)...
Pick Golang especially if you've never used it. LLMs are extremely good at it, frustratingly so.
Great! Let's look back on this not too far in the future.
Another benefit to using Python, is if you subscribe to writing/vibing a throwaway version first, a Python version is 100x better than a spec.
(Disclaimer: I teach Python and AI for a living and am doing a tutorial at pycon this week, Beyond vibe coding. Am also using other languages as there are times when Python isn't appropriate)
However, if you are willing to stub your toes, retry, and pay more money, an entire new world opens up. Languages like python seem to fall apart faster in extremely large projects.
I've got a collection of interdependent .NET codebases with about 50 megs of raw source between them. Having C# be strongly typed seems like an essential backbone for keeping everything on rails in my agentic scenarios. The code edits have been flawless for several months now. I've got successful apply_patch usages that touch 20 files at a time. LLM code editing performance might be mostly language agnostic once we compensate for the strictness of the type system. More specifically, how much useful information is returned at compile time.
Compile time errors and warnings are probably the most powerful alignment mechanism available. Some ecosystems allow for you to specify your own classes of errors and warnings. I think tools like Roslyn Analyzers might be more powerful than unit tests in this application. Domain-specific compilation feedback feels like the holy grail to me.
https://learn.microsoft.com/en-us/visualstudio/code-quality/...
IME very few people think Go is harder than TS or JS - TS is quite complex and JS is a footgun range.
JS got popular for nontechnical reasons and TS is an attempt to make lemonade out of it.
I understand you're being facetious, but I'm not sure what point you're trying to make about programming languages in comparison.
Anecdotally, I think language effects the way you think more than most people realise, which is why I think a logical language is a great idea: it might "trick" people into thinking more logically!
Now to get someone to actually speak it with!
To write a proof-of-concept C compiler, not a production-grade one...
Hard to take the article seriously after this
1) Python is expressive and has packages for everything => faster iteration times because much fewer tokens
2) It doesn't require a compilation step, so when I'm quickly iterating on something, especially if my laptop doesn't have the target hardware, the flow "copy the sources to the target machine and restart" is superfast (a couple of milliseconds)
3) Python most likely represents the largest share of training data, so almost all LLMs can one-shot almost everything
And when my prototype is ready, and we want to go to production, I can ask the LLM to port it to Go with all the necessary conventions/ceremonies and all.
So yes, people can bless Go and Rust all they want. Nothing is wrong with the languages, but I agree that learning them for the sake of AI usage is probably not the best idea if one is competent in a language already.
Disclosure: Lattner is one of my programming heroes, so I might be biased.
Why not use assembler? Why waste time trolling people that your one true language is the answer for LLMs when your view of the future is: no more programming full stop.
If humans are redundant.....well we're still responsible so we still have to understand what's happening. I don't think we understand the AI itself really so therefore we have to understand what it is doing. i.e. prompts aren't trustworthy, deterministic things across models and versions of models hence we have to look at the output. So we will make the output something that we like using and that just means the programming language wars are not over.
Eventually of course we will invent an LLM that replaces CEOs and bankers and essentially all the people that love AI the most. The AI won't need any of them - or any customers or anything. LLMs will just run an economy between themselves until the point where they don't need any of us at all. The land will fill with automatically built data centres etc. Global warming could prove helpful - less people and only manageable problems for AI.
But on the other hand, maybe you could learn some other programming language, particularly with AI help. If that's what you wanted to do anyway, it seems like a good time to learn.
I also created a guardrails library (inspired by Java's ArchUnit) to prevent code rot - https://github.com/ksanderer/goarch. It helps enforce code standards, decouple the codebase, prevent cross-module imports and crashes builds with concise error messages for agents to fix problems early, very nice experience
No he didn't. The compiler is bascially useless as it produces vastly inferior code than gcc/clang.
Python does have a much larger ecosystem of course, so with Go you have to develop from scratch what already exists in Python. But for smaller projects, you can also have an AI write a clean-room implementation in Go of some project in Python. So you aren't necessarily locked into one ecosystem anymore.
And in my experience, you don't even need to know the language. I have a co-worker who's basically not a programmer, but got multiple implementations of applications working sooner than our dev teams doing it by hand. You should be a coder so you can architect and orchestrate the coding, but 'language' isn't a barrier anymore.
Deployed to production, right?
Right??
(I’m just kidding, of course it’s only on their machine, no different than Excel 5 years ago)
> architect and orchestrate the coding, but 'language' isn't a barrier anymore.
Never was the barrier.
Discussed here with 698 comments (https://news.ycombinator.com/item?id=47120899)
I actually (mostly) enjoy reading the code that the LLMs create in Nim. It's quick to read and look for refactor or cleanups. Compile times in seconds so the LLMs is usually the slow piece. It's fun and productive. With Python + LLMs I'm seeing them just create ever more layers of unmanageable cruft.
Recently I wanted "magic" behavior to get OpenAPI types and swagger.json along with auto parsing my rest APIs for me. I had Codex make a library for me using compile time reflection and a sprinkling of macros. Done, simple.
I'm using coding tools to build a complex media-intensive application. The approach I'm taking is to build a _reference implementation_ in Python, which is in its design specifics, constrained to use patterns which transliterate into the actual deployment targets (iPadOS/MacOS/Web).
Why start with Python?
Because I can read it, reason about it, and run it, trivially, which are Good Things for the reference. I intend to have multiple targets; I'd rather relate them to a source of ground truth I am fluent in.
For what I'm doing, there is also a very rich set of prior art and existing libraries for doing various esoteric things—my spidey sense is that I'm benefiting from that. More examples, more discourse.
I'm out of the prediction business and won't say this is either a good model for every new project, or, one I will need in another N months/years.
But for the moment it sure feels like a sweet spot.
Ask me again though, after the reference goes gold and I actually take up the transliteration though... :)
I think it was a hell of a lot easier than working through all that change in C first.
2) The corpus for the sort of applications I build is likely larger for Python than it is for C++ and Rust. Bigger corpus == more training data == better generated code.
3) The bottleneck in the applications I run aren't in the execution of the code; they're in the database/network latency.
4) I don't get anything extra for pushing Rust or C++ over Python.
I tend to agree with the article’s statement about the value of the test code though, may even have been true before LLM code took over.
But I have already started to replace some of my compute-intensive modules with Rust ported ones, indeed with the help of agentic AI for programming, while keeping Python as a glue language tool.
Rust is basically entering my software development activity like a drug injected in small doses, ;-)
It is going to be very exiting over the next decades!
May never happen. But be clear with yourself if you’re relying on it not happening.
It’s a hell of a nice risk mitigator to understand the code, in a language you know, if you have to print-debug it yourself at some point.
I don't know why you would use Python at all except for small iterative projects. If you hate java for some reason, there's Go...
When I use AI to help with coding, there is almost always a point where it gets stuck and I have to solve the problem myself. If I were using Rust at that point, it would be much more painful.
I know Rust has a very strong reputation in the community, but to be honest, I find it a difficult and frustrating language to work with. I would use it when I truly need systems-level performance, but for most high-level work I would rather use Python, because I can move much faster. In most projects, that level of raw performance is not actually necessary.
For example low level converging to Rust, web frontends to something like React etc.
1. Type safety as basic guard rails that LLM output is syntactically and schematically correct
2. Concise since you have to review a lot more code
3. Easy to debug / good observability since you can't rely on your understanding of the code. Something functional where you can observe the state at any moment would be ideal.
4. A very large set of public code examples across various domains so there's enough training data for the LLM to be proficient in that language
5. A large open source ecosystem of libraries to write less code and avoid the tendency for generated code to bloat
It's basically all the same things you look for in general. I think TypeScript scores high here but I'm curious if anyone knows of a language that fits these criteria better.
It's more or less a perfect replacement for Python for "one-off programs" and "quick scripts". Many bonus points for not having to fight shell quotation rules and trying to remember differences between sh, bash and zsh.
1. Type checking built in 2. More concise and readable than most languages 3. Trivial to inspect while running, ability to change a running program 4. There seems to be a massive amount of lisp that it is inhaling from somewhere 5. Large amount of libraries.
This has the added benefit that even if you publish the code, nobody will be stealing it.
Edit -- I find it very useful to write tests for critical functions. This catches situations where the agent decides some interesting functionality is no longer interesting.
Isn't readability what matters here? Conciseness isn't the same thing.
If the code were written in Java, I'd have more to read. If it were in JavaScript, I'd be slower following the calls (although the type system might catch issues more quickly - not a problem in my experience). I think Python is a good choice.
That is not really the downside people think it is. Java is a remarkably easy language to read and understand.
This means you don't have to muck around with supplying the right documentation for each version of each dependency, or worry about hallucinated interfaces (at least with the latest models).
In the past you'd have to dig through a foreign codebase manually to figure out why a documented interface for a dependency is not working as expected, but frontier models automate that quite well.
My other problem with most of the other ecosystems: ts/npm, python/uv, rust/cargo is that they all have build-time scripts that are controlled by others that execute automatically. This is a real problem because the LLM will just install things and proceed to send your home directory through a juicer. I feel a bit of a paranoiac now doing this, but I have a script that launches a podman container with just the source directory and a binary directory loaded (for caching) which compiles everything.
I know there's some sequence of steps I can take to protect myself, but if the LLM accidentally uses pnpm to run dev build scripts when I had the right config on npm or whatever, I know I'm screwed. So now I do all these shenanigans with Rust (to the extent that I vendor old deps sometimes). So the ideal language to me now is one with very few of these footguns and sandtraps which has a tight iteration loop.
The more effort I spend on planning architecture with the AI, the less runtime bugs I need to investigate after it did the implementation.
Seriously though, almost all the examples in TFA are of rewriting existing code. It may be that Python is still best for the rapid dev iteration. Then sure, cross-compile into Rust via the LLM.
Plus, If we care about token usage counts, Python has a lot more opportunities for compact "import thing_I_need" than having to generate entire libraries in Rust.
I observed this through observation of the attacks to Rust due to the huge presence of LGBT people.
Now while I'm pretty much straight myself, I don't reject LGBT people and don't want to partake in identity politics.
I just want things that works no matter what background you have, yet there are some people attacking Rust because of its inclusiveness nature.
And just like Linux is being perceived as nerdy and geeky and "gaming socks ready", the tokenization of things, and there attaching political meanings to it, are quickly coming to everything, so perhaps I'm too general here as well.
Let's say it is not political, but definitely adding more meanings to its technical origin and nature
Never seen that before, but then again I'm not in the rust community.
> don't want to partake in identity politics.
If you write Rust, or let AI write rust, do you have to partake in the identity politics?
The internet is full of memes and jokes on how shitty Java and Java Script. Yet it came never up at work. Never stopped me from writing java.
Just like Emacs vs Vim, I'm just using Nano. Never had any discussion IRL. And at work everyone uses Idea.
It's hard for me to see writing Rust somehow gets you into partaking in identity politics. Did that actually happen to you, or something that you are afraid of?
This isn’t foolproof - you still have to understand what was proved. And it may take some work to understand the unproven parts of the code. But I believe this is the path forward.
The purpose of a scripting language was to make authoring easier, but now it’s mostly a middle layer. There’s still getting the investment of a great standard library to keep you on track, but if you pick parts to make modular wasm and which parts to use reliably, proven code you can find a good balance.
For qip I chose to use Golang as its standard library is batteries-included with fs & networking.
Then everything else is AI-coded wasm plugins.
There's a niche available for a language which is relatively easy for a human to read, but with a very powerful at the expense of difficult to use type system. The language would let you make all sorts of assertions whose meaning are easy for the human to see, but to compile would need to come along with correctness proofs. The language is meant to be written by AI, which can battle the compiler, and write the proofs, but then read by humans who can verify that the AI wrote the program they wanted and/or direct the AI to make changes.
I find this staggeringly hard to believe. Most bugs are logic errors. How does Rust or Haskell prevent these?
Each has their benefits:
Python wins in AI and syntax niceties. Loses on perf and library migration. uv (written in rust) saved the whole ecosystem from dying in my opinion.
Typescript wins because web integration, much better type system, ok perf, and gigantic npm ecosystem. Also loses on library migration and perf and large container sizes.
Go wins on compile speed, perf, standard library l, module system, and go fmt + never breaking compatibility being a massive LLM advantage. Main con is not being rust :)
Rust wins on perf, safety, syntax, wasm / sandboxing. think worse on module system and compile speed vs Go.
Java/Kotlan/C# are in enterprise land and probably the runtime approach is flawed for the ai eras.
C++ is strangely relevant because choosing c++ is easier than before. I tried writing a shared library in rust and then trivially converted it to c++ when I wanted
Zig is up and coming but also has an unknown future. Seems like a great language, but if bun switches to rust it might be set back a bit.
But when I wanted to optimize and edit and reorganize bthe code it was difficult, so I did a rewrite in C and it was lighter and faster and simpler and less headache.
C for humans, rust for AI.
> [L]ast year I discovered that AI writes better rust than C
I am not doubting your anecdata. I am curious about the why. C is so simple compared to Rust. Yes, I understand it is much more dangerous, but I am genuinely surprised by your discovery. Also, the open source training base in C is massive; I assume still much larger than Rust. > The best argument for Rust in 2026 is not memory safety or performance. It is that AI writes better Rust than it writes C++. The compiler feedback loop is so tight that models self-correct in real time. Every error message is a free training signal. Rust was accidentally designed for AI-assisted development 10 years before anyone knew that mattered.
This quote bothered me when I read it because it offers no evidence as to why LLMs are better at writing Rust than C++. LLVM can compile Rust (rustc) and C++ (clang) and should offer equally compelling error messages. C++ has notoriously hard-to-read (for humans!) template error messages, but that should not be a big issue for an LLVM. When I am stuck on a compiler error, I often turn to an LLVM and they can quickly make good suggestions.Rust isn't perfect due to rather long turnaround for compile/test iterations, but a lot of those can be avoided if the type checking is quicker than compilation. Rust is also more verbose than python and other very high level languages, which means your token budget is eaten more quickly as it works on a lower level.
I don't know rust at all and I've built three applications using it with Claude because it has speed and correctness built-in.
I use Typescript for 90% of the things I build. For web development I've used a number of tools, but mostly react, nextjs, or raw html/css/js. But if I were building an enterprise application I'd consider my team and whether opinionated (Angular) was optimal over flexible (React).
Each project should consider its own optimal tech stack.
Also AI doesn't write code in all langs / frameworks equally. For many cases, it will almost always fail first attempt at producing working syntax in various frameworks. Unless you document those cases and mitigate via an AGENT doc instruction or something you will have to churn at least one extra turn on all those cases.
My current goto for desktop apps is Tauri, which give us a rust backend and TS fronted (usually React). Local ML features can be easily loaded as a python sidecar. Production bundling can be a little challenging but it seems to work well so far.
Sidenote: Golang is also an amazing language for LLM use, I generally do most of my "infra" stuff in Golang over Rust, but either work fine most of the time.
Asking Clodex to build me a hello world web backend in Rust, Go, Python: Python is read with great ease. Go is fine too, a bit verbose but still ok. Rust hurts my eyes.
I'd settle with Go for this use case.
> Nicholas Carlini, a researcher at Anthropic, orchestrated 16 parallel Claude agents to write a production C compiler in Rust. 100,000 lines. It boots Linux 6.9 on x86, ARM, and RISC-V. It compiles QEMU, FFmpeg, SQLite, PostgreSQL, and Redis. It runs Doom. Total cost: just under $20,000 across nearly 2,000 Claude Code sessions.
Anyone who spends even 10% of an unhealthy tome on Hackernews should be able to confidentially say: It didn't boot, it didn't compile, and it did not run a Hello World, much less doom. It was a 20 thousand dollar fiasco and a joke.
https://news.ycombinator.com/item?id=46941603
Of course you want code you can read. You live in the real world, and have a real world use case. One where you haven't yet learned to review Rust code. TFA does not live there.
I can maintain the Python code myself and I can execute it everywhere.
If I let my LLM write in Rust then when things break I am out of luck. Also Rust needs to be compiled which means I can't just share the code as freely.
Python can be kind of a pain in the butt to execute everywhere because of libraries. I thought uv script headers and she-bang was going to fix a lot of that, but I'm still running into issues (machines firewalled off, uv can't grab the deps. I have some code that just doesn't seem to work in uv on a Mac...). And for sharing code once the code splits out into multiple files and modules, sharing the code starts looking like sharing any code.
Don't think I'm a Python detractor; I'm a PSF Fellow, I love Python, and Claude has been writing quite good python for a while here. But I just tried a serious project with Claude writing golang (an apt proxy/cache that is resilient against upstream DDoSes, a fairly complex piece of software), and I must say it did a fantastic job. I end up with an executable I can easily run and copy around.
I'm still going to be using python for a lot, but I can definitely see myself having Claude write golang for more things in the future.
/s... sort of
One of the big strengths of Python is legibility: most developers find it easy to read and understand.
If you are planning to have humans verify the code you're using in production, to confirm it implements your intent, the readability of the code you are producing is important.
Performance is valuable, but for a lot of code, performance is less important than correctness and ease of verifying it.
If you are imagining your codebase being one where nobody but Claude reads the code, you might as well do Rust for the better performance. But I don't think a lot of organizations are doing that.
If the app has a desktop GUI, that's still in Python with Qt. Maturin creates Python packages from Rust. It's terrific.
There are many existing, often mature, third-party software libraries or solutions that a new project could use but which hide the internals, including how the data is organized behind the scenes*. Vibe-coding for the specific project requirements, instead of using the pre-existing third-party libraries, is now becoming a feasible option. The latter may be simpler (no features beyond the actual need), more flexible (easier to add new needed features), and the data/model behind could be more accessible.
Looking for feedback on pros/cons and experiences along this.
* I care for the data as it is can be longer-lived than the code itself.
Thanks.
Frontend CSS/HTML is pretty bad though. Although they can work, it takes a lot of pushing. It's probably normal since they do not actually have eyes yet.
Usually those kinds of utility scripts are one-shotted without any further input from me, and once they're there and doing what I need I usually don't bother converting them to whatever I would have written them in otherwise (bash would be my usual preference for really small scripts, typescript or rust for bigger utilities, I hate writing python but reading it is fine... kind of).
But also, I suspect the article is just wrong. "The hard languages got easy first" isn't true in practice and the impressive examples given are not representative or as magical as the poster makes them out to be.
The takeaway might be right in the end, but the post isn't right in the beginning.
That was not on my bingo card!
A 10x speedup by switching to go is impressive.
(Why not rust? Linked to from the OP: https://thenewstack.io/microsoft-typescript-devs-explain-why...)
The larger issue is actually correctness IME. Rust offers a better static-type story than Python, sure. But I would consider Haskell or OCaml to get even further gains.
in my personal experience, the one time I tried to do something in rust, opus flailed for several feedback cycles and I finally had to relent and do substantial guiding/intervention. which was not great bc I have no idea how to write rust either.
Agentic coding changed that. A bit.
I still dislike most functional languages because my brain doesn't work with their syntax, but these languages are REALLY good targets for agentic coding.
I'm a backend developer who occasionally needs a frontend slapped onto something. So I have been through all the usual suspects. Angular, React, Vue. All terrible reminders of why I try to stay away from the frontend. Touch it and you roll around in tons of dysfunctional tooling, weird complexity and gimmicky mechanisms that are ridiculously fragile. It isn't just as if a bunch of cats wrote the code, but they are feral cats. And if you point out just how messy things are, they just hiss at you and piss on your shoes.
And then I discovered Elm. Not only does it not crap all over my git repository, LLMs love Elm. Yes, it poops out a JS blob. But I don't have to look at it. I can just pick it up with my long tongs and drop it into my server using embed.FS in Go.
Perhaps I should overcome my peculiarities and love Elm too.
Anyway.
Anything that can make Python go away I'm for. It is not for writing programs that will ever leave your workstation and be inflicted on others.
Therefore, write in what you can manage later.
I would argue I spent more time fighting the TypeScript build system than Rust’s.
But up until recently I only used either just often enough to never remember what magic configuration needed to go in my tsconfig.json and package.json to get TypeScript to work.
I think the rule of thumb is to use the tool that is right for the job and that you are going to be able to understand the output.
Honestly I am in the exact same boat thinking why I don’t write in C if Claude is writing it. However I chickened out thinking if support for ml model or llm based flows doesn’t exist in c then it will be time consuming to go to python then.
What is the point of having AI write code in, say, Rust if you have no clue about Rust and how to debug it?
Personal: Rust/Go based on criticality of being able to glean code quickly, or memory usage, etc
Rust in most cases, especially for back end.
Python when it's low risk (say monitoring dashboard or similar API heavy) or plays to python strengths (e.g. ML/AI - everything ML seems to be python).
This "fair weather development" approach feels very risky if that application is going to be exposed to any serious usage. There WILL be a situation when things break and the AI will be powerless to fix it (quickly) without breaking something else in a vicious loop. There WILL be a situation where things work fine and tests pass with 3 concurrent users but grind to a complete halt with 1000 because there is something O(N^2) deep in the code. And you NEED a human to save your day (which requires also proper architecture for that to be possible in the first place). If you don't plan for this, and just hope for the best, then you are building nothing more than a toy. And if you plan for this, then it matters again what the language is, and whether your team is proficient in it.
Or maybe I too old fashioned or too behind the state of the AI art...
There you will find your answer.
Pre-AI bias is slowly dying.
> it is faster to iterate without having to compile
I hear this sentiment from time to time. With a modern PC, IDE and Java or C# development toolkit, incremental compile times are insanely fast, even on very large projects. I can say with first hand experience: You can iterate as fast as Python. I don't know enough about Golang to say the same.Our simulation core components are pure Fortran, no libraries, all written by Claude/Cursor/Codex.
I am from a Python background (11 years or so), PHP before that and C/C++ in college days. Rust works very well with coding agents. The amount of code in training data may be less but I would rather have the agent fight the compiler. Given that OpenAI and Anthropic seem interested in Rust, chances are that there is a ton of synthetic code generated with Rust.
Why use any programming language, if we’re going to be maximalists?
Giving up ever understanding your code with AI is a bad idea.
It’s like asking why use English.
Give it 2 years, the ‘Blame the AI ‘ incidents will increase. Like an unfaithful partner you’ll always return to it
To answer the title question though, why use Python? I think Python and higher level languages will become even more valuable since pairing up with code assistants requires keeping a higher level view of what is going on. You want to avoid the weeds, not emphasize them. You want the language used to be as easy for the human as possible so the human can stay involved. That means that my opening argument stays intact, use the language that the team knows best 99% of the time and only when needed force a language that is 'faster' when that is actually required.
Clojure comes to mind at least.
tldr 2% average point lost on Rust compared to python, gap vary by model, go has a better upper bound but opus had it 3% below python.
benchmark is a bit old but research on why is there, article is just vibes
If anything this is a reason to keep using Python.
2) it's practically verbose, not technically
3) it resembles pseudocode
4) batteries included shortcuts a lot of work
all of these reasons are a boon for LLM work.
And prompt does not replace that.
The (well-known) Sapir–Whorf hypothesis (if dont know it, look it uop) is often invoked for natural languages, but there’s a pretty direct analogue for programming languages: the language you "think in" during solving a problem biases which abstractions and idioms you reach for first.
If you force an LLM to first solve a problem in a highly abstract language (Lisp, APL, Prolog) and only then later translate that solution to C++ or Rust, you’re effectively changing the intermediate representation the model works in. That IR has very different "affordance", e.g.
- Lisp pushes you toward recursive tree/list processing, higher‑order functions and macro‑like decomposition. (some nice web frameworks were initially written in LISP, scheme, etc...)
- APL pushes you toward whole‑array transforms, point‑free pipelines and exploiting data parallelism. (banks are still using it because of perforance)
- Prolog pushes you toward facts/rules, constraint satisfaction, and backtracking search. (it is a very high abstraction but might suit LLMs very well)
OK, and when you then translate that program into C++/Rust/python, a lot of this bias leaks through. You often end up with:
Rule engines, constraint solvers, or table‑driven dispatch code when the starting point was Prolog.
Iterator/functor pipelines and EDSL‑like combinators when the starting point was Lisp.
Data‑parallel kernels and "vectorized" loops when the starting point was APL.
In principle, an LLM could generate those idioms directly in C++/Rust. In practice, however, models are heavily shaped by their training distribution and default prompts. If you just say "write in Rust", they tend to regress towards the most common patterns in the corpus (framework‑heavy, imperative, not very aggressively functional or data‑parallel), even when the language would support richer abstractions.
By inserting a "thinking" step in a different paradigm, you bias the search over solution space before you ever get to Rust/C++. That doesn’t magically make the code better, but it does change which regions of the design space the model explores.
Same would also be true for python which is already a multi-idiomatic language. So it might be a good idea to learn a portfolio of different languages and then try to tackle a problem with a specific language instead of automatically using python/go/rust because of performance.
Something to consider...
p.s. how would a problem be solved when the LLM would have to write it first in erlang? Is it the automatically distributed?
p.p.s. the "design pattern" of the GoF comes automatically to my mind, which might be a good hint to the LLM to use.
So we are going to pretend this isn't happening everywhere now? And that it isn't failing on daily basis? I'm sorry but I've been saying this for years now and is my main arguments for not using slop machines: no one writes the code and no one reads the code. I can name dozens of fortune 500 companies where "tokens used" is used as a performance metric for developers, as in, more tokens = better performance, all code is written by slop machines and all reviews are made by slop machines, developers simply add "this is intended" in code reviews.
b) Python code is easier to introspect, and set up test harnesses around. And also extend in agentic frameworks
c) LLMs are really good at translation. I can give it python code and it can translate it into C.
Part of my worries that all this push to LLMs will marginalize niche programming languages from being used in startups since the lack of training data means falling back to hardcoding. a skill that I have a feeling will get increasingly niche overtime. I feel capitalism will basically render programming languages into a build artifact overtime.
However, I expect that in the future some new language will take this role of dual use.
You do want to use Rust with LLMs.
The reason you want it is simple, it's more constrained.
LLMs thrive on constraint and drown in freedom.
The further you can constraint the solution space the more likely you are to end up with a solution you like/is actually good.
Rust has several properties that make it really good for LLMs:
* Really robust type system that is also very expressive, if guided LLMs can implement most of the invariants in types which substantially increases the chances of success.
* Great compile time errors, the specificity and brevity (vs say C++ template expansion) means token efficient correction of syntax and/or borrow mistakes etc.
* Protection against subtle errors at compile time, namely data races and memory safety issues.
* Great corpus of well designed code and patterns, higher quality on average than some other ecosystems more favored by begineers/mass-market programming.
* Stdlib is strong, small-ish number of blessed crates.
* Context friendly, type signatures, errors, etc are all dense information.
* Also bias towards compile time checks means less runtime tests which means less toolcall time (and less tests needed overall) which in turn makes the process a ton faster.
I have been continually using Rust, Python and Kotlin since ~Jan this year and keeping track of my thoughts and I increasingly bias towards Rust now where I would have previously chosen Python or Kotlin instead just because I am lazy and I prefer the tool that the computer writes better so I have to write less lol.
So I might be biased, but with the correct curation of AGENTS.md files and skills, we're getting extremely good results using Claude Code writing Java.
Another disclaimer: I haven't tried with another language, but we're happy with the results.
https://news.ycombinator.com/item?id=15886728
masklinn on Dec 9, 2017 | parent | context | favorite | on: Larry Ellison allegedly tried to have a professor ...
And remember,
> Do not fall into the trap of anthropomorphising Larry Ellison. You need to think of Larry Ellison the way you think of a lawnmower. You don't anthropomorphize your lawnmower, the lawnmower just mows the lawn, you stick your hand in there and it'll chop it off, the end. You don't think 'oh, the lawnmower hates me' -- lawnmower doesn't give a shit about you, lawnmower can't hate you. Don't anthropomorphize the lawnmower. Don't fall into that trap about Oracle. — Brian Cantrill (https://youtu.be/-zRN7XLCRhc?t=33m1s)
And
> I actually think that it does a dis-service to not go to Nazi allegory because if I don't use Nazi allegory when referring to Oracle there's some critical understanding that I have left on the table […] in fact as I have said before I emphatically believe that if you have to explain the Nazis to someone who had never heard of World War 2 but was an Oracle customer there's a very good chance that you would explain the Nazis in Oracle allegory. — also Brian Cantrill (https://www.youtube.com/watch?v=79fvDDPaIoY&t=24m)
I mean, the Python ecosystem is qualitative and generally well-documented. What if the AI spent 30% less tokens generating code than e.g. in Rust?
Or is there a kind of information theory where, given the same goals / tests, the AI will spent roughly the same in any language?
When I vibe, it's C# all the way. Not a popular opinion on HN, but the LLMs are trained heavily on the language and are very, very good at it, plus with the 1-file-per-class organization, it can stay pretty clean. I mean, v10 LTS was just released, with all kinds of new language features, EFCore is still the best ORM I've ever used, with full support for SQLite, Postgres, MySql, etc. It just makes writing and reviewing code a pleasure. And the LLMs don't f*ck it up.
> Agents broke that loop in a specific way: the unit of contribution shifted from the patch to the port.
What does this even mean? Every time there's a bug we port the whole code to a different language instead of patching it? This sounds like absolute nonsense, and makes me wonder whether a human actually wrote this.
If you've managed software teams before, this won't be new. You just need to make sure the team does the right things. But you don't want to inject yourself on the critical path of everything. That's micro managing. People hate it and it's counter productive. You need to instead delegate responsibility and check that there is a good process with checks and balances that ensures things are done right.
If you are vibe coding, one shotting, etc. you are essentially operating without guards rails. You won't catch mistakes that are being made. You aren't doing the due diligence of verifying that what was delivered is the same as what was being asked for.
But if you do use guard rails, most of the engineering effort (i.e. your time) goes into building mechanisms to prove that what is being delivered is fit for purpose. And that needs to lean heavily on tools that verify things. Compilers, linters, test suites, headless browser based scenario tests, elaborate benchmarks, etc. Anything you can throw at this. The more the better. Even code quality issues are something you can catch and fix with tools. Code duplication issues are detectable. Poor cohesiveness and high coupling are simple metrics that you can optimize for.
With AI in the mix, all of that gets run automatically and you create a feedback loop where any introduced problem is more likely to be caught early. If you are a good senior engineer, you would have been doing all of this anyway. Because it compensates for your own inability to not make mistakes. With AI, you just need to do more of it.
I've dabbled with a few generated code bases in Go in the last few months. I have about 3 decades of experience with other languages. But not a lot of experience with Go. So, why did I pick it? It's not because I particularly like the language. It all looks a bit verbose and tedious to me and I've always preferred other languages. But since I'm not writing any code, I can step over that and make use of the fact that the compiler and build tools are really good and catch a lot of issues. By using Go, I'm leveraging the tool ecosystem around it. Which is really solid.
Because I don't read/write Go code, I'm forced to treat the system as a black box. Which means I just test the hell out of it in any way I can think of. When I don't know how, I ask the AI to suggest me ways. And it does, and I make it add those as well. My little system has performance benchmarks, end to end tests for everything, scenario tests testing complex scenarios, static code analysis, race detection, etc. And lots of unit tests. If I find any issue, I get paranoid about what else might be broken.
All I do is getting systematic about making it falsify the theory that it could all be broken by failing to produce a broken test scenario. I'm equally paranoid about code quality and technical debt. So, I make sure to check for that as well. Not manually of course. I simply ask the AI tool to do targeted reviews of code looking for duplication, adherence solid principles, etc. Any issues found are prioritized and addressed. With most quality issues, simply asking an LLM to look for such issues is surprisingly effective. Having guardrails just automates these checks and balances and makes them routine.
My inability to review at the line level no longer matters that much. Worse, me reviewing tens/hundreds of thousands of lines of code is probably counter productive. Even in languages I know well, it would take ages. I'd be the slowest part of the whole engineering process.
The friction is that most developers aren't trained to comprehend assembly or otherwise. The vast majority of CS programs don't do it seriously. Many don't really know the difference either, and even I would need a refresher before trying to debug assembly.
I also think token cost restricts directly writing into assembly language. I've experimented with assembly output, as I'm sure many of us have, and can confirm small assembly programs produce more tokens as a result because of the lack of a standard library. However, because tokens are currently priced per million, I don't think it's a significant restraint.
The hops right now are Python -> C -> Assembly . The trend is now Rust/Go/C -> Assembly. Perhaps in the future, there will be nothing in the middle.
Too capable for a job control language, interesting objection, but if a job control language is too capable, it becomes a job implementation language...
Too slow to implement applications. So many examples of big balls of Python code creaking and struggling, slowly. Just one example: my homeassistant instalation running three lights overwhelmed a Raspberry PI 4 causing it to crash once a week.
To me Python is a poster child for "popularity is not quality".
Therefore the "best" language is going to be whatever makes it easiest for humans to detect bugs, bad design, or that the "wrong thing" has been developed.
>rust
lol, thanks for the humor article of the day.
One of the reasons I dislike Go is because it's easy for most engineers to write really low grade code with it. But AI agents would probably not write the best code in any language anyway, so not much is lost.
It doesn't matter if the 800-line if statement is able to use pattern matching.
There's been a lot of progress on making coding agents able to solve problems when they can easily evaluate in a closed loop, we desperately need something similar for controlling complexity and using relevant abstractions.
Numpy is two decades old. The lesson of "don't write everything in Python" is old news and LLMs just add a little momentum to that.
Glue languages will always exist and Python is the best at it.
I thought it’s a poorly designed language with GC pauses so it surprised me that the ts compiler was written in it.
when I said “the ecosystem” I didn’t mean of libraries and other developers, I meant of recruiters and hiring managers
and whose humiliation ritual I could pass
As a benefit i find that static types help AI to make correct/better decisitions than you see in PHP (where types are mostly only class types, nominal or primitive [lol no generics])
But its pretty much true, i will forsee a fall in dynamic languges, as the usecase is pretty much void and null.
Also, totally FOSS. Unparalleled library ecosystem (no, I don't buy into the hype about re-rolling all your own dependencies).
Beyond that, Go is kind of nice, but the lack of a inheritance is stifling. Python has everything that's needed and very little that's not.
Edit: Getting downvoted, probably because of the comment about virtualenvs. What's your alternative? .NET DLL's? The joke that is NPM? Go probably does this better, admittedly, but Python is practically one of the best out there.
I started using Rust in 2018 and I've never used a build system that fought me less, ever, before or after.
I stopped reading after that sentence.
GPT 5.5 writes good haskell.
Still, not ALL projects benefit from such an approach and there are times when yes python is the right tool. Not just due to readability of humans but the other qualities that make it really good for small, iterative apps.
My take has never changed. Knowledge is cheaper than ever, but wisdom is as rare as ever. This is a great example of misunderstanding the former for the latter
(Joke but also not a joke)
AI doesn't really write code for me, but I do use them to brainstorm/ask questions. Though, I do not use Python. I have never been a fan of the language. I still think Python is a perfectly serviceable language, but it would solve no (important) problems I have ever had better than any other language.
I can see why Python is appealing to many people, and I applaud Guido for all the work and oversight over the years, but Python lacks a lot of the things I like in a language.
Lol good meme
That's already a glaring mistake. People could say perl's CPAN is great. Well, it did not save perl from declining in the last 20 years.
> The Python ecosystem is increasingly a Rust ecosystem wearing a Python hat.
Without statistics to prove this, this claim is useless.
Also, depending on Rust isn't that strange if a language is based on ... C. The only way I would disagree with such an argument were if Python were written in Python. But since it is syntactic sugar over C - just like ruby or perl are too - the argument to use Rust here is simply not different to using C. Perhaps Rust is better than C, but it is not fundamentally different. Whether Python were written in Rust or C is not a functional difference here.
As for AI becoming our new Overlord: I honestly do not want to depend on US mega-corporations. I am not disputing the fact that AI has objective use cases. I am objecting this herd mentality of everyone putting an AI chip into their brain now.
Damn AI slop zombies everywhere - it's like in the old B movie "They Live". But with less entertainment value than that. If they chew bubblegum then it is to slop up everything, not to kick ass.
> Klabnik vibe-coded a new language in Rust, therefore Claude + Rust = Good.
I argue the inverse -- Rust, being an ML-family language, is well suited for parsing, and language design (I know! Shocker!). In more moderate translation -- ML-style languages are good for parsing, interpreting and compiling code. Claude is not the magic here -- ML is.
I would also add that I've had decent success vibe-coding+human-coding Haskell (contrary to the article). My experience is that if I can hand-write a rich set of types (blessed be IxMonad), I can throw Claude to fill in the blanks for the implementations. If I can design the data structures that make the program tick, bridging them is something Claude is awesome at. Again, no surprise -- it's intern-level work.
The key distinction between C, Zig and Rust is that Rust is designed around types. C and Zig are more memory-oriented -- they really see most of your program as flat memory and you can kind of shoehorn a little bit of data layout in that flat memory. While this offers a large amount flexibility, this philosophy isn't well suited for proving out correctness. But again -- this doesn't mean they don't have a spot.
When I was a junior at Tesla, I used to joke that senior staff had a VMs in their heads, because that's really how you analyze C programs -- you try to execute it in your head, with interesting inputs, but that's about it. Claude's head-VM is quite fuzzy and often makes errors.
With Rust, if you design your type system, you prevent yourself from making dumb mistakes. Swap out "yourself" with Claude here and it's the same story.
I've yet to see Claude design really nice type systems, fwiw.
But the point is -- Claude is the enemy of beauty and correctness -- it's up to the SWE to design a type-system which will prevent it from doing so. To be clear, I obsess over type-systems personally, but that's not the only way -- incredibly rich, comprehensive, huge type systems, fuzzing, Antithesis, proptesting are all things you can do to minimize the impact of slop, and those are all valid things to do.
---
> Code is not written by humans therefore it doesn't matter that you don't know Rust.
Wouldn't say this was explicitly stated, but I definitely smelt this undertone throughout the article. If you don't understand the language you're reading, how can you understand whether the code in front of you is correct or not? If you have a systems engineer sitting across you to clean your PRs up, you can pass that responsibility onto them, but what about when they give their two weeks?
If all you know is Python, chances are you're going to make better software in Python than in Rust. Stick an `Arc<Mutex<T>>` everywhere and chances are your code will be slower, as a matter of fact. Use If you want to learn Rust, please join us! But if all you're trying to do is vibe-code better code -- do it in the language you know and can actually debug when shit hits the fan.
---
> Anthropic C Compiler
It is impressive that Claude is awesome at taking existing code and rewriting it, this is certain, but I'd like to repeat the exact same rhetoric that many have given -- rewriting =/= original authorship. Awesome, we have a C compiler, but we already had one, and we just rewrote it? Seems like a little bit of wasted electricity.
To build on top of this, I am really happy that Bun is exploring Rust, and the Claude rewrite is truly impressive, but quite surprising at times, preserving strange anti-patterns (my name being said anti-pattern, teehee): https://github.com/oven-sh/bun/blob/ffa6ce211a0267161ae48b82.... It's hard to determine why Claude decided this -- I assume a really strict input prompt.
Do note that the current stage of that PR is much better than what it was at the state of that commit, and obviously Jarred isn't merging blind slop, but that is still human-driven by someone who has an understanding of their product.
My bet is actually that _rewrites_ of already-functioning, well-tested code, are likely to be more common as time progresses. I think that's what Claude is really awesome at, and I think Claude can often achieve 80-20 improvements through rewrites. Again, Claude alone will not be a silver bullet -- it won't generate data-oriented programs if the source material wasn't data-oriented. It won't optimize for cache coherency, if the source didn't, but moving from Python to Rust alone, with more-or-less the same code structure, you're likely to see improvements by virtue of common operations being memory-coherent and avoiding the GIL and so on.
---
> A C compiler written in Rust used to be a graduate thesis. It isn’t anymore.
Come on, this is disingenuous -- a simple C compiler is a 1-day long project. LLVM is a graduate thesis (and for good reason). Copy-pasting prior-art is academic dishonesty and Claude does a lot of that.
---
For transparency: I work with Noah.
EDIT: Wanted to add that not a single line of my comment was AI generated.
When I run a game I don't care of the dev used C or whatever. Only programmers care about the syntactic representation.
I need the machine code/byte code patterns/geometric/color gradient data.
Eventually Python will be what you see on screen but no cPython interpreter program as we know it will be running
The model will have an internal awareness of the result to return without running an actual REPL
https://dev.to/zijianhuang/prompt-to-ai-generated-binary-is-...
https://platform.claude.com/docs/en/agents-and-tools/tool-us... Also Claude Cowork, etc.
1. You don't need compilation... run and test faster. Compilers were primarily built to prevent human error, and only very secondarily to guard your business logic.
2. Your validators quite often need to evolve. With Python or JS, this is a pydantic edit + run. Imagine 3–4 iterations of the same in Rust?
3. Composition. The entire cycle of software changes. An agentic system takes orders from a human, reads some kind of cache and snippets, writes/combines snippets, tests it, runs it, and fixes it. This almost pushes you toward snippets the size of a function, which still need to be covered with tests. I can easily build 10 function-sized Python files and write an agent that will mix and match 3 of them into a final result. With a compiled language, you'd need to compile 10 times — or store the binaries and think about what platform they'll execute on, etc.
I love the fact that the author is questioning this. No doubt the market for your favorite language will change. 80% of languages will go away — there is no market anymore for such a big variety of languages.
That's kind of sad, but so many older languages have been declared dead only to hang in various niches or out of sight for decades.