Just use Go. LLMs have seen a ton of it, they write it well, it compiles practically instantly, and it has all the advantages of a typed compiled language.
I created a big Python codebase using AI, and the LLM constantly guesses arguments or dictionary formats wrong. Unit tests and stuff like pydantic help, but it's better to avoid that whole class of runtime errors altogether.
That’s what I’ve settled on. Python is so flexible that there are a million ways to organize code, pass arguments, etc. If you already have a code base to work from, an LLM can make new code in the style of the old code. But a fresh project? Once you get to a certain level of complexity it quickly can turn into write once, read never code.
This is where I’ve found that a compiled, strongly typed, language (any one really) works well with an LLM. With the little bits of friction with a language like Go and the LLM can produce pretty decent (and readable) code.
But what is the selling point for Go? I get that it is allegedly hailed to be a simple language with basically no batteries included, but why is that a selling point? Does Go excel at anything no other language does?
That would matter if we were asking the AI to generate code open-loop: someone probably already wrote something close to what you asked for in Python. But if the agent generates code, tries to compile it, sees the detailed error messages and acts on those messages to refine the code, it's going to produce a higher quality result. rustc produces really good diagnostics. And there's a lot of Rust code online now, even if there's so much more Python and Javascript/Typescript.
LLMs don't actually semantically parse the error messages. They will generate the most likely sequence resulting from the error message based on their training data, so you're back to the training data argument.
Admittedly, I have very little experience with LLM-assisted Python. However, based on the severe degradation in output quality I have seen from an LLM working with plain JavaScript as opposed to TypeScript, I can't imagine choosing to start a project in Python at the moment.
Training data can't be the whole answer. LLMs are really good at translating to different programming languages. This makes sense, given that they are derived from text translation systems. I'm getting great results in languages with comparatively small bodies of freely available code. The bigger hurdle is usually that LLMs tend to copy common idioms in the target language and if it is an "enterprise-y" language like Java or C#, the amount of useless boilerplate can skyrocket immediately, which creates a real danger that the result grows beyond the usable context window size and the quality suffers.
Dimensionality gets bizarre in 1000-D space. Similarity and orthogonality express themselves in strange ways and each dimension codes different semantic meaning.
Therefore, if the training data is highly consistent you are by definition reducing some complexity and/or encoding better similarity.
> LLMs are really good at translating to different programming languages.
...for which ample training data is available.
> This makes sense, given that they are derived from text translation systems.
...for languages with ample training data available.
Yes, LLMs can combine information in novel ways. They are wonderful in many respects. But they make far more mistakes if they can't lean on copious amounts of training data. Invent a toy language, write a spec, and ask them to use it. They will, but they will have a hard time.
I have a language I wrote for processing data pipelines. I’ve used it for years, but I can count the number of users on one hand. I wrote it partially to learn about writing a scripting language, partially because Nextflow didn’t exist yet. I still use it now because it works much better for my way of processing data on HPC clusters.
The only code that exists on the internet for this is test data and a few docs in the github repo. It’s not wildly different from most scripting languages, from a syntax point of view, but it is definitely niche.
Both Codex and Claude figured it out real fast from an example script I was debugging. I was amazed at how well they picked up the minor differences between my script and others. This is basically on next to zero training data.
Would I ask it to produce anything super complex? Definitely not. But I’ve been impressed with how well it handles novel languages for small tasks.
That might be an argument for not using a novel homebrew programming language. But it's not an argument against, like, any top-100 or even top-1000 programming language, which will be adequately represented in the training data.
It is if more training data results in better performance. In which case, GP will continue to use the language that is likely to have the most training data available.
They are also good at generating plausible code. The kind that has no obvious bugs in it. I wouldn’t be surprised if humans in the loop over report success with these tools. Combined with decision fatigue… it’s not a good recipe for humans making good decisions.
An experienced Rust developer is going to be in a better position to drive an agent to generate useful Rust code than a Python programmer with little or no Rust experience. Not sure I agree with the author that everyone should just generate reams of Rust now.
At least if your get paged at 3am to fix the 300k AI-generated Django blog you’ll have a chance at figuring things out. Good luck to you if Claude is down at the same time. But still better than if it was in Rust if you have no experience with that language.
Also, every single interpreter error has an entire corpus of StackOverflow-esque fix suggestions alongside it, and the model has been fine-tuned to minimize such errors on the first try. This hasn't been done for more obscure languages. You'll likely take more turns, on average, to get a working output, even if your problem is fully verifiable via test input/outputs - and if it's not verifiable, you don't want the "attention" of the model focused on syntax rather than the solution.
Well, go on and do the experiment! Perhaps LLMs can right code as well in BF as Python but I don't recommend it because hallucinations are really hard to notice in BF.
If you are going to worry about high level computer languages and AI, you are going to have to start with getting to grips with machine code and assemblers and that. Once you know how say some Python code ends up being processed by your laptop CPU(s), then you will know when BF might be best!
> Frontier models score ~90% on Python but only 3.8% on esoteric languages, exposing how current code generation relies on training data memorization rather than genuine programming reasoning.
I find if I ask most LLMs to write a self contained script/utility, even in codebases that are 90-100% written in some other language most will default to using python for it, or sometimes bash.
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).
Yes language arts are the only constructive uses of our brains.
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.
Python has a much more mature ecosystem than Rust, especially for AI/ML stuff. I ran into a rust crate that purported to do a certain ML algorithm but did not do it correctly. I managed to write a replacement with Claude though.
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.
I almost always pick Rust. Recently I wrote a plugin for something that was written in Go. I could have used Rust, but Go for one felt right because if the thing turned out well, others would surely find more value in having one toolchain.
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.
i think the enforcing the type system is good with AI for a couple reasons:
- (speculating) typed language have faster/better LSPs that can be used to more efficiently modify code with tool use.
- when a human DOES need to step in and start investigating/modifying the code, the strong typing makes it much easier to get oriented within their spaghetticode
Yeah, I mean, if I'm going to step away from the Python ecosystem and let AI manage/polyfill my dependencies, I might as well shift the whole way to OCaml/F# rather than Rust.
Definitively something to be said for AI/ML library support. I find myself going with Rust / TS for a ton of my backend work lately though, even though I'm a huge Django fan for backend.
Bit off topic but why in the world are people still posting on medium? The reading experience is abhorrent; I couldn’t even finish reading this article before a full screen popup literally blocked the sentence I was reading.
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.
> 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.
They don't ship it because of greed. They only want your attention because of greed. They only infest their website with ads because of greed.
> The browser is a delivery mechanism,
http is a delivery mechanism. The browser is a user agent. It's supposed to display content according to the preferences of the user. If your browser isn't doing that for you it's time to find a new browser or beat the one you have into submission until it behaves. "reader mode" is a useful compromise.
> It's supposed to display content according to the preferences of the user.
That's right, the original idea was exactly about that, but like I said - in practice that is no longer a thing.
Using the editor for reading any content is enormously underrated. Check this out - this entire thread opens in my editor as an outline with nested structure. Meaning that all the regular outline operations are available to me - folding, imenu (interactive TOC), narrowing, quick search, contextual search, pattern-based search, sparse-tree search.
Extracting all the URLs on the page while ignoring HN-internal ones is a single keypress for me - there's a link to a YT video - I can watch it, controlling the playback directly from my editor, I can extract transcript and summarize it with an LLM request - all without opening new tabs, without switching focus.
I can narrow on the sub-thread, or select a region and export only that part to a pdf. The possibilities are virtually unlimited.
It seems like it's just the latest evolution of the writer-friendly blogging platform; easier than Wordpress to package into a newsletter, and also easier to monetize with a paid tier.
Insofar as AI is great at accidentally deleting your production and backup Wordpress databases, and forcing you to start from scratch with something else.
Yep, Medium was free and everyone donated content... then it put up reading paywalls and conned everyone, I'm also surprised when I see people writing on there.
You can of course use any language but here is my advice: you should use the language that you know best to make your life as uncomplicated as possible when you want to understand what the LLM was creating.
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...)
Previously in my life as an IC, I wrote a lot of Golang. I worked on the larger end to end encrypted video calling service.
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.
AI's are really good with Python. Quick turnaround. Easy to read. Tons of training data/examples. Many of the same reasons we wrote Python before.
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)
Problem with Python and other non-strict typed languages is that if you let an LLM to write some stuff, you cannot truly be confident that nothing has broken. Even if your tests all pass. The LLM could have broken some path that only gets run in production in a very specific case. At least with strongly-typed languages you get a compiler error. In big codebases is non-negotiable
Dynamically typed languages just add one more type of bug that can’t be caught at compile time. That’s not helpful, sure, but it’s one type of bug among many.
The issue you mention, execution paths not hit by test cases, is made worse by having more complicated code. Duck-typing can help reduce the number of paths.
Static vs dynamic… I don’t see an obvious winner here.
Python has had type hinting for quite a while, and adding validation with mypy/pyright/ty as a step in CLAUDE.md (as well as having it as part of your CI pipeline) can emulate static type checking pretty well.
My take is that I can never be confident that anything an LLM produce will not be broken. Since I will have to check everything it produces anyways, why not write it in a human friendly language, i.e. python? C and rust may have better strictness, but the amount of boiler code to set up that system takes up a lot of mental space that could be better used to architect the problem at hand.
Haskell is a good language for LLMs! Claude knows it really well, and the type system catches so many mistakes. Just make sure to tell it to model the domain in the type from the start.
Also, Haskell can be really performant and lots level, while still keeping the benefits of typing. With the C foreign function interface you can really do anything in Haskell!
so it just boils down to strictness even when we're talking LLMs?
I agree with you about fast failure being a nice feature , but I also think that if you're TDDing a bunch of stuff and it fails in some categorical way , well then the test suite was lazy.
> so it just boils down to strictness even when we're talking LLMs?
The article describes what I've been doing for the past few months - I did small python projects in the past because of the ecosystem: I couldn't possibly write a ton of the stuff required for the things I wanted to do, so I leaned into python because someone already wrote it for me. Quality of deps was mostly ok for the happy paths, but always a chore to patch the broken ones.
Nowadays I tell Claude what I want to build and I always ask it whether rust is a good choice for it. It'll pick up the right crates or choose whether it should DIY, do all the plumbing, nail all the logic, and in ~30m I'll have something very solid that would have taken me 3+ weeks of part-time evening coding in python. I think the article is right and rust is the closest to the "best language" we have for LLM coding at the moment: the strict typing and the tooling dramatically reduce the output space for LLMs, and 99% of errors have a clear, precise explanation that is actionable, and the compiler helps you a lot there too.
I think it also boils down to the fact that you cannot reliably and quickly answer "why is this arg None?" in languages like python without figuring out the call graph and evaluating possible states and inputs/outputs. Rust makes all that explicit and forces you handle it, which I feel dramatically cuts the time an LLM needs to spend figuring out why it's broken or what to do next. EDIT: The fact that you get memory safety on top of all this and it's handled by the compiler is yet another advantage for LLMs: the logic that gets written is simpler to reason about, because if you try to mutably access the same variable in two different places, the compiler will feed this back to the LLM at build time. In other languages that would be a "code smell" or would require static analysis.
Strictness is a quality for software and a chore for humans, and of course the stricter you are at representing your logic and your state machine, the less ways a program can break. LLMs writing in rust give you the strictness without the chore part, and it's a very good deal from my point of view.
My anecdotal (sample size 1) experience is not consistent with this. I code fast. Refactor fast. My stuff doesn't break. But my methodology isn't the same as other's.
Yeah, last year I discovered that AI writes better rust than C, so I switched to rust and it made some quick good code that it couldn't do in C.
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.
Better question is why use any code? Generate random functions and select based on measuring the distribution of output of these functions against metrics of interest. A pure black box of instruction that is more performant than any verbose code or algorithm we could come up with, because all we select for is performance above all. Directed evolution essentially of the codebase, generated through mutation and selection, just like everything else on planet earth.
Isn’t the answer usually - because the same ai said python is the right language for it?
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.
I had agents code up an app for me in Swift a while back and the entire experience was so much better than your typical Python experience. The agents took full advantage of the compiler and static typing. There were far fewer bugs than expected.
Many here propose replacing Python with more performant, but less familiar languages - mostly Rust, Go. But I find the argument that the AI - HUMAN interface is the most important. A simple version of this is “no, stick with Python if that’s what you know”. A more interesting version is “use this new found AI leeway to move up the abstraction level”, “try something more expressive and human oriented”, “make a DSL and parser that suits the domain (and focuses the AI)”. Despite being a minority language, Raku is ideal for these aspects (esp with built in Grammars and general kitchen sink repartee) and works surprisingly well with most popular LLMs.
Claude writes java pretty well, and faster than Rust. It's a great middle ground for some projects. I've switched back from Rust to Java for some things.
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...
I still use TypeScript because I know it best. When AI makes a mistake, I can find the bug much faster. For me, the speed of writing code doesn't matter as much as the speed of fixing it.
As others have said, the main benefit with Python over Rust is library support especially with ML features. The other gap as I see it with Rust is the lack of native flexible UI support. The nice thing about Rust though is it can serve as a very fast and stable core for an app and offload specifics to TS and Python as their strengths allow, so you get the best of all worlds.
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.
This post resonates. I recently built a little web service to scratch an itch I've been having and after discussing the options with Claude we settled on Go, and honestly it's been fantastic. Highly performant, native threading, dead simple to deploy with containers. And I don't even know how to read or write Go.
Oh man... I like go because it is compiled, performant, strong and statically typed. But "fun" is not something I would say about it. The ergonomics of error handling, lack of ternary operator and other stuff that compiled 30yo languages already had ...
I did go through the Go tutorial many many years ago, but it's been so long I don't remember anything. I do remember it was an enjoyable process though, and I'd love to pick it up again.
Never mind cleaning up, you also have to understand the language just to judge and review the LLMs output. How else are you to separate good design and implementation from a bad one?
I know a couple languages fairly well: C, Perl, Python, Bash. I never formally learned Go, but as a test of AI coding, I started some vibe coded projects in Go. It worked very well: the code is minimal, there's few dependencies, and it compiles down to a static app. But most importantly, I can actually read the Go code and understand basically what it's doing. I can also use LLMs to critique the code if I'm uncertain. The big benefit of Go is the simpler language and "batteries included" standard library. This leads to fewer dependencies and less lines of code, which improves overall AI output. In theory, AI should be able to write better code faster in Go than in another language like Rust.
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.
Here's the kicker: The devs spent nearly 5 months on a solution, and it ended up being so crap it was abandoned. The multiple vibe-coded solutions were all better.
Of course language was the barrier, that's part of why it was always hard to hire people. It takes years to get good at a particular language, and most people are idiots from bootcamps who learned a single framework.
Really agree. Python is popular because it's easy for human to implement. But now if the coder becomes AI, then Rust would be preferable for agent, just like Python for human. In addition, it brings better performance.
This seems sort of like asking whether a chatbot should answer you in English or Japanese. Obviously, it should use whichever language you understand. If you understand Python best, why not write code in Python?
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 think it is an interesting question what kind of programming language one needs for an era of agents. It is clear that the programming language that was designed for humans is not necessarily the best for AI-driven software development. I guess the qualities one would want is some formal correctness guarantees, high performance. A question is whether this language is Rust or it is possible to design a better new language.
Assuming you are thinking about software architecture and looking under the hood, you are likely to be reading much more code than before. Python is really nice on the eyes and you can easily get a good grasp of what the code is doing. Plus, it's dynamically but strongly typed, so what you see in the code is usually what you get.
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.
Lately I just have Claude build most things in Rust, it's really amazing. I tried Go, but I found it wasn't as good--Rust really does to
me feel like Python. That said, it still struggles with the same class of errors of building complex systems. I've tried using TLA+, Alloy, and other things but haven't found the trick yet. The best I've found is reimplementing all external systems in memory and e2e testing everything extensively, without reimplementing the tests become unusably slow, and Claude can rewrite huge surface areas with ease--it's somewhere between mocking and literally just reimplementing the external systems.
I have been wondering on a similar thing; am looking for feedback:
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.
Great question. And I don’t think that Python, Ruby and PHP have a good answer. Scripting languages cater to human weaknesses. The 10-100x perf cost was never really worth it but now it’s impossible to justify.
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.
I'm still not sure. Would love thoguhts on this.. but in this new ai world we are in... is it better to go fullstack typescript? or go with proven mature frameworks? .net, ruby, django, etc? Seems TS is moving fast but maybe its time to not reach for the shiny object and stick with proven tech? or in 5 years will we regret it?
If you're using GenAI, you should go through the process of selecting an optimal tech stack for each solution, but also take into consideration that Claude and other services probably the most knowledge of python, javascript, and typescript with go, rust, java, and c# following closely behind. Consider what you're building and what elements of the tech stack is optimal for your problem-space.
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.
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.
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.
Golang. People trash it for being verbose on errors but it's an extremely readable language and it's almost like bash, only much stronger typed and with a very rich stdlib (so it's not likely you'll need a library for a quick script).
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.
In a world where AI supposedly can write in any language, Go is much better choice than TypeScript. Imagine contemplating for more than a few seconds a choice between simple, fast, cross-compilable language, and a TypeScript -> JavaScript -> Interpreter -> JIT stack.
If you don't know Go, it's more efficient to learn it than to waste the hardware resources of thousands to stay within JavaScript.
Well, these days a small CLI program in Java (say, ls) starts up cold, runs, and terminates in ~70ms, not 1500ms, but yeah, sometimes 70ms is too long to wait for a script.
People never believe me when I say it but I start noticing scripts needing 75-100ms to start. Modern hardware is ultra fast; I want my programs to make full use of it. I got no patience for tech or people who keep insisting "it's not much, it'll not kill you". Well duh, obviously it will not but that's not the point and never was. I want stuff to work between my blinking my eyes and I have achieved that hundreds of times over the course of my career.
That's perfectly fine, and I totally understand people who don't want to sit and wait 70ms for their script to finish running (that 70ms is not the time it takes to start), but let's not turn a <40ms startup into 1.5s. Now, it is true that if you want to launch a minimal HTTP server in Java you may need to wait ~100ms, which may be too long for you, but is also a far cry from 1.5s.
It is, but I am still quoting what I saw before, it was not a fantasy. I don't deny it's likely better nowadays, sure, but I remain moderately skeptical because JVM is still a runtime that needs to boot.
Then again, Golang has one as well, though it does manage to start it up faster it seems.
This is just Kotlin. Strongly typed, more concise than Java or Go (and probably Typescript), less likely to blow up at runtime than Typescript, epic tooling, plenty of public code, and a library for basically anything because JVM.
Java runs a Hello World, cold, in a packaged JAR, in about 40ms. What you've seen isn't JVM startup but programs that do a lot at initialisation (like MS Word), as many Java programs like to do (because they often expect to run for a long time, so they don't care about startup time).
I have not worked with Java in a long time but I seem to remember that most Java programs also accrue a good amount of dependencies and some of them have their own init routines.
That adds up, fast. No idea how is it nowadays, admittedly. Maybe a ton of optimization work was done.
C. At least with Gemma 4 it does a fine job. Writes good error checking. Writes memory management. Mostly straightforward and easy to read. A lot of libraries. Runs everywhere.
There are two working LLM axes. Critic strength: how much the language catches before runtime. Sensor strength: how good the empirical feedback loop is. LLMs benefit from both, but the sensor axis often is undervalued.
Type safety is great, but you can't just quietly disregard the benefits some dynamically typed languages provide; that would be completely ignoring that different tasks weight the two axes differently.
Systems code, performance-critical code, code where correctness across all cases matters more than exploration: parsers, compilers, network protocols, data structures - statically typed languages (like Rust) give you an edge here. The compiler's depth pays for the verbosity, and exploration is less of the work because the problem shape is known up front.
For stuff like building a web scraper, or rapidly prototyping, or exploratory scripts, something like Rust would be actively bad. You cannot poke at a live browser (you can with Clojure). Async Rust adds another layer of type complexity. The signal-to-noise for "figure out what is on the page" collapses entirely.
If I were picking a single language for general LLM-assisted work, weighted across task types, it would be Clojure (or Elixir), with OCaml as the most interesting alternative if the ecosystem were stronger.
Using Clojure and Elixir and LLMs are fantastic with both. Sure, if I get to a super-stable situation then maybe I'd consider moving to Rust (or Jank?), but for now I'm just so happy with Clojure and Elixir in this new world. I'm solving new problems with fully bespoke architecture so the flexibility is key. Clojure for business logic and most DB. With Elixir, it's the actor model and hand-holding as I'm using it for the web layer. I bet Ruby on Rails would also shine for some cases, prob most CRUD for example.
I think there's something key you get at in terms of the combo of dynamic environment + type safety maximising both. With a dynamic environment, the LLM can do a lot of interrogation to understand the problem space on the fly. I've witnessed agents sort out pretty complex issues through `python -c "..."`, `groovy -e "..."`, executing snippets of code with Node etc which is much less accessible if they have to compile it first. They can also inject logging code that interrogates the runtime as well (what type do we really have at line 1003?) etc which works better with runtimes that have deep introspection capabilities.
I assume this is why things like PyO3 are popping up? If so, sort of a fascinating way to compartmentalize new rust code into legacy .py code in lieu of a refactor, or at least, a way to do a staggered refactor and eat the elephant in bites :)
Devs still have to maintain this code, the Python devs can definitely get the LLM to write (some kind of) Rust, but when it goes wrong and you hit the wall with the LLM then they can will have to learn Rust which might take a while, this sounds like a bit of a project risk.
For me, whether it's AI or my own handcrafted artisanal code, the choice of language comes down to what has the least friction. This means I turn to vite/react for a lot of frontend requirements, and that the backend will be in nodejs or python, because those are easier for me to debug than writing an equivalent application in C++ or Rust.
The article applies to a narrow case of a totally green field application that's going to be completely vibecoded. This is the only case where you reasonably can be indifferent to what the language is, and so you can abandon familiar Python and go with unfamiliar Rust. (If you _are_ familiar with Rust, the point of the article is moot.)
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...
For the utilities I write it is faster to iterate without having to compile. When I get to the point where I'm done adding changing features, and performance is an annoyance I can always ask the AI to "rewrite this in Go". (I've never gotten to that point.)
This point only makes sense if you ship AI code without reviewing it. And if you're shipping AI code without reviewing it, you're going to run into much bigger problems than Python performance limitations.
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 dislike Go but I have to admit, it's a great language for AI generated code. Simple enough, it compiles quickly and it performs meh-well enough for most applications.
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.
you still need to look at the code oneday so id say c++ still would be a preferred target language even for ai. i know i hear a lot about rust but im still getting the idea its a niche language overall. i know people love it and point out its advantages, but sometimes good enough is good enough (i.e. c++)
A somewhat contrarian/pessimistic view: The hardest thing in any future of LLM generated code is going to be the verification step, and especially types of verification that require humans which are going to be the most expensive.
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.
The LLMs just churns out non-idiomatic slop in any language.
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.
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.
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... :)
Let's go through some of the arguments, in no particular order:
> 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.
Read the first few comments and surprised I didn’t see it, but training data. The voluminous amount of Python in the training data.
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?
I built a programming language, and LLMs can code phenomenally well in it.
I don't think the training set matters that much.
Programming languages have a lot in common. Python is kind of odd when it comes to languages.
Surprisingly, LLMs are actually much worse at reasoning in Python than other common programming languages for agentic coding tasks.
Data here: https://gertlabs.com/rankings?mode=agentic_coding
Just use Go. LLMs have seen a ton of it, they write it well, it compiles practically instantly, and it has all the advantages of a typed compiled language.
I created a big Python codebase using AI, and the LLM constantly guesses arguments or dictionary formats wrong. Unit tests and stuff like pydantic help, but it's better to avoid that whole class of runtime errors altogether.
That’s what I’ve settled on. Python is so flexible that there are a million ways to organize code, pass arguments, etc. If you already have a code base to work from, an LLM can make new code in the style of the old code. But a fresh project? Once you get to a certain level of complexity it quickly can turn into write once, read never code.
This is where I’ve found that a compiled, strongly typed, language (any one really) works well with an LLM. With the little bits of friction with a language like Go and the LLM can produce pretty decent (and readable) code.
But what is the selling point for Go? I get that it is allegedly hailed to be a simple language with basically no batteries included, but why is that a selling point? Does Go excel at anything no other language does?
That would matter if we were asking the AI to generate code open-loop: someone probably already wrote something close to what you asked for in Python. But if the agent generates code, tries to compile it, sees the detailed error messages and acts on those messages to refine the code, it's going to produce a higher quality result. rustc produces really good diagnostics. And there's a lot of Rust code online now, even if there's so much more Python and Javascript/Typescript.
LLMs don't actually semantically parse the error messages. They will generate the most likely sequence resulting from the error message based on their training data, so you're back to the training data argument.
Perhaps the training data about what compiler diagnostics mean is particularly semantically rich training data.
Admittedly, I have very little experience with LLM-assisted Python. However, based on the severe degradation in output quality I have seen from an LLM working with plain JavaScript as opposed to TypeScript, I can't imagine choosing to start a project in Python at the moment.
Training data can't be the whole answer. LLMs are really good at translating to different programming languages. This makes sense, given that they are derived from text translation systems. I'm getting great results in languages with comparatively small bodies of freely available code. The bigger hurdle is usually that LLMs tend to copy common idioms in the target language and if it is an "enterprise-y" language like Java or C#, the amount of useless boilerplate can skyrocket immediately, which creates a real danger that the result grows beyond the usable context window size and the quality suffers.
In higher dimensional vector space, yes it can.
Dimensionality gets bizarre in 1000-D space. Similarity and orthogonality express themselves in strange ways and each dimension codes different semantic meaning.
Therefore, if the training data is highly consistent you are by definition reducing some complexity and/or encoding better similarity.
In Go the statement
Is almost always going to be followed by In a highly dynamic language you may not get Unless explicitly asked for.Very true.
I have to steer models hard for C++. They constantly suggest std::variant :P
is that bad?
Godbolt got a 2x speed improvement switching from what he thought was a good fast impl to std:variant
https://www.youtube.com/watch?v=gg4pLJNCV9I
> LLMs are really good at translating to different programming languages.
...for which ample training data is available.
> This makes sense, given that they are derived from text translation systems.
...for languages with ample training data available.
Yes, LLMs can combine information in novel ways. They are wonderful in many respects. But they make far more mistakes if they can't lean on copious amounts of training data. Invent a toy language, write a spec, and ask them to use it. They will, but they will have a hard time.
I have a language I wrote for processing data pipelines. I’ve used it for years, but I can count the number of users on one hand. I wrote it partially to learn about writing a scripting language, partially because Nextflow didn’t exist yet. I still use it now because it works much better for my way of processing data on HPC clusters.
The only code that exists on the internet for this is test data and a few docs in the github repo. It’s not wildly different from most scripting languages, from a syntax point of view, but it is definitely niche.
Both Codex and Claude figured it out real fast from an example script I was debugging. I was amazed at how well they picked up the minor differences between my script and others. This is basically on next to zero training data.
Would I ask it to produce anything super complex? Definitely not. But I’ve been impressed with how well it handles novel languages for small tasks.
That might be an argument for not using a novel homebrew programming language. But it's not an argument against, like, any top-100 or even top-1000 programming language, which will be adequately represented in the training data.
It is if more training data results in better performance. In which case, GP will continue to use the language that is likely to have the most training data available.
> It is if more training data results in better performance.
Sure. But given the relation with translation systems, it seems far more likely that there are diminishing returns to larger volumes of training data.
They are also good at generating plausible code. The kind that has no obvious bugs in it. I wouldn’t be surprised if humans in the loop over report success with these tools. Combined with decision fatigue… it’s not a good recipe for humans making good decisions.
An experienced Rust developer is going to be in a better position to drive an agent to generate useful Rust code than a Python programmer with little or no Rust experience. Not sure I agree with the author that everyone should just generate reams of Rust now.
At least if your get paged at 3am to fix the 300k AI-generated Django blog you’ll have a chance at figuring things out. Good luck to you if Claude is down at the same time. But still better than if it was in Rust if you have no experience with that language.
> I could write in brainfuck with ai, but I presume, wouldn’t get the same results than if going with python.
https://esolang-bench.vercel.app/
and this sums it up right here.
I wrote about the meta thesis of programming languages in the training data here
https://jry.io/writing/use-boring-languages-with-llms/
Please distill instead of having me navigate off site. Include link for additional info.
edit: side -> site
Also, every single interpreter error has an entire corpus of StackOverflow-esque fix suggestions alongside it, and the model has been fine-tuned to minimize such errors on the first try. This hasn't been done for more obscure languages. You'll likely take more turns, on average, to get a working output, even if your problem is fully verifiable via test input/outputs - and if it's not verifiable, you don't want the "attention" of the model focused on syntax rather than the solution.
The LLMs are actually worse at generating Python than other langs, hypothesized due to quality of training data lol.
I still read the generated code, so I'm not quite willing to give up on Python yet though.
"I could write in brainfuck with ai"
Well, go on and do the experiment! Perhaps LLMs can right code as well in BF as Python but I don't recommend it because hallucinations are really hard to notice in BF.
If you are going to worry about high level computer languages and AI, you are going to have to start with getting to grips with machine code and assemblers and that. Once you know how say some Python code ends up being processed by your laptop CPU(s), then you will know when BF might be best!
> Frontier models score ~90% on Python but only 3.8% on esoteric languages, exposing how current code generation relies on training data memorization rather than genuine programming reasoning.
https://news.ycombinator.com/item?id=48100433#48102985
There's enough training data on the other langs.
No if that mattered you'd write everything in html and css. Because that has way more training data.
Those are not programming languages.
I find if I ask most LLMs to write a self contained script/utility, even in codebases that are 90-100% written in some other language most will default to using python for it, or sometimes bash.
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).
If AI writes your articles, why use brain?
^ Underrated reply.
Yes language arts are the only constructive uses of our brains.
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.
> A shipped app, in a language nobody on the team knew
Great! Let's look back on this not too far in the future.
Yeah that's probably the only thing in the world that could be scarier than the electron app they were replacing
Python has a much more mature ecosystem than Rust, especially for AI/ML stuff. I ran into a rust crate that purported to do a certain ML algorithm but did not do it correctly. I managed to write a replacement with Claude though.
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.
I almost always pick Rust. Recently I wrote a plugin for something that was written in Go. I could have used Rust, but Go for one felt right because if the thing turned out well, others would surely find more value in having one toolchain.
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.
What peers are speaking Mojo? I’m not aware of any place it’s penetrated enough to be a “lingua Franca”
i think the enforcing the type system is good with AI for a couple reasons: - (speculating) typed language have faster/better LSPs that can be used to more efficiently modify code with tool use. - when a human DOES need to step in and start investigating/modifying the code, the strong typing makes it much easier to get oriented within their spaghetticode
Yeah, I mean, if I'm going to step away from the Python ecosystem and let AI manage/polyfill my dependencies, I might as well shift the whole way to OCaml/F# rather than Rust.
Then I get the benefits of GC and strong typing.
Definitively something to be said for AI/ML library support. I find myself going with Rust / TS for a ton of my backend work lately though, even though I'm a huge Django fan for backend.
Bit off topic but why in the world are people still posting on medium? The reading experience is abhorrent; I couldn’t even finish reading this article before a full screen popup literally blocked the sentence I was reading.
Is there some incentive I’m not seeing?
check out Scribe an alternative medium frontend that's why better: https://scribe.rawbit.ninja/@NMitchem/if-ai-writes-your-code...
https://sr.ht/~edwardloveall/Scribe/ https://libredirect.github.io/
> The reading experience is abhorrent
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.
> 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.
They don't ship it because of greed. They only want your attention because of greed. They only infest their website with ads because of greed.
> The browser is a delivery mechanism,
http is a delivery mechanism. The browser is a user agent. It's supposed to display content according to the preferences of the user. If your browser isn't doing that for you it's time to find a new browser or beat the one you have into submission until it behaves. "reader mode" is a useful compromise.
> It's supposed to display content according to the preferences of the user.
That's right, the original idea was exactly about that, but like I said - in practice that is no longer a thing.
Using the editor for reading any content is enormously underrated. Check this out - this entire thread opens in my editor as an outline with nested structure. Meaning that all the regular outline operations are available to me - folding, imenu (interactive TOC), narrowing, quick search, contextual search, pattern-based search, sparse-tree search.
Extracting all the URLs on the page while ignoring HN-internal ones is a single keypress for me - there's a link to a YT video - I can watch it, controlling the playback directly from my editor, I can extract transcript and summarize it with an LLM request - all without opening new tabs, without switching focus.
I can narrow on the sub-thread, or select a region and export only that part to a pdf. The possibilities are virtually unlimited.
> 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?
Because that’s an enormous pain in the ass. Not scalable at all.
I beg to differ. You clearly misinterpret what I'm talking about. Please expand on "scalable", what do you mean by that?
They have made an honest attempt to pay writers. It's a different model than substack, but that's why.
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.
My best guess is momentum. Some people are very, very brand loyal and have to do things in relation to what/how others do things.
In reality it doesn't matter where something is posted, just give us a url, but some people don't operate that way.
It seems like it's just the latest evolution of the writer-friendly blogging platform; easier than Wordpress to package into a newsletter, and also easier to monetize with a paid tier.
But don't we have AI to deal with the complexity of Wordpress? :-)
Insofar as AI is great at accidentally deleting your production and backup Wordpress databases, and forcing you to start from scratch with something else.
Yep, Medium was free and everyone donated content... then it put up reading paywalls and conned everyone, I'm also surprised when I see people writing on there.
Nice perspective on languages in the AI era. I think AI should be used to build best performing and highly scalable software systems.
You can of course use any language but here is my advice: you should use the language that you know best to make your life as uncomplicated as possible when you want to understand what the LLM was creating.
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...
Previously in my life as an IC, I wrote a lot of Golang. I worked on the larger end to end encrypted video calling service.
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.
[1] https://jry.io/writing/use-boring-languages-with-llms/
AI's are really good with Python. Quick turnaround. Easy to read. Tons of training data/examples. Many of the same reasons we wrote Python before.
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)
Problem with Python and other non-strict typed languages is that if you let an LLM to write some stuff, you cannot truly be confident that nothing has broken. Even if your tests all pass. The LLM could have broken some path that only gets run in production in a very specific case. At least with strongly-typed languages you get a compiler error. In big codebases is non-negotiable
Dynamically typed languages just add one more type of bug that can’t be caught at compile time. That’s not helpful, sure, but it’s one type of bug among many.
The issue you mention, execution paths not hit by test cases, is made worse by having more complicated code. Duck-typing can help reduce the number of paths.
Static vs dynamic… I don’t see an obvious winner here.
Python has had type hinting for quite a while, and adding validation with mypy/pyright/ty as a step in CLAUDE.md (as well as having it as part of your CI pipeline) can emulate static type checking pretty well.
My take is that I can never be confident that anything an LLM produce will not be broken. Since I will have to check everything it produces anyways, why not write it in a human friendly language, i.e. python? C and rust may have better strictness, but the amount of boiler code to set up that system takes up a lot of mental space that could be better used to architect the problem at hand.
This is why you should use Haskell.
Haskell is a good language for LLMs! Claude knows it really well, and the type system catches so many mistakes. Just make sure to tell it to model the domain in the type from the start.
Also, Haskell can be really performant and lots level, while still keeping the benefits of typing. With the C foreign function interface you can really do anything in Haskell!
so it just boils down to strictness even when we're talking LLMs?
I agree with you about fast failure being a nice feature , but I also think that if you're TDDing a bunch of stuff and it fails in some categorical way , well then the test suite was lazy.
> so it just boils down to strictness even when we're talking LLMs?
The article describes what I've been doing for the past few months - I did small python projects in the past because of the ecosystem: I couldn't possibly write a ton of the stuff required for the things I wanted to do, so I leaned into python because someone already wrote it for me. Quality of deps was mostly ok for the happy paths, but always a chore to patch the broken ones.
Nowadays I tell Claude what I want to build and I always ask it whether rust is a good choice for it. It'll pick up the right crates or choose whether it should DIY, do all the plumbing, nail all the logic, and in ~30m I'll have something very solid that would have taken me 3+ weeks of part-time evening coding in python. I think the article is right and rust is the closest to the "best language" we have for LLM coding at the moment: the strict typing and the tooling dramatically reduce the output space for LLMs, and 99% of errors have a clear, precise explanation that is actionable, and the compiler helps you a lot there too.
I think it also boils down to the fact that you cannot reliably and quickly answer "why is this arg None?" in languages like python without figuring out the call graph and evaluating possible states and inputs/outputs. Rust makes all that explicit and forces you handle it, which I feel dramatically cuts the time an LLM needs to spend figuring out why it's broken or what to do next. EDIT: The fact that you get memory safety on top of all this and it's handled by the compiler is yet another advantage for LLMs: the logic that gets written is simpler to reason about, because if you try to mutably access the same variable in two different places, the compiler will feed this back to the LLM at build time. In other languages that would be a "code smell" or would require static analysis.
Strictness is a quality for software and a chore for humans, and of course the stricter you are at representing your logic and your state machine, the less ways a program can break. LLMs writing in rust give you the strictness without the chore part, and it's a very good deal from my point of view.
If you are using TDD with any recent model and even local models (qwen3.5+), you alleviate most of the issues mentioned.
Note that:
Writing code, then tests
Is not equivalent to:
Writing tests, then code
My anecdotal (sample size 1) experience is not consistent with this. I code fast. Refactor fast. My stuff doesn't break. But my methodology isn't the same as other's.
i have bad news
Lay it on. I love to collect other's anecdotes and see where they align (or disagree)
I figure a big part of it is that SWE-Bench is the target benchmark for programming and it's all python.
Python being the language LLMs are best at predates SWE-Bench by years.
Yeah, last year I discovered that AI writes better rust than C, so I switched to rust and it made some quick good code that it couldn't do in C.
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.
Better question is why use any code? Generate random functions and select based on measuring the distribution of output of these functions against metrics of interest. A pure black box of instruction that is more performant than any verbose code or algorithm we could come up with, because all we select for is performance above all. Directed evolution essentially of the codebase, generated through mutation and selection, just like everything else on planet earth.
Isn’t the answer usually - because the same ai said python is the right language for it?
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.
Yes, and wondering why all the AI tooling is written in node.
AI/ML stuff: Python
Personal: Rust/Go based on criticality of being able to glean code quickly, or memory usage, etc
I had agents code up an app for me in Swift a while back and the entire experience was so much better than your typical Python experience. The agents took full advantage of the compiler and static typing. There were far fewer bugs than expected.
Many here propose replacing Python with more performant, but less familiar languages - mostly Rust, Go. But I find the argument that the AI - HUMAN interface is the most important. A simple version of this is “no, stick with Python if that’s what you know”. A more interesting version is “use this new found AI leeway to move up the abstraction level”, “try something more expressive and human oriented”, “make a DSL and parser that suits the domain (and focuses the AI)”. Despite being a minority language, Raku is ideal for these aspects (esp with built in Grammars and general kitchen sink repartee) and works surprisingly well with most popular LLMs.
First one to vibe code a language for LLMs, by LLMs, wins a cookie?
Well don't ask us. If AI writes your code, why not ask it? You could probably make it write a whole article for ya.
Claude writes java pretty well, and faster than Rust. It's a great middle ground for some projects. I've switched back from Rust to Java for some things.
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...
It certainly makes sense to use python for ML or data science.
I still use TypeScript because I know it best. When AI makes a mistake, I can find the bug much faster. For me, the speed of writing code doesn't matter as much as the speed of fixing it.
As others have said, the main benefit with Python over Rust is library support especially with ML features. The other gap as I see it with Rust is the lack of native flexible UI support. The nice thing about Rust though is it can serve as a very fast and stable core for an app and offload specifics to TS and Python as their strengths allow, so you get the best of all worlds.
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.
This post resonates. I recently built a little web service to scratch an itch I've been having and after discussing the options with Claude we settled on Go, and honestly it's been fantastic. Highly performant, native threading, dead simple to deploy with containers. And I don't even know how to read or write Go.
Go is fun, you should actually learn it
Oh man... I like go because it is compiled, performant, strong and statically typed. But "fun" is not something I would say about it. The ergonomics of error handling, lack of ternary operator and other stuff that compiled 30yo languages already had ...
I did go through the Go tutorial many many years ago, but it's been so long I don't remember anything. I do remember it was an enjoyable process though, and I'd love to pick it up again.
Because AI creates unmaintainable messes in any language, and ergonomic ones help humans clean up.
Never mind cleaning up, you also have to understand the language just to judge and review the LLMs output. How else are you to separate good design and implementation from a bad one?
I know a couple languages fairly well: C, Perl, Python, Bash. I never formally learned Go, but as a test of AI coding, I started some vibe coded projects in Go. It worked very well: the code is minimal, there's few dependencies, and it compiles down to a static app. But most importantly, I can actually read the Go code and understand basically what it's doing. I can also use LLMs to critique the code if I'm uncertain. The big benefit of Go is the simpler language and "batteries included" standard library. This leads to fewer dependencies and less lines of code, which improves overall AI output. In theory, AI should be able to write better code faster in Go than in another language like Rust.
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.
> I have a co-worker who's basically not a programmer, but got multiple implementations of applications working sooner than our dev teams
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.
Here's the kicker: The devs spent nearly 5 months on a solution, and it ended up being so crap it was abandoned. The multiple vibe-coded solutions were all better.
Of course language was the barrier, that's part of why it was always hard to hire people. It takes years to get good at a particular language, and most people are idiots from bootcamps who learned a single framework.
Really agree. Python is popular because it's easy for human to implement. But now if the coder becomes AI, then Rust would be preferable for agent, just like Python for human. In addition, it brings better performance.
This seems sort of like asking whether a chatbot should answer you in English or Japanese. Obviously, it should use whichever language you understand. If you understand Python best, why not write code in Python?
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 think it is an interesting question what kind of programming language one needs for an era of agents. It is clear that the programming language that was designed for humans is not necessarily the best for AI-driven software development. I guess the qualities one would want is some formal correctness guarantees, high performance. A question is whether this language is Rust or it is possible to design a better new language.
So you have a chance to be able to read the absurdly barroque code AI produces.
you mean baroque-n code?
… because model tool calls is non-standard, so Python as the only tool call available works wonders
(Joke but also not a joke)
Assuming you are thinking about software architecture and looking under the hood, you are likely to be reading much more code than before. Python is really nice on the eyes and you can easily get a good grasp of what the code is doing. Plus, it's dynamically but strongly typed, so what you see in the code is usually what you get.
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.
This idea is already being taken to the next step in labs; why generate code?
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-...
Didn’t Tencent do a study comparing AI performance across about 20 languages showing that Elixir was the top performer?
Once you are over a certain threshold it’s more about the average quality of training data than the quantity.
Lately I just have Claude build most things in Rust, it's really amazing. I tried Go, but I found it wasn't as good--Rust really does to me feel like Python. That said, it still struggles with the same class of errors of building complex systems. I've tried using TLA+, Alloy, and other things but haven't found the trick yet. The best I've found is reimplementing all external systems in memory and e2e testing everything extensively, without reimplementing the tests become unusably slow, and Claude can rewrite huge surface areas with ease--it's somewhere between mocking and literally just reimplementing the external systems.
I have been wondering on a similar thing; am looking for feedback:
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.
Great question. And I don’t think that Python, Ruby and PHP have a good answer. Scripting languages cater to human weaknesses. The 10-100x perf cost was never really worth it but now it’s impossible to justify.
Python is rather a UI for human logic comprehension. A mathematical notation of logics. Not a code to drive computer.
And prompt does not replace that.
Interesting question.
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.
I'm still not sure. Would love thoguhts on this.. but in this new ai world we are in... is it better to go fullstack typescript? or go with proven mature frameworks? .net, ruby, django, etc? Seems TS is moving fast but maybe its time to not reach for the shiny object and stick with proven tech? or in 5 years will we regret it?
The main risk-of-regret is: How will you feel when/if the $20/month plan costs $2000/month?
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.
If you're using GenAI, you should go through the process of selecting an optimal tech stack for each solution, but also take into consideration that Claude and other services probably the most knowledge of python, javascript, and typescript with go, rust, java, and c# following closely behind. Consider what you're building and what elements of the tech stack is optimal for your problem-space.
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.
The ideal language for AI coding:
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.
I use Lisp for my projects
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.
Golang. People trash it for being verbose on errors but it's an extremely readable language and it's almost like bash, only much stronger typed and with a very rich stdlib (so it's not likely you'll need a library for a quick script).
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.
In a world where AI supposedly can write in any language, Go is much better choice than TypeScript. Imagine contemplating for more than a few seconds a choice between simple, fast, cross-compilable language, and a TypeScript -> JavaScript -> Interpreter -> JIT stack.
If you don't know Go, it's more efficient to learn it than to waste the hardware resources of thousands to stay within JavaScript.
Absolutely. And in this same thread I am noticing people offering Java (lol). Yeah, we all need 1.5s startup time for one-off scripts, surely.
Well, these days a small CLI program in Java (say, ls) starts up cold, runs, and terminates in ~70ms, not 1500ms, but yeah, sometimes 70ms is too long to wait for a script.
People never believe me when I say it but I start noticing scripts needing 75-100ms to start. Modern hardware is ultra fast; I want my programs to make full use of it. I got no patience for tech or people who keep insisting "it's not much, it'll not kill you". Well duh, obviously it will not but that's not the point and never was. I want stuff to work between my blinking my eyes and I have achieved that hundreds of times over the course of my career.
That's perfectly fine, and I totally understand people who don't want to sit and wait 70ms for their script to finish running (that 70ms is not the time it takes to start), but let's not turn a <40ms startup into 1.5s. Now, it is true that if you want to launch a minimal HTTP server in Java you may need to wait ~100ms, which may be too long for you, but is also a far cry from 1.5s.
It is, but I am still quoting what I saw before, it was not a fantasy. I don't deny it's likely better nowadays, sure, but I remain moderately skeptical because JVM is still a runtime that needs to boot.
Then again, Golang has one as well, though it does manage to start it up faster it seems.
This is just Kotlin. Strongly typed, more concise than Java or Go (and probably Typescript), less likely to blow up at runtime than Typescript, epic tooling, plenty of public code, and a library for basically anything because JVM.
And needs the JVM to start for 1.5s before you get any results. Sure.
Golang or just shell scripts.
The JVM takes tens of milliseconds to boot up, not a second and a half.
Obviously it depends on a bunch of factors but -- not on my machines. They are all with Intel and AMD CPUs and I don't use M-series Macs.
Never saw an instantly starting JVM in my life though.
Java runs a Hello World, cold, in a packaged JAR, in about 40ms. What you've seen isn't JVM startup but programs that do a lot at initialisation (like MS Word), as many Java programs like to do (because they often expect to run for a long time, so they don't care about startup time).
I have not worked with Java in a long time but I seem to remember that most Java programs also accrue a good amount of dependencies and some of them have their own init routines.
That adds up, fast. No idea how is it nowadays, admittedly. Maybe a ton of optimization work was done.
> Concise since you have to review a lot more code
Isn't readability what matters here? Conciseness isn't the same thing.
C. At least with Gemma 4 it does a fine job. Writes good error checking. Writes memory management. Mostly straightforward and easy to read. A lot of libraries. Runs everywhere.
I’d also argue it needs to compile fast/ have fast static analysis. Feedback loops like this are super helpful for agents
Type safety feels like the big one; anything you can shift to static/compile-time regimes benefits agents immensely.
There are two working LLM axes. Critic strength: how much the language catches before runtime. Sensor strength: how good the empirical feedback loop is. LLMs benefit from both, but the sensor axis often is undervalued.
Type safety is great, but you can't just quietly disregard the benefits some dynamically typed languages provide; that would be completely ignoring that different tasks weight the two axes differently.
Systems code, performance-critical code, code where correctness across all cases matters more than exploration: parsers, compilers, network protocols, data structures - statically typed languages (like Rust) give you an edge here. The compiler's depth pays for the verbosity, and exploration is less of the work because the problem shape is known up front.
For stuff like building a web scraper, or rapidly prototyping, or exploratory scripts, something like Rust would be actively bad. You cannot poke at a live browser (you can with Clojure). Async Rust adds another layer of type complexity. The signal-to-noise for "figure out what is on the page" collapses entirely.
If I were picking a single language for general LLM-assisted work, weighted across task types, it would be Clojure (or Elixir), with OCaml as the most interesting alternative if the ecosystem were stronger.
Using Clojure and Elixir and LLMs are fantastic with both. Sure, if I get to a super-stable situation then maybe I'd consider moving to Rust (or Jank?), but for now I'm just so happy with Clojure and Elixir in this new world. I'm solving new problems with fully bespoke architecture so the flexibility is key. Clojure for business logic and most DB. With Elixir, it's the actor model and hand-holding as I'm using it for the web layer. I bet Ruby on Rails would also shine for some cases, prob most CRUD for example.
What made you use Clojure for business logic and DBs rather than using Elixir for everything? The JVM?
> Critic strength .... Sensor strength
that's a nice breakdown
I think there's something key you get at in terms of the combo of dynamic environment + type safety maximising both. With a dynamic environment, the LLM can do a lot of interrogation to understand the problem space on the fly. I've witnessed agents sort out pretty complex issues through `python -c "..."`, `groovy -e "..."`, executing snippets of code with Node etc which is much less accessible if they have to compile it first. They can also inject logging code that interrogates the runtime as well (what type do we really have at line 1003?) etc which works better with runtimes that have deep introspection capabilities.
Java?
Was thinking the same. Modern Java is similar or at least quite a bit closer to many other less verbose languages. Not like your dad's Java anymore.
I assume this is why things like PyO3 are popping up? If so, sort of a fascinating way to compartmentalize new rust code into legacy .py code in lieu of a refactor, or at least, a way to do a staggered refactor and eat the elephant in bites :)
Devs still have to maintain this code, the Python devs can definitely get the LLM to write (some kind of) Rust, but when it goes wrong and you hit the wall with the LLM then they can will have to learn Rust which might take a while, this sounds like a bit of a project risk.
Because LLMs fuck it up near-constantly and I need to review it
For me, whether it's AI or my own handcrafted artisanal code, the choice of language comes down to what has the least friction. This means I turn to vite/react for a lot of frontend requirements, and that the backend will be in nodejs or python, because those are easier for me to debug than writing an equivalent application in C++ or Rust.
The article applies to a narrow case of a totally green field application that's going to be completely vibecoded. This is the only case where you reasonably can be indifferent to what the language is, and so you can abandon familiar Python and go with unfamiliar Rust. (If you _are_ familiar with Rust, the point of the article is moot.)
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...
You’re behind the state of the art. I’m not exaggerating when I say AI can diagnose and solve those issues for you too.
For the utilities I write it is faster to iterate without having to compile. When I get to the point where I'm done adding changing features, and performance is an annoyance I can always ask the AI to "rewrite this in Go". (I've never gotten to that point.)
So I can fix it when it breaks. I don’t understand anyone shipping real code without human review.
Give it 2 years, the ‘Blame the AI ‘ incidents will increase. Like an unfaithful partner you’ll always return to it
https://arxiv.org/pdf/2508.09101
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
1) python is one of the foremost trained upon languages
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.
Because I can understand and edit that code by hand if I need to.
I stay for the libraries
This point only makes sense if you ship AI code without reviewing it. And if you're shipping AI code without reviewing it, you're going to run into much bigger problems than Python performance limitations.
1) I still have to comprehend it.
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.
If all the libraries are rust as the article claims having the top layer in Python probably makes even less difference.
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.
I dislike Go but I have to admit, it's a great language for AI generated code. Simple enough, it compiles quickly and it performs meh-well enough for most applications.
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.
Clojure is better. REPL + immutable defaults.
you still need to look at the code oneday so id say c++ still would be a preferred target language even for ai. i know i hear a lot about rust but im still getting the idea its a niche language overall. i know people love it and point out its advantages, but sometimes good enough is good enough (i.e. c++)
Also easier to ship a binary like a cli
A somewhat contrarian/pessimistic view: The hardest thing in any future of LLM generated code is going to be the verification step, and especially types of verification that require humans which are going to be the most expensive.
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.
The LLMs just churns out non-idiomatic slop in any language.
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.
One thing to consider:
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.
As always, "it depends."
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... :)
Why use any general programming language at all? Just write it in assembly or binary. Skip the middleman bro
a) Python (and Node) comprise the largest training set for all the models, so you are likely to get way better accuracy, especially with local models
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.
Let's go through some of the arguments, in no particular order:
> 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.
MISSION
FUCKING
ACCOMPLISHED
Perl might just be the most token efficient language