Item — HN Reader

I really liked this part:

In December 2024, during the frenzied adoption of LLM coding assistants, we became aware that such tools tended—unsurprisingly—to produce Go code in a style similar to the mass of Go code used during training, even when there were newer, better ways to express the same idea. Less obviously, the same tools often refused to use the newer ways even when directed to do so in general terms such as “always use the latest idioms of Go 1.25.” In some cases, even when explicitly told to use a feature, the model would deny that it existed. [...] To ensure that future models are trained on the latest idioms, we need to ensure that these idioms are reflected in the training data, which is to say the global corpus of open-source Go code.

7 comments

PHP went through a similar effort a while back to just clear out places like Stackoverflow of terrible out of date advice (e.g. posts advocating magic_quotes). LLMs make this a slightly different problem because, for the most part, once the bad advice is in the model it's never going away. In theory there's an easier to test surface around how good the advice it's giving is but trying to figure out how it got to that conclusion and correct it for any future models is arcane. It's unlikely that model trainers will submit their RC models to various communities to make sure it isn't lying about those specific topics so everything needs to happen in preparation of the next generation and relying on the hope that you've identified the bad source it originally trained on and that the model will actually prioritize training on that same, now corrected, source.

This is one area where reinforcement learning can help.

The way you should think of RL (both RLVR and RLHF) is the "elicitation hypothesis[1]." In pretraining, models learn their capabilities by consuming large amounts of web text. Those capabilities include producing both low and high quality outputs (as both low and high quality outputs are present in their pretraining corpora). In post training, RL doesn't teach them new skills (see E.G. the "Limits of RLVR"[2] paper). Instead, it "teaches" the models to produce the more desirable, higher-quality outputs, while suppressing the undesirable, low-quality ones.

I'm pretty sure you could design an RL task that specifically teaches models to use modern idioms, either as an explicit dataset of chosen/rejected completions (where the chosen is the new way and the rejected is the old), or as a verifiable task where the reward goes down as the number of linter errors goes up.

I wouldn't be surprised if frontier labs have datasets for this for some of the major languages and packages.

[1] https://www.interconnects.ai/p/elicitation-theory-of-post-tr...

[2] https://limit-of-rlvr.github.io

I believe you absolutely could... as the model owner. The question is whether Go project owners can convince all the model trainers to invest in RL to fix their models and the follow up question is whether the single maintainer of some critical but obscure open source project could also convince the model trainers to commit to RL when they realize the model is horribly mistrained.

In Stackoverflow data is trivial to edit and the org (previously, at least) was open to requests from maintainers to update accepted answers to provide more correct information. Editing is trivial and cheap to carry out for a database - for a model editing is possible (less easy but do-able), expensive and a potential risk to the model owner.

I know Claude will read through code from Go libraries it has imported to ensure it is doing things correctly, but I do have to wonder for other languages and those small libraries, if we'll start seeing things like AGENT_README.md a file that describes the project, then describes what functionality is where in the code, and if necessary drills on a source file by source file basis (unless it's too massive - context limits are still limits). In any regard, I could see that becoming more common. Especially if you link to said file from the README.md for the model to go to. ;)

I think this can be fixed more generally by biasing towards newer data in model outputs and putting more weight on authoritative sources rather than treating all data the same. So no one needs to go in and specifically single out Go code but will instead look at new examples which use features like generics from sources like Google who would follow best/better practices than the rest of the codebase.

They're particularly bad about concurrent go code, in my experience - it's almost always tutorial-like stuff, over-simplified and missing error and edge case handling to the point that it's downright dangerous to use... but it routinely slips past review because it seems simple and simple is correct, right? Go concurrency is so easy!

And then you point out issues in a review, so the author feeds it back into an LLM, and code that looks like it handles that case gets added... while also introducing a subtle data race and a rare deadlock.

Very nearly every single time. On all models.

> a subtle data race and a rare deadlock

That's a langage problem that humans face as well, which golang could stop having (see C++'s Thread Safety annotations).

For Go, there is https://pkg.go.dev/gvisor.dev/gvisor/tools/checklocks. There are some missing things from C++ Thread Safety annotations, but those could be added.

Go has a pretty good race detector already, and all it (usually) takes to enable it is passing the -race flag to go build/test/run/etc.

No language protects from dead lock.

Not true! There are a fair number of them, and they're even reasonably general-purpose, e.g. https://www.ponylang.io/

Most that can recall achieve this by simply not having any locks at all. That's feasible with some careful design.

Outside proof-oriented languages though, I'm not aware of any that prevent livelocks, much less both. When excluding stuff that's single threaded but might otherwise qualify, e.g. Elm. "Lack of progress" is what most care about though, and yeah that realm is much more "you give up too much to get that guarantee" in nearly all cases (e.g. no turing completeness).

I probably agree that they don't protect you from all deadlocks, but some languages protect you from some dead locks.

You should be using rust... mm kay :\

Doing concurrency in Rust was more complex (though not overly so) than doing it in Golang was, but the fact that the compiler will outright not let me pass mutable refs to each thread does make me feel more comfortable about doing so at all.

Meanwhile I copy-pasted a Python async TaskGroup example from the docs and still found that, despite using a TaskGroup which is specifically designed to await every task and only return once all are done, it returned the instant theloop was completed and tasks were created and then the program exited without having done any of the work.

Concurrency woo~

Yeah, I'd really have liked to see something like [Trio](https://trio.readthedocs.io/en/stable/) gain mid more attention in Python. Structured approaches can prevent a huge amount of those issues, and Python code is absolutely riddled with concurrency problems in my experience. Much more so in practice than other languages except maybe javascript (when including async ordering mistakes). It makes it a real nightmare to try to build anything actually reliable.

So you are trying to explain concurrency to the folks who implemented CSP in both Plan9 and Go. Interesting. I should return "cspbook.pdf" back.

Good use case for Elixir. Apparently it performs best across all programming languages with LLM completions and its concurrency model is ideal too.

https://autocodebench.github.io/

This is the exact opposite of my experience.

Claude 4.6 has been excellent with Go, and truly incompetent with Elixir, to the point where I would have serious concerns about choosing Elixir for a new project.

Shouldn't you have concerns picking Claude 4.6 for your next project if it produces subpar elixer code? Cheapy shot perhaps, but I have a feeling exotic languages will remain more exotic longer now that LLM aided development is becoming the norm.

We live in different realities.

Opus and Sonnett practically writes the same idiomatic elixir (phoenix, mind you) code that I would have written myself, with few edits.

It's scary good.

I have run into that a lot which is annoying. Even though all the code compiles because go is backwards compatible it all looks so much different. Same issue for python but in that case the API changes lead to actual breakage. For this reason I find go to be fairly great for codegen as the stability of the language is hard to compete with and the standard lib a powerful enough tool to support many many use cases.

The use of LLMs will lead to homogeneous, middling code.

Middling code should not exist. Boilerplate code should not exist. For some reason we're suddenly accepting code-gen as SOP instead of building a layer of abstraction on top of the too-onerous layer we're currently building at. Prior generations of software development would see a too-onerous layer and build tools to abstract to a higher level, this generation seems stuck in an idea that we just need tooling to generate all that junk but can continue to work at this level.

But Go culture promulgates this practice of repeating boilerplate. In fact this is one of the biggest confusion points of new gophers. "I want to do a thing that seems common enough, what library are you all using to do X?". Everyone scoffs, pushes up their glasses and says, "well actually, you should just use the standard library, it's always worked just fine for me". And the new gopher is confused because they really believe that reinventing the wheel is an acceptable practice. This is what leads to using LLMs to write all that code (admittedly, it's a fine use of an LLM).

LLMs have always been great at generating code that doesn't really mean anything - no architectural decisions, the same for "any" program. But only rarely does one see questions why we're needing to generating "meaningless" code in the first place.

This gets to one of my core fears around the last few years of software development. A lot of companies right now are saddling their codebases with pages and pages of code that does what they need it to do but of which they have no comprehension.

For a long time my motto around software development has been "optimize for maintainability" and I'm quite concerned that in a few years this habit is going to hit us like a truck in the same way the off-shoring craze did - a bunch of companies will start slowly dying off as their feature velocity slows to a crawl and a lot of products that were useful will be lost. It's not my problem, I know, but it's quite concerning.

The "LLMs shouldn't be writing code" take is starting to feel like the new "we should all just use No-Code."

We’ve been trying to "build a better layer" for thirty years. From Dreamweaver to Scratch to Bubble, the goal was always the same: hide the syntax so the "logic" can shine. But it turns out, the syntax wasn't the enemy—the abstraction ceiling was.

Where are the amazing no-hassle, no-boilerplate tools from last generation? Or the generation before that? Give me a break: it's easy to post this but it's proven very hard to simply "pick the right abstraction for everyone".

It does. I’ve been writing Go for long enough, and the code that LLMs output is pretty average. It’s what I would expect a mid level engineer to produce. I still write code manually for stuff I care about or where code structure matters.

Maybe the best way is to do the scaffolding yourself and use LLMs to fill the blanks. That may lead to better structured code, but it doesn’t resolve the problem described above where it generates suboptimal or outdated code. Code is a form of communication and I think good code requires an understanding of how to communicate ideas clearly. LLMs have no concept of that, it’s just gluing tokens together. They litter code with useless comments while leaving the parts that need them most without.

I am also of the opinion that LLMs are still pretty bad at what's called "Low level design" - that is structuring functions and classes in a project. I wonder if a rule like torvalds' "No more than 4 levels of indentation" might make them work better.

Do LLMs generate code similar to middling code of a given domain? Why not generate in a perfect language used only by cool and very handsome people, like Fortran, and then translate it to once the important stuff is done?

This might work if Fortran were portable, or if only one compiler were targeted.

middling code, delivered within a tolerable time frame, budget, without taking excessive risk, is good enough for many real-world commercial software projects. homogeneous middling code, written by humans or extruded by machines, is arguably even a positive for the organisation: lots of organisations are more interested in delivery of software projects being predictable, or having a high bus-factor due to the fungibility of the folks (or machines) building and maintaining the code, rather than depending upon excellence.

You might even say that LLMs are not capable of understanding a brilliant language but we want to use them to build good software. So, the language that we give them has to be easy for them to understand and easy to adopt.

This is a rob pike reference for the uninitiated: https://news.ycombinator.com/item?id=30688969

I'm not sure if that's a criticism or praise - I mean, most people strive for readable code.

LLM generated code reminds me of perl's "write-only" reputation.

Does it really? Because I see some quite fine code. The problem is assumptions, or missing side effects when the code is used, or getting stuck in a bad approach "loop" - but not code quality per se.

In all honesty I've only used LLMs in anger with Go, and come away (generally speaking) happy with what it produced.

For a few years, yeah. Eventually it will probably lead to the average quality of code being considerably higher than it was pre-LLMs.

I'd prefer we start nuking the idea of using LLMs to write code, not help it get better. Why don't you people listen to Rob Pike, this technology is not good for us. Its a stain on software and the world in general, but I get it most of ya'll yearn for slop. The masses yearn for slop.

I totally agree. I read threads like this and I just can’t believe people are wasting their time with LLM’s.

The masses yearn to not have to fiddle with bs for rent and food

I definitely see that with C++ code Not so easy to "fix", though. Or so I think. But I do hope still, as more and more "modern" C++ code gets published

battle of my life. several times i’ve had to update my agent instructions to prefer modern and usually better syntax to the old way of doing things. largely it’s worked well for me. i find that making the agents read release notes, and some official blog posts, helps them maintain a healthy and reasonably up-to-date instructions on writing go.