Hacker News Reader

Comment by bigupthewhole | Hacker News Reader

While this was true 1+ years ago, the agents don't write bad code anymore. They might not write according to your preferences, but you can also have it adjust to your style.

Also, humans didn't evolve to write or read code, even the best programmers are terrible at it from an objective point of view, if we compare it to what is theoretically possible.

View on HN

The agents still write bad code unless my colleagues and the people who submit PRs to the repos I am responsible for somehow use agents from last year.

And, no, it is not a question about style or preferences, the code is messy and contains to me as a maintainer obvious bugs.

>still write bad code unless my colleagues and the people who submit PRs

It's interesting you feel comfortable to comment with such authority on the capability of a tool you don't even use... Lisp is a bad programming language because my 10 year old brother wrote some really bad stuff, btw!

I use them. The top 3 models still write terrible code, regardless of harness. They’ll litter the code with safety checks for scenarios we’ve already guaranteed won’t happen. And they’ll solve problems you didn’t ask them to, and they’ll come up with a working, but convoluted solution. They’ll sometimes write really weird code if you’re using them in an obscure language.

Basically, they still need a human guiding them pretty closely unless you’re happy to just ship a mountain of vibe coded mush.

That said, I haven’t tried Astra in earnest yet. We’ll see in a few weeks whether or not it is a step change.

I'm fully convinced that people who think agents write decent code now simply don't read the code. Their bar is "it runs, so it's fine".

Just as they produce writing that isn't structured well for humans, they do the same with code. Sure, if you're willing for that code to be write-only and for agents to be the sole interface through which you interact with it, that works fine, but it is not better than what a professional, skilled human could produce. It's often full of pointless crap and design decisions that over index heavily on local structure that make very little sense when global system structure is considered.

In my experience, the people who are enamored with LLMs for programming fall into one of three categories. Shills, people who could never program decently to begin with, and people who don't care about the quality of their work. Everyone I know who actually is a good programmer and cares about doing a good job has found that LLMs make too many mistakes to be trusted, and avoid them if they can.

Most people who are good at their job understand the greatest value comes from deep understanding and internalization. Some people have been in roles where alienation from the code is institutionalized, and they've been playing that role for so long that the idea of having a strong understanding and near prescient mental map of a 2 million line codebase sounds absurd.

I think the question is less about the capabilities of LLMs and their outputs (because obviously that can change), but because they represent a workflow which codifies exactly that role I mention. You fulfill the role of a manager more than a programmer. You have a vague idea of the codebase, but lack any kind of intuitive mechanical understanding. There are many cases in which I can see an issue, identify the problem, and have it fixed faster than you could prompt an LLM to investigate. After a certain point of familiarity, all reasoning is local reasoning. That's the power of mechanical sympathy.

Many, many large shops don't give a shit. They've long since galloped into the dark woods of bureaucratic know-nothing hell. Let the LLM figure that shit out. Those shops are already optimizing for maximal output and churn, so the value proposition seems obvious. I've been smugly saying it for years, but LLMs have me saying it quite a lot more these days: "Sure looks like it sucks down there!"

So why doesn't Anthropic write a good Claude Code in Rust already? Why don't we get a Codex with the polish level of Sublime text editor?

These companies have many billion dollar investments and yet their main product is a bloated meh.

They don’t hand write it because it would be an embarrassment. At the same time, Claude Code doesn’t really need to be any better than it is while Anthropic has great models no one else has.

Didn't evolve to write or read code? Obviously! Humans designed and created programming languages to be human readable and writable. It didnt exist to be evolved toward.

> Also, humans didn't evolve to write or read code

That's a weird way of looking at it. Programming lives at the end of a well-worn path uncountably many millenia long, defined entirely by our biology. We didn't "evolve to do it" in the same way a duck didn't evolve to swim in a specific river.

Reply on news.ycombinator.com