Hacker News Reader

Comment by hollowturtle | Hacker News Reader

> Yes, at work, where you are paid to produce value for a business, the economics of purely hand-written code have changed

I'd love to genuinely love these blog posts but they are all the same! They're all like "now that machine are better than us $xyzargument". I'm so pissed off when I read so blantant lies like AI "writing code better than me" maybe better than you! But in general it's just a big lie.

> the economics of purely hand-written code have changed

who are you to know? I still hand code and produce a lot of value because I do care and polish everything I do. There are some pieces of the software that I wrote that I never touched again, they work perfectly, no errors no customer complaints

View on HN

I don't know that it's a lie. The code that I can get Claude Sonnet to produce is pretty much what I would and that's two iterations behind. We've vibecoded scripts recently that genuinely are better than what I would have written and I don't think I'm terrible.

There are a lot of mediocre programmers aren't there, I wouldn't be surprised if on average Claude is doing better than what exists. It's not a "lie", anyways, just a point of view.

> We've vibecoded scripts recently that genuinely are better than what I would have written and I don't think I'm terrible.

I'm ok with vibecoding scripts, I love programming the machine and making the end user happy and not dealing with bash! I don't judge scripts quality as long as they are tools in the workflow, I dont care about that, unless they will somehow be part of the production. The name Script implies that, there isn't any architecture and Claude Opus sucks at architecture

This. I hate that Claude Code is taking the fun part of my job, but the reality is that Claude Code Opus 5 can one-shot a script in 15 seconds that would have taken me 45 minutes to write.

I don't begrudge the business for making the prudent choice of making me use AI, I just sort of lament that what I spent years enjoying and getting good at was so utterly derivative that it was perfect for the LLMs. It would be insulting if it weren't just a fact.

Every time you don’t write the script you lose a little of the ability to do so.

Just like if you don’t read you will lose the ability to.

These skills are not innate to us and can be forgotten over time.

I wonder if it will get to the point where people don’t even know what to ask for. Or how to judge if what the machine produces could have been better. Or if it does what they think it does.

> Just like if you don’t read you will lose the ability to.

This is wholly untrue.

Not wholly. It’s under study. I don’t mean in the sense that you can’t read words anymore. I mean in the sense that comprehension and attention go down.

Okay but notice that “Losing the ability to” as-is is not even in the ballpark of “comprehension and attention go down”, we are not mind readers..

Joke's on you, I never knew how to write bash.

Yes, sadly, coding is a skill that can and (before long) will be forgotten by all but a niche dedicated few. It will go the way of assembly language, smith-ing, weaving, bread-baking, meat- / fish-gutting, hunting, remembering phone numbers, driving manual (and, before long, driving in human mode), hand-writing, horse-riding, wood-chopping... you get the picture. The price of progress!

...are you sure that Claude Sonnet writes code that you'd write?

For me, it writes code that mostly works, which is great. It insists on adding some of the least readable comments I've seen. It also likes to pad out the code to about double or triple what I would write.

It doesn't seem to have that little voice that says "hey, this approach sucks" so it needs the harness or the prompter to tell it to go back and refactor. Even Fable and Opus struggle with this.

> It doesn't seem to have that little voice that says "hey, this approach sucks" so it needs the harness or the prompter to tell it to go back and refactor.

Well, yes. For humans, the "this approach sucks" voice is still very separable from the thought process that just figures out the syntax and algorithms etc. The important thing is whether the overall system has this feedback and that it's having the desired effect. Better harnesses are progress in the same way that more capable models are.

And yeah, there's a wide range in the standards that programmers hold, in practice. Plus they don't at all agree on what's "better" than which.

Regarding the "this approach sucks" voice, there's an important difference between humans and AI which I've noticed and which I didn't explain well.

The general human approach to programming at most companies I worked at prior to 2024 was to head off in vaguely the right direction and refine the approach as we went. There might be a design and a plan, but at some point we just winged it and fixed our work after the fact.

This seems to be the worst way of getting an AI to do anything. They do better if they can one-shot an implementation from a clear spec then only need to fix bugs. Better models can handle a less clear spec but they still can't really iterate in a major way. To have an AI model build a bigger thing, you one-shot increasingly detailed specs and decompose them into smaller ones.

I can only speculate, but my guess is that this limit is pretty fundamental, yeah. However capable the models get, they can't read your mind. The kind of iteration you have in mind is difficult because they don't have a basis to iterate upon; to the extent that they could pretend, that gets into what people have been calling "taste".

And even if they somehow got better taste, it wouldn't be your taste. If that were acceptable, why bother prompting anything in the first place? At that point the system might as well choose its own projects entirely. But I don't think humans would want to live in that world; even if not paperclip-maximized, we'd eventually become an underclass.

That's assuming basically unlimited capacity for the models to get more powerful. If there are limits (and I hope there are!) then it'd be even more certain that projects continue to require human guidance.

But even despite that, I hope I don't sound AI-psychotic if I suggest: the models are capable of at least assisting with that spec-writing and decomposition.

You two are talking past each other. At a lower level LLMs can write better code than humans. At the architectural level they make a mess.

You’re forgetting the “economics” part of the argument. Let’s assume that you are actually producing better code than say Opus 5.5. You’re probably slower, you very likely have less “stamina,” and you probably cost more than 200x more per month than a $20 subscription.

You’d have to be pretty unbelievably stellar to produce code 200x better than Opus. I can get a lot of business value out of not-great $20/mo coding. The people writing our checks do not care about the same things we care about.

And I say this as someone who is very upset about what this stuff is doing to my career and hobby.

I don't get this. I work in a part of the industry where "passes 80% of the test suite" is worth approximately $0/mo, which is the best opus can do without a human hand holding it.

But I read the OP's argument to mean that your responsibility as an engineer is now to understand the tasks an agent can accelerate. And in many cases that might be the task that you previously enjoyed, like working out some tricky variable length encoding serialization scheme, or tracking down some nefarious race condition by stepping through code and reasoning. (The eureka moment is a big dopamine hit for me, dunno about you.)

If you can't get Opus 5.5 to pass more than 80% of your test suite, then you're probably holding it wrong. Not "holding it wrong" as in Apple's "we designed a defective product and try to put the blame on you", but as in *"this is an incredibly sophisticated tool which takes skill and practice to use to its fullest extent".

LLMs, even in a good harness, are deceptively simple, but in reality rather complex tools. They're good at grasping intent, but they can't read your mind, they might not have access to the full context of your business, and they require learning how to use them properly.

> "this is an incredibly sophisticated tool which takes skill and practice to use to its fullest extent".

In other words, hand holding.

No, not "hand holding". If you insist on anthropomorphizing the AI, "guiding it on a high level" would be more fitting. In a team of 1 architect and 5 software developers, Opus 5.5 + Claude Code replaces the 5 developers, not the architect. You don't have to "hold its hand" anymore than you have to hold the hand of a competent software developer — though you will have to figure out a workflow that suits you personally, including configuring Claude Code to follow that workflow (through skills, for example) so you can work ergonomically with it.

I have never worked on a team with one thinker and 5 minions, so, again, more evidence I work in a different part of the industry than wherever your line of thought originates.

> one thinker and 5 minions

Your error lies in thinking Opus 5.5 is equivalent to a "minion". The "5 software developers" in my example were meant to stand in for "reasonably competent, above-average developers", and not that the architect is the only one on the team who can "think".

Of course, it may well be the case that you are working in an elite sector of the industry where even the bottom rungs are filled by the top 20% of software developers. Possible, but unlikely, so I'm still leaning towards "you're holding it wrong".

"an incredibly sophisticated tool" it's matmuls that generate logits

No, your argument is literally backwards.

Opus has to be 200x worse for them to sell that subscription. If their AI is as good as they claim, they should charge way more, like one dollar less per month than a human developer costs.

> If their AI is as good as they claim, they should charge way more, like one dollar less per month than a human developer costs.

Only if they had a monopoly, which they thankfully don't.

The industry is in a desperate race to throw away what little quality it had (could you imagine if other fields of engineering had catastrophic failures at the rate software does?) in favour of "look at all the lines of code I generated today but don't actually read them please". It's always been a fad-heavy field, but this time it's really gone fully off the deep end. The whole "it's better than me" is just re-asserted endlessly. Nobody even pretends to care about evidence. I wish the boosters would go ahead and "leave me behind" already, I've been waiting years for it to happen now. The zone has never been more flooded, I think.

> I wish the boosters would go ahead and "leave me behind" already, I've been waiting years for it to happen now.

I completely agree. The fact that the software from places that are very vocally pro-AI (microsoft, salesforce, etc...) is still so full of bugs is really contrary evidence.

> The whole "it's better than me" is just re-asserted endlessly. Nobody even pretends to care about evidence.

The evidence that the AI tool is better than oneself? I don't know about you, but I think I'm a pretty good judge of whether Claude Code is better than I am, having seen my own work and Claude's firsthand.

Then you have a skill issue :)

It depends on your definition of better. Some businesses value speed above almost all else. Especially early-stage businesses where time to market and test ideas is paramount. I think in those cases, LLMs are better than probably any human.

On the other hand, if you mean tight, performant, simple, well-factored code, then yeah. LLMs simply aren't there, and may never be. On the other hand, maybe that matters less than we thought it did. Time will tell.

> Especially early-stage businesses where time to market and test ideas is paramount

To me it's just crazy of how dumb it is to test an idea with a vibecoded shitty app. I'm not saying it should be perfect from day zero, let's pretend we are talking about cakes, how crazy it would sound you will tell people to try and judge your new idea of a cake you don't even know how it taste or what's inside?

I agree. You shouldn't test an idea with a vibecoded shitty app. That doesn't mean you shouldn't test it with an LLM-generated, refined, well-tested (by human users) app. I don't think it's a boolean. I think it's a continuum.

Those companies, and those people within otherwise decent companies, have existed since long before the advent of LLMs. Avoiding their influence is a career skill as important as the code itself. And I'm not saying you can just avoid them for your whole career, they're everywhere and are being supercharged by LLMs.

It's disappointing that the prevailing sentiment now is that architecture is less important than we thought. Software is always going to be getting more important, and all of it should be well thought out. If these people had these tools in 2004 they might've made thousands of fragmented buggy social media startups and actually slowed progress on that front.

How is this different than testing an idea with a janky, rushed, handcoded protype or MVP? That's how a lot of product development has always been done.

Well there’s at least two reasons people might not buy your cake: they don’t want cake, or the cake is bad.

For the first group, it doesn’t matter how good your cake is. If you spend a year perfecting it, you won’t sell them any more cakes. And for many products, this group is basically 100% of customers.

I think their major issue is actually naming. I don't think they have issues making well factored code because you can tell them how to factor it. You can even tell them how to name it. But that takes time. People want the functionality now without having to wait. So the well factored and well named stuff is a functoin of how much time you want to spend giving overview. I don't think you're going to have many cases where its faster to type each thing yourself. You can describe the details in a prompt which can be referenced as you review.

The issue is nobody is spending time in review because the stuff coming out "just works". So if no one is reading the code, what is the big deal with naming? The big deal is we need to know what our later sessions are talking about and we need linguistic separation of concerns.

A higher level form of programming, if you will. Linguistic precision seems to be where the specificity can be lost or won. The naming will reflect that with some degradation. But humans will also degrade the spirit of the specifications as well in many cases!

Hopefully this post makes some sense to someone.

> I don't think they have issues making well factored code because you can tell them how to factor it

I mean, who's doing the factoring then?

Great point and I think it highlights the issue with talking about this topic because the terminology is so black and white, "AI made it" or "I made it". But the truth is more like I got on a bicycle that I did not build and rode a marathon through treacherous land that I have not been through before, but I have been through other treacherous land before and I had some scar tissue that helped me through this time. Jobs was on the money when he referred to the computer as being like a bicycle. ("for the mind" ... to finish the quote)

Yes, I think of a computer as a bicycle and AI as a motorcycle. It will definitely get you there faster, but is it better for you?

Also "great point" is sadly a phrase that has been co opted by sycophantic AI so I can no longer take it seriously.

Well I’m sorry you feel that way. I don’t see “great point” used often but in terms of its agreeableness I can see how one could take it that way. (In fact I had a feeling someone would say that, but it was a concise wording for what I intended so I left it in).

I see it as a multiplier (the bicycle metaphor) and yes certainly getting across town on a bicycle or across the country on a plane is less cumbersome on your health than walking every where.

I don’t ride motorcycles, I rode a scooter for a long time doing a delivery job as a younger man and the speed and instability (of two wheeled motorized vehicles) along with lack of visibility to others creates a dangerous combination.

And yes I can see the corollaries there to AI usage especially in these days when the excitement of the acceleration is being found seductive by many and without much visibility into the longer term consequences.

> The issue is nobody is spending time in review because the stuff coming out "just works".

People used to be able to write code that was horrible by all the standard metrics but still "just worked" (or at least appeared to). From the stories I've heard, it would get into production all the time, too. But if it's a bigger problem now, then necessarily we know how to fix it, so.

I guess the excitement of the new toy is just leading to some irresponsibility.

I don‘t know that we know how to fix it so well after the fact, but rather we know how to avoid it which is by writing understandable code the first time round.

But we don’t always do that and then things get rewritten and broken etc

Welcome to life in software dev 2026, not so different from 1996

Although "lies" might be too strong a word, I agree with you that the the AI-is-the-future people haven't done nearly enough to prove their claims. But that's nothing new in our industry, really.

I would recommend "The Leprechauns of Software Engineering" by Laurent Bossavit for an insightful and entertaining overview of just how much of software "engineering" is based on "truths" that we don't question.

I read so blantant lies like AI "writing code better than me" maybe better than you! But in general it's just a big lie.

Those weren't lies. The AI does produce better code than those people. They simply weren't very good programmers to begin with.

Exactly. Someone on HN put it really succinctly earlier in the year: "LLMs look like they are good at things you are not good at."

I think this goes a long way to explain why there is such a huge difference in opinion about the quality of AI-written code: Some people say LLMs write code far better than they could, and others say the code is always terrible and requires a great deal of fixing before it's acceptable. How can two different developers say those very different things about the same product?

It might be a tough pill to swallow, but if you think your LLM is producing better code than you could produce... well, I might have some bad news for you about your own skills.

And the number of those people will grow as more people lean too heavily into AI coding.

In my case I am more than capable of producing better code than the LLM, but I have never been the fastest on my team. My development speed was acceptable because the quality and robustness of the output was worth the time that thoughtfulness took. I was also able to improve the comprehensibility and robustness of my colleagues' code during code reviews.

But in the new environment of LLM-assisted coding, the LLM produces acceptable code so much faster than I can write it. Just taking time to understand and retool the LLM's output means an order of magnitude lower productivity and it doesn't feel like that's acceptable in this era.

Folks I know are landing 10+ PRs a day and nobody is reviewing anymore, and these are not companies that would have been skipping code review 2 years ago. The whole culture is turning against taking the time to comprehend and reason about our own work.

Here's the thing, it doesn't have to produce code better than me.

There is a wide range of code that will do the same job. 15s vs 2h for code 60% as good? 10 minutes for code 80% as good? 10 minutes today and 10 minutes tomorrow to fix something that is 30% as good?

A lot of people are going to draw that line in a different place and now that people can make that trade off you will quickly find they will and it probably wont fall where you would like it to.

At the end of the day a business pays you to produce value that will earn them money / customers. They don't pay you to write good code. Sometimes those two align and both sides are happy but they don't and often wont.

> At the end of the day a business pays you to produce value that will earn them money / customers.

I wonder, given LLM's ability to "produce code faster" are they also producing value faster? I don't see it as an end user. Software is getting worse and worse, buggier and buggier, every release. I don't see it at my company either. We're producing massive quantities of code constantly, but still every project is on fire and every line of business is way behind schedule.

> are they producing value faster?

I don't know. Maybe.

Software might be getting worse but we are not far enough along yet for it to matter. Most people use less than 10 websites total, the internet has shrunk. It will take a while for it to be obvious in the other backend services that they dont know they use. People are really tolerant of buggy shit and some software has already been basically wiped out. Dev tooling is a blood bath right now, things like tailwind and all sorts of little CI tooling is fast disappearing. I wonder what happens to the open core stuff like elasticsearch / gitlab / redis where they gate some features in the enterprise offering that could probably be implemented by LLM pretty fast.

> We're producing massive quantities of code constantly, but still every project is on fire and every line of business is way behind schedule.

Right but did they start altering the schedule when AI entered the picture? We have barely just passed "tokens == productivity".

Also people tolerate quite a bit of buggy shit. Anthropic is down to two 9s (aha) and people still use them. There is a github ci failure post on here like every other day.

> Anthropic is down to two 9s (aha) and people still use them. There is a github ci failure post on here like every other day.

Anthropic and GitHub’s consumers are software teams. I wonder what would happens if something like SolidWorks was that unreliable.

The answer is no, at least so far. There's been a few big quantitative studies (CircleCI, a few others I can't remember) which show this - more code, more branches, more PRs, but actually less going to production. It's not even like a 10% increase on average or something, it's less! And yet anecdotally people say it gives you 2x, 3x, 10x speedups.

That's not even fully true - I can't draw but I can tell LLM 'art' looks like shit

Please have a look through https://www.astralcodexten.com/p/how-did-you-do-on-the-ai-ar... , and then keep in mind that that's from almost two whole years ago.

The strongest evidence for AI art being too uncanny is that people are still paying artists to draw cartoon porn.

Eh, my understanding is that the models are supposed to have guardrails against that. But also, getting better results generally requires refinement, a critical eye, and the skill to write out what's wrong. I'd guess these skills aren't common among that audience. But much more so, it likely defeats the purpose. The point is to be lazy (that's at least part of why people are willing to spend money on it).

On the other hand, I'd also expect the most degenerate among us to not care about the uncanny details. So who knows, really.

If you can't get it to output good code I'm afraid it's probably a skill or prompting issue. There are programmers that are almost certainly better than you that are using LLMs regularly to produce good code. Do you think the Linus is a bad programmer?

Linus has motive behind his position on AI. It's his operating system that's at the center of the AI revolution. The Linux Foundation is involved with several AI projects and has a lot of members that are all-in with AI.

He is not a programmer anymore. He admitted that he mostly reviews, tests and gives suggestion now. He still does a little coding, but usually small patch that hard to say if he is good or bad programmer.

I would say he is a good PM/visionary.

Wow you really think that it is a "skill or prompting issue" in a field where model capability is growing over time?

It's not a "skill or prompting issue", it's either you have a software factory or you don't and suffer the poor quality.

If it's a matter of whether the person using it is good at programming or not, then it's just the programmer programming with more steps.

I like to think I hold very high standards; I find that I can fairly easily get the LLM most of the way to where I want the code to be, but there's definitely a Pareto cutoff.

I do feel that Codex and Claude are both bad at something I'm good at: writing documentation. But it doesn't matter, because no human reads it. I'm spending a lot of time rewriting its output just for it to be piped back in again.

And that's just sorting out my self-image. Convincing someone who controls the pursestrings that I'm adding value was hard already.

>How can two different developers say those very different things about the same product?

Well the Dunning-Kruger effect could certainly explain the cohort that thinks they're skills are amazing and unbeatable.

>I'd love to genuinely love these blog posts but they are all the same! They're all like "now that machine are better than us $xyzargument". I'm so pissed off when I read so blantant lies like AI "writing code better than me" maybe better than you! But in general it's just a big lie.

Yeah and this is what pisses me off about the software factory deniers. Without the software factory to guide the agent and reject its code, how exactly is it supposed to do better than an experienced developer who already internalized the company workflows? All the Agent sees is its context window, it's not a fully formed human that can learn the workflow over time.

If I guide the Agent file by file and function by function, then review and reject the code, the code quality becomes reasonably good, but when I let it do whatever it wants, the code quality reaches all time lows and it just compounds on itself.

I also noticed that the code quality heavily depends on the underlying software ecosystem with Javascript often having the worst code quality because it presumably got trained on minified code as well.

Agent harnesses do not solve software development in a hands off way. You still have to baby sit them. Every single time I use an auto review based workflow (aka vibe coding) it just doesn't work.

Thank you! It's made me lose a lot of respect for my field how credulous everyone is to any argument that starts out with a fig leaf of pretending to be on your side.

"The AI labs are scared of their own models, surely they must be credible sources on how powerful those models are"

"They are asking the government for regulation, they must be even more responsible than the average company"

"They are claiming to be a software engineer who automated their job with AI, surely I have to buy their course"

There are many boilerplate pieces of software that an LLM can generate faster than a human, but nearly half a decade into the LLM era there is no evidence that it's leading to the apocalyptic end of software organizations. And that's all with the subsidy in place that is now disappearing. Everyone online is lying for money, and that money is coming out of the same investor-subsidized marketing budgets.

Yeah all the viral crap from LLMs are low quality rip offs of stuff that already exists. There is no innovation from what I've seen

Reply on news.ycombinator.com