My AI skeptic friends are all nuts | Hacker News Reader

My AI skeptic friends are all nuts

fly.io

2,356 pointstabletcorry2,826 comments

I think this article is pretty spot on — it articulates something I’ve come to appreciate about LLM-assisted coding over the past few months.

I started out very sceptical. When Claude Code landed, I got completely seduced — borderline addicted, slot machine-style — by what initially felt like a superpower. Then I actually read the code. It was shockingly bad. I swung back hard to my earlier scepticism, probably even more entrenched than before.

Then something shifted. I started experimenting. I stopped giving it orders and began using it more like a virtual rubber duck. That made a huge difference.

It’s still absolute rubbish if you just let it run wild, which is why I think “vibe coding” is basically just “vibe debt” — because it just doesn’t do what most (possibly uninformed) people think it does.

But if you treat it as a collaborator — more like an idiot savant with a massive brain but no instinct or nous — or better yet, as a mech suit [0] that needs firm control — then something interesting happens.

I’m now at a point where working with Claude Code is not just productive, it actually produces pretty good code, with the right guidance. I’ve got tests, lots of them. I’ve also developed a way of getting Claude to document intent as we go, which helps me, any future human reader, and, crucially, the model itself when revisiting old code.

What fascinates me is how negative these comments are — how many people seem closed off to the possibility that this could be a net positive for software engineers rather than some kind of doomsday.

Did Photoshop kill graphic artists? Did film kill theatre? Not really. Things changed, sure. Was it “better”? There’s no counterfactual, so who knows? But change was inevitable.

What’s clear is this tech is here now, and complaining about it feels a bit like mourning the loss of punch cards when terminals showed up.

[0]: https://matthewsinclair.com/blog/0178-why-llm-powered-progra...

This article does not touch on the thing which worries me the most with respect to LLMs: the dependence.

Unless you can run the LLM locally, on a computer you own, you are now completely dependent on a remote centralized system to do your work. Whoever controls that system can arbitrarily raise the prices, subtly manipulate the outputs, store and do anything they want with the inputs, or even suddenly cease to operate. And since, according to this article, only the latest and greatest LLM is acceptable (and I've seen that exact same argument six months ago), running locally is not viable (I've seen, in a recent discussion, someone mention a home server with something like 384G of RAM just to run one LLM locally).

To those of us who like Free Software because of the freedom it gives us, this is a severe regression.

One thing that I find truly amazing is just the simple fact that you can now be fuzzy with the input you give a computer, and get something meaningful in return. Like, as someone who grew up learning to code in the 90s it always seemed like science fiction that we'd get to a point where you could give a computer some vague human level instructions and get it more or less do what you want.

I'm not a skeptic, but I keep LLMs on a short leash.

This is a thoughtful article. Thanks `tptacek

My LLM use is: 1 - tedious stuff; web pages interacting with domain back end. 2 - domain discovery.

In a recent adventure, I used Claude 4 to tease out parameters in a large graph schema. This is a combination of tedium and domain discovery (it's not my graph and I'm not a domain expert). In the first day, Claude uncovered attributes and relations no other LLM or Google search uncovered. And it worked!! The next day, I allowed it to continue. After a bit, results didn't pass the sniff test.

I checked into details of Claude's thinking: it decided to start making up schema attributes and inventing fallback queries on error with more made up attributes. It was "conscious" of its decision to do so. By the time I caught this, Claude had polluted quite a bit of code. Sure, plenty of well placed git commits helped in rolling back code...but it's not quite that simple..over the many git commits were sprinkled plenty of learnings I don't want to toss. It took another two days of carefully going through the code to pull out the good stuff and then roll things back. So now I'm at day five of this adventure with cleaned up code and notes on what we learned.

I suspect continual improvements on tooling will help. Until then, it's a short leash.

I use AI every day, basically as a "pair coder."

I used it about 15 minutes ago, to help me diagnose a UI issue I was having. It gave me an answer that I would have figured out, in about 30 minutes, in about 30 seconds. My coding style (large files, with multiple classes, well-documented) works well for AI. I can literally dump the entire file into the prompt, and it can scan it in milliseconds.

I also use it to help me learn about new stuff, and the "proper" way to do things.

Basically, what I used to use StackOverflow for, but without the sneering, and much faster turnaround. I'm not afraid to ask "stupid" questions -That is critical.

Like SO, I have to take what it gives me, with a grain of salt. It's usually too verbose, and doesn't always match my style, so I end up doing a lot of refactoring. It can also give rather "naive" answers, that I can refine. The important thing, is that I usually get something that works, so I can walk it back, and figure out a better way.

I also won't add code to my project, that I don't understand, and the refactoring helps me, there.

I have found the best help comes from ChatGPT. I heard that Claude was supposed to be better, but I haven't seen that.

I don't use agents. I've not really ever found automated pipelines to be useful, in my case, and that's sort of what agents would do for me. I may change my mind on that, as I learn more.

The reaction to this article is interesting. I have found AI to be useful in software contexts that most people never exercise or expect based on their intuitions of what an LLM can do.

For me, a highly productive but boring use of LLMs for code is that they excel at providing midwit “best practice” solutions to common problems. They are better documentation than the documentation and can do a lot of leg work e.g. Linux syscall implementation details. My application domains tend to require more sophisticated solutions than an LLM can provide but they still save a lot of rote effort. A lot of software development exists almost entirely in the midwit zone.

Much more interesting, they are decent at reducing concepts in literature to code practice for which there are no code examples. Google and StackOverflow turn up nothing. For example, I’ve found them useful for generating specialized implementations of non-Euclidean computational geometry algorithms that don’t really exist in the wild that I’ve ever seen. This is a big win, it literally turns months of effort into hours of effort.

On the other hand, I do a lot of work with algorithms that don’t exist in literature, never mind public code, with extremely performance-engineered implementations. There is an important take away from this too: LLMs are hilariously bad at helping with this but so are human software developers if required to do the same thing with no context.

Knowledge for which there is little or no training data is currently a formidable moat, both for LLMs and humans.

It's fascinating how over the past year we have had almost daily posts like this one, yet from the outside everything looks exactly the same, isn't that very weird?

Why haven't we seen an explosion of new start-ups, products or features? Why do we still see hundreds of bug tickets on every issue tracking page? Have you noticed anything different on any changelog?

I invite tptacek, or any other chatbot enthusiast around, to publish project metrics and show some actual numbers.

I’m AI neutral but the writing style here is pretty dismissive, and - to match the tone of the article - annoying as fuck.

Most completely reasonable objections to LLMs were totally dismissed.

My main concern is not even mentioned in this article and there are hardly any comments here addressing it: Privacy / allowing 3rd parties to read and potentially train on your proprietary source code.

I've used LLMs to crank out code for tedious things (like generating C-APIs and calling into poorly documented libraries) but I'm not letting them touch my code until I can run it 100% locally offline. Would love to use the agentic stuff but from what I've heard it's still too slow to run on a high end workstation with a single 4080.

Or have things got better lately, and crucially is there good VisualStudio integration for running local agents / LLMs?

I love LLMs, and I really like programming with Cursor, but I never managed to get the "agents with tons of stuff in their context" mode to work for me. I use Cursor like a glorified code completer, 4-5 lines at a time, because otherwise the LLM just makes too many mistakes that compound.

If you let it run in the "write my code for me" mode, and ask it to fix some mistake it made, it will always add more code, never remove any. In my experience, in the end the code just ends up so brittle that the LLM will soon get stuck at a point that it never manages to overcome some mistake, no matter how many times it tries.

Has anyone managed to solve this?

> Does an intern cost $20/month? Because that’s what Cursor.ai costs.

> Part of being a senior developer is making less-able coders productive, be they fleshly or algebraic. Using agents well is both a both a skill and an engineering project all its own, of prompts, indices, and (especially) tooling. LLMs only produce shitty code if you let them.

A junior developer often has negative value to a team, because they're sapping the time of more senior developers who have to help train them, review code, fix mistakes, etc. It can take a long while to break even.

The raw cost of Cursor's subscription is surely dwarfed by your own efforts, given that description. The actual calculous here should be the cost to corral Cursor, against the value of the code it generated.

Hundreds of comments. Some say LLMs are the future. Others say they don't work today and they won't work tomorrow.

Videogame speed running has this problem solved. Livestream your 10x engineer LLM usage, a git commit annotated with it's prompt per change. Then everyone will see the result.

This doesn't seem like an area of debate. No complicated diagrams required. Just run the experiment and show the result.

I have one very specific retort to the 'you are still responsible' point. High school kids write lots of notes. The notes frequently never get read, but the performance is worse without them: the act of writing them embeds them into your head. I allegedly know how to use a debugger, but I haven't in years: but for a number I could count on my fingers, nearly every bug report I have gotten I know exactly down to the line of code where it comes from, because I wrote it or something next to it (or can immediately ask someone who probably did). You don't get that with AI. The codebase is always new. Everything must be investigated carefully. When stuff slips through code review, even if it is a mistake you might have made, you would remember that you made it. When humans do not do the work, humans do not accrue the experience. (This may still be a good tradeoff, I haven't run any numbers. But it's not such an obvious tradeoff as TFA implies.)

It's been so much more rewarding playing with AI coding tools on my own than through the subtle and not so subtle nudges at work. The work AI tools are a walled garden, have a shitty interface, feel made to extract from me than to help me. In my personal stuff, downloading models, playing with them, the tooling, the interactions, it all been so much more rewarding to give me stable comfortable workflows I can rely on and that work with my brain.

The dialog around it is so adversarial it's been hard figuring out how to proceed until dedicating a lot of effort to diving into the field myself, alone, on my personal time and learned what's comfortable to use it on.

Title nitpick: The amount is people who care about AI for coding assistance is a relative minority. For everyone else, there's 'AI', which has a huge branding problem. 'AI' is filling all search results with trash, and creating trash websites full of trash to fill up the rest of the search results. It's generating trash images to put at the top of every blog post. It's eating up all the server load with scraping. It's what's been fooling my dad every day on Facebook.

When people are sick of AI, this is what they are talking about. AI hype people ignore this perspective each and every time.

It doesn't matter how great your paper Mill's paper is, if you're dumping PCBs in the river, people are going to quite rightly get pissed off.

I agree with the main take in this article: the combination of agents + LLMs with large context windows + a large budget of tokens to iterate on problems can probably already yield some impressive results.

I take serious issue with the "but you have no idea what the code is" rebuttal, since it - to me - skims over the single largest issue with applying LLMs anywhere where important decisions will be made based on their outputs.

To quote from the article:

  People complain about LLM-generated code being 
  “probabilistic”. No it isn’t. 
  It’s code. It’s not Yacc output. It’s knowable. The LLM 
  might be stochastic. But the LLM doesn’t matter. What 
  matters is whether you can make sense of the result, and 
  whether your guardrails hold.

  Reading other people’s code is part of the job. If you can’t metabolize the 
  boring, repetitive code an LLM generates: skills issue! How are you handling the 
  chaos human developers turn out on a deadline?

The problem here is that LLMs are optimized to make their outputs convincing. The issue is exactly "whether you can make sense of the result", as the author said, or, in other words: whether you're immune to being conned by a model output that sounds correct but is not. Sure, "reading other people’s code is part of the job", but the failure modes of junior engineers are easily detectable. The failure modes of LLMs are not.

EDIT: formatting

The author defends mediocore code, yet wrote this piece:

https://fly.io/blog/vscode-ssh-wtf/

Where he dunks on how SSH access works in VSCode. I don't know. The code and architecture behind this feature may well be bananas, but gets the work done. Sounds like a clear case of mediocority. I wonder how does he reconcile those two articles together.

For me this is more of a clickbait. Both of the articles. With that in mind, if I am nuts for being sceptical of LLMs, I think it is fair to call the author a clickbaiter.

The argument that I've heard against LLMs for code is that they create bugs that, by design, are very difficult to spot.

The LLM has one job, to make code that looks plausible. That's it. There's no logic gone into writing that bit of code. So the bugs often won't be like those a programmer makes. Instead, they can introduce a whole new class of bug that's way harder to debug.

The argument seems to be that for an expert programmer, who is capable of reading and understanding AI agent code output and merging it into a codebase, AI agents are great.

Question: If everyone uses AI to code, how does someone become an expert capable of carefully reading and understanding code and acting as an editor to an AI?

The expert skills needed to be an editor -- reading code, understanding its implications, knowing what approaches are likely to cause problems, recognizing patterns that can be refactored, knowing where likely problems lie and how to test them, holding a complex codebase in memory and knowing where to find things -- currently come from long experience writing code.

But a novice who outsources their thinking to an LLM or an agent (or both) will never develop those skills on their own. So where will the experts come from?

I think of this because of my job as a professor; many of the homework assignments we use to develop thinking skills are now obsolete because LLMs can do them, permitting the students to pass without thinking. Perhaps there is another way to develop the skills, but I don't know what it is, and in the mean time I'm not sure how novices will learn to become experts.

There's something that seems to be missing in all these posts and that aligns with my personal experience trying to use AI coding assistants.

I think in code.

To me, having to translate the into natural language for the LLM to translate it back into code makes very little sense.

Am I alone in this camp? What am I missing?

Curious how he reconciles this:

> If you build something with an LLM that people will depend on, read the code. In fact, you’ll probably do more than that. You’ll spend 5-10 minutes knocking it back into your own style.

with Joel Spolsky's fundamental maxim:

> It’s harder to read code than to write it.

https://www.joelonsoftware.com/2000/04/06/things-you-should-...

To quote an excellent article from last week:

> The AI has suggested a solution, but the added code is arguably useless or wrong. There is a huge decision space to consider, but the AI tool has picked one set of decisions, without any rationale for this decision.

> [...]

> Programming is about lots of decisions, large and small. Architecture decisions. Data validation decisions. Button color decisions.

> Some decisions are inconsequential and can be safely outsourced. There is indeed a ton of boilerplate involved in software development, and writing boilerplate-heavy code involves near zero decisions.

> But other decisions do matter.

(from https://lukasatkinson.de/2025/net-negative-cursor/)

Proponents of AI coding often talk about boilerplate as if that's what we spend most of our time on, but boilerplate is a cinch. You copy/paste, change a few fields, and maybe run a macro on it. Or you abstract it away entirely. As for the "agent" thing, typing git fetch, git commit, git rebase takes up even less of my time than boilerplate.

Most of what we write is not highly creative, but it is load-bearing, and it's full of choices. Most of our time is spent making those choices, not typing out the words. The problem isn't hallucination, it's the plain bad code that I'm going to have to rewrite. Why not just write it right myself the first time? People say "it's like a junior developer," but do they have any idea how much time I've spent trying to coax junior developers into doing things the right way rather than just doing them myself? I don't want to waste time mentoring my tools.

One of the biggest anti LLM arguments for me at the moments is about security. In case you don't know, if you open a file with copilot active or cursor, containing secrets, it might be sent to a server a thus get leaked. The companies say that if that file is in a cursorignore file, it won't be indexed, but it's still a critical security issue IMO. We all know what happened with the "smart home assistants" like Alexa.

Sure, there might be a way to change your workflow and never ever open a secret file with those editors, but my point is that a software that sends your data without your consent, and without giving you the tools to audit it, is a no go for many companies, including mine.

The problem with LLMs for code is that they are still way too slow and expensive to be generally practical for non-trivial software projects. I'm not saying that they aren't useful, they are excellent at filling out narrow code units that don't require a lot of context and can be quickly or automatically verified to be correct. You will save a lot of time using them this way.

On the other hand, if you slip up and give it too much to chew on or just roll bad RNG, it will spin itself into a loop attempting many variations of crap, erasing and trying over, but never actually coming closer to a correct solution, eventually repeating obviously incorrect solutions over and over again that should have been precluded based on feedback from the previous failed solutions. If you're using a SOTA model, you can easily rack up $5 or more on a single task if you give it more than 30 minutes of leeway to work it out. Sure, you could use a cheaper model, but all that does is make the fundamental problem worse - i.e. you're spending money but not actually getting any closer to completed work.

Yes, the models are getting smarter and more efficient, but we're still at least a decade away from being able to run useful models at practical speeds locally. Aggressively quantized 70b models simply can't cut it, and even then, you need something like 10k tps to start building LLM tools that can overcome the LLM's lack of reasoning skills through brute force guess and check techniques.

Perhaps some of the AI skeptics are a bit too harsh, but they're certainly not crazy in the context of breathless hype.

The privacy aspect and other security risks tho? So far all the praise I hear on productivity are from people using cloud-hosted models.

Claude, Gemini, Copilot and and ChatGPT are non-starters for privacy-minded folks.

So far, local experiements with agents have left me underwhelmed. Tried everything on ollama that can run on my dedicated Ryzen 8700G with 96GB DDR5. I'm ready to blow ~10-15k USD on a better rig if I see value in it but if I extrapolate current results I believe it'll be another CPU generation before I can expect positive productivity output from properly securely running local models when factoring in the setup and meta.

The reason that I personally don't use LLMs was not addressed by the article:

I haven't found a way to use it that makes me develop faster.

The articles talks about "tedious code." If you need to generate a large static value table or something, then OK an LLM might give you a really fast result and cut through the tedium. Most of us were already writing short scripts to do that. I'm open to the possibility that an LLM can do it faster. But it's such a rare requirement that the productivity gains are truly negligible here even if they can. And in those cases, it's obvious what the repetitive task needs to be. I often find myself writing the code by hand to be quicker than coming up with a prompt to get it to write the code that I then need to review for correctness.

The article then mentions scaffolding. Things like "bookkeeping" when it comes to creating and setting up a new repo (whatever he means by that). This is why I have, historically, been a big fan of frameworks and generators. Point being, this is already a solved problem and I haven't found a way to further improve the state of this world with LLMs. LLMs might be an alternate tool that work just as well. But they haven't made my existing daily workflow any faster. Setting up new repos is also something that is done so rarely that even if an LLM netted a 100% increase in efficiency, it wouldn't really impact much.

I am an AI "skeptic" but I'm not a naysayer. I do use LLMs regularly. I just don't use them for developing code because I have yet to find a problem that they solve for me. Don't get me wrong, there are problems that they can solve... I just haven't come across any solutions to previously-unsolved problems. Meaning I can swap an existing solution for an LLM-based one... and it is a valid solution... but I don't observe any increase in productivity from doing so. The existing solution was already working fine.

I am genuinely looking forward to the day when this changes. When I identify a single existing problem without an existing solution that LLMs solve for me when developing software. I just have yet to come across one.

This happened with the introduction of smartphones too. Every slashdot post had a haughty and upvoted ‘why would i want such a thing!!!’.

It was obviously huge. You could see it taking off. Yet a lot of people proudly displayed ignorance and backed each other up on it to the point that discussion around the topic was often drowned out by the opposition to change. Now today it takes minutes of playing with ai coding agents to realise that it’s extremely useful and going to be similarly huge.

Resistance to change is not a virtue!

My 5 cents would be that LLMs have replaced all those random (e.g. CSS, regex etc) generators, emmet-like IDE code completion/generator tools, as well as having to google for arbitrary code snippets which you'd just copy and paste in.

In no way can AI be used for anything larger than generating singular functions or anything that would require writing to or modifying multiple files.

Technically you might be able to pull off having AI change multiple files for you in one go, but you'll quickly run into sort of "Adobe Dreamviewer" type of issue where your codebase is dominated by generated code which only the AI that generated it is able to properly extend and modify.

I remember when Dreamviewer was a thing, but you essentialyl had to make a choice between sticking with it forever for the project or not using it at all, because it would basically convert your source code into it's own proprietary format due to it becoming so horribly messy and unreadable.

Regardless, AI is absolutely incredible and speeds up development by a great deal, (even) if you only use it to generate small snippets at the time.

AI is also an absolute godsend for formatting and converting stuff from anything and to anything - you could e.g. dump your whole database structure to Gemini and ask it to generate an API against it; big task, but since it is basically just a conversion task, it will work very well.

Here are two routine problems I have to solve at the moment. Can any of the current LLM systems do either?

1. Input is an 256x256 pixel elevation map stored as a greyscale .png file, and a minimum and maximum elevation. A pixel value of 0 corresponds to the minimum elevation, and a pixel value of 255 corresponds to the maximum elevation. Read in the .png file and the elevation limits. Then construct a 256x256 floating point array of heights. From that array, construct a triangle mesh with X and Y dimensions 0..255. Perform a mesh reduction operation on the triangle mesh to reduce the number of triangles. Mesh reduction must not generate holes in the mesh. From the reduced mesh, generate a glTF file where the UV parameters run from 0.0 to 1.0 along the X and Y axes.

2. Given four glTF files constructed as above, corresponding to four quadrants of a larger square, construct a single 511x511 mesh which combines all four input meshes to cover a larger area. Because the input meshes are 0..255, not 0..256, there will be gaps where the four quadrants meet. Fill those gaps with reasonable triangles. Perform a mesh reduction as above. From the reduced mesh, generate a glTF file where the UV parameters run from 0.0 to 1.0 along the X and Y axes.

Rust code is preferred; Python code is acceptable.

So, what service should I sign up for?

Weird to claim the llm does all the boring learning and boilerplate for you as a selling point, but then also insist we still need to responsibly read all the output, and if you can't understand it's a "skill issue".

Also the emphasis on greenfield projects? Starting is by FAR the easiest part. That's not impressive to me. When do we get to code greenfield for important systems? Reminds me of the equally absurd example of language choice. You think you get to choose? What?

Imagine all the code these agents are going to pump out that can never be reviewed in a reasonable time frame. The noise generated at the whim of bike-shedding vibe coders is going to drown all the senior reviewers soon enough. I'll call that Cowboy Coders on Steroids. Anyone with skills will be buried in reviews, won't have time for anything else, and I predict stricter code gen policies to compensate.

> An LLM can be instructed to just figure all that shit out. Often, it will drop you precisely at that golden moment where shit almost works, and development means tweaking code and immediately seeing things work better.

Well, except that in order to fix that 1% you'd need to read and understand whatever the LLM did and then look for that 1%. I get the shills just thinking about this, whether the original programmer was human or not. I'd rather just write everything myself to begin with.

The one main claim the article makes: Senior developers should not ignore the productivity gains from LLMs.

Best use of evidence is deductive: Lots of code is tedious and uninteresting -> LLMs are fast at generating lots of tedious code -> LLMs help productivity.

Weakest part of the argument: The list of rebuttals doesn't have an obvious organization to it. What exactly is the main argument they're arguing against?

It's not stated outright but because the post is bookended by references to 'those smarter than me', I think this is an argument against the shaming of developers using (and loving) LLM tools.

Which I think is fair.

Overall, the post did not add anything to the general discussion. But the popularity of the author (and fly.io posts) may make it a beacon for some.

A couple of thoughts:

On 'just read the code' - all well and good. Of course this implies insisting on team members who can read the code, will read the code and are empowered to read the code.

Otherwise orgs will try to hire pure vibe coders who aren't interested in that and only allow time for literally just instructing agents to generate code because it sounds cheaper and execs don't understand the nuance so long as it looks like product is shipping - until it all blows up and the one standing senior developer on hand is supposed to fix a prod issue buried in millions of lines of vibe reviewed code ASAP.

On 'but it's cheaper than a junior': cloud hosted LLM systems are currently massively subsidised to an absurd degree. The cost side of things is all smoke and mirrors geared towards accelerated market adoption at all costs. It's not a profitable enterprise at the model development level.

At some point that AI economy is going to expect to make that money back, and future (especially near-future) hardware advancements don't explain where all of that is going to come from.

If you are resource constrained this article will make you sad.

I do not have unlimited funds to plug in some token and burn a bunch of money when writing code.

I am gpu poor. I'm lucky that 8gb vram can run the smallest models. But the output is so poor that I lose out to anyone using a hosted service.

If anything this article shows that building great programs is less democratized than it once was.

I'm working with python and it does not work very well, LLMs usually generates at least an order of magnitude more code than what I would write. That code often uses outdated practices, does poor design choices and does not understand hints like writing code in a way that reduce cognitive load, even when we explain that it means, i.e. keep the number of classes and functions small.

It's the complete opposite of OP's main language GO, as he says:

> Go has just enough type safety, an extensive standard library, and a culture that prizes (often repetitive) idiom. LLMs kick ass generating it.

Python is an interpreted dynamically typed language and the static type checkers are not there yet (most popular 3rd parties libraries have no type hints for example). Also it allows for many different programming styles that the LLMs struggle to choose from.

1. Every extra line of code is much more risky. 2. It's much harder to verify the LLM's code.

On the other hand I think rust will be in a good place in regards to LLMs in the next few years thanks to the robustness of the language and the quality of its diagnostic messages. Those 2 attributes should compound very well.

Let’s not conflate LLM’s with AI. LLM’s are a kind of AI that can be a software engineer’s assistant, at best. But the degree of hype is undermining belief in AI among many professionals who the hypesters claim are going to be replaced. No, this iteration is not going to replace doctors or engineers. But the degree to which the hype train wants to do so is alarming.

> LLMs can write a large fraction of all the tedious code you’ll ever need to write. And most code on most projects is tedious.

You still need to read this tedious code to verify that it actually does what you want it to do. Given this, I'd much rather prefer to write the tedious code myself than having to make sense of someone else's tedious code.

I think the key premise here is that one can effectively and efficiently audit code that the LLM is producing.

I doubt that. First, human attention and speed is very limited. Second, when I see something, I am already predisposed to assume that it is right (or at the very least, my subsequent inquiries are extremely narrow and anchored around the solution I have seen presented to me.)

A lot of people who are wary of LLMs aren’t against the technology itself, but rather the narrative surrounding it. You can take advantage of the tool while turning a blind eye to the discourse.

This 16-minute, expletive-filled, edgy-old-man-trying-too-hard-to-be-cool article could easily be dismissed as yet another AI creed that somehow found its way to the top of the HN front page.

When I made a mid career change to Cobol programming in 1981 my first manager (a suit as programmers referred to them then) pointed to a book on his desk; the title,as I can best recall was Programming Without Programmers. He "You got in too late," he said. I retired from programming in 2010. The hype has a long history. I hope I'm around long enough to see how this plays out.

I'm on the side of AI is a huge productivity booster (by my guess... 10x)

But I don't want to make the claim lightly, so I did an experiment.

I signed up for copilot pro, and have been using their 'edit' feature. This is more than just their auto complete.

I set myself a goal to create a playable web game of classic frogger.

It took 4 hours with copilot "edit" and my full attention.

I didn't write a single line of code, but I did ask it to refactor and gave it a project description.

I suspect this would have taken me 4 days full time to get to this level.

Try it out:

https://dev.zebrar.com/jd/frogger/

I will create an account just to point out that the response to

> But it is an inherently plagiarist technology

Was

> Developers frequently engage in copyright infringement, and so will I, so unless if you're a lawyer, shove it up your ass

"I am a bad person so I get to continue being bad" is not the gotcha you think it is, Patrick.

"Kids today don’t just use agents; they use asynchronous agents. They wake up, free-associate 13 different things for their LLMs to work on, make coffee, fill out a TPS report, drive to the Mars Cheese Castle, and then check their notifications. They’ve got 13 PRs to review. Three get tossed and re-prompted. Five of them get the same feedback a junior dev gets. And five get merged."

I would jump off a bridge before I accepted that as my full-time job.

I've been programming for 20+ years and I've never wanted to move into management. I got into programming because I like programming, not because I like asking others to write code on my behalf and review what they come up with. I've been in a lead role, and I certainly do lots of code review and enjoy helping teammates grow. But the last fucking thing I want to do is delegate all the code writing to someone or something else.

I like writing code. Yes, sometimes writing code is tedious, or frustrating. Sometimes it's yak-shaving. Sometimes it's Googling. Very often, it's debugging. I'm happy to have AI help me with some of that drudgery, but if I ever get to the point that I feel like I spend my entire day in virtual meetings with AI agents, then I'm changing careers.

I get up in the morning to make things, not to watch others make things.

Maybe the kind of software engineering role I love is going to disappear, like stevedores and lamplighters. I will miss it dearly, but at least I guess I got a couple of good decades out of it. If this is what the job turns into, I'll have to find something else to do with my remaining years.

There are many aspects to AI push back.

- all creatives are flat against it because it’s destroying their income streams and outright stealing their intellectual property

- some technical leaders are skeptical because early returns were very bad and they have not updated their investigations to the latest tools and models, which are already significantly ahead of even six months ago

- a tech concern is how do we mentor new developers if they don’t know how to code or develop logic. LLMs are great IF you already know what you’re doing

- talent is deeply concerned that they will be reduced and replaced, going from high paying careers to fast food salaries

We have a lot of work to balance productivity with the benefits to society. “Let them eat cake,” is not going to work this time either.

Seeing everything these days being about vibe coding, I feel a little old with my VIM setup and my LSP servers who I already thought were a nice productivity increase.

The problems I have with the stuff relating to MCP is that the tech around it is developing so fast that it's hard for outsiders to catch up with what the best working setup is, for example.

What would you do, for example, if you want to selfhost this?

- which models (qwen ai coder?)

- which api (with ollama? Bolt? Aider? Etc)

- how to integrate PRs with a local gitlab/gogs/forgejo instance? Do you need another MCP agent for git that does that?

- which hardware dependencies to run it?

I am currently trying to figure out how to implement a practical workflow for this. So far I'm using still a synchronous MCP agent setup where it basically runs on another machine in the network because I have a too unperformant laptop to work with.

But how would I get to the point of async MCP agents that can work on multiple things in my Go codebases in parallel? With the mentioned PR workflows so that I can modify/edit/rework before the merges?

The author makes a lot of claims and talks always about that their opponents in the argument are not talking about the same thing. But what exactly is the same thing, which is reproducible locally for everyone?

Machine translation and speech recognition. The state of the art for these is a multi-modal language model. I'm hearing impaired veering on deaf, and I use this technology all day every day. I wanted to watch an old TV series from the 1980s. There are no subtitles available. So I fed the show into a language model (Whisper) and now I have passable subtitles that allow me to watch the show.

Am I the only one who remembers when that was the stuff of science fiction? It was not so long ago an open question if machines would ever be able to transcribe speech in a useful way. How quickly we become numb to the magic.

Can someone comment on the cost of running agentic models? Not for a company but for an individual. I tried "vibe coding" a personal project I was struggling with and left even more frustrated because I kept running into token rate limits with Claude (used inside of Zed if it matters). Did I pick the wrong model, the wrong editor, or do I just need to not be so tight with my money?

> This was the craftsman’s ‘Golden Age’ and much time and trouble was taken over the design of tools. Craftsmen were being called upon to do more skilful and exacting work and the use of tools and the interest in development had become very widespread.

Above pulled from A Brief History of the Woodworking Plane [0]. A woodworking tool that has evolved over 2,000 years. Now there are electric planers, handheld electric planers and lots of heavy machinery that do the same thing in a very automated way. If a company is mass producing kitchen cabinets, they aren't hand planing edges on boards, a machine is doing all that work.

I feel like with AI we are on the cusp of moving beyond a "Golden age" and into an "industrial age" for coding, where it will become more important to have code that AI understands vs. something that is carefully crafted. Simple business pressure will demand it (whether we like it or not).

^ A comment I made just yesterday on a different thread.

For software developers AI is like the cabinet maker that gets a machine to properly mill and produce cabinet panels, sure you can use a hand plane to do that but you're producing a very different product and likely one that not many people will care about, possibly not even your employer when they see all the other wood shops pumping out cabinetry and taking their market share.

[0] https://www.handplane.com/879/a-brief-history-of-the-woodwor...

Everything about that is true but, and that's a big BUT, the code I write with LLM I can only iterate on it with an LLM.

My mind doesn't develop a mental model of that code, I don't know where the relevant parts are, I can't quickly navigate through it and I have to reach the LLM for every small change.

Which is why I like Copilot style editing more than agents as a working model but agents are just so much more powerful and smarter thanks to everything available to them.

I tend to agree with the gist of this, namely that the change is here and "AI" presents some huge potential to save me from having to do the tedious stuff. But we do need a bit of a sanity check. I'm in the middle of trying to convince any of the AIs that I have access to to write me some simple Lua. Specifically, I'm trying to write a query that returns a certain subset of pages from the Silverbullet V2 note tool. This isn't particularly difficult, but it's become this hilarious journey to try to get the AI to figure this problem out. Silverbullet is a niche use case, v2 even more so, but a reasonably skilled developer could peruse the API documentation and come up with a pretty good starting point. AI tools? Absolutely insane wrong answers. I finally specifically asked one or two of them not to guess with their answers and they just straight up said "nah, we don't know how to do this."

Point being: there's some real power in these tools, but if the ground is not already well-trodden, they risk creating a huge time sink that could be better served just learning to do the thing yourself.

I think the hardest part is not spending the next 3 months of my life in a cave finishing all the hobby/side projects I didn't quite get across the line.

It really does feel like I've gone from being 1 senior engineer to a team that has a 0.8 Sr. Eng, 5 Jrs. and one dude that spends all his time on digging through poorly documented open source projects and documenting them for the team.

Sure I can't spend quite as much time working on hard problems as I used to, but no one knows that I haven't talked to a PM in months, no one knows I haven't written a commit summary in months, it's just been my AI doppelgangers. Compared to myself a year ago I think I now PERSONALLY write 150% more HARD code than I did before. So maybe, my first statement about being 0.8 is false.

I think of it like electric bikes, there seems to be indication that people with electric assist bikes actually burn more calories/spend more time/go farther on an electric bike than those who have manual bikes https://www.sciencedirect.com/science/article/abs/pii/S22141....

As someone who has spent the better part of today fixing the utter garbage produced by repeated iteration with these supposedly magical coding agents, I'm neither in the camp of the "AI skeptic" (at least as defined by the author), nor am I in the camp of people who thinks these things can "write a large fraction of all the tedious code you’ll ever need to write."

Maybe I'm doing it wrong, but I seem to have settled on the following general algorithm:

* ask the agent to green-field a new major feature.

* watch the agent spin until it is satisfied with its work.

* run the feature. Find that it does not work, or at least has major deficiencies [1]

* cycle through multiple independent iterations with the agent, doing something resembling "code review", fixing deficiencies one at a time [2]

* eventually get to a point where I have to re-write major pieces of the code to extract the agent from some major ditch it has driven into, leading to a failure to make forward progress.

Repeat.

It's not that the things are useless or "a fad" -- they're clearly very useful. But the people who are claiming that programmers are going to be put out of business by bots are either a) talking their book, or b) extrapolating wildly into the unknown future. And while I am open to the argument that (b) might be true, what I am observing in practice is that the rate of improvement is slowing rapidly, and/or the remaining problems are getting much harder to solve.

[1] I will freely grant that at least some of these major deficiencies typically result from my inability / unwillingness to write a detailed enough spec for the robot to follow, or anticipate every possible problem with the spec I did bother to write. T'was ever thus...

[2] This problem is fractal. However, it's at least fun, in that I get to yell at the robot in a way that I never could with a real junior engineer. One Weird Fact about working with today's agents is that if you threaten them, they seem to do better work.

By the reply count it appears this post was excellent at trolling HN's core audience :-). The author may be right, but his preface "look I'm one of you" was in a very different tone to the rest of his "now shut up and get to work" rant.

I do use the AI tools to some extent, if for no other reason than that they are currently the path of least resistance, and google+friends have lately played themselves out of the game.

He is probably right we should get acquainted with using agents before dismissing it :-).

> But all day, every day, a sizable chunk of the front page of HN is allocated to LLMs: incremental model updates, startups doing things with LLMs, LLM tutorials, screeds against LLMs. It’s annoying!

You forgot the screeds against the screeds (like this one)

I suspect a large proportion of claims made for productivity increases are skewed by the fact that the speed at which code is produced by AI makes you _feel_ productive, but these gains are largely replaced by the effort to understand, refactor, review and clean up the code. The high that you get when something "works" tends to stick more in your memory than the time when you had to spend a day cleaning up dead code, refactoring 2k line modules into a more readable project structure etc.

I'm not saying that AI can't make you productive, it's just that these claims are really hard to verify. Even the recently posted Cloudflare OAuth worker codebase took ~3 months to release (8 Mar - 20 May), producing a single file with >2k lines. Is that going to be harder to maintain than a codebase with a proper project structure that's easily parseable by a human?

Let's just imagine we're critiquing cars or planes in about 1910. They're so flawed. No one can say with certainty whether or how they will improve.

Side note of interest, from Wikipedia: "Flying Machines Which Do Not Fly" is an editorial published in the New York Times on October 9, 1903. The article incorrectly predicted it would take one to ten million years for humanity to develop an operating flying machine.

I'm mostly skeptical about AI capabilities but I also think it will never be a profitable business. Let's not forget AI companies need to recoup a trillion dollars (so far) just to break even [1].

VCs are already doubting if the billions invested into data centers are going to generate a profit [1 and 2].

AI companies will need to generate profits at some point. Would people still be optimistic about Claude etc if they had to pay say $500 per month to use it given its current capabilities? Probably not.

So far the only company generating real profits out of AI is Nvidia.

[1] https://www.goldmansachs.com/insights/articles/will-the-1-tr...

[2] https://www.nytimes.com/2025/06/02/business/ai-data-centers-...

> but the plagiarism

This entire section reads like, oddly, the reverse of the "special pleading" argument that I usually see from artists. Instead of "Oh, it's fine for other fields, but for my field it's a horrible plagiarism machine", it's the reverse: "Oh, it's a problem for those other fields, but for my field get over it, you shouldn't care about copyright anyway".

I'm all for eliminating copyright. The day I can ignore the license on every single piece of proprietary software as I see fit, I'll be all for saying that AIs should be able to do the same. What I will continue to complain about is the asymmetry: individual developers don't get to violate individual licenses, but oh, if we have an AI slurp up millions of codebases and ignore their licenses, that's fine.

No. No, it isn't. If you want to ignore copyright, abolish it for everyone. If it still applies to everyone else, it should still apply to AIs. No special exceptions for mass-scale Open Source license violations.

> Almost nothing it spits out for me merges without edits. I’m sure there’s a skill to getting a SOTA model to one-shot a feature-plus-merge!

How does this section fit in with the agent section just after? In an agentic model, isn't the merge getting done by either the model or a tool, and the retry-loops on failures would be mostly invisible?

E.g. when using Aider + Gemini Flash 2.5, probably 90% of the changes apply cleanly from my perspective (maybe half actually apply cleanly, the other half after a couple of roundtrips of Aider telling the model that the patch didn't apply). The 10% that only apply partially I usually throw away and redo the prompt, it's really rare that I start merging the code manually.

A big problem is that you're either hearing breathless over the top insane hype (or doomerism, which is breathless over the top hype taken to a dark place) or skepticism that considers AI/LLMs to be in the same league as NFTs.

Neither of these is accurate, but I guess nuanced thinking or considering anything below surface vibes is out these days.

So far after playing with them I'm using them as:

1. A junior intern that can google really really fast and has memorized a large chunk of the Internet and the library, and can do rough first-pass research and dig for things.

2. Autocomplete 2.0 that can now generate things like boilerplate or fairly pedestrian unit tests.

3. Rubber duck debugging where the rubber duck talks back.

4. A helper to explain code, at least for a first pass. I can highlight a huge piece of code and ask it to summarize and then explain and walk me through it and it does a passable job. It doesn't get everything right but as long as you know that, it's a good way to break things down and get into it.

For those things it's pretty good, and it's definitely a lot of fun to play with.

I expect that it will get better. I don't expect it to replace programmers for anything but the most boring mindless tasks (the ones I hate doing), but I expect it to continue to become more and more useful as super-autocomplete and all the other things I listed.

Fully agree with @tptacek here:

> But AI is also incredibly — a word I use advisedly — important. It’s getting the same kind of attention that smart phones got in 2008, and not as much as the Internet got. That seems about right.

However, I just don't think the AI coding part is that interesting or future-thinking. We're seeing so much more progress in semantic search, tool calling, general purpose uses, robotics, I mean, DeepMind just won a Nobel for goodness' sake.

Don't get me wrong, I use ChatGPT to write all kinds of annoying boilerplate, and it's not too bad at recalling weird quirks I don't remember (yes, even for Rust). But hard problems? Real problems? Zero shot. Novel problems? No way.

> But I’ve been first responder on an incident and fed 4o — not o4-mini, 4o — log transcripts, and watched it in seconds spot LVM metadata corruption issues on a host we’ve been complaining about for months.

I'm going to go ahead and press (X) to doubt on this anecdote. You've had an issue for months and the logs were somehow so arcane, so dense, so unparseable, no one spotted these "metadata corruption issues?" I'm not going to accuse anyone of blatant fabrication, but this is very hard to swallow.

Listen, I also think we're on the precipice of re-inventing how we talk to our machines; how we automate tasks; how we find and distribute small nuggets of data. But, imo, coding just ain't it. Donald Knuth calls computer programming an art, and to rob humanity of effecting not just coding—but any art, I'd argue—would be the most cardinal of sins.

>but it’s bad at rust

I have to say, my ability to learn Rust was massively accelerated via LLMs. I highly recommend them for learning a new skill. I feel I'm roughly at the point (largely sans LLMs) now where I can be nearly as productive in Rust as Python. +1 to RustRover as well, which I strongly prefer to any other IDE.

One point is missing for me - you get lazy. People are less and less equiped to think about complex problems the more they RELY on this for coding. Also this article is just about coding.

Not to derail, but NFT mania (part of the opening salvo in the article) was the giant shitshow that it was -not- because the concept of unique digital bits in the possession of a single owner was a bad idea (or, the concept of unique verification of membership in a club was a bad idea) -- it was a diarrhea-slicked nightmare because it was implemented via blockchains and their related tokens, which inherently peg fluctuating fiat value to the underlying mechanisms of assigning and verifying said ownership or membership, and encourages a reseller's market .. not to mention the perverse, built-in economic incentives required to get nodes to participate in that network to make the whole thing go.

Had NFTs simply been deployed as some kind of protocol that could be leveraged for utility rather than speculation, I think the story would be a complete 180. No clue personally how to achieve that, but it feels like it could be done.. except that, too, would have been completely perverted and abused by centralized behemoths, leading to a different but terrible outcome. Can you imagine if all data became non-fungible? Convince all the big identity vendors (Google, Apple, etc) to issue key pairs to users that then get used by media companies to deliver audio and video keyed only to you that's embedded with maybe some kind of temporal steganographic signature that's hard to strip and can be traced back to your key? It's not just cracking AACS once and copying the bytes. It becomes this giant mess of you literally can't access anything without going through centralized authorities anymore. Then build more anti-patterns on top of that lol. Prolly better that it was mostly just monkey JPEGs and rug pulls.

Anyways, I'm so far off topic from what's actually being discussed -- just couldn't help myself from veering into left field.

Hmm a bit of a tone shift from: https://fly.io/docs/about/open-source/

> Fly.io builds on the work of a tremendous open source infrastructure community. We want open source authors to benefit from their work because we think a healthy, thriving open source ecosystem will help us build better products.

To the fine article:

> Meanwhile, software developers spot code fragments seemingly lifted from public repositories on Github and lose their shit. What about the licensing? If you’re a lawyer, I defer. But if you’re a software developer playing this card? Cut me a little slack as I ask you to shove this concern up your ass.

> My AI skeptic friends are all nuts

My CEO friends that overstate the efficacy of LLMs are all nuts.

(I say this as someone who has been working in "AI" as part of Digital transformation roadmaps for SEVEN years and as someone who has used LLM's daily in my workflow for 3 years).

There's a huge caveat i don't see often, which is that it depends on your language for programming. IE. AI is reallllly good at writing Next.js/Typescript apps, but not so much Ruby on Rails. YMMV

I am finding the most destructive aspect of LLM assistance to be the loss of flow state.

Most of the time I can go faster than these tools if I have confidence in myself and allow the momentum to build up over the course of 20-30 minutes. Every time I tab out to an LLM is like a 5 minute penalty over what I could have done unaided on a good day.

Getting the model prepared to help you in a realistic domain often takes a few minutes of arranging code & comments so that it is forced toward something remotely sane. I'll scaffold out entire BS type hierarchies just so I can throw a //TODO: .... line in the middle somewhere. Without this kind of structure, I would be handling unfiltered garbage most of the time.

It's not that these tools are bad, it's that we need to recognize the true cost of engaging with them. ChatGPT is like a jackhammer. It will absolutely get you through that concrete slab. However, it tends to be quite obnoxious & distracting in terms of its operational principles.

I like Thomas, but I find his arguments include the same fundamental mistake I see made elsewhere. He acknowledged that the tools need an expert to use properly, and as he illustrated, he refined his expertise over many years. He is of the first and last generation of experienced programmers who learned without LLM assistance. How is someone just coming out of school going to get the encouragement and space to independently develop the experience they need to break out of the "vibe coding" phase? I can almost anticipate an interjection along the lines of "well we used to build everything with our hands and now we have tools etc, it's just different" but this is an order of magnitude different. This is asking a robot to design and assemble a shed for you, and you never even see the saw, nails, and hammer being used, let alone understand enough about how the different materials interact to get much more than a "vibe" for how much weight the roof might support.

Not a programmer now but my real takeaway from working with developers, enterprises at various stages of adoption: AI / LLM are not magic. The hype still exists but basically AI/LLM are tools. How we use these tools is what makes the enterprises different. unfortunately the quantity hype is '+' by big boys from MS, Goog adding to the noise. Personally, I love to see the quality and most importantly some real examples going beyond coding assistance. Be it automating the QA/QC or even some of the infra

I find the Konwinski Prize to be very interesting in this context. 1 million dollars to whoever's open source LLM solves >90% of a set of novel Github issues.

https://www.kaggle.com/competitions/konwinski-prize/

Currently, the #1 spot sits at a score of 0.09, not 0.9. A far cry from being useful. I know that open source models are not as good as closed source, but still, we're a long way from LLMs being good for code on their own.

And that supports OP's point - these tools aren't AGI, they produce trash that needs evaluation, but they're still useful.

I tried to take this article seriously, but it seems more like HN ragebait than an actual attempt at discussion. The engagement bait appears to be working, given all of the comments.

It’s like half of the arguments are designed as engagement bait with logical consistency being a distant concern:

> If hallucination matters to you, your programming language has let you down.

This doesn’t even make sense. LLMs hallucinate things beyond simple programming language constructs. I commonly deal with allusions to functions or library methods that would be great if they existed, but the LLM made it up on the spot.

The thing is, the author clearly must know this. Anyone who uses LLMs knows this. So why put such a bizarre claim in the article other than as engagement bait to make readers angry?

There are numerous other bizarre claims throughout the article, like waving away the IP rights argument because some programmers pirate TV shows? It’s all so bizarre.

I guess I shouldn’t be surprised to scroll to the bottom and see that the author is a HN comment section veteran, because this entire article feels like it started as a reasonable discussion point and then got twisted into Hacker News engagement bait for the company blog. And it’s working well, judging by the engagement counts.

> Often, it will drop you precisely at that golden moment where shit almost works, and development means tweaking code and immediately seeing things work better. That dopamine hit is why I code.

Only if you are familiar with the project/code. If not, you were throw into a foreign codebase and have no idea how to tweak it.

One other thing that might feed into skepticism is that, for the average person who does not actively choose to use LLMs, the most regular LLM interaction is likely to be with AI Overviews in Google search, which are generally a pretty unfortunate experience. For example my last search with an AI Overview, for interpretation of blood pressure reading, resulted in the output: "Systolic pressure (top number): 123 mmHg, which is within the normal range of less than 120 mmHg".

The fact that LLMs can generate this kind of output is not incompatible with the OP's argument about how useful they can be when combined with agents and code review. But search is the lowest-friction way for anyone to interact with LLMs currently, and right now it's a pretty poor advertisement.

This is one of the most obnoxious things I’ve ever read. “I’m right and every vague straw man I can think of is stupid and naive.” Great. Why did you write this?

These systems do work, quite well. Especially the latest release of Claude. Palantir is using AI to process your IRS information without your consent, https://archive.is/QAY9E. This should concern everybody. Not only can this information be used to determine specific account details. They can use this to target individuals which don't align with the current administration's agenda, as well exclude hiring base of political alignment. They could also use this data to exclude and discriminate those against Peter Theil's network state agenda aka "Freedom Cities", but it seems everybody is complacent with that.

> "LLMs can’t write Rust"

This really doesn't accord with my own experience. Using claude-code (esp. with opus 4) and codex (with o3) I've written lots of good Rust code. I've actually found Rust helps the AI-pair-programming experience because the agent gets such good, detailed feedback from the compiler that it can iterate very quickly and effectively.

Can it set up great architecture for a large, complex project from scratch? No, not yet. It can't do that in Ruby or Typescript either (though it might trick you by quickly getting something that kinda works in those languages). It think that will be a higher bar because of how Rust front-loads a lot of hard work, but I expect continuing improvement.

I was skeptical at first, but after trying to use it several times for work (Cursor/Claude) I concluded that it's worse than I thought. Lot more time wasted for little to no results. Maybe for boilerplate frontend code or TODO app generation it's good or something?

Anyway in a couple of months I'll try again. Until that time, I guess the search result summaries are nice?

> pull in arbitrary code from the tree, or from other trees online, into their context windows,

I guess this presupposes that it is ok for 3rd parties to slurp up your codebase? And possibly (I guess it ostensibly depends on what plan you are on?) using that source code for further training (and generating that same code for others)?

I imagine in some domains this would not be ok, but in others is not an issue.

You know what, I can use the word “fuck” too, just like this pro coder does: I’m fucking tired of these fucking nuts who act as if there is no problem at all having AI mess with my code. It harms my relationship with my code! It reduces my agency. It threatens my accountability.

He implies that if I am talking about some other flavor of AI than he personally uses than I am just not serious— but that’s what they said six months ago and two years ago. These fucking boosters have already blown their credibility.

If you want to prove that AI agents writing code for you is safe then perform a large scale test and show your results. We’re not crazy for doubting you. You are crazy for not doubting yourself.

It's so all-or-nothing this debate. If you're drawing a benefit from using AI tools, great. If you aren't, then maybe don't use them, or try some other approach to using them.

Personally I find AI coding tools situationally useful. I certainly wouldn't use them to write all my code, but I also think I'd be a fool not to leverage them at all.

Sometimes i feel like people who really like AI have a very different experience programming then i do.

They are constantly talking about AI doing all the tedious boilerplate bullshit. Don't get me wrong, some of my code is that too and its not fun. However the pro-AI people talk as if 80% of your day is dealing with that. For me its simply a rare enough occurence that the value proposition isn't that big. If that is the killer app of AI, it just doesn't sound that exciting to me.

> As a mid-late career coder, I’ve come to appreciate mediocrity. You should be so lucky as to have it flowing almost effortlessly from a tap.

If the kind of work this person is doing requires no thought, then of course an LLM will be perfect for it.

This kind of guilt-by-association play might be the most common fallacy in internet discourse. None of us are allowed to express outrage at the bulk export of GitHub repos with zero regard for their copyleft status because some members of the software engineering community are large-scale pirates? How is that a reasonable argument to make?

The most obvious problem with this is it's a faulty generalization. Many of us aren't building large-scale piracy sites of any sort. Many of us aren't bulk downloading media of any kind. The author has no clue whether the individual humans making the IP argument against AI are engaged in piracy, so this is an extremely weak way to reject that line of argument.

The second huge problem with this argument is that it assumes that support for IP rights is a blanket yes/no question, which it's obviously not. I can believe fervently that SciHub is a public good and Elsevier is evil and at the same time believe that copyleft licenses placed by a collective of developers on their work should be respected and GitHub was evil to steal their code. Indeed, these two ideas will probably occur together more often than not because they're both founded in the idea that IP law should be used to protect individuals from corporations rather than the other way around.

The author has some valid points, but dismissing this entire class of arguments so flippantly is intellectually lazy.

This guy may be right about a lot of things said here but he's smug and it's off-putting. preaching to the choir.

I keep talking to people who've had a good bit of success using gemini or cluade to build quick prototype front ends for some applications. I think theres some questions in my head of how well the process scales when you want to keep adding features, but according to them it's not been hard getting it to make the needed changes.

My issue with it is that it gates software development behind paid services with various levels of context supported. Absolutely not the dream I have of how more software should be open source and everyone should be empowered to make the changes they need.

I was a bit skeptical at first too, but once I let AI dive into my code, run tests, refactor functions, and even catch its own bugs, it felt like I suddenly had a really competent teammate. AI has been genuinely helpful for me , it saves time and lets me focus on the things that actually matter.

The author posits that people don't like using LLMs with Rust because LLMs aren't good with Rust. Then people would migrate towards languages that do will with LLMs. However, if that were true, then Julia would be more popular since LLMs do very well with it: https://www.stochasticlifestyle.com/chatgpt-performs-better-...

> but the plagiarism […] Cut me a little slack as I ask you to shove this concern up your ass. No profession has demonstrated more contempt for intellectual property.

Speeding is quite common too, yet if you get caught — especially overdoing it — you'll have a problem.

Also, in this case, presumably everything produced with AI is fair game too? The argument being made here isn't even "it's not plagiarism", rather "it's plagiarism but I don't care" — why would anyone else respect such an author's copyrights?

> Think of anything you wanted to build but didn’t. You tried to home in on some > first steps. If you’d been in the limerent phase of a new programming > language, you’d have started writing. But you weren’t, so you put it off, for a > day, a year, or your whole career.

> I can feel my blood pressure rising thinking of all the bookkeeping and Googling and dependency drama > of a new project. An LLM can be instructed to just figure all that shit out. > Often, it will drop you precisely at that golden moment where shit almost works, and development means > tweaking code and immediately seeing things work better. That dopamine hit is why I code.

This is me . Just yesterday I finished something I have been putting off for 2 years. I maintain 2 mobile phones one because of work and i needed a way to see sms messages from the other phone. I wanted to write a simple andoird app that reads incoming messages and saves them somewhere for me to use for OTP's and other stuff. For the longest time I wanted to build something like this for myself but I have been putting it off just because of all the weird issues i had been getting with mobile development. A couple of days ago since I had some time and I fired up gemini and asked it to write me a simple flutter app for and to save the messages to supabase and voila it was amazing . With some help and some tweaking I have a fully functioning mobile app on andoird , a small tauri application on desktop to read the messages nicely formatted in a data table without much coding on my part at all. Truly remarkable.

Loved this style of writing in 2005 from Maddox on the best site in the universe or whatever.

Sorry if I don't want google and openAI stealing my or anyone else's work.

The problem is not about quality of the code, it simply almost never works. Just tried some models still in beta on the code of one of the C projects and, while some pieces were okay and could be used as is, some are complete garbage and wouldn't even compile. Maybe few years later it will not become a problem but right now they are useless. For the AI-assisted research and internet search, that's a different thing and works pretty well already.

“I work mostly in Go.”

That should have been the start of the essay.

New title: LLMs make Go programming feel good

I have been using agentic AI to help me get started writing an OpenGL-targeted game from scratch (no engine). I have almost no background experience with computer graphics code, but I understand most of the fundamentals pretty well and I have almost 13 years of software experience. It's just that the exact syntax as well as the various techniques used to address common problems are not in my arsenal yet.

My experience has been decent. I don't know that it has truly saved me much time but I can understand how it FEELS like it has. Because it's writing so much code (sometimes), it's hard to vet all of it and it can introduce subtle bugs based on faulty assumptions it made about different things. So, it will dump a lot of code at once, which will get me 90% of the way there, but I could spend an hour or two trying to nudge it to fix it to get it to 100%. And then I will probably still need to go back and reorganize it, or have it go back and reorganize it. And then sometimes it will make small adjustments to existing, committed code that will subtly break other things.

Something that has surprised me (in hindsight, it isn't surprising) is that sometimes when I feel like it misunderstood something or made a faulty assumption, it was actually me that had the misunderstanding or ignorance which is humbling at times and a good learning experience. It is also pretty good at bug hunting and DEFINITELY very good at writing unit tests.

I count myself as pretty lucky that this domain seems to be very well covered in training. Given the law of averages, most people's domains will probably be covered. I'm not sure how it would fare with a niche domain.

This article feels incredibly defensive. If you have really have a technique that makes you 100x, 50x, or even just 2x more productive, you don’t need to write an article calling people who don’t agree with you nuts.

You keep using that tool, to your advantage. I’d you’re really altruistic you post some videos of how productive you can be like DHH did with his blog in 15 minute videos.

If you’re really that much more productive, the skeptics won’t be able to keep up and it should only take 6 months or some for that to become self evident.

"If hallucination matters to you, your programming language has let you down."

But every time I try this,this happens. I've tried the same thing on multiple agents and they all make up some magic api that doesn't exist, even when I've used the right thing elsewhere in the code to do the same job with that function commented as such. Asking them to use that always fails.

This seems a massive flaw to me that is just hand waved away here.

I felt this article was a lot of strawman-ing.

Yes, there are people who think LLMs are just a fad, just like NFTs, and I agree these people are not really serious and that they are wrong. I think anyone who has used an AI coding agent recently knows that they are highly capable and can enhance productivity in the right hands.

But, as someone who gets a lot of value in AI coding agents, my issue is not with gen AI as a productivity enhancing tool - it's with the absolute torrent of BS about how AI is soon going to make coders obsolete, and the way AI has been shoved onto many engineering teams is like yet another incarnation of the latest management fad. My specific arguments:

1. As the author pretty much acknowledges, AI agents still basically suck at large, system-wide "thinking" and changes. And the way they work with their general "guess and check" method means they can churn out code that is kinda sorta right, but often leaves huge holes or outright laughable bugs.

2. Hallucinations are the worst possible failure modes - they look correct, which makes it all the more difficult to determine they're actually bullshit. I shudder to think about who will need to maintain the mountains of "vibe code" that is now being generated. Certainly not fucking me; I had a good career but I think now is definitely the right time to peace out.

3. Even if I could totally agree that there is a strong business case for AI, I can still, as an individual, think it makes my job generally shittier, and there is nothing wrong with having that opinion.

I don't think I'd be so anti-AI if I saw a rational, cautious debate about how it can enhance productivity. But all I see are folks with a vested interest overselling its capabilities and minimizing its downsides, and it just feels really tiresome.

I use LLM as a super Stackoverflow. So far it has been helpful. For code gen, 8 out of 10 cases, the code won't compile and need fixing. LLM did help a lot in learning new topics. Gemini did help tremendously recently in pointing out a new approach to a problem I was stumped on. Though the code it generated was garbage, the code did provide insight on a new approach which I adopted to make it work eventually.

I can’t even get copilot to autocomplete 5 working lines consistently. I spend hours every day arguing with ChatGPT about things it’s hallucinating. And Agents? It took me a year to convince anyone to buy me a copilot subscription. It’s not good enough now? But it was the bees knees just a year or two ago? See I hate the thing where the JS-framework tempo thing happens to the part of the software world I’m in.

I simply do not get this argument about LLMs writing tedious code or scaffolding. You don't need or want LLMs for that, you want libraries and frameworks.

I barely write any scaffolding code, because I use tools that setup the scaffolding for me.

> People coding with LLMs today use agents. Agents get to poke around your codebase on their own. They author files directly. They run tools.

I'll be damned if I give up control of my machine to a tool that hallucinates actions to take using hastily put together and likely AI-generated "agents". I still want to be the primary user of my machine, and if that means not using cutting edge tools invented in the last 6 months, so be it. I don't trust the vast majority of tools in this space anyway.

> I’m sure there are still environments where hallucination matters.

Still? The output being correct matters in _most_ environments, except maybe art and entertainment. It especially matters in programming, where a 99% correct program probably won't compile.

> But “hallucination” is the first thing developers bring up when someone suggests using LLMs, despite it being (more or less) a solved problem.

No, it's not. It's _the_ problem that's yet to be solved. And yet every AI company prefers chasing benchmarks, agents, or whatever the trend du jour is.

> I work mostly in Go. [...] LLMs kick ass generating it.

I also work mostly in Go. LLMs do an awful job generating it, just as with any other language. I've had the same shitty experience generating Go, as I've had generating JavaScript or HTML. I've heard this excuse that the language matters, and IME it's just not the case.

Sure, if you're working with an obscure and niche language for which there is less training data, I suppose that could be the case. But you're telling me that there is no good training data for Rust, the trendiest systems language of the past ~decade? C'mon. Comparing Rust to Brainfuck is comical.

I won't bother responding to all points in this article. I will say this: just as AI doomsayers and detractors deserve criticism, so does this over-the-top praising. Yes, LLMs are a great technology. But it is also part of a wildly overhyped market that will inevitably crash as we approach the trough of disillusionment. Their real value is somewhere in the middle.

His arguments are mostly right I would say. I tend to use a bit of LLMs for coding - just Copilot though - and I do like how it quite often saves me time with autocomplete spanning entire lines.

ChatGPT is also pretty usefull - especially for naming things.

However, on a global scale, I would compare the shift towards 'vibe coding' and massive LLMs use to the one we got when people doing the coding don't even know what Assembly even is. Current state of software is that it's incredibly bloated and slow (and I do love python). Doesn't matter that much since we've got massive improvements in hardware, but still. In 5 years, instead of not really realizing that we're writing slow and bloated code, we will not really realize we're writing incorrect code.

I don't know anything about web developement, but in the environements I've seen during my career - where most code is quite critical and a stupid mistake, one that an LLM could easily hallucinate, can cost a lot - I don't see it replacing humans anytime soon. It's a pretty good and usefull tool though.

I like LLMs, it's legitimately impressive to see people vibe code solutions (some times even to novel ideas). Agents are a much better use of LLMs than auto complete (which is terrible).

However I've tried coding agents, and well they suck... for me. Nearly all of the code I write these days is performance sensitive, as well as design sensitive (making abstractions for others to use), and not close to what could be considered boilerplate. The boilerplate I do write usually takes a trivial amount of time to write by hand. I'll keep trying whatever is new in AI every few months, I certainly don't want to be left behind, but so far that strategy hasn't left me optimistic.

> but the code is shitty, like that of a junior developer

It's worse than a junior. Juniors learn and grow, if they're a good junior it's often fairly rapid and self directed. Juniors can also ask for guidance or help, this is something maybe agents will be capable of in the future, but I could see it going pretty poorly for reasons tied to the first point about not learning. While LLMs contain a ton of information, more than a junior could, the junior is a better reasoning machine; I don't consider LLM "reasoning" to be anything close to the meatbag version.

> but the craft

Craft matters, it's important to enjoy and take pride in your work. If you enjoy managing AI agents that's fine, lots of programmers end up on the manager career path, but it's not for everyone.

Use of LLMs to program really doesn't make much sense. The very nature of computer software is that it's a small effort from a small number of people that is multiplied out to massive productivity gains for a large number of people. Unless the LLM is producing better output than the people would, you are saving time for a small number of people in exchange for degrading the experience of a much larger number of people. If your AI-coded app goes down for a day and you have to vibe-debug it, that is probably gonna put you in the red from a time perspective.

With the amount of supervision you need to give this stuff, I can't see it being useful for much more than auto-completing the next few lines of code as you type. You've got to read and go over the whole output carefully anyway, so you might as well have finer grained control over it.

I just don’t like using AI tools. It’s absolutely boring. If you like that, fine.

Me? I like programming.

I have a very simple counter argument: I've tried it and it's not useful. Maybe it is useful for you. Maybe even the things you're using it for are not trivial or better served by a different tool. That's fine, I don't mind you using a tool far away from my codebase and dependency tree. It has not been useful for me, and it's very unlikely it's ever going to be.

Except that's not the argument people are making. They are arguing it will replace humans. They are arguing it will do research level mathematics. They are arguing this is the start of AGI. So if you want to put your head in the sand and ignore the greater message that is plastered everywhere then perhaps some self reflection is warranted.

I think another thing that comes out of not knowing the codebase is that you're mostly relegated to being a glorified tester.

Right now (for me) it's very frequent, depending on the type of project, but in the future it could be less frequent - but at some you've gotta test what you're rolling out. I guess you can use another AI to do that but I don't know...

Anyway, my current workflow is:

1. write detailed specs/prompt,

2. let agent loose,

3. pull down and test... usually something goes wrong.

3.1 converse with and ask agent to fix,

3.2 let agent loose again,

3.3 test again... if something goes wrong again:

3.3.1 ...

Sometimes the Agent gets lost in the fixes but now have a better idea what can go wrong and you can start over with a better initial prompt.

I haven't had a lot of success with pre-discussing (planning, PRDing) implementations, as in it worked, but not much better than directly trying to prompt what I want and takes a lot longer. But I'm not usually doing "normal" stuff as this is purely fun/exploratory side-project stuff and my asks are usually complicated but not complex if that makes sense.

I guess development is always a lot of testing, but this feels different. I click around but don't gain a lot of insight. It feels more shallow. I can write a new prompt and explain what's different but I haven't furthered my understanding much.

Also, not knowing the codebase, you might need a couple attempts at phrasing your ask just the right way. I probably had to ask my agent 5+ times, trying to explain in different ways how translate phone IMU yaw/pitch/roll into translations of the screen projection. Sometimes it's surprisingly hard to explain what you want to happen when you don't know the how it's implemented.

I've only just realized how bad AI is for me, in a manner I didn't at all expect. I don't use it. I don't want to use it. I hate it. As another commenter here nicely put it, I want to do, not tell. I like writing code. I like solving problems. I like doing the work myself. I like struggling and then, the next day, finding that my capabilities have grown.

And yet I'm overwhelmingly demoralized and depressed about AI's encroachment on software engineering, so much so that I am finding it difficult to marshal cognitive resources. I'm finding it harder and harder to engage with coding problems, even though I love programming.

The last thing I expected was that AI, even kept at a distance, would rob me of the joy I find in one of the things I love most.

It so devalues what I do and my basic capabilities as a human being that I kind of want to die. I'll stick around for my daughter, my partner, and any future children, but I'm not optimistic at all about the future.

I think in a lot of cases the skeptics haven't really witnessed how good LLMs can be for a subset of software engineering tasks. I find for mid-complexity code they are far, far better than I am at coding. Even though I often tweak their solutions, they outweigh that by often identifying APIs and even whole libraries to call where I fully would have set about implementing it myself - either out of ignorance or an assumption that it would be too hard to implement it myself. They also put in much more error checking than I would normally do. Even if I do implement something myself, it's insane not to at least see what the LLM would do to get ideas. The thing is, once I do that, I nearly always come away with the conclusion that I might as well just use the code it made.

> Does an intern cost $20/month? Because that’s what Cursor.ai costs.

So then where do the junior developers come from? And then where do the senior developers come from?

> but it’s bad at rust

I gave chatgpt (the free desktop web version) my Ruby solution to one of my favorite adventofcode challenges and asked it to write in in Rust. I've no experience in Rust except reading the few first chapters of the book but with the help of the compiler I managed to fix it and it gave the right responses 100x faster than Ruby did. So chatgpt did a great job keeping the logic intact.

I would have done the same thing if LLMs didn't exist, write it in Ruby first and then in Rust. The flexibility and expressiveness of Ruby allows to actually find a working solution. Experienced Rustaceans would do it directly in Rust, but writing in Ruby should be easier and an LLM can transform it.

Where are the results? I keep hearing about how great these coding agents are. What have you built beyond the complexity of a disposable toy app? Where are the swarms of new apps built by these agents that are solving new problems and taking over markets? I’m just not seeing results. I guess I need to spend some time using cursor as an agent to see it for myself.

I am a believer that these tools will be (if they aren’t already) extremely useful. Game changing in many ways, and I worry about the future of knowledge fields. Definitely not a skeptic in the trajectory. I do wonder if that makes those of us who are on the cutting edge more valuable as AI can quickly reproduce things that are well understood.

What worries me about this is that the benefit of well practiced developers using AI to boost their productivity is there, but what happens to the new developers that piggyback off of LLM tooling too much and end up with worse understanding as a whole?

Maybe it's only me, but I just don't write that much code. I try to change less than 100ish lines per day. I try to keep codebases small. I don't want to run a codebase with hundreds of thousands of lines of code in a production environment.

Please stop this. It’s clear to me that you’re not a software engineer. You weren’t one before AI, and you are not one after. You can do whatever you want, but please don’t push this AI narrative down everyones throat. I’ve had enough of this.

I read this article completely thru, twice, and I still have no idea what the author is trying to say. All I can read here is a bunch of sarcastic dismissals of stuff that I don't believe and have no way to evaluate.

I can sum it up like this: if I could know in advance the exact right thing to build, producing the physical code, has not for a long time, been the bottleneck. I've been vibe coding long before it was cool. It's sometimes called model driven development.

For those that think only procedurally, I can see how it helps them. Because procedural first development has a lot of boilerplate logic.

For those who think model first, the AI may help them rubber duck, but ultimately the physical writing of the characters is minimal.

Most of my time is thinking about the data model. The AI writes almost all of my procedures against said data model. But that is about 20% speedup.

This post feels quite dishonest. Every concern is countered with some variation of ”you’re wrong and I am right”, zero nuance.

I use Claude for specific questions, it works great (but apparently I’m doing it wrong according to the author), code generation doesn’t, I don’t want my boilerplate code to have subtle differences in every file, that’s the worst kind of boilerplate and that’s what I get if I automate it.

There are real world implications of all the infrastructure being built for this too, not to mention the extremely short e-waste cycles during the boom we’re in right now. To me it also seems like LLMs are the wrong thing to solve for, it’s like a gasoline car, make the engine larger and more fuel consuming to go faster and no matter how much you perfect it, it will still be fossil fuel being burnt at a high rate. What’s the alternative, I don’t think we know yet, but if we were a smarter species we would see that this is not it.

I found this comment especially interesting: "The code in an agent that actually “does stuff” with code is not, itself, AI. This should reassure you. It’s surprisingly simple systems code, wired to ground truth about programming in the same way a Makefile is. You could write an effective coding agent in a weekend. Its strengths would have more to do with how you think about and structure builds and linting and test harnesses than with how advanced o3 or Sonnet have become."

I haven't used coding agents. Can someone who has (or knows about these things) elaborate, or point me to a good resource? Thanks!

Does no-one think ahead, for more than a moment?

> Professional software developers are in the business of solving practical problems for people with code.

Are we? Is that our job? You might think it is, and it might even feel like it. But it is not. Our job, in 99% of cases, is to make stock price go up. That means making software that people will buy, making people want to keep buying it, and doing whatever makes your investors happy.

So what happens as you use these tools? Your job changes from coder to "LLM manager", but let's assume you're happy with that. You're more "productive", which is an artificial metric and nothing to do with your real job as described above. That's fine for now, but what happens down the road? Now your company is paying more and more money to these LLM services, and paying you less and less.

> "but they take-rr jerbs" So does open source.

This is a gross misunderstanding of open source. Most open source developers are paid to be, full time, by companies. Why? Warm fuzzy feeling? Goodness of their heart? No. It's because it's good for business. These open source techs are a baseline, there' not worth being competitive on. Build the open base in the direction you want, and build the "business value" ontop for profit.

> We used to pay good money for databases.

We still do. Have you looked at managed PostgreSQL prices lately? Also if you haven't noticed, Oracle hasn't gone anywhere.

> LLMs really might displace many software developers. That’s not a high horse we get to ride. Our jobs are just as much in tech’s line of fire as everybody else’s have been for the last 3 decades.

This isn't the same thing as replacing punch cards, or moving away from Java 6, or replacing artisans with factory machines. This is centralisation and rent-seeking. As companies become more and more dependant on a handful of companies with the upfront capital to build massive server farms... what do you think is going to happen? Once they've convinced enough people that LLMs are essential to their work it's going to cost a lot more than $20/month. Once these companies are entrenched they will wring every last dollar out of every company that's grown to depend on them, many will simply fold.

Maybe we are in a seismic shift in our industry. But the beneficiaries are not going to be whole-hearted LLM adopters, no amount of saying "you need to change how you work" is going to save any of us. The only people that will benefit are going to be the people that own the data centres.

> If you were trying and failing to use an LLM for code 6 months ago †, you’re not doing what most serious LLM-assisted coders are doing.

Translation: Drop your work and current tools every few months to pick up the shiny new thing or you aren't cool or even sane.

Reality: Your favorite cool new tool will be integrated into development work as it is proven, documented, and time allows. This idea that development must leap onto the latest shiny new thing every few months is extremely corrosive to reasonable business practice. The needs of the many outweigh the preferences of the lead developer.

What this boils down to is an argument for slop. Yeah, who cares about the quality, the mediocrity, the craft... get the slop, push it in, call it done. It mostly works in the golden path, it's about 6 or 7 orders of magnitude slower than hand-written software but that's ok, just buy more AWS resources, bill the client, whatever.

I can maybe even see that point in some niches, like outsourcing or contracting where you really can't be bothered to care about what you leave behind after the contract is done but holy shit, this is how we end up with slow and buggy crap that no one can maintain.

Without AI, software is getting worse every year. My question is: will AI resolves that problem, or will it make it worse?

My feeling is that AI will keep the software industry in the same trend: produce more software of lower quality while being more profitable.

People sometimes say "perfect is the enemy of good". But this implies that putting less effort is "good enough". I believe it goes the other way: why would a company put efforts into selling you something good, if it can make you pay the same for something worse? It's all about maximising profit, and AI is no exception there.

This: "Important caveat: I’m discussing only the implications of LLMs for software development. For art, music, and writing? I got nothing. I’m inclined to believe the skeptics in those fields. I just don’t believe them about mine". I want to emphasize this. AI seems like a great 'first draft', but not a 'finished product'. In the hands of a programmer, a first draft is a great start. But pushing AI to frontline workers or customer support fails, as Klarna learned, because a first draft is not good enough.

When I ask chatGTP to generate the code for a M4F MCU that implements the VirtIO code for accessing the GPIO through RPMSG using FreeRTOS it produces two answers that are both incomplete and incorrect.

I wonder if it comes down to there being many kinds of programming and this stuff is good for some but not all.

The author describes a world where most code is boring grunt work and most collaborators are mediocre. This world is not familiar to me. I can see LLMs being useful in such a world, but it’s not where I work.

The comment about buying a table instead of building it really stuck with me. We’re the ones who design the table! Saying we should buy one from the store is like saying farmers should buy their food from the store instead of putting in all that work to grow it.

Not a very bright move from fly to publish this kind of condescending post on the company blog.

Making fun of people who care about copyright (implying that fly abuses GPL for instance) is just icing on the cake.

"If you’re making requests on a ChatGPT page and then pasting the resulting (broken) code into your editor, you’re not doing what the AI boosters are doing. No wonder you’re talking past each other."

This is in fact what I am doing (though I am not trying to convince anyone not to use LLMs). How can I get into this other mode? Any pointers? Do I need to pay up for a service? I do not want to carry out vibe coding, I want to use LLMs to solve physics, math and logic problems/puzzles. I'd appreciate any recommendations.

What an arrogant jerk.

It sounds like someone taking a condescending tone to say that "Fast fashion is great! The clothes are really fine, and being able to wear a t-shirt once and then throwing it away instead of washing it is more convenient. This is most definitely progress. Sustainability? Hmm I can't say, that's not my field of expertise, therefore it's probably not important".

A well articulated blog, imo. Touches on all the points I see argued about on LinkedIn all the time.

I think leveling things out at the beginning is important. For instance, I recently talked to a senior engineer who said "using AI to write programming is so useless", but then said they'd never heard of Cursor. Which is fine - but I so often see strong vocal stances against using AI tools but then referring to early Copilot days or just ChatGPT as their experience, and the game has changed so much since then.

Damn. Well I'll spend a few bucks trying it out and I'll ask my employer if they're okay with me using agents on company time, but

But I'm not thrilled about centralized, paid tools. I came into software during a huge FOSS boom. Like a huge do it yourself, host it yourself, Publish Own Site, Syndicate Elsewhere, all the power to all the people, borderline anarchist communist boom.

I don't want it to be like other industries where you have to buy a dog shit EMR and buy a dog shit CAD license and buy a dog shit tax prep license.

Maybe I lived through the whale fall and Moloch is catching us. I just don't like it. I rage against dying lights as a hobby.

At 0 temperature an LLM is a Map<String,String> - a string input (key) will give you the same string output (value) every time. Hypothesis: there exists a key whose value is a complete, working, fully-tested application which meets your requirements 100% and fulfills your business need. This key is the smallest, most complete description of what your application does. It is written in natural language and represents a significantly compressed version of your application code.

My part-time obsession over the last few months has been trying to demonstrate this and come up with a method for finding these magic keys (I even tried to get the LLMs to search for me, lol). What I really want is to give the latest thinking models (200k input, 100k output) a 5-6 page design doc (4k words, 5k tokens) and have them produce a complete 5kloc (50k tokens) microservice, which would show a 10x compression. It's hard, but I haven't seen any reason to think it wouldn't work.

For better or worse, I think this will be close to what IC jobs will be like in few years. Fundamentally, our jobs are to try work with other functions to agree to some system that needs to exist, then we talk to the computers to actually implement this. If we switch kotlin+compiler for design doc+llm, it still going to be somewhat the same, but far more productive. Agents and such are somewhat of a stop-gap measure, you don't want people giving tasks to machines, you want to accurately describe some idea and then let the computers make it work. You can change your description and they can also figure out their own tasks to evolve the implementation.

I've been using LLM's pretty regularly for just over a year (my PyCharm AI license just renewed). My experience before that was pretty mixed and fell into the "sort of worth it" category.

It's improved to the point where the summary that I would email a junior developer is a pretty good prompt that produces code pretty close to want I want.

I had a weird project recently where I needed a stand alone event handler that would send an email receipt to the customer, another email to customer service, log the event in an accounting system via an API. I wrote up a proposal for the rest of the management team and, once accepted, pasted the proposal into ChatGPT (and a few others, but ChatGPT gave the best results).

I basically got a full recipe for setting up AWS API Gateway, Lambda, and SES, the Python Lambda function with api calls to SES, suggestions on using AWS SAM, etc. etc. Tweaked the prompt a couple of times and then took over from there. I had this into the hands of the testing team in less than a day.

I'm pretty sure this would reduce my team size in ANY role I've had in my career. Maybe not halve the teams, but it would eliminate a lot of the junior or journeymen developers.

Everything works right until it doesn't. LLMs are trained on things that have worked. Let's revisit in 2027 when things are insanely faster, but not much better.

I think I figured out a big problem I have with AI, and that is that I just hate using other people's software. The vast majority of software is just garbage. It all tries to take way too much of my attention. It doesn't help that all these AI tools are new, so they are inextricably linked to some SV startup that wants to be the next unicorn, so they have some bullshit account-generation/data-retention/terms-of-service/acceptable-use-policy/remarketing-email-campaing song and dance that is all stuff I have to pay attention to. All I really want to do is sit down and think on my own. I'm constantly inundated with distractions at work, so when I get a few minutes to myself, I sure as hell don't want an AI bot messaging me about reviewing its PR. I don't even want to review the human-written PRs; just push your code already, if it breaks you get to fix it. I certainly don't want a 500th unread email in my inbox.

You know what doesn't constantly nag for my attention? Dead text files on disk. They just sit there, waiting for me to grace them with my presence. That's the way it should be. Computers serve me, not the other way around.

I’m gaining more confidence AI is unable to take any of our jobs if this level of marketing is needed to try to convince us it will.

I think for most people that do not get value out of AI assisted coding, its a skill or learning issue. I don't care how good of a developer you consider yourself to be. If in 2025, these tools do not provide you significant (to almost absurd) value, you just do not know how to use them, or are to stubborn to push past the initial learning curve. Alternatively, maybe... programming is changing so much with AI, that you just might not see what's happening right infront of your eyes because its so foreign to all the other paradigms in the past.

In 2025 I am about 800-900 hours deep into using Cursor, blowing through 500 credits each 2 days (500 credits = 20$). I also organize a local "vibe code" workshop and meetup, with people from all sorts of experience levels, from no coding experience, to twice my senior. And what I am seeing is that what separates those from those who are good at using AI and those who are not, generally is a level of rigid stubbornness.

I can't wait for time to do its course here and those people who now are "all nuts" to slowly see programming of the past become an old relic.

> Does an intern cost $20/month? Because that’s what Cursor.ai costs.

And who the fuck will train those interns, now they've been replaced by machines?

LLM fanatics throw me nuts, instead of learned helplessness, at least try to understand what the token predictor outputs and be honest when you used LLM to generate your code dump when asking for help. Otherwise you are just wasting time and being simply a middle man for a machine and person who expects an interaction with actual human being with their valuable time.

The premise of “lol, I’m just a dumb developer and just wait until the smart developers catch on” falls extremely flat coming from Thomas.

It kinda invalidates the entire premise because it builds on this false foundation of “look how useful it is to a mere plebeian” and the skills actually required to sift through LLM garbage are what puts you in the top 5% of programmers.

Most senior developers cannot conduct useful code review and spot errors. In every company I’ve been at, there were only 1 in 20 people that would give the useful in-depth reviews required to pick up LLM bullshit.

This style of engagement bait is brutal and it severely overshadows other nuggets of wisdom of how LLMs can be useful.

So the reaso we should not use Rust (a memory safe language that gets rid of categories of bugs) is because some new AI tools are not read?

Using AI for over a year now on a daily basis, most AI tools do not struggle with Rust, they struggle to provide value. If I do not instruct them to keep the output to the bare minimum, I usually end up with 10x the output what would be a simple solution to my problem.

FWIW with proper MDC/ rules I’ve found LLM programming agents excellent at rust. There’s a lot of complex and tedious minutiae in rust that I know but forget to apply everywhere it’s helpful while a SOTA LLM agent does well, especially with proper rule guidance to remember to use it.

Generally though I find LLMs have a pretty rapidly diminishing return on what you can expect out of them. They’re like a 3-5 year senior programmer that has really learned their domain well, but doesn’t have the judgement of a principal engineer. You get to a point where you need to reach in and right things and really pay attention, and at that point the diminishing returns set it rapidly and you’re better off just doing the rest yourself. Refactors and stuff can be delegated but that’s about it.

I find this true regardless of the language. None the less, I’ve been able to improve my overall velocity dramatically completing several projects in the last few months in the span of one typically. If tooling improves I hope to continue that but I’m already getting close to the limit of how fast I can conceive of useful creative things.

There’s a fine line between being an AGI skeptic and an AI skeptic.

While I don’t believe the “AI will take all jobs” VC tales, I also think not using LLMs in software development workflows is similar to not wanting to use Excel as an accountant in the 1990s.

AI won’t replace all jobs, but companies that cleverly integrate it into their processes will be faster and more efficient compared to those that don’t.

I believe that AI is very useful in software development but I don't buy the narrative that AI is responsible for layoffs over the past few years (at least not most of them). I think that narrative is a convenient cover for systemic misallocation which created a need to contain inflation. I think big tech execs understood that, beyond increasing their company stock prices, they also need to work together to keep the monetary system itself under control. This is why they've been firing people whilst having record profits. They've reached such scale and the system has reached such fragility that they have to think and act like economists to keep the thing going. The economy itself has become the responsibility of big tech.

But who knows, maybe AI will accelerate so rapidly that it will fix the economy. Maybe we'll have robots everywhere doing all the work. But I worry about the lack of market incentives for people to adapt AI to real world use cases.

For example, I'm an open source developer who likes to tinker but I've been booted out of the opportunity economy. I can't afford to program robots. People like me are too busy using AI to parse spreadsheets and send targeted ads to even think about automating stuff. We work for companies and have no autonomy in the markets.

If things had worked out differently for me, I'd probably own a farm now and I'd be programming robots to do my harvest and selling the robots or licensing the schematics (or maybe I'd have made them open source, if open source had worked out so well for me). I don't have access to such opportunity unfortunately. The developers who worked for big tech are good at politics but often disconnected from value-creation. Few of them have the skills or interest to do the work that needs to be done now... They will just continue leveraging system flaws to make money, so long as those flaws exist.

A lot of the doubters are basically saying: "If it's so great, show me proof." While that is usually the right stance, I think it is short-sighted in times of revolutionary change.

To be sure, you definitely should not blindly trust the people who have a stake in selling AI products. However, you probably should trust the people who have no stake in selling AI, but are using it to accelerate themselves. The problem is, these people are probably too busy building things to spare any time convincing you. And in fact, it is a competitive advantage for them if others don't figure it out.

Here's my take: Everyone needs to figure out how AI works for them. The thing with AI is that it is a technology that behaves like people. It has the speed, efficiency and scale of computers but the fallibility and quirks of humans. This is why it's so confusing to discuss, and why folks have such varying experiences. The right approach to working with AI is like working with people: understand what they're good at and where they are weak, and then work with them to achieve your goals.

This will require using it "in anger" on many non-trivial tasks over a significant period of time, which will take persistence and patience. Yes, the hype is so over-the-top that you might not want to invest your valuable time on it. But you owe it to yourself to ensure you can fully capitalize on the sea change that is happening.

If it helps, I (like the OP) have no stake in selling AI and I posted this comment about my experiences on the other thread about the AI coding -- https://news.ycombinator.com/item?id=44163543 -- The upshot is, AI has enabled me to produce innovative (probably cutting edge) work in domains I had little prior background in. And I've not even used agents yet!

Nah, I’m not nuts, you’re just terminally brain damaged by the amount of ketamine you’ve huffed thinking AI was such a great tool.

It is going to be so much fun when all AI companies suddenly double prices, because what are you going to do, hire juniors?

I so strongly agree with this.

I've never been a completionist. I'm distracted. I like novelty. I make half-form experiments, but mostly provide deeply immersive conversations for the benefit of a small group of people around me, some of whom go on to make/do things that I proudly feel my influence on. The final 10% of work often defeated me.

Now, I just make it happen. The pain is gone. Something else can carry the last 10%, and my ideas get realized by me (and AI).

I can't even conceive how this will change the world en masse. Knowledge domains (not just tech, but science, art, etc) have been built of the things that the type-A completionists knew how to see. The others didn't finish, and what they saw of the world has largely been marginalized.

The mind of ADHD and creative types is about to get much more represented in the output products of civilisation. This is going to be a big recalibration, and I think a re-balancing toward something more healthy tbh

> but the craft - Do you like fine Japanese woodworking? All hand tools and sashimono joinery? Me too. Do it on your own time.

That is not my objection, though some seem to think it is.

I got into coding because I found it interesting and fun. I've avoided managing others because I hate that sort of thing. Working with an LLM, prompting it to do work then assessing that work and adjusting the prompts or cleaning/completing the output myself, is far too much like outsourcing work to a 3rd party junior, and I don't want to do that whether said junior is a human in a far off land (or just down the street, for all I know!) or some LLM run by MegaCorp-ᴙ-Us.

I was already seriously disliking much of the modern programming industry (remote teams, the shit-show that client-side web work continues to devolve into, and much else) anyway. If it gets to the point where I can't do my job (or can't do it competitively enough) without using LLMs on top of dealing with the other things I don't care for, I'll be retraining for something else entirely. Heck, I've got the mortgage paid, I don't have or plan to have a family beyond the cats, I could just about survive on shelf-stacking…

I know it isn't a passing fad, though I'm convinced the current and near-future iterations aren't going to be quite as significant in the longer term as many believe. I don't object to others using it (aside from the matter of “MegaCorp-ᴙ-Us”s of the world suddenly deciding that piracy is just fine because they stand to benefit from it, in fact are already benefitting from it, in this case). But I don't plan to be nagged into using it myself.

> Do it on your own time.

If/when that career shift happens, or I survive as I am long enough to retire, I will probably do exactly that. As it stands I spend more than enough of my life on tech in DayJob, my own time is increasingly filled with unrelated hobbies and people.

Can we get a video of a workday conducted by these people?

Unless there's a significant sense of what people are working on, and how LLMs are helping -- there's no point engaging -- there's no detail here.

Sure, if your job is to turn out tweaks to a wordpress theme, presumably that's now 10x faster. If its to work on a new in-house electric motor in C for some machine, presumably that's almost entirely unaffected.

No doubt junior web programmers working on a task backlog, specifically designed for being easy for juniors, are loving LLMs.

I use LLMs all the time, but each non-trivial programming project that has to move out of draft-stage needs rewriting. In several cases, to such a degree that the LLM was a net impediment.

> All this is to say: I write some Rust. I like it fine. If LLMs and Rust aren’t working for you, I feel you. But if that’s your whole thing, we’re not having the same argument.

Yes we are, because the kind of work you need to do in C++ or Rust is probably entirely different from the work this person manages to get the LLM to do in Go.

The most important thing in this article in my mind is in the level setting section - if you are basing your perspective on the state of AI from when you tested it 6mo+ ago, your perspective is likely not based on the current reality.

This is kind of a first though for any kind of technology. The speed of development and change here is unreal. Never before has a couple months of not being on top of things led to you being considered "out of date" on a tool. The problem is that this kind of speed requires not just context, but a cultural shift on the speed of updating that context. Humanity just isn't equipped to handle this rate of change.

Historically in tech, we'd often scoff at the lifecycle of other industries - Airlines haven't changed their software in 20 years?? Preposterous! For the vast majority of us though, we're the other industry now.

>pull in arbitrary code from the tree, or from other trees online, into their context windows, run standard Unix tools to navigate the tree and extract information, interact with Git, run existing tooling, like linters, formatters, and model checkers, and make essentially arbitrary tool calls (that you set up) through MCP.

....for the vast majority of my career, anyone who suggested doing this - much less letting code that no one in the world (much less the company) truly understands the logic flow of do this - would be fired.

If yesterday we were sued for singing the happy birthday song in public, and today we see corporations with market caps the size of countries pretending that IP ownership is not much of a thing, the alarm bells should be deafening. Control over intellectual property has given us fantastic things—Linux, Blender, etc.—and you can’t have GPL without an ability to defend copyleft. IP laws provide that ability.

One does not need to be a skeptic about machine learning and its potential as technology to refuse to engage with its practical applications when they are clearly ethically corrupt. The ends do not justify the means. Just like refusing to buy products of forced labour or disproportionate environmental impact, or to invest in war and bloodshed, everyone makes this decision for themselves.

Seriously? Is this argument in all earnestly "No profession has been more contemptuous therefore we should keep on keeping on"? Should we as an industry not bother to try and improve our ethics? Why don't we all just make munitions for a living and wash our hands of guilt because "the industry was always like this".

Seems a bit ironic against the backdrop of <https://news.ycombinator.com/user?id=tptacek>:

(although perhaps this is tongue-in-cheek given the last year)

> You’ve always been responsible for what you merge to main.

I've never merged anything with an unclear copyright to main, let alone if I positively suspected that it had been plagiarized.

> Cut me a little slack as I ask you to shove this concern up your ass. No profession has demonstrated more contempt for intellectual property.

Every place I've been employed was nearly paranoid about copyright, with rules in place about never copying and pasting anything copyrighted into the company code base.

All the major open source distros and projects, likewise.

There are some bad apples out there, like, oh, Sam Altman, who wants massive plagiarism to be fair use.

So, au contraire, you can shove your contemptuous blog up your ass, buddy.

> I like moving the code around and chuckling to myself while I delete all the stupid comments. I have to read the code line-by-line anyways.

I'm glad I got out when I did. :) This simply isn't something I want to do for a living. I'm glad there are people who do, though.

This op ed suggests that it’s easier to audit a huge amount of code before merging it in than is to write the code from scratch. I don’t know about anyone else, but I generally find it easier to write exactly what I want than to mentally model what a huge volume of code I’ve never seen before will do?

(Especially if that code was spit out by an alien copypasta that is really good at sounding plausible with zero actual intelligence or intent?)

Like, if all I care about is: does it have enough unit tests and do they pass, then yeah I can audit that.

But if I was trying to solve truly novel problems like modeling proteins, optimizing travel routes, or new computer rendering techniques, I wouldn’t even know where to begin, it would take tons of arduous study to understand how the new project full of novel algorithms is going behave?

I feel silly asking, but is this reality or hyperbole?:

> An agent can occupy itself for hours putzing with your tests in a VM and come back later with a PR.

> Kids today don’t just use agents; they use asynchronous agents. They wake up, free-associate 13 different things for their LLMs to work on, make coffee, fill out a TPS report, drive to the Mars Cheese Castle, and then check their notifications. They’ve got 13 PRs to review. Three get tossed and re-prompted. Five of them get the same feedback a junior dev gets. And five get merged.

I'm working better and harder with Roocode in Code mode (and somtimes Architect). I've only been doing it for a few weeks, and it reads to me like I'm already using an 'old approach'.

Finally, a proper good take from someone (seemingly) approaching it rationally and objectively, rather than blindly sitting on either side of the hype fence. LLMs are good tools, and even I, a decidedly-not-software-engineer, have been increasingly tinkering with locally-run tools to create applications and tooling that meets my own needs. The one thing that’s held me back from publishing has been the IPR angle, but I quite like this line on that topic:

> But if you’re a software developer playing this card? Cut me a little slack as I ask you to shove this concern up your ass. No profession has demonstrated more contempt for intellectual property.

The OP isn’t wrong. Some of the most brilliant minds in computer science have built tools to shatter DRM, enable widespread piracy, and repeatedly lift code wholesale from public repos if it helps them move a product forward. The modern crop of technologists (myself included) have long had a hypocritical view on IPR: “rights for me, not for thee” to oversimplify things. The entirety of the modern public cloud is built by trodding on the backs of open source projects while wearing lawn-aerating spikes. We steal shit from others so our Founders and Masters can reap billions in profit; to decry plagiarism now feels incredibly dishonest and insincere, at the very least.

Look, I’m an AI skeptic myself. On the whole, these tools are bad. They’re taking jobs, polluting the planet, and dumbing down a society that still can’t try rebooting something when it fails, let alone identify where their WiFi ends and the internet begins. They’re toxic walled gardens that can reauthor reality to the whims of the VCs and Billionaires funding them, locking you into subscriptions forever.

But even I will admit that modern LLMs, when it comes to writing code, are actually really good for the majority of C-tier devs and below. Are they as helpful to bleeding-edge engineers on the cusp of the future? Nah, not really, but if you’re just writing a checklist app for iOS to help you stay on task, or a Packer manifest to lay out some VM templates, they’re kinda awesome.

I think of programming languages as an interface between humans and computers. If anything, the industry expanded because of this abstraction. Not everyone has to learn assembly to build cool shit. To me AI is the next step in this abstraction where you don't need to learn programming languages to potentially build cool projects. The hard part of software engineering is scale anyways. My bet is that this will expand the industry in unprecedented ways. Will there be contraction of traditional programming jobs? Absolutely. The growth in tech jobs over the last 20 years weren't more assembly programmers. They were abstraction experts. I'm sure the next wave will be even bigger, professional prompting will explode in size.

The author mentions that smart programmers kick off agents, then go for a coffee and wait to be notified.

But agents like Claude Code and Cursor run on a user’s machine right? So how can they work in the background like that? are there agents that can work async?

He forgot "but the 15 year old behemoth codebase that doesn't fit in any context window and isn't well documented and changing of which requires changing tons of ancillary files not easily connected to each other logically".

That's what a lot of the highest paid programmers work on. I work in such a firm and the agent they have for us is incapable. And it's the latest and greatest from whoever coders love to rave about online.

I love that AI written tests and other generally greenfield things work well and can be vetted quicker than writing and getting the code oneself. But for most more complex tasks it's not there. I'm ready when it is though

Fun article, although I'd say I'm maybe slightly more on the skeptical side.

I've found LLMs can't really write a whole program there isn't already a template for. They're good enough at writing individual functions, and as the author mentioned, their breadth of knowledge is better than any individual.

Honestly, I see them as a supercharged version of googling, without all the out of context answers you get from Google since they also place ads and whatnot. I find an LLM can get me an answer far quicker than a regular internet search, even if it's imperfect and I'm probably not going to put it straight into my code.

Yeah, and it’s progressing so fast. Singularity is definitely on the table.

Whoever says otherwise should read their own comments from 2 years ago and see how wrong they were about where AI is today.

Not saying singularity will happen for sure, but is it a possibility? Hell yeah.

I know this is a website for programmers and I am not one. I like LLMs for quick answers and simple tasks like proofreading (just to fix errors, not to add or change anything) and simple text formating. I absolutely hate for everything else. I am a writer and LLMs prose may be correct and sometimes whimsical but it has soulless quality to it that makes me forget about what I just read after five seconds. My friends use Meta AI on our chat groups on WhatsAppp and I hate it so much, I delete every single AI message as soon as I see it.

Will LLMs ever break the “you have to treat it like an intern/junior” ceiling?

>It’s getting the same kind of attention that smart phones got in 2008, and not as much as the Internet got. That seems about right.

Are you kidding me? Mosaic was released in 1993. In 1998 Apple released the iMac on a theory, based on some marketing research, that people might be really interested in "getting online." Five years. And that's about a decade before we realized we might want a touch phone. More than a decade before an App Store.

This is very much still the top the first. AI will make even the internet seem small. Which makes sense — the entire whole internet is merely one input into this technology.

For some reason, every hardcore AI proponent I hear from resorts to gaslighting and namecalling to convince everyone it's so useful for everything.

When that doesn't work, they start inventing hyper-specific scenarios and graphs with grounding that exists only in their imagination to push the topic further. If the AI can't solve a problem in a domain for you, well, it's that domain that's the problem, not AI. You simply need to re-invent the paradigm around your entire problem - easy stuff.

Finally, when they've failed to convince anyone with a modicum of rationality, they resort to some version of "well, fuck you, it's the future and I'll just be leaving you behind!"

Interestingly, this article has all three.

My current workflow with Codex is (coding environment from OpenAI):

(1) Ask to write an implementation plan for a specific change or a feature. It will go through the source code, look up references, make notes and produce a plan

(2) Review the plan. Point out missing things, or stuff that needs improvement.

(3) Once I'm satisfied with the plan - ask to draft PR. Launch a few attempts in parallel and pick the one that I like the most.

(4) While drafting PR, Codex will run unit tests (even can run E2E tests in its container), linting and type checkers at every single step. This helps a lot with the stability.

(5) I review the code and merge the PR if I like it. Ask to cleanup - if not.

This feels like working with a remote team - very patient and diligent at that.

Ultimately, I get to get more features done per day. But I also feel more tired by the end of the day due to a higher level of cognitive load. There are more decisions to make and less idle time (e.g. no more hours spent tidying up the code or doing relaxing and pretty refactoring).

TLDR; this AI thing works really well at least for me. But it comes with trade-offs that might slow down its adoption by companies en masse.

I do use LLMs for coding and the newer models have definitely been a blessing. I don't know about using coding agents (or agentic coding) though. I personally do not find this better than chatting with the llm, getting the code back and then copy / pasting it and grokking / editing it. The author of this seems to suggest that.. there is one correct flow, his flow (which he doesn't entirely detail) and everything else is not appropriate. He doesn't go into what his process is when the LLM hallucinates either. Not all hallucinations show up in static analysis.

Yes! It's amazing how even in a field that tends to lean more early adopter than average you still get a lot of the default knee-jerk dismissal and cynicism - even when it's something clearly amazing and useful as thinking machines.

We're in the middle of a major shift - there will benefits to those that adapt first. People outside the field have no idea what's coming, even those of us in the field are underestimating the shift.

There were a few outliers in the 60s who understood what the computing revolution meant and would mean, but most did not. This is likely an even bigger change than that.

"If you build something with an LLM that people will depend on, read the code. In fact, you’ll probably do more than that."

Thus sucking up all the time you "saved" by having the LLM write the code you saved. The machine made the easy parts of coding easier while making the hard parts harder.

"You’ll spend 5-10 minutes knocking it back into your own style."

As if style is the only thing that matters?

"LLMs are showing signs of adapting to local idiom, but we’re not there yet."

First step fallacy. Ladders are showing signs of being tall enough to reach the moon, but we're not there yet.

"If you’re making requests on a ChatGPT page and then pasting the resulting (broken) code into your editor, you’re not doing what the AI boosters are doing"

I am actually doing this the whole day long. For example i have setup today a fresh new debian vps for some interns. U had to provide them with a docker system, support for go, nginx stuff and i made a quick hello world app in angular with a go backend. I could have done it myself. But i asked chatgpt to provide me with all the commands and code. No idea how an agent could do this for me. I got everything running in like 30 minutes.

If you’re leaning out, spend two weeks leaning in.

I did, and learned a ton, and likely not going back to how I was before, or how I used it a week ago.

The comments in the article about not reading the agent is good but it’s more than that…

Vibe coding is for non-coders. Yet, you get a feel for the vibe of the AI. With windsurf, you have two or three files open, and working in one. It starts smashing out the multi, interspersed, line edits and you know with a flutter of your eyes, it’s got your vibe and correctly predicted your next ten lines. And for a moment you forgive it for leading you astray when you read what it said.

LLMs are a tool, a very precious one if you know how to use it.

I've seen Antirez use it for hard to find bugs in the Redis codebase (videos are on YouTube, albeit Italian), I've seen Cursor refactor in a matter of minutes code that would've taken hours, correctly, I've seen security researchers leverage it to find complex edge cases, etc

My 2 cents is that developers that won't learn to use the tools will be left behind (albeit not for long imho), but also, that sales people promoting the idea that devs can be replaced equally dangerous.

> LLMs can write a large fraction of all the tedious code you’ll ever need to write.

For most software projections, they cannot even remotely do that - at least not today and as far as I can tell.

And that is to say, it's not that they can write a not-so-large fraction. They cannot write any significant fraction of it. They can often write you a draft that helps you get there faster, though. They can also facilitate web search result consolidation etc. which can also help. But - they can also lead you down wrong paths and dead-ends, on occasion.

As a junior, I feel most complexity in software is manufactured. LLMs simplify that mess for me, making it easier to get things done. But I’m constantly hit with imposter syndrome, like I’m less skilled because I rely on AI to handle the tricky stuff. And Gemini is better than me!

I want to skip all the debate and just have someone tell me what setup I need to get to here:

> Professional software developers are in the business of solving practical problems for people with code. We are not, in our day jobs, artisans. Steve Jobs was wrong: we do not need to carve the unseen feet in the sculpture. Nobody cares if the logic board traces are pleasingly routed. If anything we build endures, it won’t be because the codebase was beautiful.

I think this is my key point of disagreement. Professional software developers often don't see themselves as artisans, because Steve Jobs's unseen feet are implemented through our implicit standards of ourselves. I don't go out of my way to be a perfectionist on the job, but neither will I try to push 1000 lines of spaghetti to solve a minor problem that seems like it ought to be a simple fix.

I've worked with people in the past who don't have this instinct, and it's my experience that their productivity is quite low if not negative, even though they often perceive themselves to be 10x developers who can handle all kinds of edge cases super fast. I suspect that most AI-generated code today ends up in this niche.

Here's what to do: Show me a video of LLM fixing four filed issues in the KiCad codebase.

If you do that, I'll swallow my AI skepticism.

I would love to have an LLM that I can turn loose on an unfamiliar codebase that I can ask questions of. I would love to have an LLM that will fill in my Vulkan boilerplate. etc.

I use emacs and Mercurial. You can demonstrate magic to me and I can be convinced even if it's not mainstream.

Rewriting Javascript slop to StackOverflow standards is not convincing me.

Get to it.

(The OAuth stuff posted earlier certainly moved my needle, but the fact that they needed a gaggle of reviewers as well as hand holding when the LLM got stuck mutes the impact significantly.)

> LLMs can write a large fraction of all the tedious code you’ll ever need to write.

But, you still have to read it:

> Reading other people’s code is part of the job...I have to read the code line-by-line anyways.

So instead of writing the tedious code, I only have to read it. Oh but don't worry, I don't have to read it too carefully because:

> Agents lint. They compile and run tests. If their LLM invents a new function signature, the agent sees the error

But remember...

> You’ve always been responsible for what you merge to main.

So now I have to oversee this web of agents and AI ontop of coding? Am I doing more now for the same pay? Am I just speedrunning myself toward lower pay? Is AI adoption a prisoner's dilemma toward lowing my wages hardest?

Because is good at coding compared to many other disciplines (e.g. math), it makes the internal AI politics among programmers more of an issue. Add fuel to that fire baby!

i am no clearer on what setup the author is using? are you hand coding all these non-code agents you mention? You talk about zed, then about cursor - don't they both do the same thing, or are you using 2 IDEs to write your code? What tooling is used for these async agents that are doing things while i eat my breakfast?

I really want to like AI assisted coding, but I guess i am just stuck as to what to do.

I have tried co-pilot, both 2 years ago, and recently. I have tried some other AI plugin on VScode (i thought that was cursor, but I don't see it in the extensions when i now search, so it was probably something else). If anything, I have found with these solutions that they are great for the first few tasks - generate these models, scaffold this, add oauth, etc - but after that they lack all consistency with the code they themselves generated (like if i add a new model, the output is wildly different to what it already created) and it takes more time to rewrite that output than to copy-paste and change whats already there.

@op - could you explain exactly what tooling you are using, maybe give an example of these async agents, etc? Would be most appreciated for us luddites who would love to be more productive.

I agree wholeheartedly with this piece, but I think it might've been taken better if it wasn't written in the style of an HN comment. That being said, definitely doesn't deserve so much flaming in the comments, sorry to see that, because it's basically correct, with some small things I disagree with here and there. Feels exciting to be on the precipice of having a Star Trek style spaceship computer at my fingertips.

I've observed that AI excels at creative work, but not exacting sciences. For example, If I say, "Write the code to put the button in the upper right corner," I'll almost certainly get a button in the upper right.

Conversely, if I say, "Put the button in the upper right corner exactly like the existing button in the upper left," I'll get a button in the upper right, but not exactly like the upper left button.

I don't doubt that LLMs will have a large impact on software and programming. But I do wonder whether it will be a positive impact. I think one of the biggest problems the software industry has is that it's far too rare for the people building software to have deep explanatory knowledge about what they have built. LLMs threaten to make this problem vastly worse than it already is.

Here's an idea... imagine you are just starting, you know nothing... an AI can and will help you.

Tell me you won't use it and explain properly, without appeal to "legacy" or "human values" or other abstracts which don't self validate.

Now note that this is POV for every child on earth.

The new discipline is provider of LLM/AI guard rails, and absolutely requires knowing what is/not correct.

AI makes me sad. When I started my CS degree, I didn't even know what silicon valley was. I was unaware of what the SWE job landscape was like. I went to school in a no-name town.

Computer science was an immensely fun subject to learn. I moved to one of the big cities and was bewildered with how much there was to learn, and loved every second of it. I gradually became good enough to help anyone with almost anything, and spent lots of my free time digging deeper and learning.

I liked CS and programming - but I did not like products built by the companies where I was good enough to be employed. These were just unfortunate annoyances that allowed me to work close enough to what I actually enjoyed, which was just code, and the computer.

Before LLMs, those like me could find a place within most companies - the person you don't go to for fast features, but for weird bugs or other things that the more product-minded people weren't interested in. There was still, however, an uncomfortable tension. And now that tension is even greater. I do not use an LLM to write all my code, because I enjoy doing things myself. If I do not have that joy, then it will be immensely difficult for me to continue the career I have already invested so much time in. If I could go back in time and choose another field I would - but since that's not possible, I don't understand why it's so hard for people to have empathy for people like me. I would never have gone down this path if I knew that one day, my hard-earned-knowledge would become so much less valuable, and I'd be forced to delegate the only part of the job I enjoyed to the computer itself.

So Thomas, maybe your AI skeptic friends aren't nuts, they just have different priorities. I realize that my priorities are at odds for the companies I work for. I am just tightly gripping the last days that I can get by doing this job the way that I enjoy doing it.

This is the same as every other transformative tool which has come to market. You've got a certain segment of the population who will embrace it and push the boundaries of what is possible with the current tools and those in the future. And you've got the curmudgeons who won't let go of the one tool and one non-cloud hosting solution they so desperately cling to and deny the utility of anything else.

> Extraordinarily talented people are doing work that LLMs already do better, out of spite.

So what, people should just stop doing any tasks that LLMs do subjectively better?

I think he's pretty much right, but it just leaves me managing an AI and hopefully having fewer offshore developers and faster turnaround. So something less than a 50% boost, and kind of variable, but still significant. I don't really think it's world shattering because most software is deeply constrained by business processes (deciding what to build, etc.) that this doesn't help.

At first I didn’t think LLMs were that impressive, but ever since I started using them, I’ve pretty much stopped writing for loops by hand. It’s not that I forgot how. I just don’t feel like doing it anymore. The time it saves me on boilerplate lets me focus on other things. Sometimes the code isn’t perfect and I need to fix it, but it’s still way easier than starting from scratch.

This blog post makes me actively want to use something besides fly. Putting aside the massive condescension in the article…

… most of the code you write is tedious boilerplate? What do you do all day? I can honestly say the most tedious code I write is migrations, because I have to edit three files. I do this like once a week. It would take way longer to explain to an ai what to do than just do it myself.

… you run the ai on your procrastination and as a result do heads down work? Seriously? 1, programming with AI is like playing a slot machine. It keeps you hooked just on the verge of winning the jackpot for far longer than the tedious task would have distracted you. And let’s say a good agentic case - it runs for 30 minutes before it’s done. That’s not even enough time to “boot up” a serious problem, and your already having to put it aside to go check the work you yourself are saying is mediocre? And you’re advocating to run multiple of these? Even the most ardent vibe covers I know talk about babysitting the agents. How is this unlocking deep work time in any way shape or form?

… craft is relagated to hobbies? I guarantee you, this person loves Linear. It was probably written on a MacBook. This whole “we solve real problems” BS is just a cop out because being great is hard.

… mediocre code is good code? All code is a liability. If you generate 10 times the code, and are admitting it’s the lowest bar of quality possible, it is at a minimum 10x the liability on your shoulders.

… letting LLMs refactor your tests? Serious question, who actually does this? Because I would LOVE to have LLMs write tests for me. I have tried every. single. time. I need to write something that needs a lot of tests to ask LLMs for tests for TWO YEARS, and not once have I ever used a single test it generates. It generates worse tests than it does code, and the last thing I want is 42 asserts in 8 tests across 600 LoC that nobody understands that poorly tests what should have been 2 asserts in 2 tests in 12 lines of code.

It depends on what level of nuts you mean. Some are AGI skeptics about LLMs, theyre probably right, there is likely more breakthroughs required before true AGI. But AGI isn't required to completely disrupt a ton of good, well-paid professions. That is the more worrying scenario. AI is already widening the wealth gap irreparably and with more progress it will only continue.

"Reading other people’s code is part of the job. If you can’t metabolize the boring, repetitive code an LLM generates: skills issue! How are you handling the chaos human developers turn out on a deadline?" Good point! Treat AI generated code as if somebody else had written it. It will need the same review, testing and refactoring as that.

> but it’s bad at rust

Is it really? I scraped all HN comments with the words "Rust" and "LLM" from the last 6 months, put it into Gemini and asked it to surface quotes of people saying they had a good time, versus quotes of the opposite. It was about 60% positive and 40% negative.

Does anyone have recent personal experience with this they can talk about?

tptacek, curious question: what agent / stack do you currently use?

This sounds very cool - almost science-fiction in parts. I want to know more details about the specific technology, whether it could handle a 30k C++ codebase, how much work you would need to do to make it learn how to run your tests, analyse their output, interface with all your internal software development services etc.

Try getting your fav LLM to do something useful with PowerCLI. That's the VMware PowerShell thing. It is quite popular but not enough, so that most prompts will end up with bollocks, masquerading as code.

When a LLM has been tamed to the point that I can define and request a PowerCLI programme, I can be fairly sure that AI is a solved thing.

> We’re not East Coast dockworkers; we won’t stop progress on our own.

we could choose to be

of course if you're a temporarily embarrassed billionaire like ptacek, you certainly don't want the workers doing this

I'll be a happy and ignorant nut as long as these models are trained on publicly scraped code without their creators' consent, and the models fail to cite the licenses and sources of the parts the answer is composed from.

Oh, GH Copilot was adding citing capabilities which was going to be in production around this time? Any progress on that front?

Thanks.

This article doesn’t address literally THE argument I have against LLM usage: that it will degrade major human skills like reading, writing, and yes, coding

But I’m at least grateful for LLMs making me more aware of how important it is for me to focus on my own skills building. I will never use it because I value having those skills sharp

I was quite happy and entlohntes when I dropped an HTML table Formates with Emojies into my prompt and told to turn it into a terraform module.

But how do you quality check this part? I am proofreading all these transformations and I am not sure if this is really faster than actually recording a VIM macro, which will do the steps I need.

Ask me again in 15 years. Assuming the world hasn’t already entered a war for the remaining resources on this planet.

>Important caveat: I’m discussing only the implications of LLMs for software development. For art, music, and writing? I got nothing. I’m inclined to believe the skeptics in those fields.

What a strange disclaimer. Image generation was a (career)killer feature long before vibe coding was halfway useful.

Can someone explain to me what this means?

> People coding with LLMs today use agents. Agents get to poke around your codebase on their own. They author files directly. They run tools. They compile code, run tests, and iterate on the results. ...

Is this what people are really doing? Who is just turning AI loose to modify things as it sees fit? If I'm not directing the work, how does it even know what to do?

I've been subjected to forced LLM integration from management, and there are no "Agents" anywhere that I've seen.

Is anyone here doing this that can explain it?

Adding fuel to the fire. I'm extremely disappointed to see such an inflammatory article on fly.io. I wouldn't want this on my personal blog, let alone on my employers, but I guess they're fine with it.

I've been using Zed and Claude Sonnet 4 (and sometimes trying Opus) heavily over the past weeks. For small edits where I have lots of unit tests, the results were great. So great that they worry me with regards to job security. For exploring a new programming domain it was also somewhat useful. I work a lot with the Typescript compiler API right now, and it has almost no documentation. Since the AI can see into every GitHub repository out there, it's much better, and more efficient, at learning APIs based on code from other folks. On the other hand it means I don't do that job, and I am forced to rely 100% on how the AI presents the Typescript compiler API to me. Are there better methods I could use? Who knows.

Where it's abysmal is code architecture. Sometimes it's almost comical: it adds an if statement to handle one highly specific edge case in a program that only makes sense if it solves the general case. This didn't happen often thought.

The hardest part was to force it to reuse existing code from the same file. My use case is transforming a Typescript AST into a GraphQL AST. The code is one big switch statement with lots of recursive calls. The AI would often add 300 lines of code that duplicate some logic which already exists somewhere else.

In the end I rewrote the whole thing from scratch. At around 900 lines of code the AI was starting to really struggle. When I wanted to take over, I realized that I didn't have the in-depth knowledge to do so. And trying to understand the code the AI had written proved futile.

Ultimately that's on me, I should have been more diligent reviewing the dozens of 300 line of code changes the AI throws at me over the course of a day. But I wasn't, because reviewing is really, really hard. For many reasons. And AI makes it even harder.

Am I therefore nuts? I find this whole article extremely one sided. Surely, based on the sheer amount of both positive and negative press, the answer is somewhere in the middle.

This is also where I am, and I guess it has been a source of mild and growing consternation since I first blagged an OpenAI GPT account when they were private, in an attempt to get ahead of what was coming -- both the positive and negative sides of the advances. Most people either ignored the advances, or quickly identified and connected to the negative side, and effectively filtered out the rest.

As somebody who comes from a politically left family, and was also around in the early days of the Web, let me tentatively note that this issue has a particular political slant, too. The left has strong roots in being able to effectively critique new developments, economic and social, that don't come from its own engines of innovation which have revolved around those critiques, plus solidarity, organization, and sociopolitical action.

The movement's theorists work far more slowly on how to integrate the effect of those changes into its vision. That means when something like this comes along, the left's cultural norms err on the side of critique. Which is fine, but it makes any other expression both hard to convey, and instantly suspect in those communities. I saw this in the early Web, where from a small group of early adopters of all political slants, it was the independents, heterodox leftists, and the right, -- and most vocally, the libertarians -- who were able to most quickly adapt to and adopt the new technology. Academic leftists, and those who were inspired by them took a lot longer to accomodate the Net into their theses (beyond disregarding or rejecting it) and even longer to devise practical uses for it.

It wasn't that long, I should say -- a matter of months or years, and any latent objections were quickly swamped by younger voices who were familiar with the power of the Net; but from my point of view it seriously set back that movement in practicality and popularity during the 80s and 90s.

I see the same with AI: the left has attracted a large generational of support across the world from providing an emotionally resonant and practical alternative to the status quo many people face. But you quickly lose the mandate of heaven if you fail to do more than just simplistically critique or reject a thing that the average person in the world feels they know better, or feels differently toward, than you do. This is something to consider, even if you still strongly believe yourselves to be correct in the critiques.

I never used AI, microservices, blockchain, Mongo, and used only 2 js frameworks in my carrer.

> “hallucination” is the first thing developers bring up when someone suggests using LLMs, despite it being (more or less) a solved problem.

i don't get it. It keeps making stuff up or gives me wrong stuff that merely compiles when i tell it signature doesn't exist.

Every time I read one of these it feels like I'm reading an AI generated sales pitch for AI.

I tried Copilot many times and found it lacking. I like discussing concepts (maybe with pseudocode) with Gemini. I find this useful but not game changing.

So genuine question for AI Boosters, what should I be using?

I would like the 10x productivity gains we were promised :)

What is the current meta on a fully agentic workflow as he describes it? Wake up, write a vague prompt, go for coffee, and get a PR to review on your phone? I'm genuinely interested to see what going all in on this looks like.

This is very well written article. I love the style: aggressive, youthful, humorous, informed, experienced.

I think it could do with a little bit more friendly, but… it’s got its own charm.

It also convinced me I should be using agents rather than raw dogging Gemini 2.5 pro

I've yet to see an LLM produce anything but disastrous bugs. Remember they don't know they're writing code that can destroy companies and livelihoods. They are just spitting out plausible but always wrong text.

For writing, I do a back-and-forth about what I'm trying to do with Claude before asking it to write anything. Produces much less "that's not exactly what I was looking for..."

One thing that really bothered me that the author glossed over (perhaps they don't care, given the tone of the article) is where they said:

> Does an intern cost $20/month? Because that’s what Cursor.ai costs.

> Part of being a senior developer is making less-able coders productive, be they fleshly or algebraic.

But do you know what another part of being a senior developer is? Not just making them more productive, but also guiding the junior developers into becoming better, independent, self-tasking, senior coders. And that feedback loop doesn't exist here.

We're robbing ourselves of good future developers, because we aren't even thinking about the fact that the junior devs are actively learning from the small tasks we give them.

Will AI completely replace devs before we all retire? Maybe. Maybe not.

But long before that, the future coders who aren't being hired and trained because a senior dev doesn't understand that the junior devs become senior devs (and that's an important pipeline) and would rather pay $20/month for an LLM, are going to become a major loss/ brain drain domestically.

Eric S. Raymond is now doing 100% of his coding with AI assistance, and claims he is much faster and more productive.

I think we've passed the threshold beyond which not using AI is just doing things the hard way for no appreciable gain.

>People complain about LLM-generated code being “probabilistic”. No it isn’t. It’s code. It’s not Yacc output. It’s knowable

okay wtf is this weird jab at yacc supposed to be? I generally prefer to write my own LR parsers because it's really not that difficult but I never felt like yacc code was difficult to read or not "probabilistic".

AFAIK most serious compilers don't use yacc either because they need to be able to parse context-dependent grammars, and also reinventing wheels is often worth the effort when the wheel is 50 years old and forces you to make compromises because it also has to support use-cases you are not interested in. But again, it's not bad, it does the thing it is designed to do and it does save a lot of effort if you're fine dealing with its limitations.

Only serious complaint I have about Yacc is that the basic posix form has frustrating limitations such as not being able to have two separate yacc instances because they'll declare the same symbols, and I'm fairly certain it's not thread-safe either. Bison fixes some of these while still remaining "probabilistic".

Anyways I'd rather use Yacc+Lex (preferably Bison+flex) to generate my parser than have to go read everything your "agent" wrote with a fine-toothed comb to make sure it didn't introduce any crippling memory corruptions or dangerous ROPportunities.

https://www.nytimes.com/2025/06/02/opinion/ai-coders-jobs.ht...

> Some of the smartest people I know share a bone-deep belief that AI is a fad — the next iteration of NFT mania.

> The median dev thinks Star Wars and Daft Punk are a public commons. The great cultural project of developers has been opposing any protection that might inconvenience a monetizable media-sharing site. When they fail at policy, they route around it with coercion. They stand up global-scale piracy networks and sneer at anybody who so much as tries to preserve a new-release window for a TV show.

So it starts with a humble strawman, and then the author illustrates how dumb they think their fellow developers really are if they have the make different decisions based on different values.

Most developers I interact with who are what you might call 'AI skeptic', are most concerned with how this tech will be used, especially since its creation is rooted in mass scale disregard for ownership of anything.

> "The great cultural project of developers has been opposing any protection that might inconvenience a monetizable media-sharing site.... They stand up global-scale piracy networks",

IMO the author is here projecting or something, cause I literally never met someone with this opinion, and I've also been in tech/developer circles for over 20 years. Personal use of pirated content is very different from commercial distribution or making money of share sites. Not everyone's values are so rooted in making money with complete disregard to the impact of their actions in doing so.

I get it, the author wants to keep using LLMs are for people to stop trying to make them feel bad but trying to make a case for how their arguments are just dumb. But the author completely missed the 'why' behind the arguments. In the end, LLMs are a tool. Use them or not is up to the individual. But that doesn't give a blanket social license to use them in any way people, or more importantly, companies want.

> You’ll spend 5-10 minutes knocking it back into your own style

You lost me here. I have often found it to be far more than a 10-minute style issue, but fundamental misunderstanding of the code purposes that I need to fix.

If they’re regurgitating what’s been learned, is there a risk of copyright/IP issues from whomever had the code used for training? Last time I checked, there’s a whole lotta lawyers in the us who’d like the business.

Someone at work said something to the effect of "Oh, this would be great for an intern but we don't have one this year" to which I responded "sure you do, its name is Copilot".

Which AI agent is actually good nowadays? I've tried cursor but i didn't want a fork replacing vscode just to be able to call an LLM. And I'd preferably want pay as you go, not a subscription model

I am in the middle, neither a dogmatic skeptic nor a full-blown prompt engineer, but I lost it when the author compared a junior developer (a human) to a SaaS subscription.

Tells you what you need to know about the AI culture.

That was a stellar read. I've had (at least parts of) many of these thoughts floating around my head over the past few weeks / months, but it'd have been beyond my ken to write them down as lucidly.

What is the agent setup that I should try?

What I tried is indeed "set up Cursor" or "copy-paste ChatGPT", and it was all bad, I have no idea what is this "agent" and "MCP" stupp.

Whilst generative AI is great for writing business letters and essays and all that, going straight for coding, one of, if not the hardest problem is a bit much.

This is a come back in 5 years when it works sort of thing.

I'm a recent convert. I _was_ a sceptic until I wasn't (https://news.ycombinator.com/item?id=44059847).

Don't be pennywise and pound foolish...

The LLM coding agents might help you fire some swe to save a penny or two, but soon enough, it will also help your customers or suppliers to replace your business model.

This isn't about the post itself, but can you please put the author's name at the top? It's written in first person and we have no idea who's talking until the end of the post.

Seems to me that AI itself will take some jobs but the cheap workable software the AI creates will take even more jobs, especially if the AI can create workable software for automation robots.

I think LLMs are great and patient teachers.

But code is something I like to write myself, it's not about speed of producing code. It's keeping it such you still fully understand it.

"You don’t believe in IPR. Then shut the fuck up about IPR."

As a software engineer, I need to call out this rhetoric. We are not all morally bankrupt. Many of the people with the most ethical backbone I know are also software engineers.

I don't use "AI" for the same reason I don't use amphetamine. I'm sure I could get stuff more stuff done faster if I used it, but I don't, because the manufacturing and the distribution of it is completely unethical.

The blatant stealing of intellectual property is only the tip of the shitberg. As bad are the giant ecological footprint[1], the colonialism of the grossly underpaid RLHF, and the conceding of even more of our digital lives to trillion dollar corporations.

We can and must ethically evaluate our work, and that includes our tools. Nobody's perfect, but doing one immoral thing does not mean you get a free pass to skip morality altogether.

1. https://www.greenpeace.de/publikationen/environmental-impact...

Sometimes AI model generates technical debt. It happens for instance if AI generates code with dependancies that to old to be automatically be updated to the latest version.

This article is ridiculous

Q: but the code that AI generates is incoherent gibberish

Guess what, life’s tough. Fixing incoherent gibberish is part of the job you wuss. Honestly, I bet you write incoherent gibberish. Would you rather be fixing incoherent gibberish that an AI wrote or incoherent gibberish that you wrote?

Q: but I have to re-do everything it touches. Everything takes me ten times as long!

What are you, some kind of vibe-coding YouTuber? Everything taking ten times as long is part of the job, cupcakes. What’s important is that now it takes ten times as long automatically. I like the way Zed does it, where it tells you to leave while it’s changing stuff.

Ironically all nuts friends seem to be talking less, less defensive than other nuts friends who keep talking about other people not buying what they believe nuts.

I can’t wait for the day when people no longer manually write text messages to each other, but instead just ask LLMs to read and respond from a few prompted words.

this is one of those fascinating cases where i agree with none of the arguments, but vehemently agree with the conclusion… it ordinarily would give me pause, but in this case i am reminded that nonsense arguments are equally applicable to both sides of the debate. if the arguments actually had logical connection to the conclusion, and i disliked the arguments but liked the conclusion, that would be real cause for introspection.

I develop space-borne systems, so I can't use the best LLM's for ITAR/etc reasons, but this article really makes me feel like I'm missing out. This line in particular makes me wonder if my skills are becoming obsolete for general private industry:

Every once in a while I see someone on X posting how they have 10 agents running at once building their code base, and I wonder if in 3 years most private industry coders will just be attending meetings to discuss what their agents have been working on, while people working on DoD contracts will be typing things into vim like a fool

Where is the counter argument to this not being sustainable?

What's the likelihood that this was 90%+ written by an LLM? It'd be ironic, especially now that there are 2000+ engineers commenting on it.

The best I can offer skeptics is the more you work with the tools the more productive you become. Because yes the tools are imperfect.

If you've had a dog you know that "dog training" classes are actually owner training.

Same with AI tools. I see big gains for people who spend time to train themselves to work within the limitations. When the next generation of tools come out they can adapt quickly.

If this sounds tedious, thats becuase it is tedious. I spent many long weekends wrestling with tools silently wrecking my entire codebase, etc. And that's what I had to do to get the productivity improvements I have now.

> For art, music, and writing? I got nothing. I’m inclined to believe the skeptics in those fields. I just don’t believe them about mine.

I am curious about this part though

> but it’s bad at rust

It's not just Rust. It's everything that has been created post the last vacuum. You need a large corpus for LLM to get decent-ish.

AI is a tool, akin to a seamstress' sewing machine. Use it well, and with intention, and it'll help speed up your work. That's it.

Soon all coding will look like L3 support - debugging something you've never seen before, and under pressure. AI is really taking away the fun parts from everything and leaving just the drudgery in place.

Call me old but I am not using AI for coding. It's not what I expect or want an AI to do for me. I would love for it to do stuff I don't like to do. It's like the meme that was floating around: that a woman says it would have loved the AI to do the dishes so she can to art; not the other way around. Yes I am also a skeptic but that is not weird as I have seen to many movies about AI taking over the world. I hope I never have to say: "I told you so". I really do see positive points to AI but just not in coding. I would not even trust them to write tests. Might make them write documentation at some point.

I would like to point out: LLM's look really intelligent until you converse with them in a less popular language, it will make weird mistakes and interpret your questions all wrong. I am Dutch and they really suck at that.

You’re the goat, Ptacek.

Seriously, this is the most based defense of LLM that I’ve read over the years. All the yapping in the thread only proves your points.

Can someone guide me which AI tools are good? And their docs and tutorial

I want to see the hype but without paying up front. My focus area is c#, .net, js.

Very weak arguments and intentionally provocative.

>but the code is shitty, like that of a junior developer >Does an intern cost $20/month? Because that’s what Cursor.ai costs.

RIP Junior devs

LLM, please find all occurrences of "shut the fuck up" and suggest actual arguments for the author to use.

Otherwise, well written and persuasive.

Thanks for the article. I've been unwilling to try agentic editing for a while, and this is the kick I need to give it a shot.

There is no AI yet, it’s just automation and that’s why people hate it. It also allows corporations to consume open source code for profit.

> I think this is going to get clearer over the next year. The cool kid haughtiness about “stochastic parrots” and “vibe coding” can’t survive much more contact with reality. I’m snarking about these people, but I meant what I said: they’re smarter than me. And when they get over this affectation, they’re going to make coding agents profoundly more effective than they are today.

Yes, yes, and yes. SOOOO much yes. We are going to be in for a wild ride once everyone gets on board with the reality of coding agents.

> Does an intern cost $20/month? Because that’s what Cursor.ai costs.

The cost is a pretty obvious indication that it's nowhere near the capability of an engineer. A robotic worker that can produce code 24/7 without tiring would be worth even more than a person of equivalent skill. I don't know what kinds of goofy hiring ya'll are doing, but all of the juniors I've been involved in hiring actually worked out great.

I find it interesting that TFA claims that Go is the most LLM legible language.

I suspect C# Asp.Net Core not so much. Any thoughts?

OP opened with an ad hominem argument that his skeptic friends are all nuts.

OP therefore loses the argument.

> but the code is shitty, like that of a junior developer > Does an intern cost $20/month? > Because that’s what Cursor.ai costs.

So sick of hearing comparisons between LLMs and Juniors, first of all are we really are seniors? After almost 20 years sometimes I still think I know nothing. Second, when I was a Junior just me being younger, fresher and with an eager learner attidute brought a net positive impact on the team I joined back then, that I still hang out with those people to this day and we still have fun joining new projects together. I'm so sick of these snobbish "seniors" that keep throwing rocks at younger and well prepared people

I don’t use AI and I don’t really care for it, but the language in this article is really bad.

I expected better from Thomas.

I find the claims of 2-5x more productivity insane. I'm at least 10x more productive.

Coming around to this conclusion myself after experimenting with the tools for a few weeks.

Things have changed.

I was willing to read until I got to this part. He can shove this drivel up his own ass for all I care if that's how he thinks. What a waste of time.

> Also: let’s stop kidding ourselves about how good our human first cuts really are.

but I like writing code, rather than reviewing code somebody/thing else wrote

Really enjoyed this article and how it was written, especially the “tookr-jerbs” part.

I mean, the root cause of this is quite mundane.

People like to think that intelligence is some intrinsic force but it's far better thought of as situational. We are contextually smarter or dumber based on the emotions we're trying to process from a situation and how our motivation reasoning allows us to avoid certain negative emotions.

The subset of people the OP is trying to argue against are people who are otherwise normally intelligent but have an emotional need for AI to be unimportant that they're unable to process so they build a social community of terrible arguments.

The nature of the internet is that visibility is terribly correlated (and often anti-correlated) with popularity so OP thinks this is some large contingent of people when it's instead algorithmically served conflict generators.

Writing against that is also an algorithmic conflict generator, as evidenced by the number of upvotes here. The correct response is to realize that people putting up terrible arguments about AI don't actually care about AI, they're using this issue to process unresolved trauma from other, unrelated areas of their life and taking their ideas literally instead of seriously is a waste and diversion. There are plenty of smart people with far more nuanced views that still disagree with you where they're interested in the content of the argument.

I don't get the focus on the coding. In all the projects I worked on for 25 years, coding has never been the bottleneck in projects. Maybe the projects I work on are simple, maybe I was lucky, but still: once the requirements are clear, coding is easy. And it's literally what takes the least time. I don't want to spend less time on coding thanks to AI. Because coding is what I like, not the rest of the job. Bring me AI that will do the meetings for me please.

*sighs in aerospace*

Let me know when I can use one of these without violating ITAR/EAR.

Spot on. Current gen LLMs also has the ability to mimic your code style. I always point it to existing code, and the result is the same as what I would have written.

Yesterday I “wrote” 18k lines of code with Claude & Cursor. Mostly UI, some backend. This would have taken me a week without LLMs.

I’m sold.

>Professional software developers are in the business of solving practical problems for people with code. We are not, in our day jobs, artisans. Steve Jobs was wrong: we do not need to carve the unseen feet in the sculpture. Nobody cares if the logic board traces are pleasingly routed. If anything we build endures, it won’t be because the codebase was beautiful.

Maybe you are, but I didn't get a Master's degree to produce quick shit for shareholders

basically nails every point i've wanted to say but couldn't lol. also in the end the endless posturing gets tired and we go back to getting shit done. either it helps you are it doesn't.

"they’re smarter than me" feels like false humility and an attempt to make the medicine go down better.

1. Thomas is obviously very smart.

2. To be what we think of as "smart" is to be in touch with reality, which includes testing AI systems for yourself and recognizing their incredible power.

I have a trick to improve code quality from LLMs. I only use the ones where I have control of the system prompt.

I use a prompt that convinces the LLM they are a Staff Engineer at Google and care deeply about best practices and writing clean, quality code.

While not perfect, it does improve the code quality.

i feel like surprisingly, front end work which used to be viewed by programmers as "easier" is now more difficult of the two, because it's where LLMs suck the most

you get a link to a figma design and you have to use your eyes and common sense to cobble together tailwind classes, ensure responsiveness, accessibility, try out your components to make sure they're not janky, test out on a physical mobile device, align margins, padding, truncation, wrapping, async loading states, blah blah you get it

LLMs still suck at all that stuff that requires a lot of visual feedback, after all, you're making an interface for humans to use, and you're a human

in contrast, when i'm working on a backend ticket ai feels so much more straightforward and useful

My take is: It is OK to don't buy into the hype. There's a lot of hype, no denying that.

But if you're actively avoiding everything related to it, you might find yourself in a position where you're suddenly being left in the dust. Maybe not now, not next month, not next year, but who some time in the future. The models really are improving fast!

I've talked with devs that (claim they) haven't touched a model since ChatGPT was released - because it didn't live up to their expectations, and they just concluded it was a big nothingburger.

Even though I don't follow the development religiously anymore, I do try to get acquainted with new releases every 3 months or so.

I hate the term "vibe coding", but I personally know non-tech people that have vibe coded products / apps, shipped them, and make more money in sales than what most "legit" coders are making. These would be the same "idea people" that previously were looking for a coder to do all the heavy lifting. Something is changing, that's for sure.

So, yeah, don't sleepwalk through it.

This article is very well-written. Obviously not written by AI.

I am still using ChatGPT like a caveman and am happy with it.

I thought so about strong static typing looong time ago.

Why would anyone rather read and fix someone else code rather than writing the code themselves? I do a lot of code review for other human code and it use so much more energy than writing my own code (and surely, as I have competent colleagues, this is not even as bad as if I expected that the code that I'm reading could be totally random shit)

Both off and on topic: I got access to Claude’s voice mode today, and for my second test I asked it for its thoughts about both the linked blog post and the HN discussion here. The full transcript is at [1]. Some parts that stood out (slightly edited):

Me: So what is your assessment of the discussion at Hacker News?

Claude: The Hacker News discussion strikes me as much more grounded than the blog post itself. What I find compelling is that many commenters are sharing specific technical experiences rather than making broad ideological arguments. When someone describes spending hours fighting with Claude agents that keep refactoring code into spaghetti, that's concrete data about the current state of these tools. ...

Me: Well, what about you, Claude? You are made by a company whose president has recently made some public statements saying that, you know, 50% of white collar jobs will be replaced soon. And the company, Anthropic, also is very aggressively developing and selling tools to help with AI coding. What are your thoughts about that?

Claude: That's a really fair question to ask me directly. I think there's an obvious tension in my position here since I'm built by a company that has clear financial incentives around AI adoption, including coding tools. When I look at my own capabilities honestly, I can see both sides of this. I'm genuinely useful for certain coding tasks like explaining unfamiliar code, generating boilerplate, or helping debug specific problems. I've helped people learn new programming languages faster than they might have otherwise. But I'm also clearly not at the level where I could replace an experienced software engineer working on complex system design or novel problem solving. ...

[1] https://claude.ai/share/c934e67c-f132-43e8-8c31-ab238b5be25a

Another one, here we go :) I think I'm going to keep it short because I am not here to convince people they're wrong, just to voice my dissent from uh almost every viewpoint this article comes from.

> Professional software developers are in the business of solving practical problems for people with code. We are not, in our day jobs, artisans.

Well that's just wrong - we have failed to capture tremendous value through knowledge of our craft and to reduce it down to min-maxing effort/profit is not a pursuit that I can get behind. Nobody is talking about building Ivory towers, but we can do a lot better than most of the common denominator technologies in use today. We're failing our users most of the time (and of course people like the author say it doesn't matter because it doesn't show up in the quarterly results).

> Nobody cares if the logic board traces are pleasingly routed.

Speak for yourself.

I think it comes all down to that, do you have pride in what you do or you don’t ?

I make a wall with bricks, even if it will be covered with coating i will do my best to have regular joints and pacing.

Could make it faster, no one would notice the difference but me… i hate that feeling when you done something and you know it’s barely enough, just barely, it’s kind of shit and you really don’t want others to see it.

On the opposite side, some people will take pride in building wall twice as fast as me and won’t care it’s horrendous.

Both cases are valid, but me i know i can’t do a work I’m not proud of.

My AI enamored friends are all nuts.

I get it. You're being so productive. (Or so you claim. I don't believe you.)

But if I don't affirm your work choices, or embrace them myself, you lose your mind like a little caffeinated squirrel.

> If you can’t metabolize the boring, repetitive code an LLM generates: skills issue!

Case in point. Touch grass.

I'd be interested in the open source repositories that the author of this article created a) before "AI" and b) after "AI". I mean significant work mostly done by him.

> Level Setting

> First, we need to get on the same page. If you're [6 months out of date with the technology you think of as overhyped after having observed it for the last 4 years or so], you're [not fairly evaluating it].

Perhaps, but this is an extremely unconvincing approach to the argument right out of the gate - it's basically on the same level as "communism has never been tried".

Especially if you're going to start out by describing your interlocutors as "crazy", rather than positing that an effectively completely new thing now exists that ought to be evaluated differently from Copilot et. al.

> You’ll only notice this happening if you watch the chain of thought log your agent generates. Don’t.

"You're nuts!" says the guy with his head intentionally buried in the sand. Also way to tell me your business model is a joke without telling me your business model is a joke. Enjoy it while it lasts.

Man the redbull is oozing off this post, talk about sipping rocket fuel.

I mean a tool is a tool, nothing wrong with that - but most of the resistence stems from AI being shoved down our throats at warp speed. Its already everywhere and I can't opt out, that stinks.

As for the skepticism in terms of adoption and usefulness, its mainly a question of whether or not it will continue improving - there's no way to no what lies ahead, but if it came to a grinding halt today well then the high water mark just isn't all that impressive.

> Yeah, we get it. You don’t believe in IPR. Then shut the fuck up about IPR. Reap the whirlwind.

This is the point that matters, and I don't think everyone is on the same page that LLMs are essentially over glorified data laundering.

The industry would get just as much "value" if we declared a jubilee and wiped out all licenses and allowed unlimited plagiarism (Looking at Zuckerburg and his 10 TB of pirated data). In fact, if AI owners published their training data sets with a capable search engine, I would bet money of it out performing LLMs in most cases. Why waste all that man power reinventing Netflix again? Just copy paste the code and give everyone their time back, sheesh.

I'm in a role that is behind the times, using a bespoke in-house framework that is immune to the benefits of LLMs, so I don't get to see what you see - so as a skeptic, I'm not convinced this isn't just the illusion of speed. I have not seen convincing results, show me the amazing things being made by AI (AI tooling itself does not count) - but yes, maybe that's because its all siloed into walled gardens.

> But something real is happening. My smartest friends are blowing it off. Maybe I persuade you. Probably I don’t. But we need to be done making space for bad arguments.

Yeah all the arguments have been made, good and bad, we're all waiting to see how it plays out. But I'd rather take the side of being a skeptic - if I'm right then I'm in the right place. If I'm wrong, that's cool too, I don't mind playing catch-up. But fully embracing the hype is, IMO, tantamount to putting all your eggs in one basket, seems like a needless risk but if that's worth it to you to get ahead then by all means, slurp up the hype.

The most salient part of the article IMO:

Under the heading: "but you have no idea what the [LLM's] code is."

> Are you a vibe coding Youtuber? Can you not read code? If so: astute point. Otherwise: what the fuck is wrong with you? You’ve always been responsible for what you merge to main. You were five years go. And you are tomorrow, whether or not you use an LLM. If you build something with an LLM that people will depend on, read the code.

> People complain about LLM-generated code being “probabilistic”. No it isn’t. It’s code. [...] The LLM might be stochastic. But the LLM doesn’t matter. What matters is whether you can make sense of the result

> Reading other people’s code is part of the job. If you can’t metabolize the boring, repetitive code an LLM generates [...] how are you handling the chaos human developers turn out on a deadline?

I am just some shmoe, but I believe that devs fall into to major categories when it comes to LLMs: those with their own product ideas, and those without their own product ideas.

The prior look upon Claude Code/Cursor/Windsurf much more favorably, as they are able to ship their ideas much faster.

This is a bit of a hot take, so I would love any replies to bring me back down to earth.

I feel like we get one of these articles that addresses valid AI criticisms with poor arguments every week and at this point I’m ready to write a boilerplate response because I already know what they’re going to say.

Interns don’t cost 20 bucks a month but training users in the specifics of your org is important.

Knowing what is important or pointless comes with understanding the skill set.

Reply on news.ycombinator.com