Why LLMs can't really build software

I think it's going to require a change in how models are built and optimized. Software engineering requires models that can do more than just generate code.

You mean to tell me that language models aren't intelligent? But that would mean all these people cramming LLMs in places where intelligence is needed are wasting their time?? Who knew?

Me.

I have a solution for that, I just need a small loan of a billion dollars and 5 years. #trustmebro
- Only one billion?? What a deal! Where's my checkbook!?

The LLM worship has to stop.

It's like saying a hammer can build a house. No, it can't.

It's useful to pound in nails and automate a lot of repetitive and boring tasks but it's not going to build the house for you - architect it, plan it, validate it.

It's similar to the whole 3D printing hype. You can 3D print a house! No you can't.

You can 3D print a wall, maybe a window.

Then have a skilled Craftsman put it all together for you, ensure fit and finish and essentially build the final product.

I hate the simulated intelligence nonsense at least as much as you, but you should probably know about this if you’re saying you can’t 3d print a house: https://youtu.be/vL2KoMNzGTo
- Yeah I've seen that before and it's basically what I'm talking about. Again, that's not "printing a 3D house" as hype would lead one to believe. Is it extruding cement to build the walls around very carefully placed framing and heavily managed and coordinated by people and finished with plumbing, electrical, etc.
  It's cool that they can bring this huge piece of equipment to extrude cement to form some kind of wall. It's a neat proof of concept. I personally wouldn't want to live in a house that looked anything like or was constructed that way. Would you?

To those who have played around with LLM code generation more than me, how are they at debugging?

I'm thinking of Kernighan's Law: "Debugging is twice as hard as writing the code in the first place. Therefore, if you write the code as cleverly as possible, you are, by definition, not smart enough to debug it." If vibe coding reduces the complexity of writing code by 10x, but debugging remains just as difficult as before, then Kernighan's Law needs to be updated to say debugging is 20x as hard as vibe coding. Vibe coders have no hope of bridging that gap.

They're not good at debugging. The article is pretty spot on, IMO - they're great at doing the work; but you are still the brain. You're still deciding what to do, and maybe 50% of the time how to do it, you're just not executing the lowest level anymore. Similar for debugging - this is not an exercise at the lowest level, and needs you to run it.
- deciding what to do, and maybe 50% of the time how to do it, you’re just not executing the lowest level anymore
  And that's exactly what I want. And I don't get it why people want more. Having more means you have less and less control or influence on the result. What I want is that in other fields it becomes like it is in programming now, so that you micromanage every step and have great control over the result.
Definitely not good. Sometimes they can solve issues but you gotta point them in the direction of the issue. Other times they write hacky workarounds that do the job for the moment but crash catastrophically with the next major dependency update.
- I saw an LLM override the casting operator in C#. An evangelist would say "genius! what a novel solution!" I said "nobody at this company is going to know what this code is doing 6 months from now."
  It didn't even solve our problem.
The company I work for has recently mandated that we must start using AI tools in our workflow and is tracking our usage, so I've been experimenting with it a lot lately.
In my experience, it's worse than useless when it comes to debugging code. The class of errors that it can solve is generally simple stuff like typos and syntax errors — the sort of thing that a human would solve in 30 seconds by looking at a stack trace. The much more important class of problem, errors in the business logic, it really really sucks at solving.
For those problems, it very confidently identifies the wrong answer about 95% of the time. And if you're a dev who's desperate enough to ask AI for help debugging something, you probably don't know what's wrong either, so it won't be immediately clear if the AI just gave you garbage or if its suggestion has any real merit. So you go check and manually confirm that the LLM is full of shit which costs you time... then you go back to the LLM with more context and ask it to try again. It's second suggestion will sound even more confident than the first, ("Aha! I see the real cause of the issue now!") but it will still be nonsense. You go waste more time to rule out the second suggestion, then go back to the AI to scold it for being wrong again.
Rinse and repeat this cycle enough times until your manager is happy you've hit the desired usage metrics, then go open your debugging tool of choice and do the actual work.
- maybe its just me but I find typos to be the most difficult because my brain and easily see it as correct so the whole code looks correct. Its like the way you can take the vowels out of sentences and people can still immediately read it.
- we must start using AI tools in our workflow and is tracking our usage
  Reads to me as "Please help us justify the very expensive license we just purchased and all the talented engineers we just laid off."
  I know the pain. Leadership's desperation is so thick you can smell it. They got FOMO'd, now they're humiliated, so they start lashing out.
- As it seems to be the case in all of these situations, AI fails hard at tasks when compared to tools specifically designed for that task. I use Ruff in all my Python projects because it formats my code and finds (and often fixes) the kind of low complexity/high probability problems that are likely to pop up as a result of human imperfection. It does it with great accuracy, incredible speed, using very little computing resources, and provides levels of safety in automating fixes. I can run it as an automation step when someone proposes code changes, adding all of 3 or 4 seconds to the runtime. I can run it on my local machine to instantly resolve my ID10T errors. If AI can't solve these problems as quickly, and if it can't solve anything more complicated reliably, I don't understand why it would be a tool I would use.
My first level of debugging is logging things to console. LLMs here do a decent job at "reading your mind" and autocompleting "pri" into something like "println!("i = {}, x = {}, y = {}", i, x, y);" with very good context awareness of what and how exactly it makes most sense to debug print in the current location in code.
I use it extensively daily.
It cannot step through code right now, so true debugging is not something you use it for. Most of the time the llm will take the junior engineer approach of "guess and check" unless you explicitly give it better guidance.
My process is generally to start with unit tests and type definitions, then a large multipage prompt for every segment of the app the llm will be tasked with. Then I'll make a snapshot of the code, give the tool access to the markdown prompt, and validate its work. When there are failures and the project has extensive unit tests it generally follows the same pattern of "I see that this failure should be added to the unit tests" which it does and then re-executes them during iterative development.
If tests are not available or if it is not something directly accessible to the tool then it will generally rely on logs either directly generated or provided by the user.
My role these days is to provide long well thought out prompts, verify the integrity of the code after every commit, and generally just kind of treat the llm as a reckless junior dev. Sometimes junior devs can surprise you, like yesterday I was very surprised by a one shot result: asking for a mobile rn app for taking my rambling voice recordings and summarize them into prompts, it was immediately remarkably successful and now I've been walking around mic'd up to generate prompts.
Working just fine. It one shot a kodi tv channel addon for me last week end. Used it to integrate kofax into docusign. Building 2 blazor apps one new one an upgrade. Used it to create a stack of mc servers for the kids with a dashboard of statuses and control switches. My son is working on his own mc mod with it. Use it almost daily for random file organization and management scripts. Using it to clean uo my media library meta data. Anytime i have to do something to more than 5 or so files i pull it up and ask for a script.
Its a tool like any other. There will be people who adapt and people who fail to. Just like we had with computers the internet. It zeems to be long forgotten now but literally ALL of these anti ai arguments were made against computers and the internet 30_50 years ago. Very similar ones were made when books and writing became common place as well.
- "Some random people were wrong about something in the past so nobody is allowed to speculate that any technology isn't as revolutionary as it's hyped to be ever again" is not a useful or compelling argument.
- Apparently I'm not up to date. I've been impressed by some things and turned off by others. But I haven't seen any workflows or setups that enabled access to my file system. How is that accomplished, and are there any safeguards around it?

Clearly LLMs are useful to software engineers.

Citation needed. I don’t use one. If my coworkers do, they’re very quiet about it. More than half the posts I see promoting them, even as “just a tool,” are from people with obvious conflicts of interest. What’s “clear” to me is that the Overton window has been dragged kicking and screaming to the extreme end of the scale by five years of constant press releases masquerading as news and billions of dollars of market speculation.

I’m not going to delegate the easiest part of my job to something that’s undeniably worse at it. I’m not going to pass up opportunities to understand a system better in hopes of getting 30-minute tasks done in 10. And I’m definitely not going to pay for the privilege.

I don't use one, and my coworkers that do use them are very loud about it, and worse at their jobs than they were a year ago.
If my coworkers do, they’re very quiet about it.
Gee, guess why. Given the current culture of hate and ostracism I would never outright say IRL that I like it or use it a lot. I would say something like "yeah, I think it can sometimes be useful when used carefully and I sometimes use it too". While in reality it would mean that it actually writes 95% of code under my micromanagement.
- Wut. At software shops the prevailing atmosphere is that you should use it and broadcast it as much as possible. This person's experience is not normal
https://survey.stackoverflow.co/2025/ai/
47% daily use
- 47% daily use
  That is NOT what that says. It says 47% of STACK OVERFLOW RESPONDENTS REPORT using AI. That does not represent 47% of devs.
  If you go to 4chan and poll of chuds, you're going to get a high percentage of respondents affirming your query. You went to stackoverflow and asked about AI. Think about the user base.
I have been using it a bit, still can't decide if it is useful or not though... It can occasionally suggest a blatantly obvious couple of lines of code here and there, but along the way I get inundated with annoying suggestions that are useless and I haven't gotten used to ignoring them.
I mostly work with a niche area the LLMs seem broadly clueless about, and prompt driven code is almost always useless except when dealing with a super boilerplate usage of a common library.
I do know some people that deal with amazingly mundane and common functions and they are amazed that it can pretty much do their jobs, but they never really impressed me before anyway and I wondered how they had a job...

Good article, I couldn't agree with it more, it's exactly my experience.

The tech is being developed really fast and that is the main issue when taking about ai. Most ai haters are using the issues we might have today to discredit the while technology which makes no sense to me.

And this issue the article talks about is apparent and whoever solves it will be rich.

However, it's interesting to think about the issues that come next.

It's true, the tech will get better in the future, we just need to believe and trust the plan.
Same thing with crypto and NFT's. They were 99% scam by volume, but who wouldn't love moving their life savings into a digital ecosystem controlled by a handful of rich gambling addicts with no consumer protections? Imagine, you'll never need to handle dirty paper money ever again, we'll just put it all in a digital wallet somewhere controlled by someone else coughmastercardcough.
And another thing, we were too harsh on the Metaverse. Sure, spending 8 hours in VR could make you vomit, and the avatars made ET for the Atari look like Uncharted 4, but it was just in its infancy!
I too want to outsource all my critical thinking to a chatbot controlled by an wealthy insular narcissist who throws Nazi salutes. The technology just needs time to mature. Who knows, maybe it can automate the exile of birthright citizens for us too!
/s
- That's exactly the hyperbole I was talking about. Your post is full of obvious fallacies, but the fact that you are pushing everything to the absolutes is the silliest one.

I love it how article baits AI-haters to upvote it, even though it's very clearly pro-AI:

At Zed we believe in a world where people and agents can collaborate together to build software. But, we firmly believe that (at least for now) you are in the drivers seat, and the LLM is just another tool to reach for.

How is that pro-AI? It clearly very neutrally says it’s just a tool, which you can also hate.

The idea of the mental model CAN be done by AI.

In my experience, if you get it to build a requirements doc first, then ask it to implement that while updating it as required (effectively it's mental state). you will get a pretty good output with decent 'debugging' ability.

This even works ok with the older 'dumber' models.

That only works when you have a comprehensive set of requirements available though. It works when you want to add a new screen/process (mostly) but good luck updating an existing one! (I haven't tried getting it to convert existing code to a requirements doc - anyone tried that?)

I tried feeding ChatGPT a Terraform codebase once and asked it to produce an architecture diagram of what the code base would deploy to AWS.
It got most of the little blocks right for the services that would get touched. But the layout and traffic direction flow between services was nonsensical.
Truth be told it did do a better job than I thought it would initially.
- The trick is to split up the tasks into chunks.
  Ask it to identify the blocks.
  Then ask it to identify the connections.
  Then ask it to produce the diagram.