3 ms·

I have yet to run into any serious project in the wild that is using LLMs for development. I have seen vibecoded intern prototypes that took half a day to vet a

by tovej 7mo ago

I have yet to run into any serious project in the wild that is using LLMs for development. I have seen vibecoded intern prototypes that took half a day to vet and dismiss because they were completely useless. I'm sure your experience is different, but you can't _seriously_ claim we're "past the point" of not using LLMs for programming. Vibecoding is a fundamentally different kind of activity than actual programming. It's a pure delusional dopamine rush, compared to the deliberate engineering required to build quality software.

ok_dad 7mo ago
It’s juvenile to consider all LLM assisted coding as vibecoding. I’m not going to expand here because this topic is about as much fun to discuss as politics, but coding assistant tools are just tools. If you give a regular person a race car, they will crash it about as fast as their vibecoded app crashes. Give the same race car to a pro age it’s a different story. I still think this was the right decision by the programming mods there. Talking about tools is pretty boring, and you need to train to use something like an LLM assistant. No one who can’t program a language should be using an LLM to learn it unless they know about 2-3 other languages already, IMO.
- apsurd 7mo ago
  Nah I think it really is more nuanced than that. It is true that a non-technical person's vibe-coded side-hustle is completely different than how a professional developer may ship genAI code, but we're willfully glossing over the real problem that professionals are pushing out TONS of genAI code that's closer to vibes than it is to the pre-AI expectations on pushing to prod.
  - ok_dad 7mo ago
    Devs that are pushing crappy code using these tools are 1) devs who would have pushed shut code anyways 2) getting stress from above to push code faster with AI Good devs will learn to use these tools effectively.
apsurd 7mo ago
For CRUD apps though, the intern closing the ticket literally 30 minutes after it's created is really hard to battle against. Especially when those tickets were created by suits. I generally agree that while I think vibe-coding is here to stay, it's different from designing useful products and systems, and I don't know how to convince colleagues that we should uhh be careful about all this code we're pushing. I fear all they see is the guy aging out.
NietTim 7mo ago
> I have seen vibecoded intern prototypes that took half a day to vet and dismiss because they were completely useless. They weren't useless, they proved if the direction that the prototype was exploring was worthwhile. I've personally made many completely shit code prototypes in the years before we had LLM's, of course they weren't magically production ready, that's not the point of a prototype.
- tovej 7mo ago
  The example I'm thinking of did not. It was just the dumbest way to execute an idea we knew was easy to execute (scanning a parameter space).
altmanaltman 7mo ago
> I have yet to run into any serious project in the wild that is using LLMs for development. How about Claude Code? 100% of it was vibe-coded according to its creator.[1] Google and Microsoft also claim a lot of their internal code is AI-generated now. [2] [3] Naturally, none of the big tech companies will just release a pure vibe-coded project due to structural reasons, but you also _seriously_ can't claim that serious projects don't use LLMs as well these days. Maybe in your limited experience, it isn't true, but that doesn't generalize to what's actually happening. 1. https://www.reddit.com/r/Anthropic/comments/1pzi9hm/claude_code_creator_confirms_that_100_of_his/ https://www.reddit.com/r/Anthropic/comments/1pzi9hm/claude_c... 2. https://fortune.com/2024/10/30/googles-code-ai-sundar-pichai/ https://fortune.com/2024/10/30/googles-code-ai-sundar-pichai... 3. https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-a...
- 59nadir 7mo ago
  > How about Claude Code? 100% of it was vibe-coded according to its creator. Agents are trivial to make. I don't know whether that means they're not "serious", but it's exactly the type of thing you can make yourself in a very short time, and exactly the type of thing even LLMs can't fuck up too bad. With regards to the overall point, I think the existence of projects using LLMs to do development doesn't really lend credence to the idea that they're somehow preferable or desirable. People tend to use hyped things, imagine they're useful even when presented with evidence to the contrary, and generally be very resistant to sobering realities. It took years before people stopped running hadoop clusters to do things that a single linux box could get done 10x faster with some basic pipes. I'm sure there are still people who have "serverless backends" that work terribly in every regard in comparison to literally just a linux VM somewhere. People in software development tend to find these types of things every once in a while and adopt them wholesale. This cycle is helped by the fact that the field has been growing constantly and a lot of the adoption comes from kids who don't know any better. Every piece of shit technology that comes and goes has meat for the grinder coming straight out of university. Would I put LLMs in the same category as these previous (nearly) useless things? Probably not... But you should never trust peoples' perception of usefulness when it comes to almost anything in software development.
dwb 7mo ago
Ok well I have plenty of serious, production-level professional experience that says otherwise. Not “vibe coding” - we certainly review the code. It’s a tool that has downsides and failure modes, of course, but it’s at the point where it’s definitely speeding us up and we are using it a lot. Trust me, I’d prefer a world, on balance, where this wasn’t true – I don’t like many of the aspects and uses of the technology – but its utility in programming is undeniable now and the capitalists aren’t taking “no” for an answer.
- tovej 7mo ago
  I'm curious, where are you using the tools? What programming language? what domain? Are you willing to share the projects you're working on?
  - dwb 7mo ago
    TypeScript and Go on a 1.5 million SLOC production codebase; a complex SaaS tool for financial planning and analysis. Quite far from being “just CRUD”. Before Anthropic Opus 4.5 I was trying out Claude Code and wasn’t all that impressed, but since then it’s definitely helped. The project I wrapped up before Christmas would have gone into the new year without it. You’ve still got to keep a close eye on it; whenever I’ve got lazy with review, trusting it too much, I’ve always regretted it. It’s never one-shotted anything, even with plan mode and all that. I’m a natural skeptic on this stuff and was very actually skeptical for most of last year. But I’m very confident there’s a large net productivity gain now.
    - tovej 7mo ago
      So is this a web app with a go backend? I don't know where you got the "just CRUD" quote from. I never mentioned CRUD. But this sure sounds like it would be CRUD with some additional models in the backend. What makes it not just CRUD? Is it using some complex model for forecasting?
      - dwb 7mo ago
        I wasn't quoting you, they are scare quotes. It is a web app with a TypeScript and Go backend, but to call it a CRUD app would be misleading (despite the fact that yes once you boil it all down, everything is a CRUD operation) because it's a complex and flexible web app. It's the kind of thing that a bunch of people would prefer to be a native app; it's spreadsheet-like.
xboxnolifes 7mo ago
What do you consider "serious", as that seems to be the main differentiator here. I know plenty of serious (multiple years of development and users, and began prior to LLMs) projects that have devs using LLMs for development.
- tovej 7mo ago
  That makes me curious, would you like to list them?
  - xboxnolifes 7mo ago
    They aren't anything particularly special, most projects I interact with have at least 1 dev who uses LLMs in some capacity. The two I was thinking of when I posted are the Minecraft modpack GregTech: New Horizons, and an Old School Runescape plugin. https://github.com/gtnewhorizons https://github.com/gtnewhorizons https://github.com/osrs-reldo/tasks-tracker-plugin https://github.com/osrs-reldo/tasks-tracker-plugin
    - tovej 7mo ago
      I see. I would not consider these to be serious projects. I'm sure they give people joy, and contain some decent problem solving, but they are both small extensions to existing software, and are not really solving interesting new problems or creating their own systems.
      - xboxnolifes 7mo ago
        The first one is hardly a small extension. It's millions of lines of code, which is multiple times more than the version of Minecraft it extends.
        tovej 7mo ago
        The size of the project is not what matters. It's the systems it includes and how complex they are. Actually, that raises a question: are the more complex systems in this collection of projects vibecoded or programmed by a human? Which parts are you aware of that invlude LLM output?
tarsinge 7mo ago
How can you still can't distinguish between using LLMs as tools and a non technical person vibe coding? I have yet to run into any serious software engineer that had to dive into a legacy codebase or an unknown tech stack and found no value in e.g. Claude Code for general understanding and refactoring. Not even talking about coding, just the capacity of generating custom contextualised documentation and examples tailored to your constraints and skills on the fly is ridiculously helpful.
- tovej 7mo ago
  The tool being useful sometimes does not support the statement that we are "past the point" of not using LLMs.

by tovej 7mo ago

ok_dad 7mo ago

It’s juvenile to consider all LLM assisted coding as vibecoding. I’m not going to expand here because this topic is about as much fun to discuss as politics, but coding assistant tools are just tools. If you give a regular person a race car, they will crash it about as fast as their vibecoded app crashes. Give the same race car to a pro age it’s a different story. I still think this was the right decision by the programming mods there. Talking about tools is pretty boring, and you need to train to use something like an LLM assistant. No one who can’t program a language should be using an LLM to learn it unless they know about 2-3 other languages already, IMO.

apsurd 7mo ago
Nah I think it really is more nuanced than that. It is true that a non-technical person's vibe-coded side-hustle is completely different than how a professional developer may ship genAI code, but we're willfully glossing over the real problem that professionals are pushing out TONS of genAI code that's closer to vibes than it is to the pre-AI expectations on pushing to prod.
- ok_dad 7mo ago
  Devs that are pushing crappy code using these tools are 1) devs who would have pushed shut code anyways 2) getting stress from above to push code faster with AI Good devs will learn to use these tools effectively.

apsurd 7mo ago

For CRUD apps though, the intern closing the ticket literally 30 minutes after it's created is really hard to battle against. Especially when those tickets were created by suits. I generally agree that while I think vibe-coding is here to stay, it's different from designing useful products and systems, and I don't know how to convince colleagues that we should uhh be careful about all this code we're pushing. I fear all they see is the guy aging out.

NietTim 7mo ago

> I have seen vibecoded intern prototypes that took half a day to vet and dismiss because they were completely useless. They weren't useless, they proved if the direction that the prototype was exploring was worthwhile. I've personally made many completely shit code prototypes in the years before we had LLM's, of course they weren't magically production ready, that's not the point of a prototype.

tovej 7mo ago
The example I'm thinking of did not. It was just the dumbest way to execute an idea we knew was easy to execute (scanning a parameter space).

altmanaltman 7mo ago

> I have yet to run into any serious project in the wild that is using LLMs for development. How about Claude Code? 100% of it was vibe-coded according to its creator.[1] Google and Microsoft also claim a lot of their internal code is AI-generated now. [2] [3] Naturally, none of the big tech companies will just release a pure vibe-coded project due to structural reasons, but you also _seriously_ can't claim that serious projects don't use LLMs as well these days. Maybe in your limited experience, it isn't true, but that doesn't generalize to what's actually happening. 1. https://www.reddit.com/r/Anthropic/comments/1pzi9hm/claude_code_creator_confirms_that_100_of_his/ https://www.reddit.com/r/Anthropic/comments/1pzi9hm/claude_c... 2. https://fortune.com/2024/10/30/googles-code-ai-sundar-pichai/ https://fortune.com/2024/10/30/googles-code-ai-sundar-pichai... 3. https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-as-30percent-of-microsoft-code-is-written-by-ai.html https://www.cnbc.com/2025/04/29/satya-nadella-says-as-much-a...

59nadir 7mo ago
> How about Claude Code? 100% of it was vibe-coded according to its creator. Agents are trivial to make. I don't know whether that means they're not "serious", but it's exactly the type of thing you can make yourself in a very short time, and exactly the type of thing even LLMs can't fuck up too bad. With regards to the overall point, I think the existence of projects using LLMs to do development doesn't really lend credence to the idea that they're somehow preferable or desirable. People tend to use hyped things, imagine they're useful even when presented with evidence to the contrary, and generally be very resistant to sobering realities. It took years before people stopped running hadoop clusters to do things that a single linux box could get done 10x faster with some basic pipes. I'm sure there are still people who have "serverless backends" that work terribly in every regard in comparison to literally just a linux VM somewhere. People in software development tend to find these types of things every once in a while and adopt them wholesale. This cycle is helped by the fact that the field has been growing constantly and a lot of the adoption comes from kids who don't know any better. Every piece of shit technology that comes and goes has meat for the grinder coming straight out of university. Would I put LLMs in the same category as these previous (nearly) useless things? Probably not... But you should never trust peoples' perception of usefulness when it comes to almost anything in software development.

dwb 7mo ago

Ok well I have plenty of serious, production-level professional experience that says otherwise. Not “vibe coding” - we certainly review the code. It’s a tool that has downsides and failure modes, of course, but it’s at the point where it’s definitely speeding us up and we are using it a lot. Trust me, I’d prefer a world, on balance, where this wasn’t true – I don’t like many of the aspects and uses of the technology – but its utility in programming is undeniable now and the capitalists aren’t taking “no” for an answer.

tovej 7mo ago
I'm curious, where are you using the tools? What programming language? what domain? Are you willing to share the projects you're working on?
- dwb 7mo ago
  TypeScript and Go on a 1.5 million SLOC production codebase; a complex SaaS tool for financial planning and analysis. Quite far from being “just CRUD”. Before Anthropic Opus 4.5 I was trying out Claude Code and wasn’t all that impressed, but since then it’s definitely helped. The project I wrapped up before Christmas would have gone into the new year without it. You’ve still got to keep a close eye on it; whenever I’ve got lazy with review, trusting it too much, I’ve always regretted it. It’s never one-shotted anything, even with plan mode and all that. I’m a natural skeptic on this stuff and was very actually skeptical for most of last year. But I’m very confident there’s a large net productivity gain now.
  - tovej 7mo ago
    So is this a web app with a go backend? I don't know where you got the "just CRUD" quote from. I never mentioned CRUD. But this sure sounds like it would be CRUD with some additional models in the backend. What makes it not just CRUD? Is it using some complex model for forecasting?
    - dwb 7mo ago
      I wasn't quoting you, they are scare quotes. It is a web app with a TypeScript and Go backend, but to call it a CRUD app would be misleading (despite the fact that yes once you boil it all down, everything is a CRUD operation) because it's a complex and flexible web app. It's the kind of thing that a bunch of people would prefer to be a native app; it's spreadsheet-like.

xboxnolifes 7mo ago

What do you consider "serious", as that seems to be the main differentiator here. I know plenty of serious (multiple years of development and users, and began prior to LLMs) projects that have devs using LLMs for development.

tovej 7mo ago
That makes me curious, would you like to list them?
- xboxnolifes 7mo ago
  They aren't anything particularly special, most projects I interact with have at least 1 dev who uses LLMs in some capacity. The two I was thinking of when I posted are the Minecraft modpack GregTech: New Horizons, and an Old School Runescape plugin. https://github.com/gtnewhorizons https://github.com/gtnewhorizons https://github.com/osrs-reldo/tasks-tracker-plugin https://github.com/osrs-reldo/tasks-tracker-plugin
  - tovej 7mo ago
    I see. I would not consider these to be serious projects. I'm sure they give people joy, and contain some decent problem solving, but they are both small extensions to existing software, and are not really solving interesting new problems or creating their own systems.
    - xboxnolifes 7mo ago
      The first one is hardly a small extension. It's millions of lines of code, which is multiple times more than the version of Minecraft it extends.
      - tovej 7mo ago
        The size of the project is not what matters. It's the systems it includes and how complex they are. Actually, that raises a question: are the more complex systems in this collection of projects vibecoded or programmed by a human? Which parts are you aware of that invlude LLM output?

tarsinge 7mo ago

How can you still can't distinguish between using LLMs as tools and a non technical person vibe coding? I have yet to run into any serious software engineer that had to dive into a legacy codebase or an unknown tech stack and found no value in e.g. Claude Code for general understanding and refactoring. Not even talking about coding, just the capacity of generating custom contextualised documentation and examples tailored to your constraints and skills on the fly is ridiculously helpful.

tovej 7mo ago
The tool being useful sometimes does not support the statement that we are "past the point" of not using LLMs.