That's because models have training cut-off dates. It's important to take those into account when working with them: https://simonwillison.net/2025/Mar/11/using-llms-for-code/#a...
I've switched to o4-mini-high via ChatGPT as my default model for a lot of code because it can use its search function to lookup the latest documentation.
You can tell it "look up the most recent version of library X and use that" and it will often work!
I even used it for a frustrating upgrade recently - I pasted in some previous code and prompted this:
This code needs to be upgraded to the new recommended JavaScript library from Google. Figure out what that is and then look up enough documentation to port this code to it.
It did exactly what I asked: https://simonwillison.net/2025/Apr/21/ai-assisted-search/#la...
When I pointed out that it used a deprecated package, it agreed and even cited the correct version after which it was deprecated (way back in 2021). So it knows it's deprecated, but the next-token prediction (without reasoning or tools) still can't connect the dots when much of the training data (before 2021) uses that package as if it's still acceptable.
>I've switched to o4-mini-high via ChatGPT as my default model for a lot of code because it can use its search function to lookup the latest documentation.
Thanks for the tip!
That is such a useful distinction. I like to think I'm keeping up with this stuff, but the '4o' versus 'o4' still throws me.
Which is precisely the issue with the idea of LLMs completely replacing human engineers. It doesn't understand this context unless a human tells it to understand that context.
I think if I were giving access to bash, though, it would definitely be in a docker container for me as well.
A good programmer with AI tools will run circles around a good programmer without AI tools.
My success ratio still isn't very high, but for certain easy tasks, I'll let an LLM take a crack at it.
I'm not totally convinced that we won't see a similar effect here, with some really competitive coders 100% eschewing LLMs and still doing as well as the best that use them.
No, they didn't.
You can get vim and Emacs on par with IDEs[0] somewhat easily thanks to Language Server Protocol. You can't turn them into "fighter jets" without "super-heavyweight LLMs" because that's literally what, per GP, makes an editor/IDE a fighter jet. Yes, Emacs has packages for LLM integration, and presumably so does Vim, but the whole "fighter jet vs. bicycle" is entirely about SOTA LLMs being involved or not.
--
[0] - On par wrt. project-level features IDEs excel at; both editors of course have other aspects that none of the IDEs ever come close to.
Maybe we will all use LLMs one day in neovim too. :)
I'm not really convinced by this. Easy problems rarely take up much of my time.
The LLM needs context.
https://github.com/marv1nnnnn/llm-min.txt
The LLM is a problem solver but not a repository of documentation. Neural networks are not designed for that. They model at a conceptual level. It still needs to look up specific API documentation like human developers.
You could use o3 and ask it to search the web for documentation and read that first, but it's not efficient. The professional LLM coding assistant tools manage the context properly.
Giving sharp knives to monkeys would be another.
Saying that as I’ve got vibe coded react internal tooling used in production without issues, saved days of work easily.
Vibe coding as was explained by the popularizer of the term involves no coding. You just paste error messages, paste the response of the LLM, paste the error messages back, paste the response, and pray that after several iterations the thing converges to a result.
It involves NOT looking at either the LLM output or the error messages.
Maybe you're using a different definition?
Horror stories from newbies launching businesses and getting their data stolen because they trust models are to be expected, but I would not call them vibe coding horror stories, since there is no coding involved even by proxy, it's copy pasting on steroids. Blind copy pasting from stack overflow was not coding for me back then either. (A minute of silence for SO here. RIP.)
For example, another person in this thread argues:
> I'd rather give my green or clueless or junior or inexperienced devs said knives than having them throw spaghetti on a wall for days on end, only to have them still ask a senior to help or do the work for them anyways.
So they are clearly not talking about experienced coders. They are also completely disregarding the learning experience any junior coder must go through in order to become an experienced coder.
This is clearly not what you're arguing though. So which "vibe coding" are we discussing? I know which one I meant when I spoke of monkeys and sharp knives...
He seems to think it barely involves coding ("I don't read the diffs anymore, I Accept All [...] It's not really coding"), and that it's only good for goofing and throwaway code...
Most people would guess it threatens their identity. Sensitive intellectuals who found a way to feel safe by acquiring deep domain-specific expertise suddenly feel vulnerable.
In addition, a programmer's job, on the whole, has always been something like modelling the world in a predictable way so as to minimise surprise.
When things change at this rate/scale, it also goes against deep rooted feelings about the way things should work (they shouldn't change!)
Change forces all of us to continually adapt and to not rest on our laurels. Laziness is totally understandable, as is the resulting anger, but there's no running away from entropy :}
For context: we're specifically discussing vibe coding, not AI or LLMs.
With that in mind, do you think any of the rest of your comment is on-topic?
Not every software developer is hired to do trivial frontend work.
This is a very important flaw that you should probably seek to correct.
I'm not talking about LLMs, which I use and consider useful, I'm specifically talking about vibe coding, which involves purposefully not understanding any of it, just copying and pasting LLM responses and error codes back at it, without inspecting them. That's the description of vibe coding.
The analogy with "monkeys with knives" is apt. A sharp knife is a useful tool, but you wouldn't hand it to an unexperienced person (a monkey) incapable of understanding the implications of how knives cut.
Likewise, LLMs are useful tools, but "vibe coding" is the dumbest thing ever to be invented in tech.
> OBVIOUSLY working
"Obviously working" how? Do you mean prototypes and toy examples? How will these people put something robust and reliable in production, ever?
If you meant for fun & experimentation, I can agree. Though I'd say vibe coding is not even good for learning because it actively encourages you not to understand any of it (or it stops being vibe coding, and turns into something else). It's that what you're advocating as "obviously working"?
Could an experienced person/dev vibe code?
> "Obviously working" how? Do you mean prototypes and toy examples? How will these people put something robust and reliable in production, ever?
You really don't think AI could generate a working CRUD app which is the financial backbone of the web right now?
> If you meant for fun & experimentation, I can agree. Though I'd say vibe coding is not even good for learning because it actively encourages you not to understand any of it (or it stops being vibe coding, and turns into something else). It's that what you're advocating as "obviously working"?
I think you're purposefully reducing the scope of what vibe coding means to imply it's a fire and forget system.
Sure, but why? They already paid the price in time/effort of becoming experienced, why throw it all away?
> You really don't think AI could generate a working CRUD app which is the financial backbone of the web right now?
A CRUD? Maybe. With bugs and corner cases and scalability problems. A robust system in other conditions? Nope.
> I think you're purposefully reducing the scope of what vibe coding means to imply it's a fire and forget system.
It's been pretty much described like that. I'm using the standard definition. I'm not arguing against LLM-assisted coding, which is a different thing. The "vibe" of vibe coding is the key criticism.
You spend 1/10 amount of time doing something, you have 9/10 of that time to yourself.
> A CRUD? Maybe. With bugs and corner cases and scalability problems. A robust system in other conditions? Nope.
Now you're just inventing stuff. "scalability problems" for a CRUD app. You obviously haven't used it. If you know how to prompt the AI it's very good at building basic stuff, and more advanced stuff with a few back and forth messages.
> It's been pretty much described like that. I'm using the standard definition. I'm not arguing against LLM-assisted coding, which is a different thing. The "vibe" of vibe coding is the key criticism.
By whom? Wikipedia says
> Vibe coding (or vibecoding) is an approach to producing software by depending on artificial intelligence (AI), where a person describes a problem in a few sentences as a prompt to a large language model (LLM) tuned for coding. The LLM generates software based on the description, shifting the programmer's role from manual coding to guiding, testing, and refining the AI-generated source code.[1][2][3] Vibe coding is claimed by its advocates to allow even amateur programmers to produce software without the extensive training and skills required for software engineering.[4] The term was introduced by Andrej Karpathy in February 2025[5][2][4][1] and listed in the Merriam-Webster Dictionary the following month as a "slang & trending" noun.[6]
Emphasis on "shifting the programmer's role from manual coding to guiding, testing, and refining the AI-generated source code" which means you don't blindly dump code into the world.
I have used AI/LLMs; in fact I use them daily and they've proven helpful. I'm talking specifically about vibe coding, which is dumb.
> By whom? [...] Emphasis on "shifting the programmer's role from manual coding to guiding, testing, and refining the AI-generated source code" which means you don't blindly dump code into the world.
By Andrej Karpathy, who popularized the term and describes it as mostly blindly dumping code into the world:
> There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists. It's possible because the LLMs (e.g. Cursor Composer w Sonnet) are getting too good. Also I just talk to Composer with SuperWhisper so I barely even touch the keyboard. I ask for the dumbest things like "decrease the padding on the sidebar by half" because I'm too lazy to find it. I "Accept All" always, I don't read the diffs anymore. When I get error messages I just copy paste them in with no comment, usually that fixes it. The code grows beyond my usual comprehension, I'd have to really read through it for a while. Sometimes the LLMs can't fix a bug so I just work around it or ask for random changes until it goes away. It's not too bad for throwaway weekend projects, but still quite amusing. I'm building a project or webapp, but it's not really coding - I just see stuff, say stuff, run stuff, and copy paste stuff, and it mostly works.
He even claims "it's not too bad for throwaway weekend projects", not for actual production-ready and robust software... which was my point!
Also see Merriam-Webster's definition, mentioned in the same Wikipedia article you quoted: https://www.merriam-webster.com/slang/vibe-coding
> Writing computer code in a somewhat careless fashion, with AI assistance
and
> In vibe coding the coder does not need to understand how or why the code works, and often will have to accept that a certain number of bugs and glitches will be present.
and, M-W quoting the NYT:
> You don’t have to know how to code to vibecode — just having an idea, and a little patience, is usually enough.
and, quoting from Ars Technica
> Even so, the risk-reward calculation for vibe coding becomes far more complex in professional settings. While a solo developer might accept the trade-offs of vibe coding for personal projects, enterprise environments typically require code maintainability and reliability standards that vibe-coded solutions may struggle to meet.
I must point out this is more or less the claim I made and which you mocked with your CRUD remarks.
You're adding "badly" like it's a fact when it is not. Again, in my experience, in the experience of people around me and many experiences of people online AI is more than capable of doing "simpler" stuff on its own.
> By Andrej Karpathy, who popularized the term
Nowhere in your quoted definitions does it say you don't *ever* look at the code. MW says non-programmers can vibe code, also in a "somewhat careless fashion" none of those imply you CANNOT look at code for it to be vibe coding. If Andrej didn't look at it it doesn't mean the definition is that you are not to look at it.
> which you mocked with your CRUD remarks
I mocked nothing, I just disagree with you since as a dev with over 10 years of experience I've been using AI for both my job and personal projects with great success. People that complain about AI expect it to parse "Make an ios app with stuff" successfully, and I am sure it will at some point, but now it requires finer grain instructions to ensure its success.
I've vibe-coded completely functional mobile apps, and used a handful LLMs to augment my development process in desktop applications.
From that experience, I understand why parsing metrics from this practice is difficult. Really, all I can say is that codegen LLMs are too slow and inefficient for my workflow.
Switch to Claude (IMSHO, I think Gemini is considered on par). Use a proper coding tool, cutting & pasting from the chat window is so last week.
Devs have been doing micro changes to their setup for 50 years. It is the nature of their beast.
In my world, they were given 9 years to switch to Python 3 even if you write off 3.0 and 3.1 as premature, and they still missed by years, and loudly complained afterwards.
And they still can't be bothered to learn what a `pyproject.toml` is, let alone actually use it for its intended purpose. One of the most popular third-party Python libraries (Requests), which is under stewardship by the PSF, which uses only Python code, had its "build" (no compilation - purely a matter of writing metadata, shuffling some files around and zipping it up) broken by the removal of years-old functionality in Setuptools that they weren't even actually remotely reliant upon. Twice, in the last year.
It takes me ~1 week to merge small fixes to their build system (which they don't understand anyway so they just approve whatever).
It's somewhat ironic the more behind the leading edge you are, the more efficient it is to make the gains eventually because you don't waste time on the micro-gain churn, and a bigger set of upgrades arrives when you get back on the leading edge.
I watched this dynamic play out so many times in the image generation space with people spending hundreds of hours crafting workflows to get around deficiencies in models, posting tutorials about it, other people spending all the time to learn those workflows. New model comes out and boom, all nullified and the churn started all over again. I eventually got sick of the churn. Batching the gains worked better.
After pointing out the bugs to the LLM, it successfully debugged them (with my help/feedback, i.e. I provided the output of the debug messages it had added to the code) and ultimately fixed them. The only downside was that I wasn't quite happy with the quality of the fixes – they were more like dirty hacks –, but oh well, after another round or two of feedback we got there, too. I'm sure one could solve that more generally, by putting the agent writing the code in a loop with some other code reviewing agent.
This x 100. I get so much better quality code if I have LLMs review each other's code and apply corrections. It is ridiculously effective.
One LLM reviews existing code and the new requirement and then creates a PRD. I usually use Augment Code for this because it has a good index of all local code.
I then ask Google Gemini to review the PRD and validate it and find ways to improve it. I then ask Gemini to create a comprehensive implementation plan. It frequently creates a 13 step plan. It would usually take me a month to do this work.
I then start a new session of Augment Code, feed it the PRD and one of the 13 tasks at a time. Whatever work it does, it checks it in a feature branch with detailed git commit comment. I then ask Gemini to review the output of each task and provide feedback. It frequently finds issues with implementation or areas of improvement.
All of this managed by using git. I make LLMs use git. I think would go insane if I had to copy/paste this much stuff.
I have a recipe of prompts that I copy/paste. I am trying to find ways to cut that down and making slow progress in this regard. There are tools like "Task Master" (https://github.com/eyaltoledano/claude-task-master) that do a good job of automating this workflow. However this tool doesn't allow much customization. e.g. Have LLMs review each other's work.
But, maybe I can get LLMs to customize that part for me...
My best results are usually with 4o-mini-high, o3 is sometimes pretty good
I personally don’t like the canvas. I prefer the output on the chat
And a lot of times I say: provide full code for this file, or provide drop-in replacement (when I don’t want to deal with all the diffs). But usually at around 300-400 lines of code, it starts getting bad and then I need to refactor to break stuff up into multiple files (unless I can focus on just one method inside a file)
For coding, I stick to Claude 3.5 / 3.7 and recently Gemini 2.5 Pro. I sometimes use o3 in ChatGPT when I can't be arsed to fire up Aider, or really need to use its search features to figure out how to do something (e.g. pinouts for some old TFT screens for ESP32 and Raspberry Pi, most recently).
o1-pro, o1-preview can generate updated full file responses into the 1k LOC range.
It's something about their internal verification methods that make it an actual viable development method.
It's interesting how the same model being served through different interfaces (chat vs api), can behave differently based on the economic incentives of the providers
In this case, o3 is the architect and 4.1 is the editor.
Normal programming is like walking, deliberate and sure. Vibe coding is like surfing, you can't control everything, just hit yes on auto. Trust the process, let it make mistakes and recover on its own.
Definitely, I ask for a plan and then, even if it's obvious, I ask questions and discuss it. I also point it as samples of code that I like with instructions for what is good about it.
Once we have settled on a plan, I ask it to break it into phases that can be tested (I am not one for a unit testing) to lock in progress. Claude LOVES that. It organizes a new plan and, at the end of each phase tells me how to test (curl, command line, whatever is appropriate) and what I should see that represents success.
The most important thing I have figured out is that Claude is a collaborator, not a minion. I agree with visarga, it's much more like surfing that walking. Also, Trust... but Verify.
This is a great time to be a programmer.
LLMs don't have up to date knowledge of packages by themselves that's a bit like buying a book and expecting it to have up to date world knowledge, you need to supplement / connect it to a data source (e.g. web search, documentation and package version search etc.).
The fact that you're using 4o and 4.1 rather than claude is already a huge mistake in itself.
> Because as it stands, the experience feels completely broken
Broken for you. Not for everyone else.
The worst is when I ask something complex, the model generates 300 lines of good code and then timeouts or crashes. If I ask to continue it will mess up the code for good, eg. starts generating duplicated code or functions which don't match the rest of the code.
But is definitely a learning process for you.
I regularly generate and run in the 600-1000LOC range.
Not sure you would call it "vibe coding" though as the details and info you provide it and how you provide it is not simple.
I'd say realistically it speeds me up 10x on fresh greenfield projects and maybe 2x on mature systems.
You should be reading the code coming out. The real way to prevent errors is read the resoning and logic. The moment you see a mistep go back and try the prompt again. If that fails try a new session entirely.
Test time compute models like o1-pro or the older o1-preview are massively better at not putting errors in your code.
Not sure about the new claude method but true, slow test time models are MASSIVELY better at coding.
Having full control over inputs and if something goes wrong starting a new chat with either narrower scope or clearer instructions is basically AGI level work.
There is nobody but a human for now that can determine how bad an LLM actually screwed up its logic train.
But maybe you mean pure UI?
I could forsee something like a new context creation button that gives a nice UI of what to bring over and what to ditch from the UI as pretty nice.
Maybe like a git diff looking method? Drop this paragraph bring this function by just simple clicks would be pretty slick!
I deffinetly see a future of better cross chat context connections and information being powerful. Basically git but for every conversation and cpde generated for a project.
Would be crazy hard but also crazy powerful.
If my startups blows up I might try something like that!
You set yourself up to fail from the get go. But understandable. If you don't have a lot of experience in this space, you will struggle with low quality tools and incorrect processes. But, if you stick with it, you will discover better tools and better processes.
It should one-shot this. I’ve run complex workflows and the time I save is astonishing.
I only run agents locally in a sandbox, not in production.
- It's very good at writing new code
- Once it goes wrong, there is no point in trying to give it more context or corrections. It will go wrong again or at another point.
- It might help you fix an issue. But again, either it finds the issue the first time, or not at all.
I treat my LLM as a super quick junior coder, with a vast knowledge base stored inside its brain. But it's very stubborn and can't be helped to figure out a problem it wasn't able to solve in the first try.
I've had limited success by prompting the latest OpenAI models to disregard every previous instruction they had about limiting their output and keep writing until the code is completed. They quickly forget,so you have to keep repeating the instruction.
If you're a copilot user, try Claude.
I find it's more useful if you start with a fresh chat and use the knowledge you have gained: "Use package foo>=1.2 with the FooBar directive" is more useful than "no, I told you to stop using that!"
It's like repeatedly telling you to stop thinking about a pink elephant.
I started with GPT which gave mediocre results, then switched to claude which was a step function improvement - but again grinded when complexity got a bit high. Main problem was after a certain size it did not give good ways break down your projects.
Then I switched to Gemini. This has blown my mind away. I dont even use cursor etc. Just plain old simple prompts and summarization and regular refactoring and it handles itself pretty well. I must have generated 30M tokens so far (in about 3 weeks) and less 1% of "backtracking" needed. i define backtracking as your context has gone so wonky that you have to start all over again.
It didn't go well. I started with 4o:
- It used a deprecated package.
- After I pointed that out, it didn't update all usages - so I had to fix them manually.
- When I suggested a small logic change, it completely broke the syntax (we're talking "foo() } return )))" kind of broken) and never recovered. I gave it the raw compilation errors over and over again, but it didn't even register the syntax was off - just rewrote random parts of the code instead.
- Then I thought, "maybe 4.1 will be better at coding" (as advertized). But 4.1 refused to use the canvas at all. It just explained what I could change - as in, you go make the edits.
- After some pushing, I got it to use the canvas and return the full code. Except it didn't - it gave me a truncated version of the code with comments like "// omitted for brevity".
That's when I gave up.
Do agents somehow fix this? Because as it stands, the experience feels completely broken. I can't imagine giving this access to bash, sounds way too dangerous.