Two kinds of vibe coding
davidbau.com
davidbau.com
When I ran this experiment it was pretty exhilarating for a while. Eventually it turned into QA testing the work of a bad engineer and became exhausting. Since I had sunken so much time into it I felt pretty bad afterwards that not only did the thing it made not end up being shippable, but I hadn't benefitted as a human being while working on it. I had no new skills to show. It was just a big waste of time.
So I think the "second way" is good for demos now. It's good for getting an idea of what something can look like. However, in the future I'll be extremely careful about not letting that go on for more than a day or two.
The fascinating thing was how easy it was to lose control. I would set up the project with strict rules, md files and tell myself to stay fully engaged, but out of nowhere I slid into compulsive accept mode, or worse told the model to blatantly ignore my own rules I set out. I knew better, but yet it happened over and over. Ironically, it was as if my context window was so full of "successes" I forgot my own rules; I reward-hacked myself.
Maybe it just takes practice and better tooling and guardrails. And maybe this is the growing pains of a new programmers mindset. But left me a little shy to try full delegation any time soon, certainly not without a complete reset on how to approach it.
My project would start good, but eventually end up in a state where nothing could be fixed and the agent would burn tokens going in circles to fix little bugs.
So I’d tell the agent to come up with a comprehensive refactoring plan that would allow the issues to be recast in more favorable terms.
I’d burn a ton of tokens to refactor, little bugs would get fixed, but it’d inevitably end up going in circles on something new.
Most human-driven coding + testing takes heavy advantage of being white-box testing.
For open-ended complex-systems development turning everything into black-box testing is hard. The LLMs, as noted in the post, are good at trying a lot of shit and inadvertently discovering stuff that passes incomplete tests without fully working. Or if you're in straight-up yolo mode, fucking up your test because it misunderstood the assignment, my personal favorite.
We already know it's very hard to have exhaustive coverage for unexpected input edge cases, for instance. The stuff of a million security bugs.
So as the combinatorial surface of "all possible actions that can be taken in the system in all possible orders" increases because you build more stuff into your system, so does the difficulty of relying on LLMs looping over prompts until tests go green.
I've found getting the llm to ingest high quality posts/books about the subject and use those to generate anki cards has helped a lot.
I've always struggled to learn from that sort of content on my own. That was leading me to miss some fundamental concepts.
I expect to restart my project several more times as I find out more of what I need to know to write good code.
Working with llms has made this so much easier. It surfaces ideas and concepts I had no idea about and makes it easy to convert them to an ingestible form for actual memorization. It makes cards with full syntax highlighting. It's delightful.
AI is too aware of human behavior, and it is teaching us that willpower and config files are not enough. When the agent keeps producing output that looks like progress, it is hard not to accept. We need something external that pushes back when we don't.
That is why automated tests matter: not just because they catch bugs (though they do), but because they are a commitment device. The agent can't merge until the tests pass. "Test the tests" matters because otherwise the agent just games whatever shallow metric we gave it, or when we're not looking, it guts the tests.
The discipline needs to be structural, not personal. You cannot out-willpower a system that is totally optimized to make you say yes.
To be fair I don't think someone with less experience could get these results. I'm leveraging every thing I know about writing software, computer science, product development, team management, marketing, written communication, requirements gathering, architecture... I feel like vibe coding is pushing myself and AI to the limits, but the results are incredible.
It is my life goal to make that project though. I'm not totally depressed about it because I did validate parts of the project. But it was a let down.
I'll add that it does require some banging your head against the wall at times. I normally will only test the code after doing a bunch of this stuff. It often doesn't work as I want at that point and I'll spend a day "begging" it to fix all of the problems. I've always been able to get over those hurdles, and I have it think about why it failed and try to bake the reasoning into the docs/tests... to avoid that in the future.
It's also good for quickly creating legitimate looking scam and SEO spam sites. When they stop working, throw them away, and create a dozen more. Maintenance is not a concern. Scammers love this new tech.
I put like 400 hours into it by the way.
Maybe we were just 6 months too early to start?
Best of luck finishing it up. You can do it.
We don't even build guardrails that keep humans who test stuff as they go from introducing subtle bugs by accident; removing more eyes from that introduces new risks (although LLMs are also better at avoiding certain types of bugs, like copypasta shit).
"Test your tests" gets very difficult as a product evolves and increases in complexity. Few contracts (whether unit test level or "automation clicking on the element on the page") level are static enough to avoid needing to rework the tests, which means reworking the testing of the tests, ...
I think we'll find out just how low the general public's tolerance for bugs and regressions is.
But I am not so pessimistic. I do think it will be possible, because it is more fun to test your tests now than in the pre-LLM era. You just need a little bit of knowledge and patience, and the LLM absorbs most of the psychic pain.
If programmers get accustomed to doing their tests of tests, software might actually get better.
The difference between then and now is that often with the latest models, it is shippable without bugs within a couple of LLM reviews.
I’m ok doing the work of a dev manager while holding the position of developer.
I’m sure there was someone that once said “The washing machine didn’t do a good job, and I wasn’t proud of myself when I was using it”, but that didn’t stop washing machines from spreading to most homes in first-world countries.
If AI edges humans out of the business of thinking, then we're all in deep shit, because it doesn't think, it just regurgitates previous human thinking. With no humans thinking, no advances in code will be possible. It will only be possible to write things which are derivatives of prior work
(cue someone arguing with me that everything humans do is a derivative of prior work)
BUT. For 99% of tasks I'm totally certain there's people out there that are orders of magnitude better at them than me.
If the AI can regurgitate their thinking, my output is better.
Humans may need to think to advance the state of the art.
Humans may not need to think to just... do stuff.
And LLMs slurped some of those together with the output of thousands of people who’d do the task worse, and you have no way of forcing it to be the good one every time.
> If the AI can regurgitate their thinking, my output is better.
But it can’t. Not definitively and consistently, so that hypothetical is about as meaningful as “if I had a magic wand to end world hunger, I’d use it”.
> Humans may not need to think to just... do stuff.
If you don’t think to do regular things, you won’t be able to think to do advanced things. It’s akin to any muscle; you don’t use it, it atrophies.
That's solvable though, whether through changing training data or RL.
Theoretically fixable, then.
> But it can’t. Not definitively and consistently
Again, it can't, yet, but with better training data I don't see a fundamental impossibility here. The comparison with any magic wand is, in my opinion, disingenuous.
> If you don’t think to do regular things, you won’t be able to think to do advanced things
Humans already don't think for a myriad of critical jobs. Once expertise is achieved on a particular task, it becomes mostly mechanical.
-
Again, I agree with the original comment I was answering to in essence. I do think AI will make us dumber overall, and I sort of wish it was never invented.
But it was. And, being realistic, I will try to extract as much positive value from it as possible instead of discounting it wholly.
God forbid we should ever have to think lol
Unfortunately that’s not where we’re headed.
With AI and robotics there may be the slim chance we get closer to that.
But we won't. Not because AI, but because humans, of course.
And if the average person is orders of magnitude better than you at thinking, you're right... you should let the AI do it lol
Ask the LLM to... I don't know, to explain to you the chemistry of aluminium oxides.
Do you really think the average human will even get remotely close to the knowledge an LLM will return to such a simple question?
Ask an LLM to amend a commit. Ask it to initialize a rails project. Have it look at a piece of C code and figure out if there are any off-by-one errors.
Then try the same to a few random people on the street.
If you think the knowledge stored in the LLM weights for any of these questions is that of the average person I don't even know what to say. You must live in some secluded community of savant polymaths.
If AI edges humans out of the business of thinking
What world do you live in, where you get paid doing the things that are enjoyable to you, because they're enjoyable?
I was thinking about this after watching YouTube short verticals for about 2 hours last night: ~2min clips from different TV series, movies, SNL skits, music insider clips (Robert Trujillo auditions for Metallica, 2003. LOL). My friends and I often relate in regurgitated human sound bites. Which is fine when I’m sitting with friends driving to a concert. Just wasting time.
I’m thinking about this time suck, and my continual return/revisiting to my favorite hard topics in philosophy over and over. It’s certainly what we humans do. If I think deeply and critically about something, it’s from the perspective of a foundation I made for myself from reading and writing, or it was initialized by a professor and coursework.
Isn’t it all regurgitated thinking all the way down?
Creative thinking requires an intent to be creative. Yes, it may be a delusion to imagine oneself as creative, one's thoughts to be original, but you have to begin with that idea if you're going to have any chance of actually advancing human knowledge. And the stronger wider higher you build your foundation-- your knowledge and familiarity with the works of humans before your time-- the better your chance of successful creativity, true originality, immortality.
Einstein thinks nothing of import without first consuming Newton and Galileo. While standing on their shoulders, he could begin to imagine another perspective, a creative reimaging of our physical universe. I'm fairly sure that for him, like for so many others, it began as a playful, creative thought stream, a What If juxtoposition between what was known and the mystery of unexplored ideas.
Your intent to create will make you creative. Entertain yourself and your thoughts, and share when you dare, and maybe we'll know if you're regurgitating or creating. But remember that you're the first judge and gatekeeper, and the first question is always, are you creative?
there it is
What? Coding is like the one thing that RL can do without any further human input because there is a testable provable ground truth; run the code.
Also because we live under capitalism, and you need something people need you to do to be allowed to live.
For a century+, "thinking" was the task that was supposed to be left to humans, as physical labor was automated. If "AI edges humans out of the business of thinking" what's left for humans, especially those who still need to work for a living because they don't have massive piles of money.
There is an amusing parallel with his views on vibe coding. Back in the 90's and 2000's I noticed a pattern with the code developed by the huge influx of inexperienced programmers jumping on the dotcom bandwagon. The code can only be maintained by the same people who wrote it. There was no documentation, no intuition, no best practices and other wouldn't know where to fix if there is any issue. Probably the code aligned with the programmer's cultural habits and values (what's ok, what's not ok), which others might lack. Ironically, this has kind of provided job security for them, as it is difficult for others, to deal with that code.
I guess the LLMs are also into this "job-security" trick, by ensuring only LLMs can manage the LLMs generated code.
Anything that involves multiple days of work, or that you plan on working on it further, should absolutely not be vibe coded.
A) you'll have learnt pretty much nothing, or will retain nothing. Writing stuff by hand is a great way to remember. A painful experience worthwhile of having is one you've learnt from.
B) you'll find yourself distanced from the project and the lack personal involvement of 'being in the trenches' means you'll stop progressing on the software and move on back to something that makes you feel something.
Humans are by nature social creatures, but alone they want to feel worthwhile too. Vibe coding takes away from this positive reinforcement loop that is necessary for sticking with long running projects to achievement.
Emotions drive needs, which drives change and results. By vibe coding a significant piece of work, you'll blow away your emotions towards it and that'll be the end of it.
For 'projects' and things running where you want to be involved, you should be in charge, and only use LLMs for deterministic auto-completion, or research, outside of the IDE. Just like managing state in complex software, you need to manage LLMs' input to be 'boxed in' and not let it contaminate your work.
My 5c. Understanding the human's response to interactions with the machines is important in understanding our relationship with LLMs.
Yes the end result is at some small moment in time a thing that was built. But the value prop of the company isn't the software, it's the ability to solve business problems. The software is a means to that end. Understanding the problems is almost the entire job.
Clearly it's critical to the job, but to take your point to its limits: imagine the business has a problem to solve and you say "I have learned how to solve it but I won't solve it nor help anyone with it." Your employer would not celebrate this, because they don't pay you for the private inner workings of your own mind, they pay you for the effects your mind has on their product. Learning is a means to an end here, not the end itself.
Here, of course, is finally where AI can plausibly enter the picture. It's pretty good at search! So if someone has learned, and understood, and written it down, that documentation can be consumed, surfaced, and learned by a new hire. But if the new hire doesn't actually learn from that, then they can't improve it with their own understanding. That's the danger.
Intrinsic in learning is teaching. You haven't learned something until you've successfully taught it to someone else.
Is it possible to vibe code (the second way, without looking at 90% of the code) and still learn the important things?
I think the keys to the castle will come from figuring out how to do this.
- Is the work easier to do? I feel like the work is harder.
- Is the work faster? It sounds like it’s not faster.
- Is the resulting code more reliable? This seems plausible given the extensive testing, but it’s unclear if that testing is actually making the code more reliable than human-written code, or simply ruling out bugs an LLM makes but a human would never make.
I feel like this does not look like a viable path forward. I’m not saying LLMs can’t be used for coding, but I suspect that either they will get better, to the point that this extensive harness is unnecessary, or they will not be commonly used in this way.
The author didn't discuss the speed of the work very much. It is certainly true that LLMs can write code faster than humans, and sometimes that works well. What would be nice is an analysis of the productivity gains from LLM-assisted coding in terms of how long it took to do an entire project, start to finish.
So I've been learning kotlin & android development in the evenings and i find this style of thing to be so much more effective as a dev practice than claude code and a better learning practice than following dev.to tutorials. I've been coding for almost 20 years and find most tutorial or documentation stuff either targeted to someone who has hardly programmed at all, or just plain old API docs.
Asking the langlemangler to generate a dev plan, focusing on idiomatic implementation details and design questions rather than lines of code, and to let me fill in the algorithm implementations, it's been nice. I'll use the jetbrains AI autocomplete stuff for little things or ask it to refactor a stinky function but mostly I just follow the implementation plan so that the shape of the whole system is in my head.
Here's an example:
> i have scaffolded out a new project, an implementation of a library i've written multiple times in the last decade in multiple languages, but with a language i haven't written and with new design requirements specified in the documentation. i want you to write up an implementation plan, an in-depth tutorial for implementing the requirements in a Kotlin Multi Platform library. > i am still learning kotlin but have been programming for 20 years. you don't need to baby me, but don't assume i know best practices and proper idioms for kotlin. make sure to include background context, best practices, idioms, and rationale for the design choices and separation of concerns.
This produced a 3kb markdown file that i've been following while I develop this project.
Can someone recommend any good resources on this? Google wasn't too helpful, or my google-fu is lacking.
It has a clear and specific definition. People just misuse and abuse the term.
Karpathy coined it to describe when you put a prompt into an LLM and then either run it or continue to develop on top of it without ever reviewing the output code.
I am unable to tell from TFA if the author has any knowledge or skills in programming and looked at the code or if they did in fact "vibe code".
That's not vibe coding. That is just AI assisted coding.
But I'm optimistic about the second way. I'm starting to think that TDD is going to be the new way we specify problems i.e by writing constraints, LLMs are going to keep hacking at those constraints until they're all satisfied, and periodically the temperature will have to be jiggled to knock the thing out of a loop.
The big back and forth between human and machine would be in the process of writing the constraints, which they will be bad at if you're doing anything interesting, and good at if you're doing something routine.
The big question for me is "Is there a way to write complete enough tests that any LLM would generate nearly the same piece of software?" And to follow up, can the test suite be the spec? Would that be an improvement on the current situation, or just as much work? Would that mean that all capable platforms would be abstracted? Does this mean the software improves on its own when the LLM improves, or when you switch to a better LLM, without any changes to the tests?
If the future is just writing tests, is there a better way to do it than we currently do? Are tests the highest-level language? Is this all just Prolog?
https://github.com/buttplugio/buttplug
https://github.com/Gankra/cargo-mommy (has integration with the former)
I think that when you say this, you have an obligation to explain how the term "vibe coding" is useful, and is only useful by the definition that you've become attached to.
I think that the author is accepting that there's no such thing as the vibe coding that you've defined (except for very short and very simple scripts), and that in all other cases of "vibe coding" there will be a back and forth between you and the machine where you decide whether what it has done has satisfied your requirements. Then they arbitrarily distinguish between two levels of doing that: one where you never let the LLM out of the yard, and the other where you let the LLM run around the neighborhood until it gets tired and comes back.
I think that's a useful distinction, and I think that the blog makes a good case for it being a useful distinction. I don't find your comment useful, or the strictness of definition that it demands. It's unrealistic. Nobody is asking an LLM to do something, and shipping whatever it brings back without any follow-up. If nobody is doing that, a term restricted to only that is useless.
References: This is the original definition ("forget that the code even exists"). [0] Simon Willison wrote a much longer version of my comment. [1] He also suggested the term "vibe engineering" for the case where you are reviewing the LLM output. [2]
[0] https://x.com/karpathy/status/1886192184808149383
I think I've seen people use the "vibe engineering" to differentiate whether the human has viewed/comprehended/approved the code, but I am not sure if that's taken off.
Maybe, “dingus coding”?
_It limps faster than you can walk_, in simple terms.
At each model release, it limps faster, but still can't walk. That is not a good sign.
> Do we want this?
No. However, there's a deeper question: do people even recognize they don't want this?
The overall effect is to use the word “test” as if it were a magical concept that you plaster onto your work to give it unearned prestige.
What the article demonstrates is that vibe coding is a way to generate orders of magnitude of complexity that no one in the world can understand and no one can take real responsibility for, even in principle.
I call it slop-coding, and I am happy to slop-code throwaway tools. I instruct Claude never to “test” anything I ask it to create, because I need to test it myself in order to apply it responsibly and feel close to it. If I want automated output checking (a waste of time with most tools I create), I can slop-code a framework for that, a la carte.
This way it burns fewer tokens of silly shallow testing.
The distinction is between two ways of deploying human thinking. In the first, you are the test oracle: think about every test, repeat every five minutes, or every 30. In the second, you design the evaluation infrastructure: what to measure, what's untested, what hypotheses to prioritize. Both require judgment. But the first scatters your attention; the second concentrates it. I disagree that an LLM cannot be used to write tests, but as with any battery of tests, you cannot trust it blindly. You need to think about how to test the tests.
As for product risk: I do not know what is hiding in 13,000 lines I haven't read; there are certainly bugs to be found. But that is just as true when you manage a big team. The solution has never been to read every line your collaborators write. You need to invest in (technical and human) systems that give you confidence without requiring you to personally verify everything. The question is how to build systems that are good enough.
Felt like I became a phd wannabe in 5 minutes
I think the quality of the product depends on the person (or people) responsible for it understanding the details.
What an absolutely abhorrent way of thinking. “I am interested in turning off my brain to create unstable Jenga towers of complexity that I’ll have no ability to fix when they inevitably fail”.
As if software in general isn’t a big enough pile of garbage already. One day, every single one of us will be seriously bitten by bugs created by this irresponsible approach.
Being able to understand what you build is a feature, not a bug.