All vibe coding tools are selling a get rich quick scheme(varunraghu.com) |
All vibe coding tools are selling a get rich quick scheme(varunraghu.com) |
I regularly use Claude code (and Claude) to write code for
- throwaways data processing or visualization
- connecting to an API I have some documentation for
- demoing some concept
I think it's great for data science and product stuff, and on those merits will be a multi-billion dollar industry. I would never let it near a real codebase.
edit: typo
The greater scam here is that rugged individualism can lead to positive societal outcomes.
Just more free market capitalism, bro! We just need more freedom for the tech bros, the finance bros, bro!
This AI shit is yet another oh look - increase in rugged individual 'productivity' is going to lead to positive societal outcomes, trust me, bro!
It's all bullshit but what else have the western retarded elite got? Or non-western elites for that matter? USA is USSR 2.0 - falling apart from the retarded short-sighted incompetence and selfishness of its sociopathic 'elites'. China will be USSR 3.0 in a generation or two.
It's all so terribly tedious, boring and cruel - the monumental amounts of wasted opportunity and resources because everyone's too busy being 'productive' to ever spend a decade or two to actually think through anything and come up with potential real solutions to real problems. The average 'elite' doesn't even know what the real problems are or how to begin finding out what they might be - they probably think it's climate change or more likely - how to make sure they retain their billions. Oh well, so it goes.
Imagine an LLM service company making contributions to climate science.
It's the scorpion ferrying the frog across the water. No different.
1 it's not the meaning Karpathy used originally, which was for creating software through prompts WITHOUT looking at the source code
2 it's not the meaning people outside programming circles mean, either, which is identical to Karpathy's original definition
the title of the article is talking about 2 and is completely correct.
YOU, on the other hand, are talking about an upgrade to IntelliSense. Use a different term. You're describing regular programming, with a new IDE tool.
If it's not going to replace programmers and allow regular people to create software without looking at the code, it's not vibe coding. Full stop.
If you’re talking about AI-assisted development, then an AI agent being able to do 80% of the work is a fantastic win. The developer can pick up the remaining 20% and come out massively ahead.
If you’re talking about vibe-coding, then an AI agent being able to do 80% of the work is useless. What’s a non-developer going to do with something that still has 20% of the development left to do? It’s not 80% useful, it’s 0% useful.
You make the framework and it colours in the lines. If you don't draw the pattern first you're going to get a kindergarten mess. There are times I'm amazed by its suggestions but it will also cheat and "lie" and if you're not paying attention, you'll miss it.
But these days many tasks for me consist of: I need this, look how I did it here before, and follow this exact pattern through to the end and write tests, too. And the results are usually exactly what I would have done. Because it is an excellent mimic. But if I hadn't put the work into the framework, it would be awful.
I’ve ended up with tests called stuff like “foobar successfully returns impossible value that suggests programmer error” lmao
I don't necessarily think it will stay this way, though. The tools are so new, we shouldn't be so sure that the future versions won't allow non-coders to code successfully.
Reminds me of the joke:
int getRandom() {
return 4; // chosen by a dice, guaranteed to be random
}Even with humans, it’s very common to ask them for something and get something back that’s completely different than what you wanted. Short of reading your mind, AI is going to require a lot of info to get the desired result.
So someone is going to have to gather requirements and break them down into a clear, well thought out set of instructions for the computer. That’s literally what programmers already do. We’re just programming in a different language now.
Care to put some money behind this fantastical claim, in a bet?
But when I try it with a stack I don’t know, I quickly find myself needing to learn the new stack to get anything done.
Extremely experienced engineers will be able to move quickly with these tools, but specialization will still exist, and I’m not clear on how juniors are going to ramp up on all of this.
It’s funny you should mention LISP macros. Five years ago, I’d have told you that the most efficient way to build something is to get an extremely experienced dev with mastery over a flexible language like LISP, and just let them go nuts. I constantly have to hold myself back from metaprogramming tricks because my coworkers won’t be able to maintain it. But on my personal projects, I really get to flex my muscles.
It occurred to me a few months ago that that’s exactly the same spot in my toolbox where Claude Code is. It’s a tool that allows one person with mastery of a tech stack to build things incredibly quickly. Look at Cursor where a tiny team has over $100M in annualized revenue.
Big companies often find themselves spreading ownership across too many hands. You know exactly how to build what you need, but you’re not allowed to touch 80% of the code involved because it belongs to a different team. But coordinating all of that is a massive amount of overhead that will derail most projects. This is why so little happens at big companies.
But these tools enable a smaller number of experts to do a job that used to take many. I expect the speed improvements to be super-linear as a result.
Of course, that brings concerns about what happens to the employment situation for devs if that happens. I hope it means that 10,000 new startups can build more ambitious products and absorb all the people who will probably be laid off by the huge corps in the next few years.
That said I might be sort of a weird case, since I am accustomed to designing and documenting product requirements, the fact that it's just for me means architecture doesn't really matter, and I'm competent enough to help it resolve some design problems like the fact that it was obviously hitting the wrong Amazon API that was pulling a report of every single sale as a line item rather than just a report with total sales numbers, etc.
I've seen less experienced devs crash and burn trying to commit massive amounts of vibe coded slop without consideration to: how much of it was necessary? how does it conform with our code base / style? how much does it take advantage of existing code and patterns? and so on.
I think if you are using one of these tools and experienced, it should be difficult to tell you are using one at all. The code I produce looks like stuff I would write, and I understand all it wrote, I don't want to outsource to producing tons of code that I don't even understand.
This is the problem. If you couldn't have coded it slowly in the old world, you will have problems coding it in AI world.
However if you have a lot of coding experience, you can now compress the time it would have taken you be an enormous amount. My experience is that I can now make extensive changes with very little effort, and very few dead ends. I've been able to take on entire secondary projects where I was just replication existing knowledge with slightly different tools.
Just this week I had a litmus test. I had an existing database that I'm pushing huge amount of data to. I decided to try a different underlying database. This would have taken me a full week of looking at documentation and writing supporting scripts, now I've done it in the spare time I had in two days of my actual work.
And it's not like the AI just did it all unsupervised. It threatened to do down the wrong path a few times, but each time I spotted it and steered it the way I wanted. I also asked it a few questions about curiosities I discovered in the emitted code, and that led to fixes as well.
If I didn't know how to code before, I would still be coding this alternative database.
I truly enjoy programming with AI, it is my favorite hobby by a mile. I'm also happy to say that, for now, I get to capture some of the productivity increase for myself while my employer catches up to how effective it is.
It is sad that there is no space to discuss these new techniques that isn't full of opportunists and clout chasers.
I am not sure what the moral of the story is but it reminds me a bit of that parable about the investor getting his shoes shined and the shoe shine kid giving him investment advice.
Partially I think the idea of an app being so valuable that it makes money without being connected to anything is just not really a thing. Some games (Flappy Bird anyone?) can be like that or some very specific type of lifestyle app, but for the most part you need a real world service connected or it will not be seen as valuable. But perhaps I am wrong.
=> You can use AI to double-check (not single check), avoid overengineering, fill knowledge gaps, scan for inconsistencies and many more things. The main benefit it has for me is not that it codes me stuff real fast. It is that my learning curve improved, drastically, whenever I have a missunderstanding. I dive down and test it. Sometimes AI is wrong and I have to go to docs, sometimes docs are wrong and I have to test and I have to open a issue.
If you combine AI capabilities with debugging skills and testing you have a lot of power as developer nowadays. Its a lot of fun.
To me it seems like most things tend to develop to: It’s more fair. If you are lazy and try to do a shorcut you’ll be punished with wasted time and a lot of frustration. If you try to do thing conscientiousnessly and take ownership for the code you push, publish and run. You’ll learn faster, improve faster, ship faster and have way less headache.
If you want to outsource your skill and thinking - that's never going to go particularly well.
It almost seems like a test for impulse control - those who can restrain themselves from taking shortcuts with it can gain tremendous capabilities.
Low effort in -> low effort out
I can actually code so no point in vibing.
To vibe means to never look at the code, where is the fun in that?
One shotting small utilities that I have no interest in writing (for me, front-end JS, chrome extensions, web scraping selectors) is wonderful. I just vibe coded a "crop a pdf" web UI and it worked with no intervention. It doesn't have to be meaningful, just useful.
Then again, I asked ChatGPT moons ago to generate a Filebeat config file, and its output contained syntax errors...
It takes significant human effort to thoroughly understand, define, and refine specifications to recursively satisfy all important feedback from multiple LLMs. Only then can the coding begin, and the actual coding then takes just a couple of minutes. Following this, the code has to be thoroughly reviewed, perhaps also slightly refactored. Lastly, sometimes some follow-up issues have to be created to improve the feature further before it is shippable.
As with anything else, the quality of the result is proportional to the level of effort you put in.
Coding is at the stage of maturing from commodity to a craft. This generation will pass it on to the next generation, and so on.
As with all crafts, there will be a great demand for quality craftsmanship. And it will command a much higher price than whatever comes out of the factory.
I generally foresee most AI generated things becoming worthless. It’s basic supply and demand. When copycats and crop up overnight in droves, whats the differentiator? The value goes way down.
I believe that winning differentiator will be quality. Uniqueness. A level of craftsmanship that an LLM can’t copy without knowing the secret sauce.
You can. You just can't do it fast.
You need to build up your tech skills first. But AI platforms are not incentivized to help you do that. So you have to be very controlled about the speed at which you code.
I learned that letting the AI drive implementation wasn't working for me. The speed is intoxicating at first, and you start trusting the output blindly. But eventually you have to face the music and fix up the mess.
So now I just use it as a glorified autocorrect and search engine, as well as a rubber duck (Ask mode in Cursor). This speed works for me. You have to find the speed that works for you.
Perhaps that's obvious to others, but it felt worth saying from my perspective. If vibe coding tools are touting themselves as a key to easy wealth, shame on the product and marketing teams creating the messaging.
I'm still vibe coding in terms of syntax and logic, but I do understand my codebase, and have made some excellent automations at work.
the actual AI companies marketing hype seems to claim:
1) you can make fun non-serious toy apps
2) any day now massive productivity increases for software companies. this is different than anyone can launch a product with no software skills.
But to be fair, this technology is still at an early stage, and we don't know it's limits.
It's scary to imagine a future where the development process in companies is fully handled by AI agents, which are the only ones who can read and maintain the code.
- "Vibe coding" is a state of action and being, a mental design paradigm for how you approach a problem.
- Generative/LLM based coding is the technology used to enable vibe coding.
I'm not going to fully dismiss the latter because I can't predict the future, and if we're being frank: no one "cares" about the underlying code when you launch a product. Only that it works. Art/Design is king.
Vibe Coding will inevitably fail once you try to scale to make bigger projects or reason with harder problems. Engineering is about understanding a problem space and figuring out a solution to it with the tools/techniques you acquire and derive; black boxing that aspect of engineering will only take you so far.
I would love to see Lovable.dev's financials, I think they're running the classic low retention, high marketing dollar consumer playbook and are burning cash fast. Base44 seems to do the same thing after getting acquired by Wix, I'm seeing more and more of their ads.
As someone technical, vibe coding makes me feel disconnected from my product and feel like I don't know what's going on. I eventually just need to dive back into the code and find many things I need to change myself.
Was it perhaps author's wishful thinking?
They'll put even less value in the people doing the implementation and double down on them being the person with the vision and that's what makes them so great.
I also just overall haven't seen a huge influx of new and useful apps and websites as you'd expect based on how tech leaders are talking about the virtues of these new tools.
I don't think AI has really found it's niche yet, and I think it's going to be much subtler than most people think, it's going to be the tools that integrate features like translations and summarization and speech to text in ways that are seamless that end up sticking - all of this other noise is just marketing hype.
Most web developers can clone twitter in a day, even before LLMs. I've seen it assigned as homework for bootcamp devs (not knocking bootcamp devs, just pointing out that it isn't a tall ask even for a junior. Most apps aren't hard to code to an MVP.
The hard part about an app based business is the business part. Marketing, billing, and getting actual revenue and customers is much harder than writing code.
They Duk-a-duk! (South Park is awesome isn't it ?)
As is with everything that is bait/clickbait/ragebait.
Eventually all the experienced people will be dead or retired.
I could still see a select few being given a foot in the door and leaping over the "argh why doesn't it compile" stage, past the "argh, I need a whole new architecture" stage to become senior in a few years. Every cohort has a few unicorns.
Alternatively, coding goes the way of the calculator. By that I mean, the people that used to be part of the engineering team, doing arithmetic. You just hire domain experts and give them the AI hoping the domain expertise is all you need to get you through the day.
Finance has historically had a lot of people who thought they could code, now perhaps they will get better. Big risk of it becoming a mess larger than back when they got VBA.
1) that the tools become more Socratic and interactive and educational and walk the engineer through what they're doing
2) juniors pair with a senior who is using the tool and see the process and the decisions being made.
I know the industry wants these things to replace us, but in fact it's more like a power drill than a spinning jenny. It augments and lets the existing craftsmen work better faster, but does not replace / automate really.
Now a junior can ask an AI like 90% of questions that would otherwise occupy a senior:
- Why does this dockerfile copy these files first? - how can I find the entrypoints to this service? - is there anything related to image processing in this entire project?
LLMs let juniors punch above their weight, and it lets seniors go faster. Of course if everybody is twice as efficient, you don’t need as many devs (debatable), but I don’t think junior devs are going anywhere. I hear a lot of CEO hype posts saying junior devs are outdated, but idk it doesn’t make any sense to me!
This has not been my experience and it is very frustrating. I've been programming almost 20 years, I'm pretty good at it
I don't know where the disconnect is, but no matter how I try (and I am trying, I don't want to get left behind) I cannot get remotely good results from LLM coding tools
They always, always, always take longer to build what I want than just doing it myself
My manager uses cursor and Claude for everything. I can always tell. He pretends like he wrote it himself even, but he obviously didn't. I end up rewriting big chunks of it, or finding all sorts of baffling quirks during review.
Idk, I'm just not at all willing to settle for "good enough," which means Claude doesn't really save me any time. If I'm going to do it, I'm going to do it right.
It's a good rubber duck, and it's good for scaffolding new projects (arguably), and it's good at converting code from one language to another or for generating a one-off bash script or something. That's about all i can get out of it.
That’s about enough for me to build out fairly complex products very fast with AI (Claude Code). Claude Code often makes some fundamental mistakes that you wouldn’t be able to catch if you didn’t have any coding experience (like today, it was trying to save large images directly in the database instead of using file storage).
This is the danger - I get why you think it's true, but it's not true. I've mentored many many bootcamp grads. The biggest danger with y'all is you don't know what you don't know.
Even if you are a really good bootcamp grad and you have good taste and ability, you do not have experience. Software experience is measured really in years and decades, not months.
QA for non-AI code is limited at best to begin with, why would AI code be any different?
Source: https://xkcd.com/221/
And I’ve realized that even when I try to stay in control, I often don’t read the output code—I just copy and paste.
Metaphorically, it’s like pilots who know how to use the autopilot, but can’t switch back to manual control. That’s the generation of coders we’re raising.
sadly, "The market can stay irrational longer than you can stay solvent". It will correct itself long term, but the damage over the last few years will linger for years, maybe even decades to come.
> You just hire domain experts and give them the AI hoping the domain expertise is all you need to get you through the day.
what domain experts? Clearly they are trying to replace everyone with AI. That's the unsettling part of the whole story; this phenomenon is happening across many sectors who want to try and skimp out on talent.
An LLM will produce a solution for you, but in most cases you need to know enough about the problem to ask it to apply the right tools. Or even have enough "known unknowns" that you can direct it to search for a better solution.
Like a trivial example: an LLM will happily write a CLI argument parser from scratch. It WILL work. It'll also be complete crap UX wise.
But if you know that a specific library (Cobra, Kong, urfave/cli) fits this specific problem better, you can ask it to use it specifically.
Same with calculators - it'll always tell you the correct answer to whatever calculation you give it. But YOU need to know that this specific problem was solved by a really smart person 100 years ago and there's a formula you can apply to solve it faster - still using the calculator, but now you're using refined algebra instead of brute-forcing a solution
The evidence isn't quite clear yet for software.
Conveniently, when it then changes the original implementation, the tests don’t fail!
I wasn't saying it will certainly reach the point where non-coders can code, I am saying we can't be certain it won't just because it can't yet
And how does a Junior in this kind of work gain the skills that transitions them to senior?
>I hear a lot of CEO hype posts saying junior devs are outdated, but idk it doesn’t make any sense to me!
well yes, they aren't making sense. Hence why I think this is all a bubble. Irrational markets and all that.
I don't know when it will burst, but even tech companis can't reject reality forever.
Um - by asking AI 90% of the time, and the senior 10% of the time. If anything, the senior can now mentor 10x the number of juniors.
"Good enough" often has AWS keys embedded into the code itself, stupid things like that
I hold myself to a higher standard