Andrew Ng Says Vibe Coding is a Bad Name For a Very Real and Exhausting Job (businessinsider.com) 78
An anonymous reader shares a report: Vibe coding might sound chill, but Andrew Ng thinks the name is unfortunate. The Stanford professor and former Google Brain scientist said the term misleads people into imagining engineers just "go with the vibes" when using AI tools to write code. "It's unfortunate that that's called vibe coding," Ng said at a firechat chat in May at conference LangChain Interrupt. "It's misleading a lot of people into thinking, just go with the vibes, you know -- accept this, reject that."
In reality, coding with AI is "a deeply intellectual exercise," he said. "When I'm coding for a day with AI coding assistance, I'm frankly exhausted by the end of the day." Despite his gripe with the name, Ng is bullish on AI-assisted coding. He said it's "fantastic" that developers can now write software faster with these tools, sometimes while "barely looking at the code."
In reality, coding with AI is "a deeply intellectual exercise," he said. "When I'm coding for a day with AI coding assistance, I'm frankly exhausted by the end of the day." Despite his gripe with the name, Ng is bullish on AI-assisted coding. He said it's "fantastic" that developers can now write software faster with these tools, sometimes while "barely looking at the code."
Re:This is him reassuring you (Score:5, Insightful)
All I know that the future will have a lot of applications that suddenly explode for no reason, because the bot vomit didn't consider obvious corner cases. But the pseudowhatevermanager that instructed the bot to deliver an application at a 100th of the price got the big bonus, so that's what matters.
I heard the same thing when the Indians took over (Score:2)
Good enough is always good enough. And you will suck it down because you gave up all the power you had when you let them destroy unions and trick you into electing politicians that wouldn't enforce antitrust law. Maybe you personally didn't but certainly your relatives did and some of your friends.
And honestly this website is so full of ol
Re: (Score:2)
I would guess the majority of people voted for this and didn't realize they were doing it.
There were no electable options which didn't support the corporatocracy. We all voted for it, and though some indeed didn't realize it, many did. You can see some of them gloating over their own impending demise here on a regular basis.
Re: (Score:2)
I would guess the majority of people voted for this and didn't realize they were doing it.
There were no electable options which didn't support the corporatocracy. We all voted for it, and though some indeed didn't realize it, many did. You can see some of them gloating over their own impending demise here on a regular basis.
Granted, but one was going to keep things more or less the same, while the other promised to completely fuck over everyone except rich assholes like himself (which he is actually attempting to do) while gaslighting the voters into thinking he was going to be "better for the economy", leaving out the part where "better" didn't include those aforementioned voters.
Re: (Score:2)
I'm with you all the way on that. I don't want to draw false equivalence. There are however unfortunate ways in which they are exactly the same, and they are not positive. The corporatocracy is based on extractive and unsustainable actions which make the world a worse place for everyone, even the upper scum. They wind up with fewer days fit for yachting.
Re: (Score:2)
The irony is that most average speccers, designers and developers consider fewer corner cases than frontier AI models do now. So many of the former have problems with thinking beyond the happy path and that isn't going to change any time soon. AI models will get even better at it, though.
Re: (Score:2)
Re: (Score:1)
Re: (Score:1)
I think that very likely we have AGI within 4 years. But I also think that you overestimate how fast society will change because of that.
Think about schools. We had the technology to make schools without teachers 20-40 years ago, but still, nothing has changed and schools are still full of teachers. There will be some who adapt fast, but majority will try to slow things down.
Re: (Score:2)
I think that very likely we have AGI within 4 years.
Frankly, we have no particular inkling of what it will take to get to AGI. Might be 6 months, might be decades. The current most popular approaches seem to hit the fundamental limits predicted by most technical people in the field, in spite of the snake oil business folks that have descended upon that subject area.
We had the technology to make schools without teachers 20-40 years ago
I've no idea what you picture as having made teachers obsolete, especially in the 80s. No, the TV cart was not a suitable replacement for a teacher.
Re: (Score:2)
I think you could manage to teach even without a TV with just a text book and practice book, but I think animations and such make it much easier to understand some areas. We know for certain that some kids can learn just from the books alone. We also know that even with a teacher in the class, some students won't learn.
Unfortunately I have not been able to find a single study that would examine how well a normal class would learn without a teacher, if they had a (good) structured self learning material on h
Re: (Score:2)
It's a very good thing you're not in change of anything remotely related to education.
Re: (Score:3)
I think that very likely we have AGI within 4 years.
Why?
Re: (Score:2)
I think we'll be working like dogs right up to the point that we are replaced.
Exactly (Score:2)
Your boss is going to fire you and replace you with an ai. That's going to happen no matter what.
Now let's say a miracle happens and you're one of the employees who doesn't get replaced there's two options.
Option number one, the AI works and your workload stays the same but your pay goes down because there's a lot more competition for work and your boss knows it.
Option number two, the AI doesn't work but you're forced to pick up its slack and your wor
Re: (Score:2)
51% of those people are going to blame woke and trans
We have to. We can't blame the Jews anymore.
antisemitic computers (Score:2)
The AI is trained on too much historical data to avoid this route.
The minute a human management looks away, the AI is going to be calling us juden and traitors and making us wear stars or whatever to identify us even though it already has facial recognition, and regardless if we're even ethnically Jewish.
Re: (Score:2)
It's also his "sometimes while "barely looking at the code."" that sounds optimistic. Pressure will cause it to be just about never looking closely at the code.
Funny thing is, I was curious enough today to search for opinions of which "AI" is best at coding and one site which tests regularly with different cases came up, and of the four that were best at it I picked Copilot as it was free to use without login.
I had it convert some C++ code snippets from my hobby forked project to C, which needed only slight
Re: (Score:3, Insightful)
it said Rust again with logical reasoning.
That's only because Rust is the most hyped up language these days, making it the one most correlated with the term "best". The AI can't have a "preference" nor do any "logical reasoning" to deduce what's better or worse.
I'm confused (Score:1)
That last quote in the summary seems to undermine his entire argument - "barely looking at the code."
Re: (Score:2)
I wager his narrative is that you have it generate some code, you fling the code at the problem blindly, then it behaves badly and you describe how it behaves badly and then it spits out new code and you fling it blindly and eventually, maybe, it will work.
Or maybe more charitably, you have a project in a supremely boilerplate heavy ecosystem. Some of these projects manage to make you enter a lot of verbosity in at least duplicate and LLMs can reasonably take a lot of that tedium out. In which case you don
Fixing the code vomited by the bot (Score:4, Insightful)
"vibe coding" is exhausting because we follow a train of logic when we develop code. The bot does not allow us to do that. It dumps a massive amount of trash on us that may or may not be correct, and you have to read whatever it vomited if it's correct or not. If it's not, then you will have to tell the bot where it's wrong, using a non-formal language, and hope that the new vomit is marginally different so that you don't have to check the whole vomit patch all over again.
Re: (Score:2, Informative)
The point is, that the bot does not follow your logic, if you don't tell it your logic.
Write "Build a note taking webapp" and you might get a good oneshot webapp, but you need to understand the logic first.
But if you lead the LLM in smaller steps like "Now add an undo function. It should use an undo stack with a configurable size, have a redo function and does not need to support branching when editing after an undo step" you have an idea what the result will look like.
You won't tell your coworker "Just bui
Re: (Score:2)
You've pretty much just described a light form of RAD tools, giving that level of detail. And they've been translating logic to code since the '90s.
Re: (Score:2)
Nothing that "improves developer productivity" benefits developers. It just raises expectations, so you wind up working just-as-hard-or-harder, but now utterly dependent on a new tool.
In fact, it makes things harder on developers because they cannot get into their zone and let the code flow out naturally. As the poster said, you have to keep operating within this broken rhythm of telling the AI what to generate (finding the right balance of detail to give) and then code-reviewing it and fixing everything
Re: (Score:2)
There is no stopping it though.
I'm not so sure, since coding AIs as they are today are absolutely unconstitutional. They take copyright works and resell them, depriving authors of their constitutional rights.
Re: (Score:1)
Everything I learned from books where books under copyright.
So I was wrong to read books?
Re:Fixing the code vomited by the bot (Score:4, Interesting)
hope that the new vomit is marginally different
The rest of your comment is basically correct, if unnecessarily negative, but this isn't. Traditional tools like diff make it very easy to see exactly what has changed. In practice, I rely on git, staging all of the iteration's changes ("git add .") before telling the AI to fix whatever needs fixing, then "git diff" to see what it did (or use the equivalent git operations in your IDE if you don't like the command line and unified diffs).
I also find it's helpful to make the AI keep iterating until the code builds and passes the unit tests before I bother taking a real look at what it has done. I don't even bother to read the compiler errors or test failure messages, I just paste them in the AI chat. Once the AI has something that appears to work, then I look at it. Normally, the code is functional and correct, though it's often not structured the way I'd like. Eventually it iterates to something I think is good, though the LLMs have a tendency to over-comment, so I tend to manually delete a lot of comments while doing the final review pass.
I actually find this mode of operation to be surprisingly efficient. Not so much because it gets the code written faster but because I can get other stuff done, too, because I mostly don't mentally context switch while the AI is working and compiles and tests are running.
This mode is probably easier for people who are experienced and comfortable with doing code reviews. Looking at what the AI has done is remarkably similar to looking at the output of a competent but inexperienced programmer.
Re: (Score:2)
This, although I do have to keep an eye on what approach it is taking, because sometimes ... it just gets bizarre. No, no, no!
Then I have to stop it, type a few lines of code or architecture and prompt it that maybe this would be better.
"Oh I see! That would be better"
Ugh.
Re: (Score:2)
Either you never tried it, or are simply bad in expressing what you want. And more importantly, what you don't want.
There is no AI in public access that produce "trash" - at least I am not aware off any.
If you want one to produce trash, you basically have to tell it to it.
Ng (Score:2)
He admits it's basically accepting the output... (Score:3)
The Stanford professor and former Google Brain scientist said the term misleads people into imagining engineers just "go with the vibes" when using AI tools to write code. "It's unfortunate that that's called vibe coding,"
Despite his gripe with the name, Ng is bullish on AI-assisted coding. He said it's "fantastic" that developers can now write software faster with these tools, sometimes while "barely looking at the code."
Which one is it? Is it in-depth assisted coding, or is it fly by the seat of pants and accept anything, style coding? That's why it has a bad name, it's not structured, careful, examined coding, all the time. It's great if you review it, and make sure it's safe, and does what you need, but if you don't? How many projects are being done that aren't being carefully watched and examined, that will have major issues because a developer accepted a block of code, that doesn't exactly do what they needed?
Obviously, the developer can write bad code, which is a different problem, and can be just as serious, if not more serious, but if AI isn't helping that, can worsen it, then why should we overly respect the concept, considering how it can be abused.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
"At this point, *all* developers are using AI in some form"
That's complete bullshit. Most programming is not app or cloud programming, most programming is embedded, fixed function, limited resource programming. AI coding is for masturbators that don't realize there is real programming out there.
Re: (Score:1)
Not sure what point you're trying to make, but the tone comes off more combative than constructive. I'm an embedded software engineer, and I agree that a huge portion of programming happens in constrained, low-level environments. That doesn't mean AI tools aren't useful—it just means they have to be used differently.
In embedded development, you can't rely on "vibe coding." You need to understand and verify every line of code. But even in this domain, AI can help with things like documentation, test
Re: (Score:3)
I'm not. My main development these days is outside of my paid employment, and my development is on personal projects related to some non-profits I'm involved in.
When it's on my own time, well development without AI is more fun - I simply opt not to use it.
I'm not alone.
Re: (Score:2)
Code reviews are a fraud and they always have been. They provide a bullet item on a checklist.
AI coding renders programmers' jobs to pretend they are adding value by doing things of no value at all. But who would know? No programmers left to tell the difference.
Re: (Score:3)
Not always. My first assignment at a new job (realtime embedded, safety critical) was to join a code review on a piece of software that needed to run at millisecond rates.
The code was written by a moderately experienced person whose previous experience was in aerospace ground systems in FORTRAN. This was C code for a very tightly constrained device, 8 bit microprocessor, with a fairly rudimentary compiler. The code being reviewed was array-rich and looked a lot like FORTRAN.
I did a BOTEC calculation very
Re: (Score:3)
Code reviews are a fraud and they always have been. They provide a bullet item on a checklist.
That is a pretty idiotic attitude.
He's not wrong (Score:3)
Doing it right is extremely tiring and taxing
Re: (Score:2)
Yep, sometimes I'll take a day or two away from a coding project so I can get the creative juices flowing again. What usually ends up happening though is I think of a batter way to accomplish what I thought was a finished task and the cycle starts over again, LOL.
Re: (Score:2)
Doing it wrong is far, far, worse in the long run. There's going to be a shit-ton of money to be made cleaning up the nightmare of "vibe" coded AI slop.
exhausted (Score:2)
Whether I'm using Copilot or not, coding all day is exhausting. If I'm in some sort of hackathon where something needs to be developed in 8-10 hours, there's no way I'm not taking a nap after that. The brain burns glucose, and a lot of it. The nice thing about Copilot is it answers directly. Humans will often not answer at all, answer tomorrow, or not answer the questions I asked. Copilot may answer wrong, but at least it's quick about it.
I would not classify anything I'm doing as vibe coding.
Re: (Score:2)
Constantly telling cursor what it is doing wrong or what it forgot is often more tiring than just doing it all yourself, no matter how many prompts you save off to copy/paste back to it.
Re: (Score:2)
This is what most competent developers figured out ages ago. It's wasting far more time that you imagine it's saving.
Hype vs Reality (Score:2)
What Andrew describes is the reality. I vibe code all day too, I admit it. It helps me reduce the amount of mundane code I otherwise have to write. But the non-mundane code, that always needs to be reviewed and often needs adjustments, etc. As a tool, it does make me more efficient, but it is a far cry from being able to replace me.
But that's not the hype. The hype is that it's generating complete apps ready for sale on the app stores, or a vibe coder gets millions of VC cash because they created an app in
Re: (Score:2)
They should be thinking how much more they'll get from their workforce by making them more productive, instead they expect to get the same level of productivity but with fewer (or better, no) people.
That would be to simple and to straight forward.
Where is the reward if you can not fire someone - with a smirk?
"barely looking at the code." (Score:2)
I believe that has been going on well before "Vibe Coding" was a thing given the state of much software out there. This is only going to make it worse.
Scam (Score:4, Interesting)
Re: (Score:2)
Re: (Score:2)
The claims are definitely overblown. At the same time, the gains are real. Just not as big as claimed.
So let me get this straight (Score:2)
I do suppose that the modern vibe coder doesn't even need to know how to code, and trying to troubleshoot bad code is pretty hard when you don't know code.
Insane.
huh? (Score:3)
sounds like he's talking about non-vibe coding, which is where you don't just go with the flow, don't just paste all the errors back in and don't accept whatever it says, and generally use the ai as a tool to help you code.
Weird.
This is I think the coined definition of vibe coding here: https://x.com/karpathy/status/1886192184808149383?lang=en-GB [x.com]
while I'm here what's with people using enshittification to mean 'become shitty'. It has a specific meaning which is more nuanced than 'become shitty'
Enshittification, also known as crapification and platform decay, is a pattern in which two-sided online products and services decline in quality over time. Coined by Cory Doctorow.
Jive Coding (Score:2)
as always....
Writing requirements (Score:2)
Anyone who has ever tried to build an app based upon a poorly written requirements document knows what kind of headache this can be. I feel sorry for the Vibe 'back end' working in Bengaluru who have to put up with this.
Vibe Coding is not the same as coding with AI (Score:4, Interesting)
Coding with AI assistance is what a software developer does who is using AI to a greater or lesser extent to help in their work. The end result is expected to be code that the developer understands (whether they wrote it themselves or not) and that meet the project's standards for code quality and such.
Vibe coding is someone using AI to build a relatively simple application that they would not have either the capability or the time to build on their own, generally to meet a specific use-case or "scratch an itch" that would otherwise go unscratched. In this case, the developer may or may not have any relevant software development skills, and does not need to understand or even read the generated code. If something doesn't work as expected, they give the AI further instructions to hopefully solve it. The end result does not meet any particular coding standards, and is likely not good quality maintainable code, but in many cases it gets the job done.
As a tool to enable non-developers to self-serve a solution for some relatively simple use-cases, i see vibe coding as somewhat analogous to excel or other applications that allow non-developers to build stuff that solves their problem, without writing code themselves. The excel formulas and automations that lurk in the giant spreadsheets that are passed around non-tech workplaces might make a software engineer scream in horror, but most of the time they get the job done.
Re: (Score:1)
Don't forget technical debt. Those excel spreadsheets and what not work fine for a while. When a version of excel makes some change, or deprecates part of the code used, or someone wants to put it in a proper system where the data isn't siloed, all hell breaks loose. And seeing as there are lots of these things around in an average enterprise, it can get to be a real headache.
I'll tell you what's challenging and exhausing (Score:2)
Put your effort into learning how to code and you'll be putting your energy in the right placed.
As pointed out by others, If you know nothing, you don't know enough to vet the output.... unless you barely look at it... which is what he said.
You either learn to make your own game, or you're trapped in someone else's game.
Vibe coding is obsolete (Score:3)
The based, living in the now people are using manifest coding. . Super simple, you perform a supplication to the Universe to manifest your project, and perfect code appears in your email, and you move on to the next manifestation.
Crystals help too.
One more thing: wtf is all this "code" for? (Score:2)
example: cranking out a report or a memo today is the same task as 30 years ago. What feature(s) have been added in 30 years that adds value to that process? Other than theGarbz, no one in their right mind is going to defend microsoft putting a ribbon in and hiding all the menus. That is a perfect example. Lots of new code, completely new interface to learn, exact same task to complete, not only no value to you, it actually removes value, makes your
what an asshole (Score:2)
"When I'm coding for a day with AI coding assistance, I'm frankly exhausted by the end of the day."
"Just like when I masturbate all day!"
"He said it's "fantastic" that developers can now write software faster with these tools, sometimes while "barely looking at the code.""
Sounds exhausting. Wonder how exhausting it is to be cashing VC checks?
Re: (Score:2)
GenAI is embarrassing and giving tech a bad name. (Score:1)
It's the cold fusion of tech, except cold fusion research knows what it is trying to achieve and serves an obvious practical purpose.
Chat GPT, Gen AI at the core of the AI hype is a sophisticated mimic, the only thing it creates is derivative based on what's known already, right or wrong.
What you could do with Agent AI requires API's everywhere and there aren't and if there were we could achieve the same thing with basic process design and engineering more reliably and predictably as we have done for years,
Wait, so reality doesn't match hype??? (Score:2)
And I was about to invest my life savings in a vibe coding company! Whew, that was close!
exhausted (Score:2)
"When I'm coding for a day with AI coding assistance, I'm frankly exhausted by the end of the day."
I can sympathize. You can cover an awful lot of ground with AI assistance but it can be some pretty heavy intellectual lifting. A much different experience from writing the code yourself, and it can be draining.
Call it what it is: outsourcing. (Score:2)
You're outsourcing or subcontracting your work to someone who needs to use a translator for everything you say to it, and everything it says to you.
Hell, we've already had recent news about "AI" companies doing your vibe-coding were a staff of hundreds of humans trying desperately to keep up with your requests.