Now the AI rainmakers are infighting about the best public relations for their hot air products.
It is encouraging though that Ng admits that dealing with AI "coding" is like wading through a swamp and leaves you exhausted.
Regarding "vibe coding": I'm waiting for the first National Security Letter that demands that a secret backdoor is inserted in the generated code. No one will notice, because we "can forget about the code".
> A company i worked for, had their MySql server unprotected on the internet for no reason at all. They still used MD5 too. Another company saved credit card information in their DB with a simple generic key in code accessable by everyone.
So your argument is that since we had incompetence before, let's have more of it?
I can imagine also a government agency asking OpenAI to have ChatGPT very subtly align people to the official discourse, whatever it would be. With its access to previous conversations and memories, it's a child's game to slowly nudge people's opinions in the "right" direction. Because something that LLMs excel at is being convincing.
> "Vibe coding," coined by OpenAI cofounder Andrej Karpathy in February, describes giving AI prompts to write code. As Karpathy puts it, developers can "fully give in to the vibes" and "forget the code even exists."
To be fair, Karpathy coined the term specifically for a playful, just-for-fun, quarter-serious approach to software development where you just tell the machine what you want and then be open for surprises. Assisted coding is not vibe coding in that sense.
We use the term as an ironic and silly thing here in the UK, at least amongst my friends, also somewhat in the pejorative. People who have no concept of what they are doing are a joke.
I have a Vibecoding sticker in the SUPREME style on my work laptop, which I think accurately summarises my opinion towards it.
Yeah - being open to surprises (code has tons of broken corner cases, security vulnerabilities, missing error handling, etc, etc) isn't exactly what comes to mind when writing real software.
I'm not even sure in what use cases, even toy ones, vibe coding works. It's been a few months since I tried using Claude to even do simple things like generate a React-based prototype for part of a web page (the sort of use case I would expect it to do well on), and even there it wasn't a "haha isn't this cool" hands-off "vibe coding" experience - I had to intervene and to do my own google research to find out why it was failing and then tell it explicity what to do.
I also have to wonder how many of these AI researchers fully realize the massive gap in complexity between what they are writing (typical ML model or prototype) and what real software development looks like. Let's see Karpathy "vibe code" a 100L-1M LOC system that needs to interact with a dozen undocumented legacy systems via proprietary interfaces, then come back and tell us how that went.
> 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.
This is exactly what is happening.
I think HE doesn't understand what vibe coding means.
Vibe coding is NOT "AI-assisted coding", it is letting the AI do everything with 0 checks beyond manually validating surface-level functionality and passing back the obvious issues to the AI.
That was the original intent, but now vibe coding is exactly synonymous with AI-assisted coding with the public at large. That's probably irreversible, and it's what Ng is highlighting.
How depressing. I sure hope it's not irreversible, but I feel it will be, because the number of novice "vibe coders" out there severely outweighs the number of experienced programmers.
I think Rick Rubin has it right that this is a kind of "punk" coding and I think vibe coding is actually a good term in that regard.
This rings especially true for me as a non-software engineer that took decades to appreciate punk music. I grew up on "virtuoso" shred guitar music and jazz. "A real musician spends hours a day in the woodshed learning their craft. Not just anyone can pick up a guitar and start making music."
I think the analogy also works because the virtuoso had a huge leg up on the untrained musician if they also decided to make simpler music.
On a longer time frame with the analogy, the next fed chairman obviously will not have a degree in clarinet from Juilliard like 99 year old Alan Greenspan.
There are a lot of real and exhausting jobs out there that make the realest and most exhausted Googler look like they sit on their arse all day poking at a keyboard. Vibe coding ain't it.
Next people will be claiming that "influencer" or "twitter shit poster" is a real and exhausting job.
have you seen how hard influencers work? we're talking 100 hour weeks with no weekends. yeah there's no traditional boss and no office to report to, instead there's a influencer house that they go live in, have a bunch of drama at. yeah it's a lot of time spent sitting on the couch poking at a phone. yeah, that's not physically lifting anything but how many jobs do, after sending all the manufacturing jobs to China? One of the most demanding jobs in the world, the President, doesn't lift any rocks or break any backs. It's a job that is just talking with people. That's a job? Just yap and sign papers? I could do that way faster and cheaper.
There's a different question to be made as to whether or not influencer is a good direction for society to strive for as a career, but that's a whole other matter.
This. One summer I worked 80 hour weeks doing construction. I did restaurants when I was young and in university. Coding is a walk in the park compared to actual exhausting work.
Well one is taxing the muscles and the other the brain, can’t both be exhausting, in different ways? Neurotransmitters are not an infinite resource, and there’s trash the brain generates in response to activity that needs clearing out every night.
This is such a BS response, first just because a job isn’t physically exhausting doesn’t mean it’s not challenging and mentally exhausting.
Second, our job in technology is to make ALL jobs easier, that’s what technology is for, not for bullshit manipulative, addictive and extractive consumer crap. The reason any of it even exists is to improve the productivity of humans.
There will always be demanding jobs, they may be demanding physically, or mentally or both, your god damn job is to figure out how to make every one of those jobs easier and LESS physically and mentally challenging.
Pointing out the obvious fact that using different metrics other jobs are harder is neither helpful, valuable nor unique.
I will however agree with you last statement, technologies abuse of people in the consumer app space is anti-social and destructive to the world, those are “jobs” we created with technology. In a sense you might say we are responsible for creating the worst jobs in the world, because as easy and valueless as being an influencer is, it destroys people mentally and turns people into shells of human beings.
So instead of trying to imply that all your fellow engineers are a bunch of whiny soft and weak complainers, you should be both simultaneously grateful that there are jobs that are physically easy and obligated to help those whose jobs still aren’t easy make as much of their jobs as easy as possible.
We live in a society we are ALL dependent on each other, specialization is what allows us to have large complex societies, without it we would all be trying to find food and build shelter. We can ONLY have our jobs because others do theirs, never forget that, that fact creates an OBLIGATION not a comparison.
hi, lets meet. imo this is bad for the whole community to bad mouth vibe coding with such negative connotations, effectively gate keeping. vibe coding is like a car for people in the age of walking. sure you like to walk, want to keep walking, even enjoy walking, but there is no comparison with how far you will get if you start using a car. you still need to walk to your car and all effectively holds.
Vibe “migrated” our docs portal to another framework the other day. Looked awesome. Only when I decided to do a quick review before switchover did I found all subtle hallucinations.
Not sure what would be faster. Manually reviewing and fixing the AI migrated docs or just give up and use sed or something to do it from scratch.
Whenever you feel ,,not sure if fixing what a huge change did is faster'' you are asking something too big and not specific enough.
It's much better to ask the specific queries that you yourself would do with sed and treat the models as smarter sed (or ask them to generate the sed statements themselves)
>It there _any_ empirical evidence for this, at all? Like, beyond the _perception_ of the users.
Someone managed to write a standards-compliant HTTP 2.0 server in just two weeks, 40k+ lines of code (including tests) with almost all of the coding done by Gemini Pro:
https://outervationai.substack.com/p/building-a-100-llm-writ... . No unassisted programmer could do that so fast (except maybe Fabrice Bellard).
How would that evidence exist? Aider writes how much code was generated in each release, and it's in the 80%s, but what you're asking for is for someone to do a timed trial on writing a bit of code unassisted, then for that same person to write the same bit of code with LLM assistance and also with forgetting the first version they wrote. Are we so caught up in proving they're not idk that we're stuck with the parachute engineering/science problem? Because, scientifically, the only way we can know parachutes are any good is to test the hypothesis by throwing the control group out of the plane without parachutes and comparing them to a group of people who jumped out of the plane with parachutes.
Yes none. However, one cannot discredit lived experience... and for many the lived experience of vibe coding is transformative and democratizing. Dare I say thats only fare given all the unwarranted "democratizing" the tech industry thought was helpful over the last 30 years...
Readit News
It is encouraging though that Ng admits that dealing with AI "coding" is like wading through a swamp and leaves you exhausted.
Regarding "vibe coding": I'm waiting for the first National Security Letter that demands that a secret backdoor is inserted in the generated code. No one will notice, because we "can forget about the code".
And you continue doing so with every other point you are making.
Just because some random people think they can vibecode real products, doesn't mean that this didn't happen before just slower.
A company i worked for, had their MySql server unprotected on the internet for no reason at all. They still used MD5 too.
Another company saved credit card information in their DB with a simple generic key in code accessable by everyone.
So your argument is that since we had incompetence before, let's have more of it?
To be fair, Karpathy coined the term specifically for a playful, just-for-fun, quarter-serious approach to software development where you just tell the machine what you want and then be open for surprises. Assisted coding is not vibe coding in that sense.
I have a Vibecoding sticker in the SUPREME style on my work laptop, which I think accurately summarises my opinion towards it.
I'm not even sure in what use cases, even toy ones, vibe coding works. It's been a few months since I tried using Claude to even do simple things like generate a React-based prototype for part of a web page (the sort of use case I would expect it to do well on), and even there it wasn't a "haha isn't this cool" hands-off "vibe coding" experience - I had to intervene and to do my own google research to find out why it was failing and then tell it explicity what to do.
I also have to wonder how many of these AI researchers fully realize the massive gap in complexity between what they are writing (typical ML model or prototype) and what real software development looks like. Let's see Karpathy "vibe code" a 100L-1M LOC system that needs to interact with a dozen undocumented legacy systems via proprietary interfaces, then come back and tell us how that went.
This is exactly what is happening.
I think HE doesn't understand what vibe coding means.
Vibe coding is NOT "AI-assisted coding", it is letting the AI do everything with 0 checks beyond manually validating surface-level functionality and passing back the obvious issues to the AI.
That was the original intent, but now vibe coding is exactly synonymous with AI-assisted coding with the public at large. That's probably irreversible, and it's what Ng is highlighting.
I think Rick Rubin has it right that this is a kind of "punk" coding and I think vibe coding is actually a good term in that regard.
This rings especially true for me as a non-software engineer that took decades to appreciate punk music. I grew up on "virtuoso" shred guitar music and jazz. "A real musician spends hours a day in the woodshed learning their craft. Not just anyone can pick up a guitar and start making music."
I think the analogy also works because the virtuoso had a huge leg up on the untrained musician if they also decided to make simpler music.
On a longer time frame with the analogy, the next fed chairman obviously will not have a degree in clarinet from Juilliard like 99 year old Alan Greenspan.
Next people will be claiming that "influencer" or "twitter shit poster" is a real and exhausting job.
Manager who sits upvoting and saying "great post" on LinkedIn all day, seems to be most middle management right now.
There's a different question to be made as to whether or not influencer is a good direction for society to strive for as a career, but that's a whole other matter.
Dead Comment
Second, our job in technology is to make ALL jobs easier, that’s what technology is for, not for bullshit manipulative, addictive and extractive consumer crap. The reason any of it even exists is to improve the productivity of humans.
There will always be demanding jobs, they may be demanding physically, or mentally or both, your god damn job is to figure out how to make every one of those jobs easier and LESS physically and mentally challenging.
Pointing out the obvious fact that using different metrics other jobs are harder is neither helpful, valuable nor unique.
I will however agree with you last statement, technologies abuse of people in the consumer app space is anti-social and destructive to the world, those are “jobs” we created with technology. In a sense you might say we are responsible for creating the worst jobs in the world, because as easy and valueless as being an influencer is, it destroys people mentally and turns people into shells of human beings.
So instead of trying to imply that all your fellow engineers are a bunch of whiny soft and weak complainers, you should be both simultaneously grateful that there are jobs that are physically easy and obligated to help those whose jobs still aren’t easy make as much of their jobs as easy as possible.
We live in a society we are ALL dependent on each other, specialization is what allows us to have large complex societies, without it we would all be trying to find food and build shelter. We can ONLY have our jobs because others do theirs, never forget that, that fact creates an OBLIGATION not a comparison.
Not sure what would be faster. Manually reviewing and fixing the AI migrated docs or just give up and use sed or something to do it from scratch.
It's much better to ask the specific queries that you yourself would do with sed and treat the models as smarter sed (or ask them to generate the sed statements themselves)
It there _any_ empirical evidence for this, at all? Like, beyond the _perception_ of the users.
Someone managed to write a standards-compliant HTTP 2.0 server in just two weeks, 40k+ lines of code (including tests) with almost all of the coding done by Gemini Pro: https://outervationai.substack.com/p/building-a-100-llm-writ... . No unassisted programmer could do that so fast (except maybe Fabrice Bellard).
Dead Comment