alt Hacker NewsThat may be true for highly critical systems, but those are a tiny, tiny, tiny minority of all software projects. I mean, how many engineers work on aviation or automotive or X-ray machine or other life-and-death code compared to pretty much anything else?
And not all "production-grade, hundred billion dollar systems" are that critical. Like, Claude Code as we all know is clearly vibe-coded and is already a 10-billion (and rapidly increasing!) dollar system. Google Search and various Meta apps meet those criteria and people are already using LLMs on that code, and will soon be "vibe coding" as I described it.
AWS meets that criteria and has already had an LLM-caused outage! But that's not stopping them from doing even more AI coding. In fact I bet they will invest in more validation suites instead, because those are a good idea anyways. After all, all the cloud providers have been having outages long before the age of LLMs.
The thing most people are missing is that code is cheap, and so automated validations are cheap, and you get more bang for the buck by throwing more code in the form of extensive tests and validations at it than human attention.
Edited to add: I think I can rephrase the last line better thus: you get more bang for the buck by throwing human attention at extensive automated tests and validations of the code rather than at the code itself.
This is you:
>> I think all coding will become vibe coding...
Nope. First of all, Let's get the true definition of "vibe coding" completely clear from the first mention of it from Karpathy. From [0]:
>> "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." [0]
>> "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." [0]
So with the true definition, you are arguing that all coding will become "vibe-coding" and that includes in mission critical software. Not even Karpathy would go as far as that and he's not even sure that he even knows that it works..."mostly".
Responsibility is what cannot be vibe-coded. The major cloud providers and the tech companies that own them have contracts with their customers which is worth billions to their revenue. That is why they cannot afford to "vibe-code" infra that causes them to lose $100M+ a hour when a key part of their infra goes down or stops working.
So:
> Like, Claude Code as we all know is clearly vibe-coded and is already a 10-billion (and rapidly increasing!) dollar system.
That is not vibe-coded anymore and it is maintained by software engineers who look at the code at all times, daily before merging any changes; AI generated or not.
> Google Search and various Meta apps meet those criteria and people are already using LLMs on that code, and will soon be "vibe coding" as I described it.
Nope. As Karpathy described it, that would never happen and human software engineers will be reviewing the agents code all the times. But that would not be vibe-coding would it?
> AWS meets that criteria and has already had an LLM-caused outage!
Are they vibe coding now after that outage? I bet that they are not.
> After all, all the cloud providers have been having outages long before the age of LLMs.
That isn't the point. Someone was held to account for the outages and had to explain why it happened.
They will lose trust + billions of dollars if they admitted that they vibe-coded their entire infra and had 0 engineers who don't understand why it went wrong.
> The thing most people are missing is that code is cheap, and so automated validations are cheap, and you get more bang for the buck by throwing more code in the form of extensive tests and validations at it than human attention.
The risk is amplified with the companies reputation on the line and it's very expensive to lose. I'm talking in the hundreds of billions annually and a 10% loss of global revenues due to constant outages can cause the stock to fall.
So you do understand the contradiction you said earlier about AWS indeed strengthens my point on the limitations on vibe coding especially on mission critical software?
[0] https://x.com/karpathy/status/1886192184808149383