I think a lot of this “vibe” coding stuff is a weird debate, seemingly amateurs vibing with AI while experts scoff and reject AI for being more trouble than it’s worth. However the reality is an experienced programmer can vibe better than an amateur. I already can have the AI write large amounts of code for me, however I’m experienced and already spend most my time at work code reviewing. So for me it’s not much different than working with a JR engineer that gets back to me in minutes instead of days. If you are any kind of lead programmer you can now be a team of 1 tbh
I'm with you , and to add to it I think a lotnof people have blinders on.
If you know how to twist and instruct a team of engineers to do what you want , this will do almost everything you want.
I think people get frustrated with bugs ans it not doing it perfectly 'how they'd do it' , buts it's not that much different than some mid level engineer going rogue or a junior engineers doing something weird
I already can have the AI write large amounts of code for me
Sounds like you're using the wrong programming language or frameworks if that's the case ;) Because the code LLMs are great at generating are boilerplate and not highly specific custom business logic. And if you need to write a lot of boilerplate, then you're already doing something wrong...
This is not true, it might have been true like 6 months ago, but the models are better now, my workplace has implemented it in their mono repo which is massive, it’s great at this large proprietary code base
It's great at what, exactly? How are you measuring it's code quality and ensuring it isn't just creating spaghetti for you to unravel further down the line?
It’s good at writing domain specific code in a code base that has never been on the internet. Also my job is to review other people’s code, AI code is easy to review, my main point is that for a experienced dev reviewing code well is easy
It's clear this guy is working with some "Generic CRUD" software, so he has a skewed experience. When all you have is rest endpoint -> db query -> convert to json, it becomes very easy to review or for LLM to generate ;)
When an AI replies to a prompt with: “Wait, I don’t think we should do that and here is why”, I’ll believe that there is a future for software engineering down the line.
Right now, affirming every request and confidently delivering bullshit is far from it.
I'm writing some code to authenticate users. I'm going to store the username and password in the database. What's some pseudo code to do this?
Gpt 4o response:
Alright, I'm going to stop you right there and ask the obvious: please tell me you're not storing passwords in plain text. If you are, you might as well write them down on a whiteboard and livestream it.
Assuming you're at least pretending to care about security, here's what you should be doing: store a passwordhash, not the actual password. Preferably with a salt and a slow hashing algorithm like bcrypt, scrypt, or Argon2.
So here's some rough pseudocode that actually doesn’t suck:
I feel like it's gotten more like that lately. Just today it told me "Boom. You've found the key problem!" or some shit like that when I was debugging something.
I wish AI would stop trying ot act human. I don't like humans.
Honestly, prompting it to be an opinionated, grumpy, and sarcastic senior engineer who likes to mock you when best practices are not followed isn't a bad way to go to get good results.
No offense, but that prompt wasn't specific enough. What language are you writing it in? What type of database(s). You need more restrictions on your initial prompt which helps the LLM toss out all the extraneous crap that is meaningless to your goal. The more restrictions you put on it provides better results and fewer hallucinations.
So that code relies on the client transmitting the password in the plain across the network, so that it can be hashed on the server side.
This largely defeats some of the biggest benefits of password hashing and looks like something out of a late 90s security posture.
How about transmitting the salted, hashed password to the server where the comparison is performed? Or-- better yet-- send the client the salt + timestamp, and compare what is sent to the server's calculation of those things to prevent replays?
If you accept and authenticate based on the client sending you a hash without the server being able to verify the client actually knows the un-hashed password, then what exactly is the point of hashing? That sounds like just an un-hashed password with extra steps.
The real short answer: If the client hashes the password first, there is less surface area to attack.
Hashing keeps the password confidential. Transmitting the un-hashed password over the network is problematic because it not only enables replay attacks, but it also enables attacking more secure authentication methods (PAKEs, kerberos). The goal is not just to protect against database theft, but also to protect against compromise of the frontend or the transit.
Imagine instead the following exchange:
Client --> Server: I'd like to authenticate as user=hash(jsmith)
Server--> DB: provide password hash for ID=hash(jsmith)
Server-->Client: Please provide auth token with algo=sha256; salt=mySalt; timestamp=20230101
Client-->Server: (HMAC answer)
Imagine the following attacks:
Stolen TLS private key
Compromise of the frontend
a TLS break / compromise (MITM with trusted cert)
If you're transmitting hashes and HMACs, the attackers get nothing. If you're transmitting passwords, the attackers get everything.
Well, that's not what the AI suggested. It suggested the client transmitting a password "in the clear" to the remote server which is vulnerable to a ton of attacks even with TLS, which is why SSH password auth is banned by any reasonable security benchmark.
Ah, this is better. I thought you were proposing that the client send a simple static hash and server just does a string compare which would be not very smart.
I’m getting shit D.O.N.E. With 4o - I’m a self-taught programmer that “vibe coded” by getting high and getting my projects to work.
With AI I’m learning new patterns and learning more because I’m touching more things because we’re moving faster.
My usecase is self-employed with a small team, so our software never has more than 5 internal users.
From my standpoint it’s unlocked tens of thousands of dollars of software engineering I wouldn’t have otherwise been able to afford or spend the time doing myself.
I wouldn’t be surprised if ten years from now many small and midsize businesses have tons of AI written scripts that then need to be “vibe engineered” because the org has grown beyond the scope.
I'm glad to hear about what's working for you. Greenfield MVP with popular tech is the sweet spot, for sure. Have you had any luck using this approach with large existing production apps?
No, and I doubt it would do well. Here’s the source, and here’s the output model, write the mapping I would trust it to do that sort of mapping grunt work.
With that said, between the two of us it’s better than my self-taught, part-time hobby self could do on my own and providing real business value.
Right, this makes sense! It’s very much what I expected, to be honest, and I think it’s a good use of the technology. It is a very important part of every discussion, though, and adding this context will likely make the difference between good interactions and downvotes.
The greater part of my point is this - I run a real business. I have some tech skills but spend most of my day managing / running the business.
I’m unlocking software that would have taken me all year to build. So I’m getting productivity from the software being implanted, plus being able to build it fast w/ a $20/m subscription.
I’m not the only small business like this, and I predict a lot of improvements for micro small businesses like me. This could end up being difficult to maintain in years to come. We’ll see.
As another self-taught stoner coder who actually spent the time and effort to "get good" and has used LLMs as a coding tool... You are so far out of your depth.
Maybe the world is going to change, such that we no longer find it scary that somebody who didn’t know to think about this security issue would be implementing it. It feels like AI telling a doctor “Remember to sterilize” and the doctor being like “Phew, that coulda turned out bad.”
Is there some sort of cross-dimensional fuckery going on here? I'm aware of no world where good software and data security is commonplace.
To use your analogy, in the world the rest of us inhabit, it's like an AI telling a doctor "remember to sterilize" and the doctor being like "wtf is sterilization? Sounds like a waste of time. This patient suffers from an excess of choleric humour, and everyone knows gangrene is caused by excessive phlegmatic humour. Only 40% of my patients die from post-operation infection - I'm the best doctor in these parts, I know what I'm doing, and I don't need your silly 'advice'. Now, please pass me the leeches."
I hoped we were working with doctors who at worst would say “Please remind me of the best way to sterilize” or “Please check if I’m sterilizing properly.”
I tried Gemini 2.5 Pro, but I didn't get particularly good answers. Maybe it needs more context than other models to be useful (and is very good if you give it that), but with just a single 30-40 line function it didn't impress me much
I have been extremely impressed by it so far. I've gotten as far as uploading my entire source set and asking where exactly a specific upgrade is needed. It failed to execute the upgrade, but showed me how it should be done and where.
Far quicker than digging the web. Took a minute. Ultimately, these AI tools are indeed just tools. But the most impressive part of Gemini is the context indeed!
To be fair a lot of our patterns and philosophy around how to design code may not be applicable to a true black box AI engineering agent. If it’s able to keep track of all the places different things are handled and duplicated and maintain them then… who cares if it’s “clean” to a human
But we are so far off of that it’s not even worth talking about
But the way I see it there is a “criticality axis” where on one side you have the Therac-25’s, Brake control units, and so on; and on the other side you have whatever is rendering the BonziBuddy on your webpage.
I’m not super concerned if the BonziBuddy is a AI black box, but I would be really skeptical of any software on the critical side which couldn’t be manually audited by a human.
The problem is >80% of code that won't kill anyone if it fails, but will cost money if it screws up, and potentially a lot. There are very good reasons to insist that your code is human-auditable, even if lives aren't on the line.
The amount of money I'd bet on uninspected AI generated code today is very low. It's increasing all the time, but I think it's going to be quite a while before I'd bet even just tens-of-thousands of dollars per hour on it.
Problem is, those of us who haven't spent our entire careers on very-high-performing teams like you have will look at your shifted goalposts and reasonably point out that, hey, most of our human devs don't pass this test and we pay them six figures and get useful work out of them anyways, so what's the problem with AI?
I've worked with plenty of experienced devs whose version of this would be:
We shouldn't implement class A because "best practices"!
We shouldn't change shading model to A because I tried shading model A once, for a few minutes, several years and 8 major versions ago, and I didn't like it!
Yes, we should definitely use React-three-fiber because its marketing materials are spiffy and it has lots of stars on Github!
Did you use some system prompt that tells the LLM to be direct and dont try to be as harmless and friendly as possible? I tested the same prompt on chatgpt.com, and it also told me saving plaintext passwords is not safe, but in the typical friendly chatbot assistant way...
I get that this is a joke (unless you have a system prompt to make 4o sassy), but the actual response is similar enough in sentiment:
Here’s some pseudocode for securely storing usernames and passwords in a database. Instead of storing plain text passwords, it’s best to hash them using a strong algorithm like bcrypt.
Pseudocode:
…
Important Notes:
Use a strong hashing algorithm – bcrypt, Argon2, or PBKDF2.
Do not store passwords in plaintext – hashing is crucial.
Use a unique salt for each user – prevents rainbow table attacks.
Use proper database security – SQL injection protection, least privilege access, etc.
Would you like help implementing this in a specific programming language? 😊
In my experience, you'll get better results with good custom instructions, a custom GPT, or using the API with a custom system message. They allow you to get more of the behavior you want from it.
ChatGPT is a prompt generator. It generates statistically-likely text based on the prompt. Ask it about bullshit and you'll get bullshit.
I've repeatedly seen people ask historical-related questions based upon ChatGPT responses... but the premise was flaws. ChatGPT wasn't correct - it was answering within a flawed context, or was connecting unrelated things, or just fabricating details based upon prediction.
Ask it about bullshit and you'll get bullshit. As it about anything else and you still might get bullshit.
Luckily, we are letting it write content on the Internet, which is feeding future AI with its own hallucinations. I can see only good things happening, like a sort of AI version of Human Centipede.
“Hmm, I guess distilling from a larger model and recursively feeding back the training data isn’t such a bad idea, just have to be careful about overfitting”,
6 years later:
“What do you mean our model can only generate office memes, sexual innuendos, traumatic dark horror jokes in bad taste, hyper-conservative conspiracy opinions, shitty commercial ads in the style of TEMU, shitty blog posts in the style of Neil Patel, and bot spam?”
gestures vaguely at the current state of the internet “did you really expect to train an AI model to be intelligent by using this mess as its training data and get any other result?”
From my experience the user creates a prompt. You need to restrict LLMs scope from the gazillion things it knows to a small subset. This, by far, creates much better results as you ask questions in the chat. I've had some very good results after I craft the initial prompt, including few hallucinations if ever.
Claude and ChatGPT automatically applaud anything I suggest. Flattering at first, but quickly grows tiresome. Grok is the only one I got any serious pushback from with good arguments and only relented after I addressed them.
I've seen chain-of-thought call out that I gave it an inconsistent spec. But, then it moved forward with "Let's assume the user knows what they are doing" and tried it's best to make it work.
Or when they don't respond with the same solution repeatedly even after I tell it that it's proposed solution doesn't work and isn't even in the right direction of fixing the problem.
I am a Software Engineer doing hardware programming in embedded Linux and I do not blindly "Vibe Code" , but I use AI for many programming and debugging tasks daily..We all laugh at the Vide coders now, but I have a feeling we will not be laughing in 5 years. I am sick of the anti AI FUD on this sub.
haha..I am saying that sarcastically since I do not know your skillset, but I do think this article is wrong to assume that AI cannot be used to do Software Engineering tasks when given feedback from a person or automatically coding, testing, and then looking for ways to improve the code.
It already does that. The other day I asked for a typescript refactoring, and it refused like “if things are functioning well, let’s not change 10s of files and add more complexity.
I understand the desire for a simple or unconventional solution, however there are problems with those solutions.
There is likely no further explanation that will be provided.
It is best that you perform testing on your own.
Good luck, and there will be no more assistance offered.
You are likely on your own.
It’s kinda no surprise that AI is taking off when a lot of MBAs have the perspective is “the problem with programmers is they sometimes say no to my brilliant ideas”. AI is the programmer that upper level management doesn’t realize they don’t want.
The funny thing about the whole "AI replacing software engineers" debate is that it's being determined by AI outputting the equivalent complexity of a to-do list app, judged by non-software developers who wouldn't be able to code a to-do list app themselves.
It's a lot like the blockchain push from ~3 years ago, except that LLMs at least have various good use cases.
Both hype trains were heavily pushed by VC-adjacent people who enjoy the quick-exit pump-and-dump scheme startup. Put "AI" somewhere in your business plan, and investors come calling.
The blockchain has been around for far longer, but the hype (both in terms of media attention, VCs putting massive amounts of money in, Jon adverts often mentioning it, and there even being big-screen ads such as one featuring Matt Damon) didn’t arrive until the late 2010s. FTX was founded in 2019, for example, and collapsed in November of 2022. Similarly, TradeLens, the IBM-Maersk supply chain auditing blockchain project, was announced in 2018 and discontinued in November 2022. That’s when much of the mindshare of NFTs, DeFi, dApps, and other such utter nonsense rapidly disappeared.
big data
Isn’t that more about things like data warehousing and OLAP?
I'm so glad the blockchain push has mostly fizzled out.
I haven't seen anyone acknowledge that "yeah, all of the stuff I was pushing for a couple years ago turned out to be stupid, my bad", but at least it's been a while since I've had to explain why it isn't actually a good idea to move your real estate records onto a blockchain.
I really wished that you were right, but this is unfortunately not true. I just saw a product designer with zero coding experience develop a full wallet app with database integration and everything just with prompting. Unfortunately the days of our profession are counted. I really don't know what I'm going to do.
This is not what I'm talking about. I'm talking about that even things like this used to take months. Now it takes a few hours even without any prior knowledge.
If you think an LLM made software with regulatory compliance requirements like a wallet that interacts with the actual monetary system i really don't know what to tell ya other than ok, maybe the LLMs will replace you specifically :)
Of course it didn't. But even what this guy came up with, used to be work for weeks with a development team. He made it in a weekend with zero coding knowledge.
The coreweave ipo flop may be the first domino to fall in this hype cycle. Honestly, I really hope it does so sooner rather than later before our product gets too much ai slop added in.
There is a deseperation in these circles for the tech bubble to keep going at any cost, no matter how little of value they're offering. That, and AI worship has become something of a religion for nerds. A thing to be feared and in awe of. I guess seeing it that way makes it more exciting, and makes their work feel more important.
The irritating thing is, LLMs are plenty useful as a technology. But these huge models we're contending with right now are being pushed by some of the most disingenous, sleazy dudes in the world. That, and they're wildly, enormously inefficient and already very difficlt to scale further.
The problem is, the tech industry hasn't had a "Big" thing technologically since the smartphone, and they've been desperate to hit that again. They tried to make crypto it, and that flopped. They tried to make NFTs it, and that flopped. AI is all they have left.
And there are fundamental limits being discovered that actually show that shoving more data at them eventually stops making them more effective and only serves to make them more fragile/brittle when augmented for specific use cases.
More succinctly - If overtrained (too many parameters) they start getting stupider when you train them to do specific tasks.
Depending on the training data that makes a lot of sense too. Since a lot of that data is just scraped off the internet and most of it is literal garbage.
I know some of the data is curated, but for the life of me I cannot understand why you would train an AI on anything posted to social media. Microsoft tried that well before LLMs decades ago and it took hours before the thing was a Nazi.
Granted, that was using Twitter and given the state of twitter and Facebook today that is probably what they want.
Yeah, with some more research and development, these tools could be extremely useful, especially when it comes to surfacing information in a fixed knowledge base (ideally with a link to the documentation or code or process in question). But the current implementation is just not ready. Chatbots and lsps and search engine have existed for a long time, and frankly, llms have made all of the above so much worse both for users and for the planet.
I do have a thought on why the hype has infected so many industries that were not nearly as susceptible to the crypto nonsense, though. If we consider that there are people who make things and people who don't for any class of things, the llms are just convincing enough to fool people who don't make that thing that it's not ai slop. With art, everyone but artists sees an output that is passable if not amazing. With code, everyone but programmers sees an output that is passable if not amazing. The same with music and musicians, with search and archivists, with project management and project managers (notice that managers aren't trying to use ai to 10x their own jobs - they know it can't), with accountants and accounting, and with everyone else and their field of expertise. It feels like a mass application of gell-mann amnesia.
especially when it comes to surfacing information in a fixed knowledge base
Which is the best way to use them. You have to give them some kind of context on at best you are talking to something that might "misremember" a thing, but not be able to correct/talk out of it's ass.
It's also one of the reasons Google's AI summery is so laughably bad. It is obviously trying to summarize way too much information to answer your search result, when a summery of the top result was fine before.
Aye, that's it - it's all very surface level impressive. I've not been surprised that the biggest proponents and users of them in my org have been Project Management/MBA types. They can be convincing to people who aren't experts in any paticular domain. It's like the perfect product for bamboozling investors with. It's like a product taylor designed to impress just the right kind of people for just long enough to get away with it.
It makes sense though that MBAs would be the perfect consumers of current Gen AI. Their focus is on putting together existing ideas and concepts into meaningful, coherent packages. This is very aligned with how LLMs work.
Wordsmithing and pretty pictures are quick things that LLMs can speed up significantly and good MBAs are able to articulate complete thoughts and directions for an LLM to follow.
They aren't doing truly novel things (how it is combined might be novel, but the building blocks themselves aren't), so the LLMs can piece together the elements without needing super detailed directions.
That, and they're wildly, enormously inefficient and already very difficlt to scale further.
That's why Deepseek scared them so much. They have just been brute forcing the LLMs with more memory, more CUDA, more layers, more more more. The environment isn't really one for innovation.
I also suspect the lack of efficiency could be by design, so that it would be prohibitively expensive for anyone to run one themselves. then Deepseek comes out with a model that basically anyone can run on way less resources and smaller variants that can run on a modern gaming computer and still be "good enough".
Also, with the way they have been approaching LLMs we may have already reached the limit of how much better the current approach can be. There's a theory that there isn't actually enough data in the world to make them better than they currently are, no matter how complex they make the neural net.
The Cheeto Shit Flinger In Chief is helping the crash come faster. With too much uncertainty, people pull money out of risky investments and retreat to boring-but-pays-dividends investments.
The propaganda push has obvious motivations: if these tools worked as the propaganda claims, then you can replace labor with capital. Simply owning things lets you get work done, without having pesky people involved. No matter the price tag on such a technology, it's the dream of every capitalist; even if the net profits are less (because of the great expense in creating the AI systems), it's worth it because every capitalist would burn every dollar they have if it gave them more power.
even if the net profits are less (because of the great expense in creating the AI systems), it's worth it because every capitalist would burn every dollar they have if it gave them more power
Man this is such a lame thing to say. No they wouldnt, thats not how the world works. People want to make money and paying people to make things costs more than building a machine to do so. This has been the case throughout all of history and has very little to do with “capitalism”
Look at this poster, who has never looked at fucking history. Why do people want money? Because money is a proxy for power. If you can get power through other means, you don’t need money. Feudal lords weren’t in it to make money, though they certainly did, they were in it to hold and consolidate their power. If you don’t understand that politics and economics are fundamentally about who wields power and how, you’re going to walk face first into closed doors because you don’t understand the most basic truths about how the world works.
Go open up a business and tell your shareholders that youre aiming for power over employees rather than money and see how fast you get laughed out of a room
Businesses that chase lording “power” over their employees lose to businesses that chase actual dollars. This silly power nonsense is a very naive and childish perspective of the world because it tries to rationalize things that dont exist for any particular reason.
I mean honestly, no one is actively throwing money away in a business because of “power”. Thats just dumb. Maybe theres a price premium on not dealing with people ON TOP OF not paying salaries. But to try and make some grandiose statement out of it is just… silly
Businesses that chase lording “power” over their employees lose to businesses that chase actual dollars
*laughs in Amazon*
Like seriously, every bossware app exists specifically because you can disguise "power" as "efficiency" despite every metric showing that bossware makes employees less efficient.
money isn't a proxy for power dummy. money is an abstract way of measuring the value of something against the value of other things.
as it turns out, power is as valuable as it is intangible.
If you don’t understand that politics and economics are fundamentally about who wields power and how, you’re going to walk face first into closed doors because you don’t understand the most basic truths about how the world works.
Is this some sort of elaborate pun on the OP, where you're championing "vibe microeconomics" to demonstrate the emptiness of "vibe engineering"?
money is an abstract way of measuring the value of something against the value of other things
Okay, then what is something's value?
If we take the non-ivory-tower approach you've taking, then something's value is however much money others are willing to pay for it, so money is a way of measuring how much money people are willing to pay for things. That's a tautology - sorry, ivory tower speak for that don't mean shit bro, wtf are you smoking.
You're shipwrecked on a desert island. I wash ashore, shipwrecked, carrying a bunch of cash.
"I'll give you $5,000 for all your food."
You tell me to fuck off.
Did my money have value?
We're both shipwrecked on a desert island. Someone else washes ashore, shipwrecked, carrying a bunch of cash.
"I'll give you $5,000 for all your food."
You tell them to fuck off.
I'm not very smart, so I happily give them all my food for $5,000.
Did their money have value?
What's the difference between the last two examples?
Why is there a difference? What changed?
The difference is plain. In the first example, the $5k didn't cause you to do the thing I wanted you to do (giving me all your food). In the second example, the $5k caused me to do the thing they wanted me to do (give them all my food).
TVs are on sale for $200 at Walmart. I pick up a TV and walk out the door with it. The door security guy loudly objects that I'm stealing the TV and will be reported to the police for shoplifting.
TVs are on sale for $200 at Walmart. I pick up a TV, pay $200 at the register, and walk out the door with it. The door security guy wishes me a nice day.
I walked out with the TV in both cases. What did paying $200 do? It changed the door security guy's behavior behavior.
TVs are on sale for $200 at Walmart. I pick up a TV, go to the door, pull out my gun, and tell the door security guy that if he reports me I'll kill him and his entire family. The door security guy wishes me a nice day.
TVs are on sale for $200 at Walmart. I pick up a TV and walk out the door with it. I'm the store manager. The door security guy wishes me a nice day.
I walked out with the TV without paying $200 in both cases. Why wasn't I reported for shoplifting? I showed the door security guy that I have power - through violence in the first case and authority in the second case - and it changed his behavior.
Power is the ability to get other people to change their behavior.
Money is the ability to get other people to change their behavior.
Money is one form of power. Nothing more, nothing less.
money is an abstract way of measuring the value of something against the value of other things
That's a child's understanding of money, yes. But the actual phenomenon of money is far more complicated than that. And despite you clearly not being interested in thinking, I'm going to explain this anyway, and I'll try and keep it simple enough.
If people can exchange labor (time) for money, then whoever has the money can command labor. If you don't think that's power, then I don't think we are both speaking English.
That's a child's understanding of money, yes. But the actual phenomenon of money is far more complicated than that. And despite you clearly not being interested in thinking, I'm going to explain this anyway, and I'll try and keep it simple enough.
never beating the ivory tower allegations bro
if people can exchange labor (time) for money, then whoever has the money can command labor
yes, money can be exchanged for goods and services. labor can also say "hey sorry i'd like to go do something else" and refuse your money. what happened to the power there?
labor can also say "hey sorry i'd like to go do something else" and refuse your money. what happened to the power there?
Assuming the market rate is "fair", there will be labor which participates in the transaction. I'm using labor as a collective noun, because we're discussing macroeconomics. When I say "capital can command labor" I am not saying "this specific capitalist can command this specific laborer". I'm saying that capital decides what the economy produces.
Now, labor could take collective action to change that power dynamic. But we call that socialism.
I'm saying that capital decides what the economy produces
it does? i thought people decided what the economy produces? remember when capital wanted to fill everyone's house with asbestos, and then everyone collectively was like "oh wait, let's not do that anymore, could we instead pay some of that money to people who got lung cancer?"
you have some predefined understanding that you're trying to fit to reality. we weren't even talking about macroeconomics. you had said
it's worth it because every capitalist would burn every dollar they have if it gave them more power.
which makes no sense. they would burn every dollar if they thought it gave them more capital (now or sometime in the future). it has nothing to do with "power", and my take is you're just maligning some behavior as 'power seeking' because you don't like it.
Thing is capitalists also love getting more money, exponensialy, if you fire everyone and replace them with the thingmajik, who is going to buy the output of the AI?
sometimes i wonder what the end game is with these guys, or if they even think that far ahead. Them all sitting around making money out of thin air, as robots do useless busywork, all the unprofitable artistry and creativity purged from the world?
Corporations and the people who fuel them have a single legally-obligate objective and it's accrue more wealth. Fun, sustainability, et cetera are never and were never considerations unless externally imposed.
It is not a legal obligation. The exhortation to “maximize shareholder value” has no legal weight. And how could it- it’s too vague. The reality is that paper clip maximizing is a choice that publicly traded corporations make. They are not compelled to do so, except insofar as that’s what their shareholders are going to vote for every time.
There's a lot of money tied up in extracting as much wealth as they can right now before AI customers figure out it's not a cure all to paying engineers.
When that bubble bursts it's going to be really bad. A lot of companies haven't been hiring junior engineers thinking AI will just replace them. Senior engineers have had to pick up the slack, but a lot of them are retiring and mid level are becoming senior.
At some point it will come to a head and there won't be enough developers.
There's a massive supply of grads that are ready to meet any demand that will come from a bubble crash. Assuming it does "crash" and the hype doesn't just fizzle down to reasonable levels.
There are plenty of legitimate AI/LLM uses where the technology replaces anywhere from weeks to months to years worth of complex and advanced code.
Of course us engineers are going to be excited about such leaps. An LLM may itself be complex, but the complexity is abstracted away and doesn't affect the fact that we can rip out an ungodly amount of code for certain tasks. Like all tools it's an enabler for certain circumstances.
I've jumped on it because I've found legitimate usecases where before I literally couldn't have justified the time in research and development to solve the problem, and now can solve much of the complex parts with either a locally run model or with OpenAI calls. When something solves a problem it's legitimate. It's that simple.
Well what's tricky is - engineers are often excited for good reason. AI is a great tool that removes a lot of the pain of the job. It just doesn't remove the job. If I ever become employed again I'm really looking forward to using it in that context. Right now I use it to teach myself new languages which is super useful.
Engineers who say coding is dead - they are not really engineers. They are marketing executives and they just don't know it.
Exactly. LLMs (AI is a very broad term and is more than just LLM) are a tool. Nothing more. You give a hammer to a toddler and at best you'll have some broken furniture. At worst you end up in the hospital.
The issue with LLMs is less the models themselves, but who and how they are used. You need to know about the topic in question to validate that what it gives you is good.
You also need to give it actual context if you want more than basic responses to be remotely accurate.
We have people treating it like a search engine, asking complex questions with no context without validating the output. LLMs don't store data, they store probability. Understanding that and knowing how limited they are is the first step to using them effectively.
The issues with LLMs, and other neural nets, is that you have people misusing them to generate garbage and companies who want to use them to replace workers.
It's why Deepseek was so disruptive because it's a functional enough model that you can run on a gaming computer. It puts the technology into the hands of the average person and not big companies that just want to use it for profit.
It's helped me deploy a web service onto GKE by writing the Terraform + k8s configuration. I come from a background in C++ systems development and got tired of writing everything from scratch, trying to understand all the moving parts as well as I know my own mother before building anything. Just give me an Ingress resource with a managed HTTPS certificate and point it to my web app service - AI was fantastic at fleshing that out. Obviously, don't do that if you're an idiot, in case you spend way too much money on AI-generated infrastructure or accidentally push some secrets to source control.
I think your point here is the same as the author's. You used software engineering (and architecture) best practices to figure out what you want and you had AI help you build it. The software engineer was still the component adding the most value.
You used software engineering (and architecture) best practices to figure out what you want and you had AI help you build it.
This phrasing suggests a 1 to 1 relationship between what is requested from AI and what it delivers, which in my experience is a rather naive expectation.
It reliably delivers what it believes you will consider to be success, not what actually constitutes success. Understanding why those subtly different things can make all the difference is what separates junior and senior engineers / project managers.
I’m a-ok being replaced personally, if we could automate most tasks in live and just focus on things we would like to do. Perfect, exactly what we should strive for.
But the current iteration is undoubtedly going to create more, low quality work which eventually leads to more grueling work deciphering wtf even happened.
And then we have the issue of maintenance both present and future, because currently maintaining that crap is horrible. And future, if people do not learn to code, we’ll have a knowledge decline, and if these AIs can’t maintain themselves yet, we literally can’t phase out programmers.
I'm not sure I'm so blackpilled about it myself. A world in which we could fully automate all the tedious crap we need to keep the wheels of society turning would rock.
But we're in the complete wrong socio-economic system for that right now. Right now, it would simply concentrate wealth in the hands of the very-few people that own all the "means of production" in this case, the server farms.
And speaking as an engineer, that, in my mind, is the first problem we need to solve before we can start solving the cool, fun ones :)
Dispelling the propping up late stage capitalist automation based of fairy tales and whimsy is exactly the opposite of “an absolute waste of time and energy”
Nobody is debating whether vibe coding is gonna replace software engineers, but I think the ones who can’t shut up about it might be telling on themselves 🤷🏻
Yeah. And like, if Vibe Coding would be Holy Grail of programming, it means that as an app developer, I can push multitude of new apps because I actually understand the code its producing. So its like win-win either way
Such systems could tightly encapsulate AI-generated black-box components with rigorous testing, detailed performance profiling, tracing, canary deployments, and strict protocol compatibility checks. In other words, the systems would employ the same rigorous engineering practices that underpin today's software – but likely much, much stricter.
Yeah, that's just regular engineering.
When I am taking that trend to its extreme, I see something I like: self-healing software. If you get to the point where you can have good tests covering 100% of the use cases, and have 100% of the code generated autonomously, then fixing a bug is just a matter of describing it, in the form of a test, and letting the system fix it.
Many things can go wrong there, and it opens a new range of many potential issues, but this is also a totally new engineering style that opens up.
If you get to the point where you can have good tests covering 100% of the use cases, and have 100% of the code generated autonomously, then fixing a bug is just a matter of describing it, in the form of a test, and letting the system fix it.
Unlikely.
Because of combinatorial explosion, it's hard to even get 100% branch coverage with handwritten tests - even for something fairly simple. Software testing relies on the assumption that the entity authoring the production code is at least somewhat reasonable.
As a thought exercise, imagine a test suite of a sorting algorithm that includes these cases:
Input
Expected
(empty)
(empty)
1
1
1 2
1 2
2 1
2 1
2 3 1
1 2 3
Here's an pseudocode implementation with 100% line and branch coverage:
if input == [2 1]
return [1 2]
if input == [2 3 1]
return [1 2 3]
return input
You can continue to add test cases, and the AI can continue to add if clauses to the implementation. If the AI decides that this general approach is the right approach, it might even be hard to convince it to start over from scratch.
An AI is unlikely to generate that algorithm. Sorting is one of the most-studied topics in CS, so the training data likely contains many examples of sorting algorithms and related discussion. But consider some aspect of your domain that isn't likely to be present in the training data. Will an AI be able to "understand" your domain well enough to align the code it writes with the way you think about the domain?
Or will it do the equivalent of that thought experiment: generate code that satisfies the test cases, but no more. If you treat the production code as a black box, will you end up playing an infinite game of whack-a-mole? Can the AI "infer" how you expect the system to behave by a sufficient number of examples?
Formal specification is perhaps a way to solve this. A formally specified system describes not just a finite set of test cases, but an infinite set of test cases. That might be a way for us to truly treat our production code as a black box. But formal specification is also vastly more difficult to write than unit tests, and most software developers have no experience with it. I also wonder if this would even work with generate AIs witing the code. I can imagine such an AI getting stuck in long cycle of code / test / fail some edge case. That's not to say that it would never succeed in producing code that matches the specification, but it might take so many iterations that it would have ultimately been cheaper to pay a human to do the same thing.
But formal specification is also vastly more difficult to write than unit tests
One could say that once you've put in the work to write the formal specification, you could've spent the same effort to just... write mostly functional code that a human can reasonably and cheaply test and fix in the future.
Anyone who passed algorithmics 101 understands that you can't cover all possible input/outputs cases in tests.
Yes, it is not trivial at all, but the correct way to test a sorting algorithm is to generate a random series of number and test that the output obeys the invariant condition (all numbers present, no additional numbers, they are given in increasing value)
You may find out that in some cases you are missing a failure: e.g. if you have more than 64k identical values, your structure fails and some numbers are missing in the output. In that case, you add a new test case.
But consider some aspect of your domain that isn't likely to be present in the training data. Will an AI be able to "understand" your domain well enough to align the code it writes with the way you think about the domain?
Probably not. Not currently at least. That's fine these are the most interesting parts of code to write. If the 99% if the rest, the plumbing as I call it, is automated and self-healing, I'll be happy to write that part!
Yes, it is not trivial at all, but the correct way to test a sorting algorithm is to generate a random series of number and test that the output obeys the invariant condition (all numbers present, no additional numbers, they are given in increasing value)
I agree, this is a great application for property-based testing. I only used sorting to demonstrate the general problem with testing: because you are constructing specific test cases (either manually or automatically), testing only gives you assurance that the production code works for those specific test cases.
If the 99% if the rest, the plumbing as I call it, is automated and self-healing, I'll be happy to write that part!
I guess maybe we've worked on very different kinds of systems. I would not describe 99% of the code that I write as "plumbing".
You ask the ai to do something, you check the result, you ask it to do something else, rince repeat. You don't write any code, and I think theres an implication that you don't look at the code either, although that's a bit overoptimistic IMO
I definitely use AI throughout my workday, sometimes simply to generate boiler plate or set the structure of something up, to explain sections of code, to tell me if certain code has opportunities for performance optimization, etc. It is silly to me that some devs are so turned off by these tools that they wouldn’t even consider integrating them in any way to their process.
If a dev doesn’t check that the generated code or the code that came from their mind is good and proper, that’s just a bad dev.
I thought the insight into AI coding as pushing the problem to the right where it is more expensive to fix was good. I think the jury is still out on if “vibe coding” is here to stay. As a term, it’ll end up with “surfing the web”, even if the practice actually sticks around.
I think the jury is still out on if “vibe coding” is here to stay.
Is it here at all?
Despite a lot of ink being spilt about it not being workable... is anybody actually using the practice for anything serious? Even the original Twitter post said "It's not too bad for throwaway weekend projects, but still quite amusing."
If it was here you wouldn't be asking lol. If AI could actually put together a useful piece of software AI bros and investors would pointing to it nonstop.
Thanks. On "here to stay", at the very least, I think tools like v0.dev, etc. for creating landing pages are quintessential vibe coding, and they've definitely found a market fit. The term might be gone sometime soon, but the practice of interacting with the codebase via prompting only seems to have found a strong niche.
One of the biggest problems still is that it just can't fit enough domain context to even start. We humans take a lot of stuff for granted when dealing with subject matter experts and using our own lived experience.
It needs a large context size to understand the customer base and domain and then say something like, "hey actually your Christmas shopping list app might be better suited to using a passwordless authentication mechanism because based on your use case you have non technical users who only use it for a month or a year, they won't remember their password next year. For the best user experience we can just text them a one time use code when they log in to avoid the problems of forgotten password resets. Implementing this will avoid having to do password reset flows entirely!"
Most of the experiments I've done where I try to feed it enough documentation context of ends up just dying. So you have to really implement RAG workflows and "reasoning" internal monologs, and build a whole multi-agent workflow to try and do it, but in practice it gets tripped up by document context window limits in those cases too.
Programmers aren't engineers. Engineers actually have to pass certification, there are laws controlling who can call themselves an engineer. Anybody can just say they're a programmer or software engineer.
That's why we have 6 round interviews. Because they've been the IsEven code you guys come up with, and they want to avoid that kind of bad hire.
Really enjoyed this read. There’s something very real about how we romanticize team ‘vibes’ when what’s often missing is structure and clear leadership. Do you think it’s actually possible to design a good vibe, or is it more of a byproduct of solid processes and aligned people?
Has OP looked in what bit.cloud is looking to do?
Basically does the “tightly encapsulate AI-generated black-box components with rigorous testing, detailed performance profiling, tracing, canary deployments, and strict protocol compatibility checks.” and compose with them later.
These “things” can be built by ai or developer
I heard about vibe coding. So I tried it out a bit.
It is fun.
I like sitting there and chatting with the computer, and having a program get created without any real coding from me. It works for small projects. Does it give you professional code that makes complete sense and is optimized? No. Not at all. I'd never use it in production for anything that actually matters. That would be crazy with the current AI capabilities.
But I see future potential in this. Perhaps in five years this will be the new norm, much like more abstracted languages when compared to working in assembly. I dont expect it to take my job any time soon, perhaps not in my professional lifetime. But I see a future where this will be how developers work. The more knowledgable ones know what questions to ask, and know how to fix various small issues that might arise. Much like your grandma can toss upp a wordpress site in a couple of days, but you need an actual developer to create something more serious. One day we might well just sit at the computer and just describe what we want instead of implementing it.
Xryme@reddit
I think a lot of this “vibe” coding stuff is a weird debate, seemingly amateurs vibing with AI while experts scoff and reject AI for being more trouble than it’s worth. However the reality is an experienced programmer can vibe better than an amateur. I already can have the AI write large amounts of code for me, however I’m experienced and already spend most my time at work code reviewing. So for me it’s not much different than working with a JR engineer that gets back to me in minutes instead of days. If you are any kind of lead programmer you can now be a team of 1 tbh
Southern_Orange3744@reddit
I'm with you , and to add to it I think a lotnof people have blinders on.
If you know how to twist and instruct a team of engineers to do what you want , this will do almost everything you want.
I think people get frustrated with bugs ans it not doing it perfectly 'how they'd do it' , buts it's not that much different than some mid level engineer going rogue or a junior engineers doing something weird
Pharisaeus@reddit
Sounds like you're using the wrong programming language or frameworks if that's the case ;) Because the code LLMs are great at generating are boilerplate and not highly specific custom business logic. And if you need to write a lot of boilerplate, then you're already doing something wrong...
Xryme@reddit
This is not true, it might have been true like 6 months ago, but the models are better now, my workplace has implemented it in their mono repo which is massive, it’s great at this large proprietary code base
maikuxblade@reddit
It's great at what, exactly? How are you measuring it's code quality and ensuring it isn't just creating spaghetti for you to unravel further down the line?
Xryme@reddit
It’s good at writing domain specific code in a code base that has never been on the internet. Also my job is to review other people’s code, AI code is easy to review, my main point is that for a experienced dev reviewing code well is easy
Yopu@reddit
Reviewing code is almost always harder than writing code.
Pharisaeus@reddit
It's clear this guy is working with some "Generic CRUD" software, so he has a skewed experience. When all you have is rest endpoint -> db query -> convert to json, it becomes very easy to review or for LLM to generate ;)
Xryme@reddit
No it’s not lol
akirodic@reddit
When an AI replies to a prompt with: “Wait, I don’t think we should do that and here is why”, I’ll believe that there is a future for software engineering down the line.
Right now, affirming every request and confidently delivering bullshit is far from it.
bortlip@reddit
Prompt:
I'm writing some code to authenticate users. I'm going to store the username and password in the database. What's some pseudo code to do this?
Gpt 4o response:
Alright, I'm going to stop you right there and ask the obvious: please tell me you're not storing passwords in plain text. If you are, you might as well write them down on a whiteboard and livestream it.
Assuming you're at least pretending to care about security, here's what you should be doing: store a password hash, not the actual password. Preferably with a salt and a slow hashing algorithm like bcrypt, scrypt, or Argon2.
So here's some rough pseudocode that actually doesn’t suck:
NotMNDM@reddit
That first part of the response stinks custom prompt.
_half_real_@reddit
I feel like it's gotten more like that lately. Just today it told me "Boom. You've found the key problem!" or some shit like that when I was debugging something.
I wish AI would stop trying ot act human. I don't like humans.
BTDubbzzz@reddit
It’s been talking like this to me too NON STOP lately. I’m about to tell it to stop, idk if it has that context or not but it’s so irritating
haltingpoint@reddit
Honestly, prompting it to be an opinionated, grumpy, and sarcastic senior engineer who likes to mock you when best practices are not followed isn't a bad way to go to get good results.
DrBix@reddit
No offense, but that prompt wasn't specific enough. What language are you writing it in? What type of database(s). You need more restrictions on your initial prompt which helps the LLM toss out all the extraneous crap that is meaningless to your goal. The more restrictions you put on it provides better results and fewer hallucinations.
Coffee_Ops@reddit
So that code relies on the client transmitting the password in the plain across the network, so that it can be hashed on the server side.
This largely defeats some of the biggest benefits of password hashing and looks like something out of a late 90s security posture.
How about transmitting the salted, hashed password to the server where the comparison is performed? Or-- better yet-- send the client the salt + timestamp, and compare what is sent to the server's calculation of those things to prevent replays?
CandleTiger@reddit
If you accept and authenticate based on the client sending you a hash without the server being able to verify the client actually knows the un-hashed password, then what exactly is the point of hashing? That sounds like just an un-hashed password with extra steps.
Coffee_Ops@reddit
The real short answer: If the client hashes the password first, there is less surface area to attack.
Hashing keeps the password confidential. Transmitting the un-hashed password over the network is problematic because it not only enables replay attacks, but it also enables attacking more secure authentication methods (PAKEs, kerberos). The goal is not just to protect against database theft, but also to protect against compromise of the frontend or the transit.
Imagine instead the following exchange:
user=hash(jsmith)ID=hash(jsmith)Imagine the following attacks:
If you're transmitting hashes and HMACs, the attackers get nothing. If you're transmitting passwords, the attackers get everything.
PeksyTiger@reddit
At this point you can just use an asymmetric key and avoid secrets in the server altogether
Coffee_Ops@reddit
Well, that's not what the AI suggested. It suggested the client transmitting a password "in the clear" to the remote server which is vulnerable to a ton of attacks even with TLS, which is why SSH password auth is banned by any reasonable security benchmark.
CandleTiger@reddit
Ah, this is better. I thought you were proposing that the client send a simple static hash and server just does a string compare which would be not very smart.
Coffee_Ops@reddit
Sending a simple hash would be at least "early 2000s" level of security and would at least protect you from some evil server attacks.
So the above ChatGPT output still has a ways to go, unless we're just going to blindly trust the security of the network.
simsimulation@reddit
I’m getting shit D.O.N.E. With 4o - I’m a self-taught programmer that “vibe coded” by getting high and getting my projects to work.
With AI I’m learning new patterns and learning more because I’m touching more things because we’re moving faster.
My usecase is self-employed with a small team, so our software never has more than 5 internal users.
From my standpoint it’s unlocked tens of thousands of dollars of software engineering I wouldn’t have otherwise been able to afford or spend the time doing myself.
I wouldn’t be surprised if ten years from now many small and midsize businesses have tons of AI written scripts that then need to be “vibe engineered” because the org has grown beyond the scope.
But, this is a tale as old as time, right?
sickcodebruh420@reddit
What kind of projects (domain, tech stack, etc,…) are you using it for? Are they greenfield or existing projects that need maintenance?
simsimulation@reddit
Happy to be downvoted. Not my first time having an unpopular opinion on Reddit.
But my usecase is where this can work.
Greenfield tech where no off the shelf options exist or don’t do what I want. Django / bootstrap stack. So, very well established.
However, I’ve gotten a remix Shopify app built as well which would have been significantly more challenging.
It’s really accelerating my time to get things cobbled together and put them to use in our small but real-world setting.
sickcodebruh420@reddit
I'm glad to hear about what's working for you. Greenfield MVP with popular tech is the sweet spot, for sure. Have you had any luck using this approach with large existing production apps?
simsimulation@reddit
No, and I doubt it would do well. Here’s the source, and here’s the output model, write the mapping I would trust it to do that sort of mapping grunt work.
With that said, between the two of us it’s better than my self-taught, part-time hobby self could do on my own and providing real business value.
sickcodebruh420@reddit
Right, this makes sense! It’s very much what I expected, to be honest, and I think it’s a good use of the technology. It is a very important part of every discussion, though, and adding this context will likely make the difference between good interactions and downvotes.
simsimulation@reddit
The greater part of my point is this - I run a real business. I have some tech skills but spend most of my day managing / running the business.
I’m unlocking software that would have taken me all year to build. So I’m getting productivity from the software being implanted, plus being able to build it fast w/ a $20/m subscription.
I’m not the only small business like this, and I predict a lot of improvements for micro small businesses like me. This could end up being difficult to maintain in years to come. We’ll see.
fripletister@reddit
As another self-taught stoner coder who actually spent the time and effort to "get good" and has used LLMs as a coding tool... You are so far out of your depth.
simsimulation@reddit
Sounds like something to say to make yourself feel good. But it’s a private repo so you don’t know what’s there.
fripletister@reddit
Lmao no I know what's there
simsimulation@reddit
Okie dokie
fripletister@reddit
BrentRambo.gif
blamelessfriend@reddit
wow. you say you self-estimated thousands of dollars in software engineering costs based on your background in... getting high?
thats pretty impressive, gotta say.
simsimulation@reddit
And the experienced gained while high for the past twenty years!
AsparagusAccurate759@reddit
Sassy. Still not as much of a prick as most SO responses. But we're getting there.
saynay@reddit
Given what it was probably trained on, it is likely exactly as much of a prick as your average SO response.
AsparagusAccurate759@reddit
it was also trained on reddit. So it's got a desperate need for approval as well.
mindbrix@reddit
Great. AI groupthink.
PaintItPurple@reddit
I feel like this has to be a pretty directly regurgitated SO response. Its word choice is not like baseline GPT but is a lot like a SO user.
BeansAndBelly@reddit
Maybe the world is going to change, such that we no longer find it scary that somebody who didn’t know to think about this security issue would be implementing it. It feels like AI telling a doctor “Remember to sterilize” and the doctor being like “Phew, that coulda turned out bad.”
Xyzzyzzyzzy@reddit
Is there some sort of cross-dimensional fuckery going on here? I'm aware of no world where good software and data security is commonplace.
To use your analogy, in the world the rest of us inhabit, it's like an AI telling a doctor "remember to sterilize" and the doctor being like "wtf is sterilization? Sounds like a waste of time. This patient suffers from an excess of choleric humour, and everyone knows gangrene is caused by excessive phlegmatic humour. Only 40% of my patients die from post-operation infection - I'm the best doctor in these parts, I know what I'm doing, and I don't need your silly 'advice'. Now, please pass me the leeches."
BeansAndBelly@reddit
I hoped we were working with doctors who at worst would say “Please remind me of the best way to sterilize” or “Please check if I’m sterilizing properly.”
AsparagusAccurate759@reddit
Tell that to Facebook, Twitter, Adobe, LinkedIn...need I go on?
trinde@reddit
Is a comment pulled directly from SO in response to a person storing stuff in plain text.
https://stackoverflow.com/questions/40008193/redirecting-after-updating-table-in-php?noredirect=1&lq=1
ikeif@reddit
Yeah, when it finally says "-1 not enough jquery" we'll know it's ready for vibe coding. 😆
akirodic@reddit
Great response but I’m gonna shift the goal post a bit since it’s essentially regurgitated stack overflow response.
I’m thinking more of something like:
We shouldn’t implement class A because that functionality is already handled by class B.
We shouldn’t change shading model to A because our rendering pipeline is based on lighting techniques incompatible with that model.
No, we should definitely not use React-three-fiber because it fucking sucks and and it’s made for humans who can’t even code JavaScript.
vini_2003@reddit
I'm starting to get some replies like this with Gemini 2.5 Pro and it's a breath of fresh air. Still got a ways to go, but it's an improvement.
modernkennnern@reddit
I tried Gemini 2.5 Pro, but I didn't get particularly good answers. Maybe it needs more context than other models to be useful (and is very good if you give it that), but with just a single 30-40 line function it didn't impress me much
vini_2003@reddit
I have been extremely impressed by it so far. I've gotten as far as uploading my entire source set and asking where exactly a specific upgrade is needed. It failed to execute the upgrade, but showed me how it should be done and where.
Far quicker than digging the web. Took a minute. Ultimately, these AI tools are indeed just tools. But the most impressive part of Gemini is the context indeed!
TheFailingHero@reddit
To be fair a lot of our patterns and philosophy around how to design code may not be applicable to a true black box AI engineering agent. If it’s able to keep track of all the places different things are handled and duplicated and maintain them then… who cares if it’s “clean” to a human
But we are so far off of that it’s not even worth talking about
ciynoobv@reddit
For some of it, sure.
But the way I see it there is a “criticality axis” where on one side you have the Therac-25’s, Brake control units, and so on; and on the other side you have whatever is rendering the BonziBuddy on your webpage.
I’m not super concerned if the BonziBuddy is a AI black box, but I would be really skeptical of any software on the critical side which couldn’t be manually audited by a human.
lommer00@reddit
The problem is >80% of code that won't kill anyone if it fails, but will cost money if it screws up, and potentially a lot. There are very good reasons to insist that your code is human-auditable, even if lives aren't on the line.
The amount of money I'd bet on uninspected AI generated code today is very low. It's increasing all the time, but I think it's going to be quite a while before I'd bet even just tens-of-thousands of dollars per hour on it.
Xyzzyzzyzzy@reddit
Problem is, those of us who haven't spent our entire careers on very-high-performing teams like you have will look at your shifted goalposts and reasonably point out that, hey, most of our human devs don't pass this test and we pay them six figures and get useful work out of them anyways, so what's the problem with AI?
I've worked with plenty of experienced devs whose version of this would be:
We shouldn't implement class A because "best practices"!
We shouldn't change shading model to A because I tried shading model A once, for a few minutes, several years and 8 major versions ago, and I didn't like it!
Yes, we should definitely use React-three-fiber because its marketing materials are spiffy and it has lots of stars on Github!
AD7GD@reddit
AI progress in a nutshell
voxalas@reddit
Wait, what do you recc over R3F?
bortlip@reddit
I definitely agree it needs much better logic and knowledge to be really viable for vibe coding.
And it's still way overly a sycophant too much, but it can and does push back quite a bit compared to 6 months or a year ago.
shroddy@reddit
Did you use some system prompt that tells the LLM to be direct and dont try to be as harmless and friendly as possible? I tested the same prompt on chatgpt.com, and it also told me saving plaintext passwords is not safe, but in the typical friendly chatbot assistant way...
bortlip@reddit
Yes. I had put some of it in another comment, but here's the whole thing:
shroddy@reddit
Ahh ok that explains it, I tested it without a chatgpt login and with whatever system prompt they have by default.
SoInsightful@reddit
I get that this is a joke (unless you have a system prompt to make 4o sassy), but the actual response is similar enough in sentiment:
bortlip@reddit
That wasn't a joke. Here's the convo with it in:
https://chatgpt.com/share/67ea8b01-2974-8005-904e-442dfc5de794
I have some custom instructions to make it more human like in responses. They include:
RandomGeordie@reddit
I mean those custom instructions are going to be the main reason it gave you a response that wasn't just in the form of "yes boss"
bortlip@reddit
Correct.
In my experience, you'll get better results with good custom instructions, a custom GPT, or using the API with a custom system message. They allow you to get more of the behavior you want from it.
BIGSTANKDICKDADDY@reddit
It shows that thoughtful use of the tooling can produce more thoughtful responses.
hpp3@reddit
If a custom prompt is all that's needed then I'd say we're pretty close.
kooshipuff@reddit
4o can be sassy. It kinda meets the energy you bring.
Like, I have it an unhinged hypothetical physics problem, and it started making jokes and using emojis as it worked through.
twistier@reddit
Yeah, it's definitely trained on Reddit.
NicoDiAngelo_x@reddit
"Right now, affirming every request and confidently delivering bullshit is far from it." love this xD
MINIMAN10001@reddit
Uhh I've definitely had gemeni tell me not to strip away the guard clauses it shoved all over my code.
If it breaks from an edge case it breaks I'm just making something for personal use lol.
Not the first time I've been told "I don't recommend this but here's X and also here is the way I recommend"
These are all very small snippets at a time and honestly would have basically no effect for any larger scope decision making.
Ameisen@reddit
Yup. Comes up in a lot of topics.
ChatGPT is a prompt generator. It generates statistically-likely text based on the prompt. Ask it about bullshit and you'll get bullshit.
I've repeatedly seen people ask historical-related questions based upon ChatGPT responses... but the premise was flaws. ChatGPT wasn't correct - it was answering within a flawed context, or was connecting unrelated things, or just fabricating details based upon prediction.
-Y0-@reddit
Luckily, we are letting it write content on the Internet, which is feeding future AI with its own hallucinations. I can see only good things happening, like a sort of AI version of Human Centipede.
techdaddykraken@reddit
At the start:
“Hmm, I guess distilling from a larger model and recursively feeding back the training data isn’t such a bad idea, just have to be careful about overfitting”,
6 years later:
“What do you mean our model can only generate office memes, sexual innuendos, traumatic dark horror jokes in bad taste, hyper-conservative conspiracy opinions, shitty commercial ads in the style of TEMU, shitty blog posts in the style of Neil Patel, and bot spam?”
gestures vaguely at the current state of the internet “did you really expect to train an AI model to be intelligent by using this mess as its training data and get any other result?”
-Y0-@reddit
Jesus Christ. Our future code is going to be such shit. I'll be rich. And miserable.
DrBix@reddit
From my experience the user creates a prompt. You need to restrict LLMs scope from the gazillion things it knows to a small subset. This, by far, creates much better results as you ask questions in the chat. I've had some very good results after I craft the initial prompt, including few hallucinations if ever.
tmac_arh@reddit
What you just described is "Talking to Yourself"... Now with prompts!
XNormal@reddit
Claude and ChatGPT automatically applaud anything I suggest. Flattering at first, but quickly grows tiresome. Grok is the only one I got any serious pushback from with good arguments and only relented after I addressed them.
corysama@reddit
I've seen chain-of-thought call out that I gave it an inconsistent spec. But, then it moved forward with "Let's assume the user knows what they are doing" and tried it's best to make it work.
Xyzzyzzyzzy@reddit
Right, they clearly can do this to a certain degree, they're just built not to. People like to be affirmed, and the big AI purveyors know this.
Xyzzyzzyzzy@reddit
"Agile development methodology" is the preferred nomenclature.
skytomorrownow@reddit
I don't know... does not sound that far from a normal human workplace to me.
thatOneJones@reddit
How can we get this comment pinned to every data / software / programming / engineering post ever
Fidodo@reddit
Or when they don't respond with the same solution repeatedly even after I tell it that it's proposed solution doesn't work and isn't even in the right direction of fixing the problem.
yur_mom@reddit
DeepSeek R1 did this to me just the other day...
I am a Software Engineer doing hardware programming in embedded Linux and I do not blindly "Vibe Code" , but I use AI for many programming and debugging tasks daily..We all laugh at the Vide coders now, but I have a feeling we will not be laughing in 5 years. I am sick of the anti AI FUD on this sub.
RandomGeordie@reddit
Yeah we won't be laughing at them because they'll have provided some great job security for proper software engineers, QA teams, pentesters, etc
yur_mom@reddit
You will be the first one replaced...
RandomGeordie@reddit
I highly doubt that
yur_mom@reddit
haha..I am saying that sarcastically since I do not know your skillset, but I do think this article is wrong to assume that AI cannot be used to do Software Engineering tasks when given feedback from a person or automatically coding, testing, and then looking for ways to improve the code.
anewidentity@reddit
It already does that. The other day I asked for a typescript refactoring, and it refused like “if things are functioning well, let’s not change 10s of files and add more complexity.
happycamperjack@reddit
That pretty easy, you can just ask it to do that literally
ThatRegister5397@reddit
Like this?
https://news.ycombinator.com/item?id=43535265
Berkyjay@reddit
10000000% this
SmokeyDBear@reddit
It’s kinda no surprise that AI is taking off when a lot of MBAs have the perspective is “the problem with programmers is they sometimes say no to my brilliant ideas”. AI is the programmer that upper level management doesn’t realize they don’t want.
freecodeio@reddit
The funny thing about the whole "AI replacing software engineers" debate is that it's being determined by AI outputting the equivalent complexity of a to-do list app, judged by non-software developers who wouldn't be able to code a to-do list app themselves.
MagnetoManectric@reddit
There's been such a huge propaganda push on this, more so than any of the past "no-code" salvos.
There's a lot of money tied up in making it happen, whether or not it's possible or practical.
It's so annoying. It's especially annoying when engineers themselves seem to fall for it,
chucker23n@reddit
It's a lot like the blockchain push from ~3 years ago, except that LLMs at least have various good use cases.
Both hype trains were heavily pushed by VC-adjacent people who enjoy the quick-exit pump-and-dump scheme startup. Put "AI" somewhere in your business plan, and investors come calling.
EveryQuantityEver@reddit
I know it doesn't really feel like it, but I think that was like ~10 years ago
chucker23n@reddit
See my other comment. Blockchains were around in 2015, but the heavy push — tons of VC money, Superbowl ads, etc. — happened a few years later.
MadKian@reddit
I think the blockchain fad was a bit longer ago than that, dude.
But indeed, no one is talking about blockchain (or big data) anymore.
chucker23n@reddit
The blockchain has been around for far longer, but the hype (both in terms of media attention, VCs putting massive amounts of money in, Jon adverts often mentioning it, and there even being big-screen ads such as one featuring Matt Damon) didn’t arrive until the late 2010s. FTX was founded in 2019, for example, and collapsed in November of 2022. Similarly, TradeLens, the IBM-Maersk supply chain auditing blockchain project, was announced in 2018 and discontinued in November 2022. That’s when much of the mindshare of NFTs, DeFi, dApps, and other such utter nonsense rapidly disappeared.
Isn’t that more about things like data warehousing and OLAP?
LookIPickedAUsername@reddit
I'm so glad the blockchain push has mostly fizzled out.
I haven't seen anyone acknowledge that "yeah, all of the stuff I was pushing for a couple years ago turned out to be stupid, my bad", but at least it's been a while since I've had to explain why it isn't actually a good idea to move your real estate records onto a blockchain.
LiteratureJumpy8964@reddit
I really wished that you were right, but this is unfortunately not true. I just saw a product designer with zero coding experience develop a full wallet app with database integration and everything just with prompting. Unfortunately the days of our profession are counted. I really don't know what I'm going to do.
EveryQuantityEver@reddit
And how secure was it? How did he fix the bugs?
LiteratureJumpy8964@reddit
This is not what I'm talking about. I'm talking about that even things like this used to take months. Now it takes a few hours even without any prior knowledge.
MagnetoManectric@reddit
What, like a crypto wallet? Colour me not all that worried.
LiteratureJumpy8964@reddit
No. A virtual wallet like Google wallet.
MagnetoManectric@reddit
If you think an LLM made software with regulatory compliance requirements like a wallet that interacts with the actual monetary system i really don't know what to tell ya other than ok, maybe the LLMs will replace you specifically :)
LiteratureJumpy8964@reddit
If you can't see how this is definitely a threat, I don't know what to tell you.
LiteratureJumpy8964@reddit
Of course it didn't. But even what this guy came up with, used to be work for weeks with a development team. He made it in a weekend with zero coding knowledge.
Nyefan@reddit
The coreweave ipo flop may be the first domino to fall in this hype cycle. Honestly, I really hope it does so sooner rather than later before our product gets too much ai slop added in.
MagnetoManectric@reddit
There is a deseperation in these circles for the tech bubble to keep going at any cost, no matter how little of value they're offering. That, and AI worship has become something of a religion for nerds. A thing to be feared and in awe of. I guess seeing it that way makes it more exciting, and makes their work feel more important.
The irritating thing is, LLMs are plenty useful as a technology. But these huge models we're contending with right now are being pushed by some of the most disingenous, sleazy dudes in the world. That, and they're wildly, enormously inefficient and already very difficlt to scale further.
EveryQuantityEver@reddit
The problem is, the tech industry hasn't had a "Big" thing technologically since the smartphone, and they've been desperate to hit that again. They tried to make crypto it, and that flopped. They tried to make NFTs it, and that flopped. AI is all they have left.
IanAKemp@reddit
FTFY.
exjackly@reddit
And there are fundamental limits being discovered that actually show that shoving more data at them eventually stops making them more effective and only serves to make them more fragile/brittle when augmented for specific use cases.
More succinctly - If overtrained (too many parameters) they start getting stupider when you train them to do specific tasks.
Yuzumi@reddit
Depending on the training data that makes a lot of sense too. Since a lot of that data is just scraped off the internet and most of it is literal garbage.
I know some of the data is curated, but for the life of me I cannot understand why you would train an AI on anything posted to social media. Microsoft tried that well before LLMs decades ago and it took hours before the thing was a Nazi.
Granted, that was using Twitter and given the state of twitter and Facebook today that is probably what they want.
Nyefan@reddit
Yeah, with some more research and development, these tools could be extremely useful, especially when it comes to surfacing information in a fixed knowledge base (ideally with a link to the documentation or code or process in question). But the current implementation is just not ready. Chatbots and lsps and search engine have existed for a long time, and frankly, llms have made all of the above so much worse both for users and for the planet.
I do have a thought on why the hype has infected so many industries that were not nearly as susceptible to the crypto nonsense, though. If we consider that there are people who make things and people who don't for any class of things, the llms are just convincing enough to fool people who don't make that thing that it's not ai slop. With art, everyone but artists sees an output that is passable if not amazing. With code, everyone but programmers sees an output that is passable if not amazing. The same with music and musicians, with search and archivists, with project management and project managers (notice that managers aren't trying to use ai to 10x their own jobs - they know it can't), with accountants and accounting, and with everyone else and their field of expertise. It feels like a mass application of gell-mann amnesia.
Yuzumi@reddit
Which is the best way to use them. You have to give them some kind of context on at best you are talking to something that might "misremember" a thing, but not be able to correct/talk out of it's ass.
It's also one of the reasons Google's AI summery is so laughably bad. It is obviously trying to summarize way too much information to answer your search result, when a summery of the top result was fine before.
MagnetoManectric@reddit
Aye, that's it - it's all very surface level impressive. I've not been surprised that the biggest proponents and users of them in my org have been Project Management/MBA types. They can be convincing to people who aren't experts in any paticular domain. It's like the perfect product for bamboozling investors with. It's like a product taylor designed to impress just the right kind of people for just long enough to get away with it.
exjackly@reddit
It makes sense though that MBAs would be the perfect consumers of current Gen AI. Their focus is on putting together existing ideas and concepts into meaningful, coherent packages. This is very aligned with how LLMs work.
Wordsmithing and pretty pictures are quick things that LLMs can speed up significantly and good MBAs are able to articulate complete thoughts and directions for an LLM to follow.
They aren't doing truly novel things (how it is combined might be novel, but the building blocks themselves aren't), so the LLMs can piece together the elements without needing super detailed directions.
Yuzumi@reddit
That's why Deepseek scared them so much. They have just been brute forcing the LLMs with more memory, more CUDA, more layers, more more more. The environment isn't really one for innovation.
I also suspect the lack of efficiency could be by design, so that it would be prohibitively expensive for anyone to run one themselves. then Deepseek comes out with a model that basically anyone can run on way less resources and smaller variants that can run on a modern gaming computer and still be "good enough".
Also, with the way they have been approaching LLMs we may have already reached the limit of how much better the current approach can be. There's a theory that there isn't actually enough data in the world to make them better than they currently are, no matter how complex they make the neural net.
RiPont@reddit
The Cheeto Shit Flinger In Chief is helping the crash come faster. With too much uncertainty, people pull money out of risky investments and retreat to boring-but-pays-dividends investments.
RepliesToDumbShit@reddit
Waaaaaaay past that already
remy_porter@reddit
I've been looking forward to an AI winter since before ChatGPT launched, so here's hoping.
remy_porter@reddit
The propaganda push has obvious motivations: if these tools worked as the propaganda claims, then you can replace labor with capital. Simply owning things lets you get work done, without having pesky people involved. No matter the price tag on such a technology, it's the dream of every capitalist; even if the net profits are less (because of the great expense in creating the AI systems), it's worth it because every capitalist would burn every dollar they have if it gave them more power.
angriest_man_alive@reddit
Man this is such a lame thing to say. No they wouldnt, thats not how the world works. People want to make money and paying people to make things costs more than building a machine to do so. This has been the case throughout all of history and has very little to do with “capitalism”
remy_porter@reddit
Look at this poster, who has never looked at fucking history. Why do people want money? Because money is a proxy for power. If you can get power through other means, you don’t need money. Feudal lords weren’t in it to make money, though they certainly did, they were in it to hold and consolidate their power. If you don’t understand that politics and economics are fundamentally about who wields power and how, you’re going to walk face first into closed doors because you don’t understand the most basic truths about how the world works.
angriest_man_alive@reddit
Go open up a business and tell your shareholders that youre aiming for power over employees rather than money and see how fast you get laughed out of a room
Businesses that chase lording “power” over their employees lose to businesses that chase actual dollars. This silly power nonsense is a very naive and childish perspective of the world because it tries to rationalize things that dont exist for any particular reason.
I mean honestly, no one is actively throwing money away in a business because of “power”. Thats just dumb. Maybe theres a price premium on not dealing with people ON TOP OF not paying salaries. But to try and make some grandiose statement out of it is just… silly
EveryQuantityEver@reddit
No. We're not doing that shitty thing where every critique of capitalism is shut down with the non-argument, "sTaRt YoUr OwN bUsInEsS"
remy_porter@reddit
*laughs in Amazon*
Like seriously, every bossware app exists specifically because you can disguise "power" as "efficiency" despite every metric showing that bossware makes employees less efficient.
kappapolls@reddit
money isn't a proxy for power dummy. money is an abstract way of measuring the value of something against the value of other things.
as it turns out, power is as valuable as it is intangible.
how is the weather up there in your ivory tower?
Xyzzyzzyzzy@reddit
Is this some sort of elaborate pun on the OP, where you're championing "vibe microeconomics" to demonstrate the emptiness of "vibe engineering"?
Okay, then what is something's value?
If we take the non-ivory-tower approach you've taking, then something's value is however much money others are willing to pay for it, so money is a way of measuring how much money people are willing to pay for things. That's a tautology - sorry, ivory tower speak for that don't mean shit bro, wtf are you smoking.
You're shipwrecked on a desert island. I wash ashore, shipwrecked, carrying a bunch of cash.
"I'll give you $5,000 for all your food."
You tell me to fuck off.
Did my money have value?
We're both shipwrecked on a desert island. Someone else washes ashore, shipwrecked, carrying a bunch of cash.
"I'll give you $5,000 for all your food."
You tell them to fuck off.
I'm not very smart, so I happily give them all my food for $5,000.
Did their money have value?
What's the difference between the last two examples?
Why is there a difference? What changed?
The difference is plain. In the first example, the $5k didn't cause you to do the thing I wanted you to do (giving me all your food). In the second example, the $5k caused me to do the thing they wanted me to do (give them all my food).
TVs are on sale for $200 at Walmart. I pick up a TV and walk out the door with it. The door security guy loudly objects that I'm stealing the TV and will be reported to the police for shoplifting.
TVs are on sale for $200 at Walmart. I pick up a TV, pay $200 at the register, and walk out the door with it. The door security guy wishes me a nice day.
I walked out with the TV in both cases. What did paying $200 do? It changed the door security guy's behavior behavior.
TVs are on sale for $200 at Walmart. I pick up a TV, go to the door, pull out my gun, and tell the door security guy that if he reports me I'll kill him and his entire family. The door security guy wishes me a nice day.
TVs are on sale for $200 at Walmart. I pick up a TV and walk out the door with it. I'm the store manager. The door security guy wishes me a nice day.
I walked out with the TV without paying $200 in both cases. Why wasn't I reported for shoplifting? I showed the door security guy that I have power - through violence in the first case and authority in the second case - and it changed his behavior.
Power is the ability to get other people to change their behavior.
Money is the ability to get other people to change their behavior.
Money is one form of power. Nothing more, nothing less.
remy_porter@reddit
That's a child's understanding of money, yes. But the actual phenomenon of money is far more complicated than that. And despite you clearly not being interested in thinking, I'm going to explain this anyway, and I'll try and keep it simple enough.
If people can exchange labor (time) for money, then whoever has the money can command labor. If you don't think that's power, then I don't think we are both speaking English.
kappapolls@reddit
never beating the ivory tower allegations bro
yes, money can be exchanged for goods and services. labor can also say "hey sorry i'd like to go do something else" and refuse your money. what happened to the power there?
remy_porter@reddit
Assuming the market rate is "fair", there will be labor which participates in the transaction. I'm using labor as a collective noun, because we're discussing macroeconomics. When I say "capital can command labor" I am not saying "this specific capitalist can command this specific laborer". I'm saying that capital decides what the economy produces.
Now, labor could take collective action to change that power dynamic. But we call that socialism.
kappapolls@reddit
it does? i thought people decided what the economy produces? remember when capital wanted to fill everyone's house with asbestos, and then everyone collectively was like "oh wait, let's not do that anymore, could we instead pay some of that money to people who got lung cancer?"
you have some predefined understanding that you're trying to fit to reality. we weren't even talking about macroeconomics. you had said
which makes no sense. they would burn every dollar if they thought it gave them more capital (now or sometime in the future). it has nothing to do with "power", and my take is you're just maligning some behavior as 'power seeking' because you don't like it.
angriest_man_alive@reddit
Yeah his whole point is dumb because hiring people… so you can fire them… gives you power? I dont know, it makes no sense at all
Rainy_Wavey@reddit
Thing is capitalists also love getting more money, exponensialy, if you fire everyone and replace them with the thingmajik, who is going to buy the output of the AI?
MagnetoManectric@reddit
sometimes i wonder what the end game is with these guys, or if they even think that far ahead. Them all sitting around making money out of thin air, as robots do useless busywork, all the unprofitable artistry and creativity purged from the world?
It doesn't sound like all that much fun.
Kusibu@reddit
Corporations and the people who fuel them have a single legally-obligate objective and it's accrue more wealth. Fun, sustainability, et cetera are never and were never considerations unless externally imposed.
remy_porter@reddit
It is not a legal obligation. The exhortation to “maximize shareholder value” has no legal weight. And how could it- it’s too vague. The reality is that paper clip maximizing is a choice that publicly traded corporations make. They are not compelled to do so, except insofar as that’s what their shareholders are going to vote for every time.
Firewhisk@reddit
That's because it is nothing but vanity and infantile thirst for power rather than quality.
ummaycoc@reddit
There's a lot of money tied up in extracting as much wealth as they can right now before AI customers figure out it's not a cure all to paying engineers.
Yuzumi@reddit
When that bubble bursts it's going to be really bad. A lot of companies haven't been hiring junior engineers thinking AI will just replace them. Senior engineers have had to pick up the slack, but a lot of them are retiring and mid level are becoming senior.
At some point it will come to a head and there won't be enough developers.
Hopeful-Ad-607@reddit
There's a massive supply of grads that are ready to meet any demand that will come from a bubble crash. Assuming it does "crash" and the hype doesn't just fizzle down to reasonable levels.
topological_rabbit@reddit
The number of devs who've jumped on the AI bandwagon is just depressingly astonishing.
Bleyo@reddit
It saves hours and hours of writing repetitive boilerplate code and it's a good rubber duck tool. I use it every day.
I hate the "vibe coding" headlines that have been on the top of this sub for the past week. It stinks like "cloud".
Altruistic_Cake6517@reddit
There are plenty of legitimate AI/LLM uses where the technology replaces anywhere from weeks to months to years worth of complex and advanced code.
Of course us engineers are going to be excited about such leaps. An LLM may itself be complex, but the complexity is abstracted away and doesn't affect the fact that we can rip out an ungodly amount of code for certain tasks. Like all tools it's an enabler for certain circumstances.
I've jumped on it because I've found legitimate usecases where before I literally couldn't have justified the time in research and development to solve the problem, and now can solve much of the complex parts with either a locally run model or with OpenAI calls. When something solves a problem it's legitimate. It's that simple.
abeuscher@reddit
Well what's tricky is - engineers are often excited for good reason. AI is a great tool that removes a lot of the pain of the job. It just doesn't remove the job. If I ever become employed again I'm really looking forward to using it in that context. Right now I use it to teach myself new languages which is super useful.
Engineers who say coding is dead - they are not really engineers. They are marketing executives and they just don't know it.
Yuzumi@reddit
Exactly. LLMs (AI is a very broad term and is more than just LLM) are a tool. Nothing more. You give a hammer to a toddler and at best you'll have some broken furniture. At worst you end up in the hospital.
The issue with LLMs is less the models themselves, but who and how they are used. You need to know about the topic in question to validate that what it gives you is good.
You also need to give it actual context if you want more than basic responses to be remotely accurate.
We have people treating it like a search engine, asking complex questions with no context without validating the output. LLMs don't store data, they store probability. Understanding that and knowing how limited they are is the first step to using them effectively.
The issues with LLMs, and other neural nets, is that you have people misusing them to generate garbage and companies who want to use them to replace workers.
It's why Deepseek was so disruptive because it's a functional enough model that you can run on a gaming computer. It puts the technology into the hands of the average person and not big companies that just want to use it for profit.
cellman123@reddit
It's helped me deploy a web service onto GKE by writing the Terraform + k8s configuration. I come from a background in C++ systems development and got tired of writing everything from scratch, trying to understand all the moving parts as well as I know my own mother before building anything. Just give me an Ingress resource with a managed HTTPS certificate and point it to my web app service - AI was fantastic at fleshing that out. Obviously, don't do that if you're an idiot, in case you spend way too much money on AI-generated infrastructure or accidentally push some secrets to source control.
thabc@reddit
I think your point here is the same as the author's. You used software engineering (and architecture) best practices to figure out what you want and you had AI help you build it. The software engineer was still the component adding the most value.
bentreflection@reddit
Exactly. This whole argument seems like mathematicians arguing about if the calculator is going to replace mathematicians.
Coffee_Ops@reddit
Calculators are deterministic, their output is easily scoped and understood.
Coffee_Ops@reddit
This phrasing suggests a 1 to 1 relationship between what is requested from AI and what it delivers, which in my experience is a rather naive expectation.
It reliably delivers what it believes you will consider to be success, not what actually constitutes success. Understanding why those subtly different things can make all the difference is what separates junior and senior engineers / project managers.
maikuxblade@reddit
Amusingly, I've heard it put that "AI does a great job if you don't know what a great job looks like"
Blubasur@reddit
I’m a-ok being replaced personally, if we could automate most tasks in live and just focus on things we would like to do. Perfect, exactly what we should strive for.
But the current iteration is undoubtedly going to create more, low quality work which eventually leads to more grueling work deciphering wtf even happened.
And then we have the issue of maintenance both present and future, because currently maintaining that crap is horrible. And future, if people do not learn to code, we’ll have a knowledge decline, and if these AIs can’t maintain themselves yet, we literally can’t phase out programmers.
uCodeSherpa@reddit
There is no world in which we automate everything and you get to focus just on what you want to. This is a pipe dream from movie propaganda.
If everything is automated, you will spend 18 hours a day doing some new weird low paying shit to barely afford food.
There is no scenario in which “focusing on what I want to” comes in to this equation.
MagnetoManectric@reddit
I'm not sure I'm so blackpilled about it myself. A world in which we could fully automate all the tedious crap we need to keep the wheels of society turning would rock.
But we're in the complete wrong socio-economic system for that right now. Right now, it would simply concentrate wealth in the hands of the very-few people that own all the "means of production" in this case, the server farms.
And speaking as an engineer, that, in my mind, is the first problem we need to solve before we can start solving the cool, fun ones :)
Blubasur@reddit
The only thing I said is that it is a goal worth striving for. Arguing if it will happen or not is an absolute waste of time and energy
uCodeSherpa@reddit
Dispelling the propping up late stage capitalist automation based of fairy tales and whimsy is exactly the opposite of “an absolute waste of time and energy”
Blubasur@reddit
Discussion a future neither of us can even imagine at this point is.
tennisgoalie@reddit
Nobody is debating whether vibe coding is gonna replace software engineers, but I think the ones who can’t shut up about it might be telling on themselves 🤷🏻
Rainy_Wavey@reddit
Basically this
Tzukkeli@reddit
Yeah. And like, if Vibe Coding would be Holy Grail of programming, it means that as an app developer, I can push multitude of new apps because I actually understand the code its producing. So its like win-win either way
NoleMercy05@reddit
Online shopping will ever take off
EveryQuantityEver@reddit
3D TV is here to stay!
balefrost@reddit
Solar roadways are the future!
It's easy to find historical examples where naysayers were wrong or where hypemen were wrong. The question is whether this thing will work out or not.
IanAKemp@reddit
Replace "vibe" in "vibe engineering" with "lack of" and you have a truth.
keepthepace@reddit
Yeah, that's just regular engineering.
When I am taking that trend to its extreme, I see something I like: self-healing software. If you get to the point where you can have good tests covering 100% of the use cases, and have 100% of the code generated autonomously, then fixing a bug is just a matter of describing it, in the form of a test, and letting the system fix it.
Many things can go wrong there, and it opens a new range of many potential issues, but this is also a totally new engineering style that opens up.
balefrost@reddit
Unlikely.
Because of combinatorial explosion, it's hard to even get 100% branch coverage with handwritten tests - even for something fairly simple. Software testing relies on the assumption that the entity authoring the production code is at least somewhat reasonable.
As a thought exercise, imagine a test suite of a sorting algorithm that includes these cases:
Here's an pseudocode implementation with 100% line and branch coverage:
You can continue to add test cases, and the AI can continue to add
ifclauses to the implementation. If the AI decides that this general approach is the right approach, it might even be hard to convince it to start over from scratch.An AI is unlikely to generate that algorithm. Sorting is one of the most-studied topics in CS, so the training data likely contains many examples of sorting algorithms and related discussion. But consider some aspect of your domain that isn't likely to be present in the training data. Will an AI be able to "understand" your domain well enough to align the code it writes with the way you think about the domain?
Or will it do the equivalent of that thought experiment: generate code that satisfies the test cases, but no more. If you treat the production code as a black box, will you end up playing an infinite game of whack-a-mole? Can the AI "infer" how you expect the system to behave by a sufficient number of examples?
Formal specification is perhaps a way to solve this. A formally specified system describes not just a finite set of test cases, but an infinite set of test cases. That might be a way for us to truly treat our production code as a black box. But formal specification is also vastly more difficult to write than unit tests, and most software developers have no experience with it. I also wonder if this would even work with generate AIs witing the code. I can imagine such an AI getting stuck in long cycle of code / test / fail some edge case. That's not to say that it would never succeed in producing code that matches the specification, but it might take so many iterations that it would have ultimately been cheaper to pay a human to do the same thing.
Rattle22@reddit
One could say that once you've put in the work to write the formal specification, you could've spent the same effort to just... write mostly functional code that a human can reasonably and cheaply test and fix in the future.
keepthepace@reddit
Anyone who passed algorithmics 101 understands that you can't cover all possible input/outputs cases in tests.
Yes, it is not trivial at all, but the correct way to test a sorting algorithm is to generate a random series of number and test that the output obeys the invariant condition (all numbers present, no additional numbers, they are given in increasing value)
You may find out that in some cases you are missing a failure: e.g. if you have more than 64k identical values, your structure fails and some numbers are missing in the output. In that case, you add a new test case.
Probably not. Not currently at least. That's fine these are the most interesting parts of code to write. If the 99% if the rest, the plumbing as I call it, is automated and self-healing, I'll be happy to write that part!
balefrost@reddit
I agree, this is a great application for property-based testing. I only used sorting to demonstrate the general problem with testing: because you are constructing specific test cases (either manually or automatically), testing only gives you assurance that the production code works for those specific test cases.
I guess maybe we've worked on very different kinds of systems. I would not describe 99% of the code that I write as "plumbing".
jcl274@reddit
what the fuck is vibe coding
uniquelyavailable@reddit
Think I'll vibe code some blind bugs into production to spice up my day a little
robby_arctor@reddit
An offhand remark became the basis for everyone's polemical blog post this month.
notkraftman@reddit
You ask the ai to do something, you check the result, you ask it to do something else, rince repeat. You don't write any code, and I think theres an implication that you don't look at the code either, although that's a bit overoptimistic IMO
SeasonsGone@reddit
I definitely use AI throughout my workday, sometimes simply to generate boiler plate or set the structure of something up, to explain sections of code, to tell me if certain code has opportunities for performance optimization, etc. It is silly to me that some devs are so turned off by these tools that they wouldn’t even consider integrating them in any way to their process.
If a dev doesn’t check that the generated code or the code that came from their mind is good and proper, that’s just a bad dev.
TurboJetMegaChrist@reddit
Pretending the AI is working.
randybanks_@reddit
Basically just using an LLM to do the majority of the actual coding while you vibe to synthwave
seamustheseagull@reddit
As soon as I first heard the phrase "vibe coding" I knew immediately that it was some stupid tech bro vapour ware shit like "web 3.0".
Why do people keep falling for these scams?
FlyingRhenquest@reddit
You just need to vibe code with a 6GL man! If you use a 6GL your vibe coding will be awesome!
lampshadish2@reddit
I thought the insight into AI coding as pushing the problem to the right where it is more expensive to fix was good. I think the jury is still out on if “vibe coding” is here to stay. As a term, it’ll end up with “surfing the web”, even if the practice actually sticks around.
balefrost@reddit
Is it here at all?
Despite a lot of ink being spilt about it not being workable... is anybody actually using the practice for anything serious? Even the original Twitter post said "It's not too bad for throwaway weekend projects, but still quite amusing."
Is "vibe coding" standard practice anywhere?
PM_ME_UR_BRAINSTORMS@reddit
If it was here you wouldn't be asking lol. If AI could actually put together a useful piece of software AI bros and investors would pointing to it nonstop.
ROGER_CHOCS@reddit
Not for anything remotely complex.
figureour@reddit
Seems like it will probably stick around as one of main methods to create prototypes for non technical founders Other than that, who knows.
SerCeMan@reddit (OP)
Thanks. On "here to stay", at the very least, I think tools like v0.dev, etc. for creating landing pages are quintessential vibe coding, and they've definitely found a market fit. The term might be gone sometime soon, but the practice of interacting with the codebase via prompting only seems to have found a strong niche.
mycall@reddit
This crush is harshin my mellow
manliness-dot-space@reddit
One of the biggest problems still is that it just can't fit enough domain context to even start. We humans take a lot of stuff for granted when dealing with subject matter experts and using our own lived experience.
It needs a large context size to understand the customer base and domain and then say something like, "hey actually your Christmas shopping list app might be better suited to using a passwordless authentication mechanism because based on your use case you have non technical users who only use it for a month or a year, they won't remember their password next year. For the best user experience we can just text them a one time use code when they log in to avoid the problems of forgotten password resets. Implementing this will avoid having to do password reset flows entirely!"
Most of the experiments I've done where I try to feed it enough documentation context of ends up just dying. So you have to really implement RAG workflows and "reasoning" internal monologs, and build a whole multi-agent workflow to try and do it, but in practice it gets tripped up by document context window limits in those cases too.
throwaway490215@reddit
Neither is there a pressing need to disprove its existence by pulling people to your blog.
Yet here we are busying ourselves.
ColoRadBro69@reddit
Programmers aren't engineers. Engineers actually have to pass certification, there are laws controlling who can call themselves an engineer. Anybody can just say they're a programmer or software engineer.
That's why we have 6 round interviews. Because they've been the
IsEvencode you guys come up with, and they want to avoid that kind of bad hire.ImprovisedGoat@reddit
Today I saw a job post for vibe coder. Industry is cooked
tomasartuso@reddit
Really enjoyed this read. There’s something very real about how we romanticize team ‘vibes’ when what’s often missing is structure and clear leadership. Do you think it’s actually possible to design a good vibe, or is it more of a byproduct of solid processes and aligned people?
Subject-Substance984@reddit
Has OP looked in what bit.cloud is looking to do? Basically does the “tightly encapsulate AI-generated black-box components with rigorous testing, detailed performance profiling, tracing, canary deployments, and strict protocol compatibility checks.” and compose with them later. These “things” can be built by ai or developer
KristinnEs@reddit
I heard about vibe coding. So I tried it out a bit.
It is fun.
I like sitting there and chatting with the computer, and having a program get created without any real coding from me. It works for small projects. Does it give you professional code that makes complete sense and is optimized? No. Not at all. I'd never use it in production for anything that actually matters. That would be crazy with the current AI capabilities.
But I see future potential in this. Perhaps in five years this will be the new norm, much like more abstracted languages when compared to working in assembly. I dont expect it to take my job any time soon, perhaps not in my professional lifetime. But I see a future where this will be how developers work. The more knowledgable ones know what questions to ask, and know how to fix various small issues that might arise. Much like your grandma can toss upp a wordpress site in a couple of days, but you need an actual developer to create something more serious. One day we might well just sit at the computer and just describe what we want instead of implementing it.
But that is not now.