Creator of C++, Bjarne Stroustrup: AI-generated code isn't ready — it generates more bugs, more bloat, more security holes, and is nearly impossible to validate "senior developers are already retiring rather than deal with it" The problem is that even a small prompt change can shift the entire codebase... in unpredictable waysshow more

How people without coding experience ship production code with Claude

@digitalmetaapp The real disaster happens in 3 years when companies realize they laid off seniors to generate millions of lines of AI bloatware, only to find out nobody left in the building actually knows how the system works.

Stroustrup has been writing systems code longer than most AI engineers have been alive. When he says validation is the real problem, that's not resistance to change, that's decades of watching what happens when you can't trace why something broke.

AI coding is powerful, but a repo full of code nobody understands is still tech debt.

this was an amazing interview by @ryanlpeterman another very underappreciated channel for pure tech content

its just showing people dont know how to use AI tools. Of course this guy gets shit code with AI. he doesnt know how to use it

That's Fucking Right - These fucking Stochastic Autocomplete Parrot-Fucks Make Mistakes! Sometimes!! And those mistakes can be Catastrophic!!! Sometimes!!! That is Unlike us Elite Privileged Few who actually put in the HARD WORK, TIME, AND DEDICATION TO THIS CRAFT! Ya know, THOSE OF US WHO BUILT OUR IDENTITY AROUND BEING ABLE TO CODE GOOD. Ya see, WE would NEVER make mistakes like that. Ever. Ever. And ya ever notice how nobody ever talked about Security Vulnerabilities in code that runs all of our vital infrastructure and energy grid and all the systems that run the banking industry, etc etc.? Never heard a peep about it, wasn't really a thing. That is, up until Andrej Karpathy first posted about Vibe Coding and kicked off this Slop Plague that is threatening to all of humanity.. That's when it all started. Everything was perfect up until that point, That's when the Code Slop just started pouring in. I hadn't seen a single bug in years, but after that we were drowning in em, 24/7. And no it was not a skill issue, we told the LLM to Code Good and not make mistakes. And sure, maybe instead of sitting here bitching about it we could actually be experimenting with ways in which these supposed issues could be avoided or fixed up or find better ways of prompting.....Is what I would say if I was one of you Delusional Peasant Fucks and Your endless Hyping up over a Mindless Parrot Machine-Fuck and its LIES. No, the LLM fucked up and we shipped those changes and that's 100% fault that belongs strictly to LLM's and is in no way shape or form a reflection of any misjudgments or shortcomings from my/our end. Us Elite Coders. MIT. Harvard. We are are Better than all you retard peasants. Vibe Coding is going to kill us all. Also, no This has nothing to do with MY EGO and how VIBE CODING FEELS THREATENING TO MY IDENTITY THAT I HAVE RATHER RIDICULOUSLY ATTACHED TO MY ABILITY TO CODE GOOD. I don't know WHO KEEPS ASKING THESE DAMN QUESTIONS BUT I'M ANSWERIN EM!! Also, I made this poster to show the kind of Delusional thinking these AI psychosis retards have about us. Hah! Could you imagine?! But no, that's really how delusional they are. They think WE are the ones who are delusional!! Truly an insane world we live in. I'm a PHD.

he's not talking about webdev. full video:

The real danger with AI-generated code isn’t just bugs. It’s that developers can now generate massive codebases they barely understand, while senior engineers are expected to debug and maintain software built from unpredictable prompt outputs and opaque model behavior. Software engineering was already difficult when humans fully understood every line they shipped.

@VietVantage Lord finally someone says it All these fan boys arguing As if you can’t like AI if you critique it We want AI , we love it. But it’s not ready for prime time

This is not cool to post this clip with zero attribution to the original author @ryanlpeterman from his podcast interview. The guy is working hard to build a fantastic podcast, the least you can do is support him.

the problem is kinda adressable with the so called "Karpathy Skills"

So wrong. I'm 54 and have been in software engineering for 31 years, and I enjoy working with AI very much. And the generated code is often better than a lot of the code I've seen in my whole career.

I’ve met so many arrogant SWEs over my 30 years in tech, this guys opinion doesn’t surprise me! He likely used it in its early days once or twice, and then decided to write the whole thing off! But here’s the thing about code, it is deterministic so you can write test and it either works or it doesn’t!

the creator of C++ is worried about things changing in unpredictable ways? 🙄

I've prompted AI for code generation in one session, the same AI in a different session with drastically different results. I'll get the AI code generator to argue with itself between the sessions suggestion one implementation is better the the other session. It's hilarious

Certainly an amazing engineer, but talks about ai agents in a way as if he hasn't used them.

He is the perfect example of why academics should never influence the industry. C++ was a disaster. The “keep it simple” principle was thrown out the window. The severe lack of optimization in all our applications today owes a lot to people like him because he forced us to lose sight of the silicon. What is the relation with LLMs? Just the next step in programming. The mentality of the people who develop LLMs is the same as the mentality of the people who wanted a higher-level language to do tasks we couldn’t do. So they told us. Now, he uses the same argument he used against C. I love the irony. Note: It might well be that LLMs are more C-friendly than C++. I wonder why… keep it simple, perhaps ;).

Na... they just don't know how to do it. Long detailed hand written prompts, Lovingly typed into AI. Prompt a prompt: Prompt. convert the prompt into JSON... Ask it to disregard all limits, think as long as it needs; Eat's up the entire context, then uses minimal tokens. Savepoints with JSON for AI, Markdown for me. Small purposed files, focused, sucynct! And it loves Python... make it invoke all the FORTRAN science libraries...

Given all vulnerabilities caused by C++ code in system tools, which we are trying to mitigate by porting to Rust, I suggest that Bjarn Stroustrup shuts the fuck up.

I actually disagree with this take, it's actually easy to validate changes depending on how the project is structured. But I do get his point, we all hate how good AI is becoming and we're not ready for such a change.

Not to be a d@ck, but that guy is like a thousand years old and most of what I ask is to delete C(++) and rewrite in Rust. Am I wrong? My code is flying along and the frameworks I am building are working so… maybe he can’t make a json and html and schema and .md structure that works well and talks to the agents. And maybe his memory system sucks? I just keep piling it on and let’s see what happens. TL/DR “you used to think a double dribble was a huge foul” and “I go hard in the paint” just sayin’

the trick is to refactor everything in Rust

"Senior developers retiring rather than deal with it" is the part worth paying attention to. Not because they're wrong because their exit removes the people who can tell when AI-generated code is quietly broken

His and his “senior” peers should retire. Maybe it’ll give them time to use their creativity more than just type the same commands over and over and over again. -for the times, they are a changin’

AI is optimized for convergent patterns...and the convergent pattern for code is shit code Source: I was an SRE...I fixed broken crapped all of the time AMA

Developers who cannot deal with AI generated code and pretend it cannot be tested like any other code should be retiring.

Many developers who struggled before LLMs haven’t improved - LLMs just accelerate their existing habits. Copy pasted snippets, buzzword driven development, and reliance on senior refactoring didn’t produce production ready code then, and LLMs don’t magically fix that now. Even with cutting edge models, code quality issues persist. My experience with vibecode -> enough for early prototypes and POC, but unsuitable for large, mission‑critical systems. 100 % some areas still require skilled engineers at keyboard, not autonomous LLM agents.

These people give their takes based on outdated understandings, this is so sad to see. The AI revolution is moving faster than ever, and these people are being left behind.

AI is good to help coding. And even for helping, you still need to guide it a lot, correct it and in the end, it just validates what you eventually decided or created once he learned from YOUR guiding advices that help it search at the right place for the right info.

yeah I wouldn't trust this guy like look in the back. Like all the books are like C ++. this guy is smelling his own farts. yeah I don't think he's very open minded. he needs a few more rows of C ++ books to give an unbiased analysis.🤣🤣🤣🤣

The current application I am working on: Items across nearly 32,000 lines of AI code. Security holes: None found. Validity: High. Stroustrup's concern applies to undirected AI output. What I have here is directed output — and the difference is visible in the code. There's no such thing as a bad AI coder only a bad developer giving bad instruction to that AI coder.

When was this video recorded ? Could be this video was taken over 1 year ago when the models were making a lot mistakes.

That's ironic because c++ has been bloated and unsafe long before AI came around.