> As opposed to the old days when people would just blindly copy/paste random shit from Stackoverflow?
Many of the people who are complaining about AI vibecoding today also didn't blindly copy/paste from StackOverflow in the past.
Many of the people who are complaining about AI vibecoding today also didn't blindly copy/paste from StackOverflow in the past.
But I often do think across adjacent abstraction levels, because abstractions are (varying levels of) leaky. Modern compilers are after many decades good enough and modern computers fast enough that it is rare that I need to dig into the assembly (but I happens, compiler explorer is in my bookmark bar in Firefox).
Other abstractions are far leakier, it is far more common that I look in wireshark to debug network issues, the application level view is often not enough.
One of the leakiest abstractions currently is LLMs. Maybe in a decade or three they will be good enough, but they aren't yet, that's for sure. At least for the hard realtime systems level programming I do. For code generation they often make enough mistakes that the time spent after review and fixes comes out in the wash, even for simple tools. Their use for bug finding, RAG and similar is however promising.
At my last job the employer paid for OpenAI access for all of us.
Baby sitting an LLM is not my idea of meaningful use of time. And reviewing code that someone else had an LLM spew out even less so.
I am not lashing out because I don’t have access to LLMs. I had access and I did try it plenty.
The change has been so rapid that I think a lot of people are having a hard time I guess wrapping their head about the lived experience of it. For a while my only access to the tools was through work. Then I ended up getting a $20/month ChatGPT account and that comes with codex and now I can't imagine sitting there Googling a problem anymore. It literally feels low tech these days. Big "I'm not paying for Cable, the antenna is good enough" energy. It saves me soooooo much time just maintaining my own local stuff. I mean it literally saves me hours and hours of personal labor.
The local models will 100% catch up. Most likely the inference I use now will be free in five years across the board and you'll be buying a cyberdeck or something with a 128G of RAM and an LLM friendly bus architecture.
But you didn't encounter those people. People said they've used the tools, and then you said it wasn't recent enough. Not because you know when they've used it last, but because their experience didn't match yours.
It's great that it's working for you, really. But since you have been posting on HN a lot about this, while simultaneously claiming you don't care what other people do, maybe consider a different approach than calling everyone luddites. If people call you a shill, and you exit the conversation claiming you don't care, but then you go to the next thread doing the same low effort stuff, don't be surprised when people see a pattern. There are lots of people communicating more effectively about this.
But for code it's a race to the bottom because it's all text and local works quite well. You can host models on LangSmith and similar and because people use those services to create chat bots the overall use of them for coding is a very tiny fraction of their overall usage. The race to the bottom is further exacerbated by the fact that as GPUs become more powerful you can host more per unit so the cost of text inference will drop precipitately. Right now people are reporting that for some of the self hosted services they are able to do everything for under $5 / month. That price WILL drop because that's how computers work.
The fact that you consider deterministic output from a compiler the same as probabilistic output from a LLM makes me think you don't know how either of those things work, even at a very superficial level.
I have a standing challenge to my co-workers that valid compiler errors will be rewarded like a birthday party, with the baked goods, alcohol, or sweets of their choice. It's only been redeemed once, and I've found less than a dozen unreported compiler bugs myself.
Currently the openbsd mailing list for port is currently going through a clang update and one of the main point is looking at all the packages that failed to build. I even took a long look at the usb stack and the audio subsystem of OpenBSD because of an issue I was having with my DAC.
So when you have a bug and a core dump, you can quickly load it in debugger, see the stack frame and then theorize a model for the bug to happen. If after verifying the source and having complete confidence that it's good, then you start looking at the assembly, most likely while single stepping with the debugger. But you rarely get to that point, because 99.99... it's your code.
That reliability is what AI tooling is lacking. It's exhausting monitoring the output because errors can be as simple as a minus character or the wrong comparison operator.
You mean a source that's been tested on billions of PCs over 45+ years?
As opposed to a LLM which outputs code that barely works on my machine™?