> in "AI" land if someone's paying for these ridiculous advertisements.

Yep, it’s like saying all products have been designed by AI for 50 years because humans have been using CAD models, CNC, and database like inventories. Without humans in the loop none of them would useful at all, just like coding with any “AI” agents so far through a single prompt. The bubble surface tension is holding in a critical amount of hot air and the upwards pressure is causing the surface to erode and evaporate even faster which is why news outlets are reporting on the AI stock sell off as fears over AI replacing all coding tasks because someone has a need they can vaguely describe. It’s a multiplier of human productivity, not the dream of end stage capitalists who want to replace and dispose of the working class.

Have you actually used Claude Code?

"anything you can do, I can do butter"

i guess we're toast

> but it is like outsourcing to junior developer, except it is much faster and cheaper

When people say, "I outsourced to a junior developer" it's a strong sign they are talking BS.

No one ever talked about outsourcing to a junior developer before AI code was a thing.

> And make inexperienced ones produce more crap. I tried it a little and it could really speed things up, but it is like outsourcing to junior developer, except it is much faster and cheaper. In general it is good at scrabing examples, produce templates code, fixing some bugs, but it makes quite a few mistakes.

Agreed. There are some things that I find pretty annoying about it that you have to tell it explicitly not to do. For example, pulling in really old versions of NPM or Python packages. It has a propensity to not know when it is correct or incorrect, or perhaps, it it is unable to share its level of uncertainty. Most of these things can be mitigated by updating and refining Claude's prompting.

I think there are some things those of us clutching our pearls at the thought of all of the AI slop ought to keep in mind:

We are not all writing air traffic control systems or medical device firmware. There is an awful lot of "enterprise software" that is not much more than database CRUD and visualization. People are not going to die if that software is not using the most efficient data types, etc.

Will somebody five years from now pick up your AI generated software and not know what to do with it? Possibly, but mostly only if you are not taking the time to generate unit and integration testing as well as documentation, and there is now basically no excuse now to skip doing either.

Senior developers should be spending more of their time thinking about the architecture and use cases of their applications, and less time thinking about whether HashMap or FxHashMap is more efficient in a given scenario.

If we are being honest, there is already a bunch of slop in our industry. Junior (or just bad) devs copying/pasting from StackOverflow... Shortcuts taken during unreasonable deadlines... Stuff we just aren't told or don't know... Software development has been in a race to the bottom for a while now.

Most of the furniture in my home is not hand-made by Amish carpenters, it is machine milled and partially assembled by cheap labor. It is not as good and will not last generations as hand-crafted furniture, and that is okay. For better and for worse, a lot of SMB executives look at software the same way (at least until it doesn't work), they want "good enough" software that works until the next merger or acquisition, so they can cash out and go on to the next thing. They are not intersted in paying for software built using hand-built assembler (SpinRite - we miss ya').

The good news is that there are things we can do, and if we get good at them, AI can do the mundane bits and software engineering will still be a thing. We get better at Specification and Test driven development. We review the hell out of the code AI generates and make sure our linting and bench-marking tools are up to snuff. We get really good at authoring prompts that keeps AI tools within the guardrails. And yes, we keep AI away from the really critical stuff, at least for now...

> Hyping 4% sounds like a scammer hyping shit coins and NFTs.

hey, it's not an ordinary 4%. it's an inflection point 4%!

> this article is pure spam/hype.

like almost every other article here. but it pays msmash's salary so we can keep enjoying this formerly venerable news site and crack jokes at the asinine nonsense that msmash gets paid for posting these days. just relax, take it all in and feel the time passing, worlds colliding in the far distance. what's not to like?

4% seems like a substantial number given how many git hosted projects there are.

> if you don't know how to code it's a god send.

I contend just the opposite. I'd be terrified to watch someone that doesn't know how to code use claude code for the very reasons you mentioned. It's going to do amazing stuff most of the time and really stupid stuff periodically. To pick up on that stupid stuff, knowing how to code and conscientiously reviewing changes with that background knowledge is the only way to get those impressive results.

> ..if you don't know how to code it's a god send.

Imagine society continuing to abuse this excuse in every other aspect of your life.

If the mechanic doesn't really know how to change the brakes on your car, it's a god send.

If your doctor doesn't really know how to do that surgery on you, it's a god send.

Oh, you'll be sending to a God alright. Perhaps quite prematurely.

> BUT it still occasionally rewrites your entire code base for to fix a button or introduces environment breaking changes for a new feature etc.

Being completely dismissive of this is what enables CxOs to assume a 12-year old that can text to the AI godz faster than every seasoned programmer, is good enough to replace them. Every t

> BUT it still occasionally rewrites your entire code base for to fix a button

I think that's how it should work in many situations. Iterating on code is not strictly necessary when you have agents that can reason about and output thousands of lines of code in minutes. Any new requirement has the potential to result in architecture-breaking changes that may be best solved through a complete rewrite. What happens with the code in the implementation may not, in a completely spec-driven scenario, be relevant at all assuming no human developers are expected to work on it.

In my experience... if I ask Claude to help with something that's my strong point (like my core coding), it's like training a junior programmer, and I can solve my problems faster myself. But... when I need something outside my core expertise, like helper programs, wrappers, or interfaces to other technologies I'm not familiar with, it's a very fast way to get 95% of the way there without wasting time climbing the learning curve myself for a one-time need.

Like any tool, you have to know when to use it and when not to, and what its strengths and weak points are.

GitHub Copilot the same. It can leverage Claude and lets you review every change. And I find that I *do* have to review every change. It often makes weird, inconsistent, and even plain wrong changes. Sometimes it wants to remove functions that are outside the scope of my prompt, even if those functions are being used. But I do find the tool very much a time-saver.