I think we have a real inflection point now. I try it a bit every year and was always underwhelmed. Halfway through this year was the first time it really impressed me. I now use Claude Code.

But Claude Code costs money. You really want to introduce a critical dependency into your workflow that will simultaneously atrophy your skills and charge you subscription fees?

It's also proprietary software running on someone else's machine. All other arguments for or against aside, I am surprised that so many people are okay with this. Not in a one-time use sense, necessarily, but to have long-term plans that this is what programming will be from here on out.

Another issue with it is IP protection. It reminded me stories where the moment physical manufacturing was outsourced to China, exact clones appeared shortly after.

Imagine investing tons of efforts and money into a startup, just to get a clone a week after launch, or worse - before your launch.

Right, we the workers are giving away control over the future of general purpose computation to the power elite, unless we reject the institutionalization of remote access proprietary tooling like this

this is why I say "AI is for idiots"

Any new useful tool must be managed in a way so that one isn’t overly dependent on it.

- google maps

- power tools

- complex js frameworks

- ORMs

- the electrical grid (outages are a thing)

- and so on…

This isn’t a new problem unique to LLMs.

Practice using the tool intelligently and responsibly, and also work to maintain your ability to function without when needed.

A year ago I could get o1-mini to write tests some of the time that I would then need to fix. Now I can get Opus 4.5 to do fairly complicated refactors with no mistakes.

These tools are seriously starting to become actually useful, and I’m sorry but people aren’t lying when they say things have changed a lot over the last year.

It might even be true this time, but there is no real mystery why many aren't inclined to invest more time figuring it out for themselves every few months. No need for the author of the original article to reach for "they are protecting their fragile egos" style of explanation.

The productivity improvements speak for themselves. Over time, those who can use ai well and those who cannot will be rewarded or penalized by the free market accordingly.

That's what it really all comes down to, isn't it?

It doesn't matter if you're using AI or not, just like it never mattered if you were using C or Java or Lisp, or using Emacs or Visual Studio, or using a debugger or printf's, or using Git or SVN or Rational ClearCase.

What really matters is in the end is, what you bring to market, and what your audience thinks of your product.

So use all the AI you want. Or don't use it. Or use it half the time. Or use it for the hard stuff, but not the easy stuff. Or use it for the easy stuff, but not the hard stuff. Whatever! You can succeed in the market with AI-generated product; you can fail in the market with AI-generated product. You can succeed in the market with human-generated product; you can fail in the market with human-generated product.

If there’s evidence of productivity improvements through AI use, please provide more information. From what I’ve seen, the actual data shows that AI use slows developers down.

The sheer number of projects I've completed that I truly would never have been able to even make a dent in is evidence enough for me. I don't think research will convince you. You need to either watch someone do it, or experiment with it yourself. Get your hands dirty on an audacious project with Claude code.

Research is the only thing that will convince me. That’s the way it should be.

https://youtu.be/1OzxYK2-qsI a 6-12% increase in pull requests per developer per month.

It sounds like you're building a lot of prototypes or small projects, which yes LLMs can be amazingly helpful at. But that is very much not what many/most professional engineers spend their time on, and generalizing from that former case often doesn't hold up in my experience.

We use both Claude and Codex on a fairly large ~10-years old Java project (~1900 Java files, 180K lines of code). Both tools are able to implement changes across several files, refactor the code, add unit tests for the modified areas.

Sometime the result is not great, sometimes it requires manual updates, sometimes it just goes into a wrong direction and we just discard the proposal. The good thing is you can initiate such a large change, go get a coffee, and when you're back you can take a look at the changes.

Anyway, overall those tools are pretty useful already.

It sounds like you're assuming I'm not a professional engineer and I only work on prototypes.

They're basing it on what you described in your previous comment. I got the same impression.

Finishing projects makes me sound unprofessional?

"sheer number" combined with "completed" sounds more like lots of small projects (likely hobbyist or prototypes) than it does anything large/complicated/ongoing like in a professional setting.

It is, at this point, rather suspect that there are mountains of anecdata, but pretty much no high quality quantitive data (and what there is is mixed at best). Fun fact; worldwide, over 200 million people use homeopathy on a regular basis. They think it works. It doesn't work.

I don't suppose anything will change your mind, but here you go.

https://youtu.be/1OzxYK2-qsI a 6-12% increase in pull requests per developer per month.

"But those diffs are AI slop and rework" you will object. Oh well, I tried.

> Fun fact; worldwide, over 200 million people use homeopathy on a regular basis. They think it works. It doesn't work.

Homeopathy works for sure. Placebo works. There are many studies confirming that.

Do you have evidence?

What does “can use” mean though. You just ask it to do things in basic English. Everyone can do that with no training.

0.1x

If only you put half as much effort into learning ai as you do trolling people who are getting gains from it...

Because it has been true for the last 3 years. Just because a saying is repeated a lot doesn't mean it's wrong.

It's gotten ok now. Just spent a day with Claude for the first time in a while. Demanded strict TDD and implemented one test at a time. Might have been faster, hard to say for sure. Result was good.