[CraigJPerry]

Ignoring the content of the post for a second (which IMO was excellent), the quality of the writing here is remarkable. This is a dry technical topic at heart and yet i enjoyed reading that entire report. It was as informative as i could hope for whilst still being engaging.

What a joy to read.

[chrismorgan]

It’s 10,000 words and a curious mixture of dense and sparse. There’s quite a bit of duplication (especially of figures), a fair bit of circumlocution in the narrative sections, and a lot of meaninglessly precise figures, half of which should have been omitted altogether. I am confident it could be significantly improved by a hard cap of 5,000 words, and suspect even 2,000 words could still be better (though 1,000 would definitely be too short to convey it all). Even apart from that, it definitely needed a table of contents, to set expectations.

As a general challenge to people: write your article, then see if you can halve its length without losing much. If it felt too easy, repeat the process! There’s a family of well-known quotes that amount to “sorry for writing a long letter, I didn’t have time to write a short letter”. Concise expression is not the easiest, but very valuable. Many a 100-page technical book can be improved by reduction to a one-page non-prose overview/cheat sheet (perhaps using diagrams and tables, but consider going more freeform like you might on a whiteboard) plus a ten page abridged version.

[Mashimo]

Mhh, I found it repeated sentences again and again. It was kinda odd to read at times.

[input_sh]

This isn't just poor writing, it's ChatGPT-padded slop.

But the same is true for the content itself, no business is paying you to actually build the same app 10x, especially so if it's something as trivial as a kanban board.

[mkl]

They'd comfortably pay for 10 AI-assisted versions. It's a trivial demo app so that implementing it 10 times is feasible - it's just to learn what to build their main app in.

[input_sh]

I wouldn't measure how good/fast/performant a library is looking at the results of the very first LLM attempt at doing a trivial task using that library. If you don't know the libraries well enough to spot some improvements the LLM missed, the only thing you're judging is either how sane the defaults are or how good the LLM is at writing performant code using that library, none of which are equivalent to how good the library is.

Also, performing well in a prototype scenario is very different than performing well in production-ready scenario with a non-trivial amount of templates and complex operations. Even the slowest SSGs perform fast when you put three Markdown posts and one layout in them, but then after a few years of real-world usage you end up in a scenario where the full build takes about half an hour.

Kinda cool that you can do that in an afternoon, but absolutely useless as a benchmark of anything.

[mkl]

I said assisted, not generated, and the author had human experts in these libraries go over the implementations.

[input_sh]

Yeah I highly doubt that. Up to five, sure. All 10? No way.

[Anon1096]

Eventually english textbooks are going to start including this isn't... it's pattern because it's so prevalent in ai slop. I close anything I read now at the first sign of it.