[cochne]

> Not every piece of code is an abstraction of course. To me, an abstraction is a piece of code that’s expressed in high-level language so that the distracting details are abstracted away. If I were to see a confusing piece of code littered with conditional logic, I wouldn’t see it and think “oh, there’s an incorrect abstraction”, I would just think, “oh, there’s a piece of crappy code”. It’s neither an abstraction nor wrong, it’s just bad code.

The wrong abstraction isn't crappy code itself. It is a reasonable looking piece of code that will force the next person into writing crappy code to accommodate it.

Edit: I think the entire project of TensorFlow is a good example of this. They built the library around a "graph" entity, and anything you did had to be shoehorned to fit that. That worked OK for some straightforward neural networks and situations for a while. As the area evolved though, it proved very burdensome. They tried to evolve it into TensorFlow 2.0 which was more forgiving, but by that point it was too late, the ecosystem became a mess. PyTorch stole the thunder because they didn't make the wrong abstraction (though I'm not sure if "duplicating" is what helped them do that)