Good abstractions work best when all backends are fundamentally doing the same job with the same constraints, but remain separate for non-technical reasons. This would include GPU APIs, OS memory allocation APIs, and window manager APIs. The details differ, but across platforms, they all serve similar technical purposes, which is why you can form a relatively clean abstraction over them. The reason the cleanest APIs in SDL can exist is because the underlying things they are abstracting over are largely equivalent. The abstractions get weirder the more you change the implementation constraints. For example, a window manager abstraction works well across systems with similar window managers, but it becomes weirder if you use that same API on a game console with a single framebuffer and no notion of windows whatsoever. This is also why, I suspect, LLMs are so good at porting. It’s ultimately an exercise in translation from one language to another, but ultimately mapping to the same thing. Usual disclaimer, that’s not useless, that’s not nothing, but it sure ain’t the same thing as “super intelligence”.