You should check out some example code for react-three-fiber (e.g. https://codesandbox.io/embed/r3f-bones-3i7iu). If you have any experience with writing raw webgl, it will become abundantly clear that something is gained by going through three.js before returning to React's more declarative style.

Most of the savings in the example are coming from three.js itself (not r3f), which is maybe the key point: three.js offers much more than a scenegraph, which is why it's worthwhile (check out the official examples to get an idea: https://threejs.org/examples/). Webgl as an API doesn't even really talk about 3d objects for the most part; the language it provides is primarily about moving data around in buffers. The difference in possible time savings in three.js vs webgl is on the level of writing in C vs an assembly language.

And once you're building an application with a non-trivial amount of state mutation (on which the view depends)—you're faced with the same dilemma as traditional web dev and the DOM, hence the desirability of react-three-fiber.

That said, I think it would be super interesting to see a three.js alternative that was 'natively' reactive/declarative, because I'll readily admit the tower of abstractions involved in writing a react-three-fiber app has its downsides. (Then again, I consider three.js to be a rare gem, know of nothing comparable in terms of simplicity/quality for building 3d apps, and would be [pleasantly] surprised to see anything like the above anywhere in the near future.)

Similarly, around 2013 threeQuery (threejs + jquery) was becoming somewhat popular too (it had a jquery "chain" api like syntax). It's good to see people experiment and attempt to improve developer efficiency. Who knows what new tricks will be discovered and what kind of benefit and new approaches will be created! However, I mostly agree with you.

I find threejs to be one of the most enjoyable libraries to work with (and its codebase is simple and beautiful too). I also highly recommend to everybody that wants to go into 3d graphics to dedicate some effort and learn the actual fundamentals (eg webgl, opengl, metal, matrix math, quaternions, etc). This way you gain domain specific knowledge that is applicable across platforms and across time. Abstractions are not future-proof, they are recycled and change according to the latest trends in development. Domain specific knowledge stays with you forever! If there is an intermediate ground on which people can meet (eg start with react-three or three.js and then dive deeper) that's a win too. Recently I have been advocating that web devs can start learning 3d graphic concepts by just playing with... CSS to familiarize with some of the concepts and then move from there. This way one can avoid all the overhead around the gl statemachines or various libs and focus on the basic concepts first

it allows you to create self-contained components. that alone will eradicate so much boilerplate and complexity. it has a real pointer event system. it takes care of managing the scene reactively, it disposes of objects it does not need any longer.

you use this for the same reason you would use react for the dom. r3f is not a wrapper that duplicates the threejs export catalogue, it is a renderer/reconciler.