A solution would be for an extension to download the entire video 2x and delete the difference.

I don't think that would work. It would be trivial for YT to put different ads in different time slots which would leave a differencing engine with no way to tell what was content and what was ad. However that thought gave me another one; the core problem is the ability to differentiate between content and ad. That problem is probably solvable by leveraging the NPU that everyone is so desperate to cram into computers today.

Nearly all of the YT content I watch, and it's a lot, has predictable elements. As examples the host(s) are commonly in frame and when they're not their voices are, their equipment is usually in frame somewhere and often features distinctive markings. Even in the cases where those things aren't true an Ad often stands out because its so different in light, motion, and audio volume.

With those things in mind it should be possible to train software, similar to an LLM, to recognize the difference between content and ad. So an extension could D/L the video, in part or in whole, and then start chewing through it. If you were willing to wait for a whole D/L of the video then it could show you an ad free version, if you wanted to stream and ran out of ad-removed buffer then it could simply stop the stream (or show you something else) until it has more ad-free content to show you.

A great way to improve this would be by sharing the results of the local NPU ad detection. Once an ad is detected and its hash shared then everyone else using the extension can now reliably predict what that ad looks like and remove it from the content stream which would minimize the load on the local NPU. It should also be possible for the YT Premium users to contribute so that the hash of an ad-free stream, perhaps in small time based chunks, could be used to speed up ad identification for everyone else.

It wouldn't be trivial but it's not really new territory either. It's just assembling existing tech in a new way.