You could probably train something like that on semi-reasonable consumer hardware. Ads often have a very distinctive style and tone, and you need only a single output - the probability of it being any given second being an ad. It would probably take a lot to run though, you better hope the people who install the extension have good PCs. And, it would probably never get 100% accurate, you'd have to put up with still seeing some ads and having to rewind when it skips over valid video.