Create a more sophisticated ad-timing algorithm
Ad timing (when in a broadcast an ad is shown, plus for how long or for how many ads) is a largely "invisible" skill in traditional media, but an important one for managing the flow of content delivery that almost no streamers have. Twitch broadcasters lack dedicated producers for this purpose, and the current algorithms that automatically deliver ads account for at most the time since the last ads a user watched. This leads to ad delivery that often interrupts streams very badly.
Twitch should develop more sophisticated algorithms that deliver ads at the least intrusive moments in streams. For example, someone playing rounds of a PVP game -> showing ads between rounds. Someone doing short/reset heavy speedruns -> show ads after 3 quick resets or 1 long one. A just chatting stream -> follow lulls in conversation.
Teach a sample of streamers the kind of producer skills that help them spot when an ad would be effective, follow up on the data of them showing those ads. Start sampling tags on the stream and the categories and information like compressed screenshots of the screen before and after an ad. Is there a prediction based on screenshots and streamer volume/speech that can tell you when chat is going to be very engaged? Use that as training data and start making a more sophisticated algorithm. That way you can avoid dropping an ad on top of what should be the highest engagement content, and put in the ads in the lulls and "victory fuzz" moments.