BitTorrent can prioritize which parts to download in which order to keep the swarm healthy (I don't know how smart it is, but in my experience it seems to go for parts with less availability first) and besides people trying to play videos/audio before it fully downloads (like popcorntime or picking "download in sequential order") this works pretty well.
With a live stream you have a few problems:
* The data you just downloaded is only relevant to the swarm for a short time
* The whole swarm wants the same data at each point in time, and the only part of the swarm that has that data is the originator of the stream
* I'm by no means an expert in this, but isn't it harder to do DHT type discovery in real time when the hashes of the content are new or are always changing? New chunks would need to broadcasted over time, similar to how livestreaming MPEG-dash or HLS works and those torrents would only be relevant to the swarm for a few minutes max.
Something that could (In my amateur/uneducated opinion/guess) work better is staggering the clients based on some sort of grouped delay. If you had X viewers and spread them out over 10 groups each staggered over 10s delay with each video chunk being 10s then the swarming could probably work better and the group with the longest delay would have the best smoothness (because 9/10s of the swarm has the chunk, and 1/10 of the swarm dedicated to only uploading that chunk) but the longest latency.
It would probably work better if the groups lower down the stack are larger too, so that each layer seeds more clients in the layer below it than it was seeded from above it.
Problems that I can see is if there is a lot of churn in the higher groups there would be latency from reconnections/seeding to those and the latency of broadcasting new chunks but I have no idea how solvable or severe those would be.
Would be interesting to build a webtorrent webapp to do something like this to see if it's feasible. If my assumptions are somewhat right it might make webrtc better for broadcast/conference use-cases without using SFUs.