Has anyone else downloaded and messed around with the AV1 source code and examples...

This is done anyway (each keyframe is a new boundary which can't be crossed by further reference frames, otherwise seeking wouldn't work with reasonable speed)
And good keyframe detection can be done easily, so it won't be a big efficiency loss.

Then your bottleneck is memory bandwidth. And entropy coding is still taking too much time.

I see. There wont be any optimization before the bitsream freeze, anyway. NONE.
Bonnet.

dead on arrival

In order to have real time 24fps encode on this hardware, they would need to get 120 times speed increase in 3 months. They've already stated hardware acceleration wont be around until the end of next year. If you've ever played around with VP9, you would know encode times were bad. The only bright side of all of this is that decoding isn't nearly as resource intensive. Otherwise AV1 would be DOA.

Any arguments to back that up?

Jewgle is greedy is one.

VP9 encode times aren't bad actually. Compare it to x265 and the only bad thing is the threading, not the actual computational power.

Common sense and doom9 lurking. Why do optimizations if you potentially will have to strip them? Plus, they're simply focused on getting gains at this point.

Compare it to something people actually use instead of a shekel-grabbing meme codec.