I think the most significant thing here is additional use/benefits of auto-labeling. The 10.4 release notes claimed a new auto-labeler led to a massive improvement in VRU detection (" precision +35%, recall +20%"), and I observed a massive improvement in no longer detecting most mailboxes as pedestrians, which resulted in a huge reduction in phantom braking on roads with mailboxes. Now, we have even further VRU perception improvements from the auto-labeler, but importantly we also see a second auto-labeler system ("new static world auto-labeler"). If they can continue developing auto-labelers for additional tasks this quickly, and continue to expand the use of the ones they already have, I expect it will dramatically accelerate progress and reduce development costs.
I'm speculating that the "Improved static world predictions" realized by the new auto-labeler will manifest in less steering wheel jerkiness approaching/executing turns, and will also reduce risk of turning into an incorrect lane.
I'm also excited to see the "Improved lateral control for creeping". Although, I really wish they had also indicated some improvements to lateral control on un-marked roads. I suppose it's possible they made some updates on that front and didn't bother including them in the release notes, but I'm probably setting myself up for disappointment thinking that.
The other thing I really enjoyed here was related to the sheer number of clips being fed into the auto-labelers. I suspect they are mostly selecting clips of times when the current software didn't perform optimally (no need to teach it what it already knows), which implies that our collective testing efforts are leading to progress in a very direct way. I know we all like to tell ourselves/others that our testing is "helping make the software better" (better in some vague way that none of us really understand), but when we don't really hear any feedback from Tesla, it sure does make you question your faith that putting up with so many manure-grade maneuvers is actually accomplishing anything. Seeing that 165k clips number makes me think that a good portion of those had to be selected from the data captured from FSD beta disengagements and snapshots, given those events are good predictors of situations where the current software performed sub-optimally.
So, here we go again on the "next rev will be better" hype cycle. I, for one, savor these moments at the peak of inflated expectations. Talk to you tomorrow from the trough of disillusionment.