From what I have seen, the biggest shift is not perfect generalization yet, but better perception and learning from real-world data. The demos look more capable, but reliability over long periods is still the hard part. It feels like progress is real, just slower than the funding hype suggests.