What happened
Researchers found that neural networks trained to match similar images work better when the difficulty of training gets adjusted automatically rather than staying fixed. In practice, this means image recognition systems can learn faster and make better matches without requiring humans to manually tweak training parameters.
Why it matters
This is a narrow optimization for a specific machine learning technique — useful for the engineers building image matching systems, but it doesn't change what's possible or affordable at scale, and it won't affect anyone outside the research community.