AI3 min read
Benchmarks
Your smart glasses remember everything but can't tell you when it happened
Meta's S-EMBER benchmark puts 9,448 hours of egocentric video through frontier AI models and uncovers a localization paradox: semantic reasoning scales with compute, but temporal grounding doesn't.
2026-07-25