Posts

Multimodal Models’ “Performance Breakthroughs” Are Splitting into Three Distinct Problems

On September 1, Google announced that Gemini’s “agentic video understanding” feature can dynamically scan video clips to reduce token consumption by up to 88% and costs by up to 66%, while improving quality by up to 7%. These figures appear to improve efficiency, price, and results simultaneously, but they measure the performance of a specific video-analysis mechanism rather than providing an overall ranking of a general-purpose multimodal model’s capabilities.