It’s about using our technology and research to enrich people’s lives. Like YouTube, its mission is to give everyone a voice and let the world know.
Working with YouTube’s product and engineering teams, we’ve optimized decision-making processes, improved safety and engagement, and improved the experience for all types of users.
Make short videos more searchable
YouTube Shorts (short videos under 1 minute) are viewed over 50 billion times a day.
Covering everything from up-and-coming K-Pop stars to local foodie guides, it’s quick to watch, quick to make, and always growing in popularity. However, because short videos are created in just a few minutes, they often don’t include descriptions or titles that are easy to find in search. Therefore, we introduced Flamingo, a visual language model that helps generate explanations.
Flamingo analyzes the first video frame and describes what it sees on the screen (e.g. “Dog balancing a bunch of crackers on its head”). This text is stored as metadata on YouTube to create clearer content categories and better match your searches with results.
YouTube is rolling out this technology across Shorts videos, with automatically generated video descriptions appearing with every new upload. Now viewers can discover and watch more relevant videos from a more diverse range of global creators.
Video compression optimization
Video has exploded in recent years, and as Internet traffic is expected to continue to grow, video compression has become an increasingly pressing issue.
We worked with YouTube to test the potential of the AI model MuZero and improve the VP9 codec, a coding format that helps compress and transmit videos over the internet. We then applied MuZero to a portion of our live YouTube traffic.
At release, we saw an average bitrate drop of 4% across various video sets. Bitrate helps determine the computing power and bandwidth needed to play and store a video, affecting everything from load times to resolution, buffering, and data usage.
By improving YouTube’s VP9 codec, we’ve reduced internet traffic, data usage, and the time required for videos to load. And by optimizing video compression, millions of people around the world can now watch more videos with less data usage.
Protecting brand safety
Since 2018, our collaboration with YouTube has helped educate creators about the types of videos that can earn ad revenue and ensure the right ads appear in the right places.
We worked with the YouTube team to develop a Label Quality Model (LQM) to more accurately label videos in line with YouTube’s advertiser-friendly guidelines. This not only improves the accuracy of ads on your videos, but also ensures that your ads appear alongside content that adheres to YouTube’s guidelines.
By improving the way we identify and categorize videos, we’ve increased trust in our platform for viewers, creators, and advertisers alike.
Auto chapter improvements
As the way we create and watch videos evolves, creators are starting to add chapters to their videos. This makes it easier for viewers to find the content they want, but the process can be time-consuming.
We worked with the YouTube Search team to develop an AI system that automatically processes a video’s transcript, audio, and visual features to suggest chapter segments and titles to YouTube creators. Auto-chapters allow viewers to spend less time searching for content and creators to create chapters for their videos.
Since the feature was introduced at Google I/O in 2022, auto-generated chapters have been applied to tens of millions of videos across YouTube (and counting).
Evolving technology and products
We’re continually looking for ways to improve Alphabet products through AI research.
Our collaboration with YouTube is already having a huge impact on people’s lives, and with more projects in the works, we’ll continue to improve the user experience.
