It's about using our technology and research to help enrich people's lives. Like YouTube – and his mission to give everyone a voice and show them the world.
Our work with YouTube product and engineering teams helped optimize decision -making processes, increase security and commitment and increase the impressions for all users.
Making more search shorts
Shorts on YouTube-short movies with a length less than a minute-watched over 50 billion times a day.
By covering everything from the emerging K-pop stars to local food guides, they are quick to watch, fast-and they become more and more popular. But because shorts are created in just a few minutes, they often do not contain descriptions and titles that make them easier to find by searching. So we introduced Flamingo, our visual language model that helps generate descriptions.
Flamingo analyzes the initial video frames and explains what is shown on the screen (e.g. “a dog balancing a pile of crackers on the head”). He writes this text as metadata on YouTube, creating more pronounced categories of content and matching users' search to better results.
YouTube introduces this technology on shorts, with automatically generated video descriptions on all new messages. Now viewers can find and watch more suitable movies from a more diverse range of global creators.
Video compression optimization
The video has exploded in recent years, and online movement is expected only in the future – video compression is an increasingly burning problem.
We have worked with YouTube to test the potential of our AI model, Muzero to improve the VP9 codec, coding format that helps compress and send video via the Internet. Then we used Muzero for some live traffic in YouTube.
During the premiere, we observed an average reduction by a 4% transmission transmission with a variety of movies. Biterate helps determine the computing ability and bandwidth needed to play and store movies – affecting everything, from charging, resolution, buffering and data use.
By improving the VP9 codec on YouTube, we helped reduce internet traffic, data use and time needed to charge movies. And by optimizing video compression, millions of people around the world are able to watch more movies when using a smaller amount of data.
Brand safety protection
From 2018, our cooperation on YouTube has helped educate creators about movies that can earn revenues from advertising, and make sure that appropriate ads appear in the right place.
We have developed a model of label quality (LQM) with YouTube team to exactly mark movies, as well as with guidelines friendly YouTube advertiser. In addition to improving the accuracy of advertising in movies, it helps to ensure that ads appear next to the content that is YouTube guidelines.
By improving the way of identifying and classifying movies, we have increased confidence in the platform for viewers, creators and advertisers.
Improving marchapters
As the video has evolved and watching and watching, the creators began to add chapters to their films. It makes it easier for them to find the desired content – but it can be a slow process.
We have worked with the YouTube search team to develop an AI system, which suggests chapter segments and titles for YouTube creators, through automatic processing of video transcripts, audio and visual functions. Thanks to the indigenous people, viewers spend less time searching for content, and the creators save time creating chapters for their films.
Because the function was introduced in Google I/O in 2022Automatically generated chapters were used up to tens of millions of films (and counting) on YouTube.
Evolutionary technologies and products
We are still looking for ways to improve alphabet products thanks to our AI research.
Our cooperation with YouTube has already had a great impact on people's lives – and with more projects we are still improving the impressions for users.
