This article was originally published by AI Front, ID: AI-front, link to the original article: t.cn/ROfPoY5
One day last year, while I was playing a video game called “Injustice 2,” I did a regular YouTube search to see how to beat a tricky part of the game. As usual, I found a video that answers my question. But the next time I checked YouTube, the site showed me more sophisticated videos of how to play the game: clips of how players play the game without being spotted by enemies; A video clip of the player killing every enemy in a masterful way; Video interviews with game makers; Brilliant satire. I just went to YouTube to search for the answer to a question, and it showed me a whole new universe.
Later, I found myself visiting YouTube several times a day. Most of the time, I don’t open the site for any particular purpose, and I’m used to passively asking the site to automatically recommend something I might like. In January, I became obsessed with a folk band called Pinegrove, and within a few weeks YouTube had recommended almost every live video of the band uploaded to its server. When I moved into a new apartment in the spring and started cooking more and more, YouTube quickly introduced me to its home-chef camp after I did a search on how to make Italian bread salad: Byron Talbott, J. Kenji Lopez-Alt of the Serious Eats channel, and Tasty, among others.
YouTube has always been useful, and since its launch in 2005 it has become one of the mainstays of the Internet. But in the last few years, YouTube has become surprisingly good for me. It began to predict with extreme accuracy what video clips I might be interested in, much better than it had done in the past. What has changed?
Over the past 12 years, YouTube has transformed itself from a search-driven site into a purpose-driven site. To reach its destination, it will require hundreds of attempts, a lot of redesign, and a giant leap in artificial intelligence. But what has really elevated YouTube is its evolution toward feeds.
It’s hard to remember now, but in the beginning, YouTube was just an infrastructure. It provides an easy way to embed videos on other sites where you are most likely to see them. As the site grew, YouTube became a place to find clips of old TV shows, keeping up with the latest late-night comedies to watch the latest viral videos. Along with Wikipedia, YouTube is perhaps the most notorious place on the Internet. Your co-worker makes a “Harlem Shake” video in the water cooler, and then you log on to YouTube all night.
At the same time, Facebook invented the standard format of our time: News Feed, a constant stream of information tailored to your interests. Feeds have taken over the entire consumer Internet, from Tumblr to Twitter, Instagram and LinkedIn. YouTube’s early efforts at things like personalization were limited, mostly letting users subscribe to channels. The idea, borrowed from television, has mixed results. In 2011, YouTube had some success with a massive push, but the average amount of time spent watching YouTube videos remained about the same, according to ComScore.
Channels no longer dominate YouTube as they once did. Now, if you open YouTube on your phone, you’ll find “channels” hidden in a separate menu. Instead, the app presents a mix of videos tailored to your interests as feeds. These videos also include videos from channels you subscribe to, of course, but they also include videos related to videos you’ve watched before.
That’s why, when I went straight to the video search for Vengeance, I started looking at recommended playthroughs and snarky reviews. YouTube has developed tools to make its recommendations not only personalized but also highly accurate, which ultimately drives viewing time across the site.
Jim McFadden, head of technology at YouTube Recommendations, said: “We know that people come to YouTube to find what they want, but we also want to satisfy people when they don’t know what they’re looking for.” McFadden has been with the company since 2011.
I first visited YouTube in 2011, just a few months after McFadden joined the company. That’s when YouTube started letting users spend more time watching its videos, and that’s still a core goal of YouTube today. At the time, things were not going very well. “YouTube.com as a home page, it doesn’t bring a lot of entertainment,” McFadden says. We thought, ok, let’s make it massively entertaining as a transformational purpose.”
The company tried everything: it bought professional camera equipment for top creators and launched “LeanBack,” which automatically ranks new videos for you as you watch them. YouTube has redesigned its home page to emphasize subscription channels rather than watching individual videos. The number of hours watched per user remained flat, but one change, which drove the upheaval of the following spring, was that their recommendation algorithms were based not on how many people clicked on the videos, but on how long people spent watching them.
Almost overnight, video creators who had benefited from misleading headlines and video downsizing saw their viewing numbers plummet. Higher quality videos, which tend to be associated with longer viewing times, began to rise dramatically. In each of the next three years, YouTube viewing time grew by 50 percent.
I subscribe to a few channels and consider myself a regular YouTube user. But for it to become a destination that can be visited many times a day, it also requires a set of new tools that have become possible in the last 18 months.
When I visited the YouTube office this month, McFadden introduced me to the roots of YouTube’s precise recommendations: Google Brain, the ARTIFICIAL intelligence division of YouTube’s parent company, Google, which YouTube has been using since 2015. Google Brain isn’t YouTube’s first attempt to use AI. YouTube has previously applied machine learning technology from Sibyl, a system built by Google, to its recommendation algorithm. Google Brain, however, introduces an approach to unsupervised learning, with algorithms that can find connections between different inputs in ways that software engineers never imagined.
“The key thing is that it’s universal,” McFadden says. “Before, if I watched a comedy video, the recommendation algorithm would say, ‘Someone else likes this video. But Google Brain’s model identifies other comedies that are similar to this, but not identical, and have more contiguous relationships. It’s able to pick up on less obvious but similar patterns.”
For example, a Google Brain algorithm suggests short videos to users of a mobile app, but longer videos to users of YouTube TV. It guessed, correctly, that recommending different lengths of video depending on platform would eventually increase viewing time. YouTube implemented more than 190 such changes in 2016 and plans to make 300 this year. “The reality is that it’s a bunch of small improvements that add up over time,” says Todd Beaupre, product manager for YouTube’s discovery group. For every improvement, you try 10 things and end up implementing one thing.”
Google Brain’s algorithm is faster than YouTube’s previous algorithm. In past years, the company said, it took several days for a user’s behavior to be integrated into future video recommendations, making it difficult to spot trends. “We have to fix this if we want to bring people in to understand what’s going on right now,” Says Beaupre. “Right now, delays are set in minutes or hours, not days.”
Integrating Google Brain into YouTube has an important impact: More than 70% of the time people spend watching videos on YouTube now comes from YouTube’s recommendation algorithm. Every day, YouTube recommends 200 million different videos to its users in 76 languages. The total amount of time people spend watching videos on YouTube’s home page is 20 times more than it was three years ago.
This is basically consistent with my own user behavior. A few years ago, I used to visit my YouTube page almost exclusively during my lunch break to watch something while I ate. But their recommendations were so good that I started watching more videos in my free time. This week, I signed into YouTube on playstation 4 so I could watch its recommended videos on my biggest screen.
This is a truly powerful personal customization Feed. The amazing thing for me is that YouTube has changed my digital life more than anything else. The Facebook Feed is based on what your friends post and what your favorite pages post. It’s useful to know who’s engaged or having a baby, but beyond those milestones, I find little pleasure in what my friends post. Twitter shows you the tweets of the people you follow and what those people have chosen to retweet. As a journalist, I have to rely on Twitter, even though there are times when my timeline seems endless and full of angst cries.
Each Feed has a length limit, although this limit was removed in 2017. No matter who you follow on Twitter, the political debate always dominates the discussion. Facebook’s short-lived enthusiasm for features like “events” and “groups” made the Feed change in shocking ways every week, leaving me feeling less connected to each of my friends. (Photo-centric Instagram seems like an oasis, so it’s no wonder the app is still growing so fast.)
Facebook, Twitter and Instagram, it seems like those feeds are all asking people to constantly perform something for them. YouTube is clearly performance-driven, but very few users upload videos to it, and YouTube has never forced users to do so. YouTube can be passively enjoyed, just as the TV channels it is trying so hard to displace do. In a crazy age like ours, it’s so calming not to be asked what we think about a news story.
YouTube’s emphasis on relevant videos you might like means its Feed is broader and more curious than other feeds. The more it seeks out different content, the more it looks like it’s escaping the model of other feeds. In a dark age, I prefer the escapism of YouTube.
In 2013, the Atlantic ran an article in which Alexis Madrigal postulated that the Feed as we know it had its peak. The future, he thinks, will be limited experiences: E-mail newsletters, Medium collections, 10-episode Netflix series. After all, the endless flow of content can be exhausting. Madrigal says: “When the order of the media universe is completely defeated, freedom will not fill the void, and a new order with its own logic will replace the old one. We find that the flow of information has shown itself to be compulsive and controlling. Faster! Much more! Faster! Much more! Faster! More!”
Four years on, YouTube’s direction only shows that the Feed model is becoming more important. An unprecedented growth in video storage, coupled with unprecedented personalization technology, will create something that is hard to resist. YouTube now surveys users on how much they like the videos they recommend, and over time, the results will make YouTube smarter, allowing more content to be consumed.
Beaupre described the process to me as crossing a chasm. “There’s the magic zone between content that is highly compatible with content you already like and content that is trending and popular.” If YouTube’s rivals cannot find a way to bridge that gap, they will find it hard to compete.
The author | Casey Newton translator | NER edit | Emily
AI Frontier provides AI enthusiasts, developers and scientists with the latest and most comprehensive technical information in the FIELD of AI, front-line industry practice cases, collation of industry technology sharing dry goods, and interpretation of the latest AI papers. Once a week, there will be an open technology sharing class to help you fully embrace artificial intelligence technology.
