A Blog by Jonathan Low


Sep 1, 2017

How YouTube Perfected the Feed

Artificial intelligence supplied by Google Brain was able to discern patterns not apparent even to experienced software engineers.

The result of constant experimentation based on AI recommendations, such as shorter videos for those using mobile phones and longer ones for those using laptops, caused audience numbers to explode exponentially. JL

Casey Newton reports in The Verge:

Google Brain employs unsupervised learning: its algorithms can find relationships between inputs that engineers never would have. Instead of basing its algorithmic recommendations on how many people had clicked a video, YouTube would base them on how long people had spent watching it. Creators who had profited from misleading headlines saw view counts plummet. Higher-quality videos, associated with longer watch times, surged. Watch time on YouTube grew 50% a year for the next three years.
Sometime late last year, as I was playing a video game named Dishonored 2, I did a routine YouTube search about how to beat a tricky section of the game. As usual, I found a video to answer my question. But on my next YouTube visit, the site offered me even more compelling Dishonored videos to watch: clips of people playing Dishonored without ever being detected by their enemies; clips where players killed each enemy in highly creative ways; interviews with the game’s creators; whip-smart satirical reviews. I had visited YouTube seeking an answer to my question, and it had revealed a universe.
Soon afterward, I found myself visiting YouTube several times a day. For the most part, I visited without having a specific destination — I had become accustomed to the site serving up something I would like, unprompted. In January, I grew obsessed with a folk-rock band named Pinegrove, and within weeks YouTube was serving me video of seemingly every live performance ever uploaded to its servers. I started cooking more once I got a new apartment this spring, and after searching for how to make a panzanella salad, YouTube quickly introduced me to its battalion of in-house chefs: Byron Talbott, and Serious Eats’ J. Kenji López-Alt, and the Tasty crew, among others.
YouTube has always been useful; since its founding in 2005, it has been a pillar of the internet. But over the past year or so, for me anyway, YouTube had started to seem weirdly good. The site had begun to predict with eerie accuracy what clips I might be interested in — much better than it ever had before. So what changed?
Over the course of 12 years, YouTube has transformed itself from a site driven by search to a destination in its own right. Getting there required hundreds of experiments, a handful of redesigns, and some great leaps forward in the field of artificial intelligence. But what really elevated YouTube was its evolution into a feed.
It can be hard to remember now, but at the beginning YouTube was little more than infrastructure: It offered an easy way to embed video onto other websites, which is where you were most likely to encounter it. As the site grew, YouTube became a place to find archival TV clips, catch up on late-night comedy, and watch the latest viral hits. Along with Wikipedia, YouTube is probably the web’s most notorious rabbit hole. Your coworkers mentioned the Harlem Shake at the water cooler, and so you went to YouTube and watched Harlem Shake videos for the rest of the evening.
Meanwhile, Facebook had invented the defining format of our time: the News Feed, an infinite stream of updates personalized to you based on your interests. The feed took over the consumer internet, from Tumblr to Twitter to Instagram to LinkedIn. YouTube’s early approach to personalization was much more limited: it involved asking users to subscribe to channels. The metaphor was borrowed from television, and had mixed results. A huge subscription push in 2011 had some success, but the average time a person spent watching YouTube stayed flat, according to data from ComScore.
Channels no longer dominate YouTube as they once did. Open YouTube on your phone today and you’ll find them hidden away in a separate tab. Instead, the app opens to a feed featuring a mix of videos tailored to your interests. There are videos from channels you subscribe to, yes, but there are also videos related to ones that you’ve watched before from channels you may not have seen.
This is why, after searching for straightforward Dishonored videos, I started seeing the recommendations for stealth runs through the game and satirical reviews. YouTube developed tools to make its recommendations not only personalized but deadly accurate, and the result has lifted watch time across the site.
“We knew people were coming to YouTube when they knew what they were coming to look for,” says Jim McFadden, the technical lead for YouTube recommendations, who joined the company in 2011. “We also wanted to serve the needs of people when they didn’t necessarily know what they wanted to look for.”
I first visited the company in 2011, just a few months after McFadden joined. Getting users to spend more time watching videos was then, as now, YouTube’s primary aim. At the time, it was not going particularly well. “YouTube.com as a homepage was not driving a ton of engagement,” McFadden says. “We said, well, how do we turn this thing into a destination?”
The company tried a little bit of everything: it bought professional camera equipment for top creators. It introduced “leanback,” a feature that queued new videos for you to watch while your current video played. It redesigned its home page to emphasize subscribing to channels over individual videos.
Videos watched per user remained flat, but a change made the following spring finally moved the needle: instead of basing its algorithmic recommendations on how many people had clicked a video, YouTube would instead base them on how long people had spent watching it.
Nearly overnight, creators who had profited from misleading headlines and thumbnails saw their view counts plummet. Higher-quality videos, which are strongly associated with longer watch times, surged. Watch time on YouTube grew 50 percent a year for the next three years.
I subscribed to some channels and counted myself a regular visitor to YouTube. But for it to become a multiple-times-a-day destination, YouTube would need a new set of tools — tools that only became available within the past 18 months.
When I visited the company’s offices this month, McFadden revealed the source of YouTube’s suddenly savvy recommendations: Google Brain, the parent company’s artificial intelligence division, which YouTube began using in 2015. Brain wasn’t YouTube’s first attempt at using AI; the company had applied machine-learning techniques to recommendations before, using a Google-built system known as Sibyl. Brain, however, employs a technique known as unsupervised learning: its algorithms can find relationships between different inputs that software engineers never would have guessed.
“One of the key things it does is it’s able to generalize,” McFadden said. “Whereas before, if I watch this video from a comedian, our recommendations were pretty good at saying, here’s another one just like it. But the Google Brain model figures out other comedians who are similar but not exactly the same — even more adjacent relationships. It’s able to see patterns that are less obvious.”
To name one example: a Brain algorithm began recommending shorter videos for users of the mobile app, and longer videos on YouTube’s TV app. It guessed, correctly, that varying video length by platform would result in higher watch times. YouTube launched 190 changes like this one in 2016, and is on pace to release 300 more this year. “The reality is, it’s a ton of small improvements adding up over time,” said Todd Beaupre, group product manager for YouTube’s discovery team. “For each improvement, you try 10 things and you launch one.”
The Brain algorithms also work faster than YouTube has before. In past years, it might have taken days for a user’s behavior to be incorporated into future recommendations. That made it difficult to identify trending subjects, the company said. “If we wanted to bring users back to find out what’s happening right now, we’ve kind of fixed that problem,” Beaupre said. “The delay, instead of multiple days, is measured in minutes or hours.”
Integrating Brain has had an immense impact: more than 70 percent of the time people spend watching videos on the site is now driven by YouTube’s algorithmic recommendations. Each day, YouTube recommends 200 million different videos to users, in 76 languages. And the aggregate time people spend watching videos on YouTube’s home page has grown 20 times larger than what it was three years ago.
That roughly matches my own behavior. Years ago I started visiting YouTube’s home page regularly on my lunch break, to have something to look at while I ate. But the suggestions were good enough that I started taking more regular YouTube breaks. This week I broke down and signed into YouTube on my PlayStation 4, so that I might watch its recommendations on the largest screen I own.
That’s the power of a truly personalized feed. And yet it’s striking to me how different YouTube’s feels from any of the others that inform my digital life. Facebook’s feed is based on what your friends post, along with posts from pages you like. It’s useful for knowing who’s gotten engaged or had a baby, and yet I find little pleasure in my friends’ posts beyond those milestone events. Twitter has tweets from the people you follow, plus anything those people have chosen to retweet. As a journalist I am all but required to live on Twitter, even though these days the home timeline is little more than an endless, anxious scream.
Each feed still has its strengths, though 2017 has diminished them. On Twitter, politics dominate the discussion no matter whom you follow. Facebook’s momentary enthusiasms for features like events and groups lead the feed to transform week to week in ways that are jarring, and leave me feeling less connected to everyone I’m friends with. (Image-heavy Instagram still feels like an oasis, and it’s little wonder the app is still growing so fast.)
Facebook, Twitter, Instagram — it seems notable that all these feeds ask you constantly to perform for them. YouTube is driven by performances, obviously, and yet a tiny fraction of its users ever upload a video — and YouTube never pressures them to. YouTube can be enjoyed passively, like the television channels it has worked so hard to replace. In a frantic age, there’s something calming about not being asked for my reaction to the day’s news.
YouTube’s emphasis on videos related to ones you might like means that its feed consistently seems broader in scope — more curious — than its peers. The further afield YouTube looks for content, the more it feels like an escape from other feeds. In a dark year, I’ll take all the escapism YouTube has to offer.
In 2013, writing in the Atlantic, Alexis Madrigal posited that the feed as we know it had peaked. The future, he suggested, would belong to finite experiences: email newsletters, Medium collections, 10-episode Netflix series. Endless streams of content are, after all, exhausting. “When the order of the media cosmos was annihilated, freedom did not rush into the vacuum, but an emergent order with its own logic,” Madrigal wrote. “We discovered that the stream introduced its own kinds of compulsions and controls. Faster! More! Faster! More! Faster! More!
Four years on, YouTube’s approach suggests the feed is only becoming more important. An ever-growing repository of videos, matched with ever-improving personalization technology, will be difficult to resist. YouTube now surveys users about how much they enjoyed the videos that are recommended to them; over time, the results will make YouTube smarter — and lead to more video being consumed.
Beaupre described this process to me as crossing a chasm. “There’s stuff that’s closely related to what you already liked, and stuff that’s trending and popular. But in between, that’s the magic zone.” And if YouTube’s rivals can’t find a way to cross that chasm, they may find it very difficult to compete.


Post a Comment