Natural language processing
This past week, at the Wired HQ at CES, I spoke with Kayvon Beykpour, the head of product at Twitter. We discussed some of the company’s new product launches, and then went deep into questions about the incentives created by the structure of the platform. And, also, whether Twitter should give every user a “troll score.” The conversation has been condensed and lightly edited for clarity.
Nicholas Thompson: You’re here at CES announcing a lot of changes involving conversations on Twitter. They’ve become the central focus. Why are you emphasizing it more now, and what have you just announced?
Kayvon Beykpour: The lifeblood of Twitter is conversations. The core atom of how people stay informed about what’s happening in the world through Twitter is through a conversation someone else starts through a tweet and the resulting discussion around that tweet. So the priority for us is really rooted in ensuring that we’re inspiring people to start conversations, inspiring them to participate in conversations, giving people the tools to have healthy discourse on the platform.
What we talked about a little earlier today with respect to conversations is really an evolution of work that we started doing in the middle of 2019, which is, we want to give people more control over the conversations that they start. So this was an example: in October, we launched globally a feature that allows you to hide replies if you want to.
NT: So if I was [Iranian foreign minister Javad] Zarif last night, and after I tweeted “Hey, we’re not going to escalate,” Trump had tweeted back, “Actually, we’re going to nuke you,” Zarif could have muted the tweet?
KB: No, because those were quote tweets. Quote tweets are I think an important mechanic that’s distinct because me yelling in your movie theater is different than you hosting your own conversation. I think that distinction is really critical actually. Because the philosophical sort of approach we took here is, when you start a conversation, as the author of the tweet, you should have a little bit more control around the replies within that tweet, which is different than me deciding to start my own conversation.
NT: So if I want to trash someone, I need to do it in a quote tweet, not a reply.
KB: I mean, you can do both, I suppose. But the mechanics of replying to someone is different, right? You're injecting yourself into their conversation to their followers, showing up in their mentions. So I think the dynamics are different there and that’s what we were interested in, sort of creating some more control around that.
NT: Let’s talk about the new conversation threading. The simplest way to explain it is that it takes Twitter conversations from incomprehensible to comprehensible.
KB: Conversations can be quite unwieldy to follow, especially if you're trying to read a conversation that has multiple people participating and thousands of replies. It can be really hard to read and see who's responding to who. Sometimes you'll see a tweet really deep in the conversation actually is from the original author, but that's not really clearly understood. Or maybe a tweet is from someone who the author originally didn’t mention, which would be important context, but it's kind of lost in the sea with how conversations look and feel. So what we've been trying to do is really reimagine the way we display a conversation to make all that easier to read. How do we make the participants, the actors in those conversations, more discernible—sort of the authoritative voices, the person who started the conversation, the person who's mentioned in it, someone you follow within that conversation, more recognizable? And all of that is packaged in this public beta app that we call Little T that we've been experimenting with the public. And we're taking the best of those annotations that we learned via research and experimentation and launching them into the Twitter app over the weeks and months to come.
NT: What is the greatest risk to the product from these changes?
KB: I think the greatest risk, personally to me, would be some of the magic of Twitter is in the replies, it’s in the unexpected back and forth you have with folks who see your tweet and respond to it, it’s in the critiques, it’s in the debates. And it’s that magic we don’t want to lose. But I think on the spectrum we have today, where participants in the conversation, particularly the author, have zero control—or the controls, I should say, that we have are very absolute. You can block, you can report, you can make your account protected. I think that there’s a more nuanced set of mechanics that we can create that allow that magic of Twitter to still be there but to provide really new use cases. If you wanted to have a fireside chat on Twitter where Bill Gates and Elon Musk were talking about the future of climate change, it’s very difficult for the two of them to have a controlled conversation without a cacophony of other people joining the conversation.
NT: All right, let's talk a little bit about how your algorithm and AI works to prioritize what we see: whether it's in the main feed or whether it’s within a conversation, prioritizing which elements of the conversation we see.
KB: There's lots of interesting vectors where this sort of ranking in machine learning is critical. One of those vectors is obviously the home timeline. What do we show people in the timeline? What is the ranking of what we show in that timeline? And another vector is within the actual conversation view itself. I think in the context of the home timeline, fundamentally, we want to show people the most relevant content that will match their interests. And so we do that by looking at a lot of things: the accounts that they follow and, as of a couple months ago, the topics that follow. And we try and show them the right tweets at the right time.
The added dimension—really, what we started doing only as recently as early 2019—is not just showing tweets in your home timeline, but showing you actual conversations. Sometimes the most interesting content might be back and forth that you and I are having that isn’t fully captured by just showing the single tweet that we started. It might be your tweet and my response to it. So that's sort of an aspect of ranking and user experience that we've tried to evolve within the home timeline, moving it beyond just tweets and into conversations.
And then finally, when you look at a conversation, like you started our conversation earlier today by saying, “Hey, I’m having this fireside chat with Kayvon, what should I ask him?” How we rank and display the replies to the conversation is really critical to providing a good user experience, especially if those conversations have hundreds or thousands of replies. And so for us, to get to your question, when we look at metrics like the relevancy that those replies have to the author, whether that be their reaching influence, or whether that be the actual proximity they have to the author—someone you follow is probably more relevant to you than someone that you don't follow—engagement metrics like, you know, the number of likes or retweets that that reply has, the likelihood that the tweet is healthy or toxic. All of these things sort of factor into how we ultimately rank the replies.
NT: The way it was done at all the social platforms three years ago was basically a bunch of levers, each of which could weigh a certain amount. You weigh the replies times X, the karma score of the user by Y, the adjacency of the person who tweeted by Z. You multiply that all together and you get the relevancy score of a particular comment. But the way it’s increasingly done is: AI optimized for relevancy, AI optimized for time on site. Which one is it? Is it a bundle of levers, or is it AI optimized for X right now?
KB: Depending on the service area, it tends to be a bundle of different approaches. I think increasingly, leveraging machine learning to try and model the behaviors that we think are most optimal for that area. So for example, we would like to show replies that are most likely to be replied to. That’s one attribute you might want to optimize for, not the only attribute by any means. You’ll want to deemphasize replies that are likely to be blocked or reported for abuse. So you bundle up all these different considerations and build machine learning around it. There isn’t any single optimization.
NT: And is there a change to one of the variables where there was an interesting trade-off, where, for example, pushing it in one direction would lead to more engagement, but pushing into another metric would lead to problems with health?
KB: That's a really good question. One of the things we've learned is that if you just optimize for ranking tweets that are most likely to get the most replies, sometimes that actually services tweets that aren’t healthy, because people will oftentimes be inclined to reply to outrageous tweets with their own replies. We've learned that through experimentation and had to adjust for that in our machine learning algorithms to make sure that just because we wanted to show replies that are more likely to get replies, we did not want that to result in a situation where we're actually surfacing more toxic replies.
NT: So that leads to a super interesting topic. If you have a system that incentivizes toxicity, you should change, right? And I'm glad that you studied that and changed it, but how do you measure toxicity? Tell me how you're using AI to parse language, and whether you’re using it to find negativity, and flipside whether you also started using it to find positivity.
KB: Yeah, this is a great question and at the heart of one of the most important things we're working on right now. Today, a very prominent way that we leverage AI to try to determine toxicity is basically having a very good definition of what our rules are, and then having a huge amount of sample data around tweets that violate rules and building models around that. Basically we're trying to predict the tweets that are likely to violate our rules. And that's just one form of what people might consider abusive, because something that you might consider abusive may not be against our policies, and that's where it gets tricky.
NT: Right, something can be almost abusive, and still be bad, but not cross the line.
KB: Correct. And I think this is where one of the most important things that we're working on really is starting to question and rethink the fundamental incentives on our service, and whether we offer the right balance of incentives. So just as an example, some prominent incentives that we have: the follower counts, the likes, the retweet, impressions. These mechanics all tend to incentivize content that gets a lot of reach and popularity. And sometimes outrage can get popularity and reach. And so that sometimes can be a very powerful thing. Things going viral on Twitter can be a really awesome thing. But oftentimes, unhealthy content can get viral more easily precisely because of those mechanics. So one of the things we've been thinking about is whether we have the right balance of incentives within the core product experience. Putting our rules aside for a moment, just as an example, there isn't really a disincentive today to being a total jerk on Twitter. And that's a product problem. Like that's something that we should think about in terms of what are the right metrics to use to tip the scales a little bit and create a disincentive for people to behave in a way that we might consider abusive.
NT: You need a red checkmark if somebody’s a total dick. There’s some line they cross, and then they get a red checkmark next to their profile.
KB: It's a funny example, but if you think about a service like Lyft or Uber, there is a disincentive to be a total jerk. As a passenger, I have a passenger rating. As a driver, I have a driver rating. And there's an understanding within the marketplace that if you behave a certain way, that your reputation will be impacted in a way that can have adverse consequences. And I think that notion exists in some capacity on Twitter, but not enough.
NT: Let's say somebody comes to you says, “You know what, I totally agree, Kayvon. Let's do this. Let's give everybody a troll score. And we'll use our AI to determine how much of a troll they’ve been. Like how much shitposting they've done, how many cruel statements, how many times they've been flagged. We'll make it one to five, it won't be the most prominent thing, but it will be next to your follower count, right?” What's wrong with that? Can we do that?
KB: It sounds like you want to be a product manager. Are you interested in doing that?
NT: If you will let me create a troll score on Twitter, I will be a product manager on Twitter tomorrow. That would be hilarious.
KB: Yeah, I think it's a good example, with the troll score as a symbol of something we could do is a good example. There's going to be account-level solutions. Then there's content-level incentives, like the likes and the retweets are mechanics that exist at the content level, not necessarily at the account level. So there isn't a single silver bullet here. But our plan is to be thoughtful about this stuff, continue doing a lot of research and experiment.
NT: So let's get more specific. Instagram has announced that they're heading toward demetrification. They're not going to show the number of likes on the story, they're going to either deemphasize or hide the number of followers you have. If you hid the number of followers that people have, you totally change the incentives of the platform. You might get less engagement, but my guess is you get more health. Why haven't you done that?
KB: Well, so actually, we have in our public experimentation app that I mentioned, Little T, we did exactly that. We deemphasized the like count, the retweet count. And when you look at the conversation, we actually don't show those metrics in the forefront. Similar to Instagram, we haven't removed them, we buried them. And we wanted to understand what the implications of that are to how the conversation unfolds. But that's a pretty minor step. Nevertheless, an interesting step. I think there are more extreme steps that we can consider.
NT: So what have you learned? Are you going to do this?
KB: That's something that we're actually moving into public experimentation with right now.
NT: So you've gone from—you've done the experimentation with Little T, you're going to do public experimentation at some point in the next year, there may be an announcement, there may not be an announcement?
KB: Yes, but not again, I think it's important to know that this isn't just limited to visually removing or—suppressing or elevating metrics is one way to do it. I personally think that are even more meaningful ways that we can introduce new incentives.
NT: So actually more like adding a troll score or changing the structure of the algorithm?
KB: No, I think examples like the troll score are super interesting, but we haven't explored as much historically.
NT: So let's go back to something I asked a minute ago, which was AI for analyzing toxicity, but also AI for analyzing sort of positivity. My understanding, and you know I've written a lot about this with Instagram, was that they had a much easier time using AI to figure out toxicity than actually figuring out whether there's something truly positive. Is the same true for you, is it easier to find the bad than the good?
KB: It's a good question and it depends on what you mean by the bad and the good. If by bad you mean policy-violating, then that's actually quite easy to define because we have policies, we have examples of tweets about their policies, and we can build really effective models. Positivity is different in the sense that it's extremely subjective. And policy prediction can also be subjective, but I think positivity is more like a relevancy question in terms of, what is the thing that I find most interesting? How do you balance interestingness with timeliness, which of course on Twitter is extremely important, balancing the most recent information that might be interesting to you versus less time-sensitive information, but nevertheless has some threshold of interest in this to you. For us, that's a particularly interesting challenge right now because we're moving beyond just a follow graph, an account follow graph, to also a topic taxonomy. So you now follow X people and Y topics. We have a larger corpus and a wider corpus of content that we can choose from to surface for you.
NT: Explain how you use AI to identify that when I tweet something, that you should then recommend a topic to follow?
KB: It starts with us creating a taxonomy around topics, which we started with around 300 topics, and we've grown to over 1,000 topics. And the fundamental challenge is being able to understand what topics tweets are about. So if you tweet something, you may be a reporter for WIRED, but you may be tweeting about a football game or about a music concert or about the Golden Globes. We ought to be able to understand semantically what your tweet is about, so that then we can do things like recommend. If I happen to be cruising through my timeline, and I see that you've tweeted about the Golden Globes and I like that tweet, I should also be able to potentially follow the Golden Globes, the topic the Golden Globes. Or machine learning as a topic because you happen to be tweeting about machine learning. So that's, I think, an important user experience step that we can make only if we understand what topics tweets are about.
NT: And when you're making product changes, clearly, there are lots of things you can weigh about the health of Twitter or engagement on Twitter. But what about the overall effect on society? Let me give you an example. I feel like one of the things that Twitter does at kind of an existential level is it consistently moves power from organizations to individuals. So the NFL has less power, but the players have more power because they can tweet. In journalism, the institutions have less influence relative and the authors have more control, which is great in lots of ways but also in some areas creates societal upheavals. When you're thinking about your product changes, are you thinking, This is what it will do to Twitter? Or are you thinking, OK, this is what's going to do to Twitter, but this is what it might do to society and how society will change and we have these values, and it might violate them?
KB: It's a great question. You know, we absolutely think about the ramifications of the work that we do outside of just the walls of Twitter. It's important to understand what the impact will be to customers, but also the world. Particularly, this is why health became such an important priority for us, you know, two-plus years back at this point. But just as an example, ensuring the integrity of our elections is paramount. And there's work that we need to do across every product surface area, across every team, to make sure that we can live up to that. So we're trying to be highly considerate about the ramifications of the work that we do.
And I think to answer your question on how we do it, fundamentally first there's a lot of defense that has to be played around just the world with health. There’s a lot within health, whether it's security, whether it's accounts compromised, whether it's coordinated, you know, malicious activity from state-sponsored actors, there's a lot that we have to do that's fundamental. But the highest leverage thing that I think is important for us to do, that we've really tried to live by, is to ensure that every product change that we make, that we are promoting health and healthy conversations through that work. We have to understand the potential negative ramifications of every product change.
NT: What is something you've rolled back after you saw an effect on the outside world, not an effect on Twitter?
KB: I don't know if rolling back is the right term because I think every product change is a revolution and a curation. Like the example I gave you around how we evolved the ranking of replies is a good example of this. We found that our initial versions of a ranking model contributed negatively to health. And we saw that through metrics, and saw that qualitatively, and we didn’t roll it back, we rapidly iterated on making it better until we felt satisfied with the results.
NT: How does the Twitter health team operate?
KB: We have core health teams that are fully focused on health and they work on a lot of things I just mentioned, whether it’s account compromised security, you know, misinformation, information operations, so on and so forth. But the highest leverage work that we're doing is not a separate health team working in a corner. It's our health teams working alongside every product team to ensure that every product launch that we make is healthy by design.
NT: There's a health person on every product team?
KB: No, it's more that every product team owns health for their product. In addition to that we have fully focused teams in certain domains within health. For example, like one of our recent priorities has been to shift to be more proactive about enforcing our rules. So having rules and enforcing them is one thing, that's been a burden on our customers to report abuses. For the last year we doubled the amount of tweets that we enforced proactively. Now over half of the tweets that we remove for being against our policies, we do so proactively. That's a dedicated team that's working on ensuring that we're proactively working on removing the most flagrant violations of our policies. But a lot of the work that we do is way more cross-functional than saying that it's just dedicated…
NT: So I want to get into one other big trade-off, which is very related to this, which is the trade-off between privacy and safety. You know, obviously, we're in an era of a privacy backlash, much bigger concern now about privacy than three years ago. What is the situation where you've had to face this trade-off?
KB: The principle that I think matters most to us is that there needs to be an equal value, at minimum equal or greater value exchange, and anytime we ask customers for data, the value that customers get for giving us that data needs to be commensurate to what they are giving up. I think customers need to fundamentally have a choice. That has not always been the case in our product experiences. And that we need to correct and ensure that it always is.
NT: What do you mean it's not always been the case?
KB: I think oftentimes in the product experiences, apps will ask for more data than they responsibly need for that circumstance.
NT: So back in the early days where by default, you would ask for location data?
KB: Yeah, you open the app and they ask for latitude and longitude even though we’d only use it if you are tweeting and wanted to share your location. So I think there's lots of examples like that where a developer can choose to go one way or another in terms of like, well, this would be convenient to collect this data, or taking the harder path and saying, “What is the customer actually getting? What problem are we solving for them? What is the data that allows us to solve that problem? Do we give them the choice? And are we sure that we’re only asking for that data when we actually need it?”
That's a mindset shift from how, I think, companies in general have operated. There are a few exceptions. Some companies have been great at this from the get-go. Apple has always been great at this. But I think we are really trying to shift our principles around how we think about privacy. Because I think only once we do that can you actually say that you're building your products with privacy by design. And that's been a shift for us. We have not always operated that way.
NT: All right, last few questions. So in that conversation thread, I asked what I should ask Kayvon, I want to pull my favorite question from them, which is, should there be a time delay in how people respond to tweets? Because often, I'll see a tweet and I'll just want to write something just furious. But if I waited three seconds, maybe I’d calm down a little bit. What about that?
KB: So do you want a time delay after you tweet something?
NT: I don't know, this was a response to the tweet I put out this morning saying what should I ask Kayvon. But I think what the questioner was asking is, if somebody tweets something and you're going to reply, as you start to reply, it sort of pauses for three seconds, or perhaps it even analyzes it and nudges you and says, “Are you sure you want to say that? That doesn't sound nice.”
KB: Yeah. I think there's two really interesting ideas here. One is the sentiment meter. Like if it detects something super aggro or angry, it warns you and says, “Are you really sure you want to say that? Consider being more nice.” Something along those lines, I think, is interesting. And then the other one that we hear a lot from customers about is the undo feature: I just tweeted something, maybe I want to take it back.
NT: How is that different from delete?
KB: It's fundamentally the same. It's a nudge to delete. It's Gmail Undo, if you've ever used that. So, again, all these have pros and cons. That's the interesting thing about the scale of Twitter. Any such decision that we make has a trade-off, but I think there are interesting aspects to both of these solutions.
NT: And then Jack tweeted, maybe two or three months ago, that Twitter is starting a project called Blue Sky, where you’re going to try to figure out whether to just be entirely open source, there's going to be an independent product team working on that. What can you add beyond the initial Jack tweets about where that is and what it means?
KB: So what I would say is, the goal of that product is not for Twitter to be entirely open source. What we’re interested in exploring is whether there is an open standard for social networking, that would make the world a better place, not Twitter a better place. We’re also interested to see if Twitter can benefit from it. But fundamentally, we believe that social networking as a protocol is an interesting concept that Twitter could benefit from. A lot of things would need to go right in order for Twitter and other people to benefit from it. But what we would like to invest in is giving that protocol a chance to exist. And so that's the work that we're doing, is researching—we're hiring a lead and a team to explore an open standard for social networking. You could imagine Twitter potentially being a client of that open standard, which is different than like, should Twitter be open source?
NT: Right. But you can also imagine, from the way you described it if I'm hearing you right, that they could come up with an open standard that is good for the world, and that could eventually create some kind of a fork in Twitter or a competitive Twitter. Is that right?
NT: And then you have to deal with that.
NT: OK, last thing. It sounds like from what you've been talking about onstage today that you want some pretty fundamental changes in how Twitter works. And the stuff that's come in the last six months has all been cool, and you've gotten a lot of credit for shipping products much faster than in the past, but there hasn't been a huge transformative change in, I don’t know like, demetrification or nudges to prevent people from shitposting. Can we expect in the next six months to a year that some big stuff is coming?
KB: So first, I would say one of our goals is to make bold changes to the product at a faster and faster pace. So hopefully you all have seen us increase our pace. And hopefully you've seen us take on bigger and bigger problems. I do think that we have taken some really big swings. Topics are a fundamental shift in the product. It's the first time in 13 years you can follow something other than an account. It's a big deal, and I think it will be transformative for the service. It's in the early days of it, obviously. But absolutely, you can expect us to continue tackling some big problems in the health space and the conversation space. There's a lot of exciting stuff that we're working on. And also we're being very deliberate about, you know, developing in public and being open and transparent about how we're thinking about something even before we launch it. Twitter's always been sort of developed in public. Some of our most interesting and iconic features like the retweet, the hashtag, you know, have come from customers essentially hacking the product in public. So I think we're trying to get into that and be transparent about that.
NT: All right, that's great. Thank you very much to Kayvon Beykpour.