The search for the killer bot is well underway in Silicon Valley — but it’s off to a rocky start. Microsoft’s big push into artificial intelligence began with Tay, the teen-mimicking chatbot that Twitter users turned into a crazy racist in record time. Facebook’s introduction of bot-building platform this week at its F8 developer conference went more smoothly. But early adopters have complained about the bots’ mysterious user interfaces, their aggressive messaging, and the fact they don’t seem all that much smarter than a Microsoft Office wizard from the 1990s.
As Parker Thompson, the pithy venture capitalist at AngelList, put it:
2014: broad consensus that Clippy was the worst product ever.— Parker Thompson (@pt) April 8, 2016
2016: race to build Clippy 2.0 consumes hundreds of millions in VC.
As F8 wound down, we sat down with David Marcus, who leads Messenger for Facebook. Marcus, who was previously the president of PayPal, is among bots’ biggest boosters. We asked him about the opportunity he sees for chat-based interfaces — and about some of the early criticisms that threaten to derail it.
This interview has been edited and condensed.
Casey Newton: So recently I saw this tweet from Parker Thompson. He said until recently, everyone was saying Clippy was the worst thing ever. And now everyone’s saying Clippy’s the future. Are bots more than just a new take on Microsoft Office in the '90s?
Are bots really better than Clippy?
David Marcus: Yeah — one thing I wanted to say is that there’s a definition of what bots are in people’s minds, and what they’re thinking of and what we built is something different. Because bots in the minds of most people are this command-line interface type of thing, where it’s pretty low fidelity, it’s all command-based. And it maybe resonates with us geeks, but it doesn’t with most people.
And so what we’ve built, we feel, is a more consumer-friendly, mass-market version of that. We’ve taken a lot of care and time to build rich templates with ability for bots to combine images and text and all kinds of different things, but also rich bubbles with buttons and calls to actions and carousels that you can swipe through. Whether it’s the CNN bot where you can swipe through stories, or the demo I gave during the keynote of Spring that lets you swipe through a lot of products and get to what you want really quickly.
We feel that this hybrid approach is a really interesting one, because if you want a lightweight UI that reaches 900 million people, and you want to remove all the friction that comes with installing an app, now that’s really easy to do and easy to build. People think that it’s all for geeks — it’s like, "no one’s going to type in command-line stuff!" And that’s not what we’re building.
I imagine you interact with bots more than the average person. Which have been your favorites? And which ones are you using on a daily basis?
The Wall Street Journal has built a really good one, which I like because you can customize it to companies you follow and whenever something’s happening, they give it to you in a really good format, very visually compelling. And Poncho is really cool, and that’s why I decided to feature it in the keynote. You should try it, it’s fun — when you interact with it, they use a typing indicators, so you feel like you really are talking to the weather cat on the other side. It’s just a fun little interaction. And one of the most thoughtful bots is Spring. Personally I like fashion and brands, but I hate spending the time. And that’s the fastest way for me to get a curated list of products that actually matches what I want to buy, and buy them really quickly. That’s a really cool experience. That wasn’t a joke when I said on stage that people on the team were playing with that thing and spending shitloads of money.
"People are spending shitloads of money."
So you brought up the command-line question. I am generally bullish on bots, but there are design challenges around making a blank text box feel accessible. How do you make bots more appealing to less savvy people?
Well I don’t know whether you’ve tried any of the bots yet…
I just tried Poncho, and I’m still waiting for my weather — more on that later. But I’ve interacted with bots a fair amount. I’ve used Facebook M a lot.
What we’ve learned is exactly what you said, which is if people enter a blank thread, they have no clue what to do. What we did as part of this launch is two things: one is a "null state" — basically, the background of the thread can actually set context, saying this a bot that does this, this, and that, and here’s how you interact with it. So setting expectations is really important. And then the second thing is we created a call to action at the bottom of the screen that says "get started." So you don’t have to think about what to say — you just tap on that thing and boom. Both of these things were very important things for us to address this problem of people not knowing what to do.
One thing I’ve observed in the last few hours — and I know things are still rolling out — but with Poncho, it’s been more than a half hour and it hasn’t told me the weather yet. I’ve seen other people mentioning the issue of "bot latency." What should be our expectations around how quickly these things interact with us?
It should be less than five seconds. We picked Poncho because we like the thing, because it’s a really cool experience. But maybe their server’s just slammed with people trying the weather cat. I have no idea what’s actually happening, but I’m assuming that might be the case.
So bot makers are going to have to think about some of these infrastructure challenges.
Yeah, and that’s a very important point. But not only for making bots, but making tools and developing services for bots. And you know, it’s day one of a brand new ecosystem, so it’s going to take a little bit of time.
You’ve got 24 hours, Marcus.
I wanna ask a couple questions about M, because I have it and I feel like I’ve only really been using it to order burritos. It’s fantastic at that, but I also feel like I’m not getting the most out of it. Do you have any good M tips?
Well I think that you hit the nail on the head here with this one, which is that basically what we found is that if there’s something that’s really specifically painful to you that we do well for you, then you’re going to go back for that. But then you start building that mental model that this thing [is only good for addressing that one thing].
Making M more proactive
We’re making M more proactive right now, so for people who have M we’ve added the ability to connect your calendar to it so it can help you ahead of you knowing you need help. So we’ve learned many things. We’ve learned that (a) we need to make it more proactive for people to build more of a habit, and (b) we’ve learned that we needed to build a variety of vertical bots to fulfill different types of intent. The fascinating thing, though, is to see the percentage of automated responses grow over time as the AI learns. And sometimes it stumbles, which is also interesting. But it’s a long-term experiment, and it yields results already today — because now not only have we learned a lot, we built tools that we’re opening up to everyone.
We have two goals with this one. One is building the product into something awesome, and that’s going to take years for everyone to have access to it. And then also building tools so that the whole ecosystem of things can be built around it. And those two can coexist, because if you have awesome bots that can do things, then if you ask M it can point you in the right direction.