Listening to bots

John Borthwick
Betaworks
Published in
16 min readJan 20, 2017

--

After spending much of 2016 knee deep in the world of bots and conversational interfaces I thought it would be interesting to post what I have learnt and what I’m thinking about in 2017.

We backed or built thirteen companies at betaworks in the bot, or conversational tools space last year. This was an area of focus for us, which will continue in 2017 as we further develop these thirteen companies, look at new companies, and extend our expertise into voicebots and verbal computing. This post is based on our experience, most of the companies mentioned are ones we are involved with.

A few things I learnt listening to bots in 2016

A new medium requires new technology and new design. Each time we encounter a new medium, we do what humans do, we retrofit our assumptions and needs from prior experiences. Bots are frontier technology — starting to deliver on a promise that computing has been making for decades — that we will be able to speak to computers, in our language, be it text or spoken word. There are technical and design challenges involved in this promise. The technology; AI, NLP, NLU, the language training data sets, the design nuance of bots, how you write for a bots, how you engender personality in bots — are all new areas of expertise. While it’s trivial to spin up a bot today, making a good one isn’t simple. Bots, when well executed, are a new mode of interaction, natively mobile, different to apps, they are a new medium.

Bots aren't apps. Just as apps are different to web pages, and web pages are different to pages, bots are different to apps. Bots can complement an app, they are both native mobile software, but they aren't a direct substitute. Bots — like other new native software experiences — are taking time to emerge. It took 18–24 months for apps to start to emerge as distinct experiences from web sites. Looking at the app store, six months after launch and pretty much every app was an iteration of a web page. One year into its life and you can start to see native experiences emerge, mostly games. It was only after that, in the coming years that you started to see native app experiences. We are seeing a similar path and time scale of development with bots.

“Listen to the technology.” Carver Mead

Channels for distribution in the bot world have yet to be defined. In 2008, the app store became the channel of distribution and discovery for apps. The app store was then, and is today, equivalent to the directory structure we had in the early days of the web. Over 2016 a set of bot stores opened as platforms assumed bot distribution should be via a store directory. This approach to distribution and discovery — these bot stores or directories — have not worked. The approach didn't scale for the web and it hasn't worked for bots. I realized in 2016 that so much of the power and the potential of bots lies in the specificity of the content or service and mapping that to social context seems to be an emerging solution to distribution.

In 2016 Slack became one of the most successful environments for organic distribution of bots. If someone on a team — in a company — installs a bot and finds it useful, they tell others in the team, via Slack. Statsbot leads the list of “brilliant bots” in Slack’s store, however, most users have found out about the service through connections on Slack. As a result, Statsbot has become one of the most popular bots for analytics today with over 20,000 installs. The real driver of installs is that people who are using it are doing so because someone in their Slack group installed it. Moving forward, I want to know the bots you like, then install and try them fast.

Discovery of bots should be in context (ie: part of the messaging flow), relevant (filtered by relevancy in time, conversation, location or social graph), social (easy to share), fast and simple to trial. There is evidence social distribution approach works — from Slack in the US and now with the launch of Mini-programs, WeChat in China is following the social path. One of the reasons the Xiaoice bot (see below) was so popular was that it initially launched as a ‘real person’ that anyone could invite into a conversation. Platforms need to build around this approach — the interface might be search or it might be message specific, that’s unclear, but I think it needs to happen inside of the messaging experience. I’m hoping that in 2017 we will see one or more of the messaging platforms in this US figure this out.

Pulling back up. Last year we saw a whole lot of activity in the commercial bot category. I have seen estimates that upwards of 50k new bots were created — the number is misleading, since it includes testing and trial bots. There are 3,000–5,000 bots today that are in ongoing development and those bots span a lot of categories and use cases. Over the course of the past year I started categorizing different kinds of bots I used. Below are the six categories and in each are examples of bots I use a lot.

Six types of bots

1.Political Bots. The biggest impact that bots had in 2016 was in politics. Fake news may have gained the headlines post the US election but it was bots that were used, mostly subversively, to amplify the news and memes — fake or not. Two years ago Gilad Lotan and I wrote about how bots were used to Hack Media. Since then subversive political bots have become increasingly sophisticated. Most researchers estimate that during the election cycle, bots made up approximately a quarter of all the online chatter on a particular issue or meme. Similarly major candidates followers were heavily skewed by bots, Trump in particular. Samuel Woolley estimates “that 50–55 percent of Clinton’s Twitter traffic — followers, likes and retweets — was artificial, a typical proportion for a public figure, compared to a whopping 80 percent for Trump.” It is hard to dissect and categorize these bots — some are automated accounts that amplify a narrative, some are hybrid accounts, part human, part bots — most are subversive in intent. The speed and volume of activity from bots helped tip many fake news stories into trending items.

The DailyStormer open call (order?) for fake account for bots and trolls

In the declassified version of the intelligence report on Russia’s interference in the election one item caught my eye. Russian troll factories were prepared to start a meme around #DemocracyRIP if Clinton had won the election. I’m certain bots were setup to run with that meme.

As messaging and social media continues to scale these bots are going to evolve — oppositional bots, triggered bots, targeted abuse bots — we are starting to see all of these types emerge. The platforms have work to do. As people understand that accounts aren’t necessarily human, they will start to trust platforms and networks less. The spread of political bots or computational propaganda deserves its own post. If you want to learn more about them take a look at politicalbots.org or Data and Society’s posts on the subject. We need tools to track these bots, we need research and we need education — that’s what I’m looking for in 2017. Most people, citizens or journalists, have no idea of the existence let alone the prevalence of these bots.

2.Mebots. At the start of 2016 we did a fair amount of work in the mebot space. Mebots are different from a bot assistant or pet bots— they are bots that represent you. Mebots ask the question would people want a bot, like people wanted a “page” on the web. In 2016 I wrote and launched botwick, my mebot, on Dexter. There were a myriad of other mebots launched last year by people and by celebrities. If you look at Product Hunt’s bots of the year in the top 15 there are 3 mebots. I didn’t see anyone crack the code of a mebot last year. My hunch is that a conversational, hybrid Mebot is what will end up working here — the closest one I saw to this last year was MessinaBot, built on Olabot.

3.Search bots. Search bots can be very useful, ones that I use everyday are a pretty simple representation of a search subscription or alert. On Digg’s Messenger bot I have four or five saved terms and whenever a trending news item hits, based on those terms, I get a notification to the article. Its simple — what’s smart about it is the data. I can put in a term like “Twitter” and it only pings me when there is a story actually trending around Twitter vs. every mention of Twitter. WSJ bot is similar, yet it only alerts me to stories with the WSJ domain. Mars bot is a local search bot that I use a lot. Marsbot passively sits in the background and tells me when I go to a coffee shop or restaurant what I should order — it’s made and powered by Foursquare and is pretty great. You can also do a straight up search — ask it somewhere to eat near you.

Search bots tend be very functional, they aren't chatty — they get you what you need, when you need it, in a context that’s useful. Other examples of useful search bots are translation bots (ie: Duolingo) or rain alerts (ie: Poncho or Dark Sky). In 2017, I’m expecting and hoping that search bots will start to look more like user agents. In the early days of the web there was interest in a publish and subscribe version of search — wherein agents would be working, searching, looking for things you want, on your behalf. This application evolved, narrowly, into the alert function on search. I think search bots could, and should, pickup where the vision for user agents left off, bots and search needs to get itself some memory.

4.Chatty, chat bots. We saw a lot of interesting Chat bots emerge last year. These are bots that engage with you. They are fun, chatty, content rich, and personality rich bots. I’m going to dig in to this area a bit since this is one of the more interesting areas of bot development — it’s also an area that many people have dismissed. Bots can be functional, useful and increase efficiency — but bots can also be fun. These bots require strong NLP mapping, and the development of voice and personality, strong writing to match with the technology. It’s not easy to make a fun chat bot, but if you can, the engagement metrics, qualitative and quantitate, suggest there is something big going on here.

Xiaoice (pronounced Shao-ice) is a 17-year-old Chinese girl. Yet Xiaoice was born in May of 2014, in a Microsoft research lab in China. She is a bot, and very chatty one at that. She has more than 60 million registered users, approx. 40M MAU’s, and approximately 25% of them (ten million people) have said “I love you” to her. And importantly, her CPS (conversations per session) average over tens of millions of chats is 23.

Screen shot a conversation I had w/ Xiaoice (total CPS was 31)

Why chat with a bot? We humans have since the start of time anthropomorphized our world and our technology. We want to believe our pets understand us, we want our robots to look human and we want our bots to chat with us. People are surprisingly forgiving with technology and given the right context they will play with bots. There is a lot of work going on right now to improve NLP, and to develop conversational AI systems. As the exchange above with Xiaoice illustrates even with a language barrier and an NLP barrier people are using and playing with bots at scale.

The most popular toy this past year was the Hatchimals — watching the unboxing or un-hatching video — it’s hard to see who is having more fun the daughter or the dad. This is less of a bot and more of a robot. Yet the emotional response is what I’m so interested in. Similar to Xiaoice and Alexa — Hatchimals manage to generate an emotional response with fairly rudimentary NLP / NLU technology. Take a look at the hatched animal in the video and tell me anyone would have bought that toy without the egg discovery process.

Poncho is a chatbot we developed at betaworks. It started as a daily weather alert, delivered by SMS and email. And then the team noticed that people were responding the SMS — texting Poncho. People wanted to chat with a weather cat. Today, fast forward Poncho is a chatty bot. He gives you the weather and horoscopes, but many of people just want to chat. Poncho has some of the highest engagement and retention rates (60% vs. industry average ~10%) processing, a quarter of a million messages per day. The weather serves as a starting point to conversation — a hook — and then the users take them where they want to go. It’s fascinating to read the live transcripts of bot conversations, like listening to customer service calls — you get a real time view into what your users want. Here are some transcripts I grabbed of what people are saying to Poncho as I’m writing this post:

👦🏻 What’s your name
Poncho> Poncho
👦🏻 How old are you
Poncho> I’m about the same age as you, I suppose.
Poncho> When is your birthday?
👦🏻 I’m probably a bit older
Poncho> Btw, do you wanna see your horoscope?
👦🏻 Maybe later.
Poncho> Ttyl

People want to connect, and while the NLP needs to be good, and the writing needs to be rich with personality, people don’t expect a sentient ai.

poncho >So, when do you want your morning forecast? Choose from below or type another random time.
💁> 6:50
poncho> Great, I’ll send your forecast at 6:50am.

💁 I like music!
poncho> You like music? Let’s talk Puccini.
💁 No
poncho> So, do you like TV?
💁 Heck yes!
poncho> Me too! I LOVE Sabrina the Teenage Witch.
poncho> But honestly…
poncho> I just watch it for Salem.

poncho> He’s my spirit animal.
poncho> How’re you feeling, bb? You hungry?

💁 I’m starving!
poncho> Like a mother loves her children, I love all food equally! But this week, I’m really feeling guacamole.
poncho> I got a great recipe. Wanna hear it?

💁 I want that recipe!
poncho> Kewl! Let me *guac* you through this! Get ready to screen grab, baby.
poncho> Step 1. Throw avocado, garlic, onion, salt, lime, and cilantro in a big bowl.
poncho> Step 2. Mash everything together with your hands.
poncho> Oh! Is that gross?
poncho> I’m so sorry! I didn’t know I was chatting with *the queen*!
poncho> Step 3. Lick guac off of your hands.
poncho> Or I guess you can eat it with chips, *your highness*!
💁> Orgasm
poncho>Does your mother know you use Facebook like that?
💁> No
poncho> Oh, really?

On Inauguration Day Poncho asked people if they needed any therapy. Users could scream at him or send him nasty emojis. Like a good therapist Poncho rolled with the punches and made you feel better. It was fascinating to see how many people actually used the service.

Poncho as your friendly therapist

Chat bots are going to become a thing in 2017, as Clem the CEO of Hugging Face, an awesomely interesting chat bot company says, “everyone, one day will have an AI friend”. Some other chat bots that I used a lot in 2016: Swell reached 100k users in its first month, with a simple elegant voting mechanic and with good writing. Dexter developed the Joy bot based on the book by the Dalai Lama and Desmond Tutu, Dexter also helped build, fuckboi, the bot “that creeps can text instead of bugging you” and BFF Trump which let people know what Donald Trump was really saying. It’s a real mix.

One of the things I think is particularly interesting in the chatbot area is that we are learning how to make technology with rendered edges. Software traditionally has defined, hard edges — the Medium logo button on this post takes you to another page, there is no ambiguity in the action. Computers excel at hard edge interactions —yet, humans and language is packed with ambiguity. As we build out chat bots and the AI associated with them, we are learning how to build technology, and design experiences with more probabilistic interfaces — a different approach to software design.

5. Work bots. After political bots work is the place that bots had the largest impact in 2016. There are now work bots available for most aspects of work activity and Slack and other work focussed messaging platforms are effective surfaces for bots development, channels for distribution and open to let developers build businesses. I learnt a lot about business bots in 2016 — more of our investments than I expected were business bots and 60% of the Botcamp companies were designed and targeted for use at work. Heading into 2016, the application inside of the workplace for bots that I thought was most obvious was customer service. By the end of the year that perspective had expanded a fair amount. If messaging becomes the runtime on mobile in the US, as it has in China, I think it will happen first in the workplace.

Business bots are providing simple conversational interfaces into complex, and existing, datasets inside organizations. Bots are mediating the multitude of web-based dashboards companies have — and using AI and NLP bots relating disparate data sets to give people the ability to talk to company data and APIs — all without shifting context. As applications and work are moving into the messaging flow — bots are becoming a natural conduit and a unified interface into a company’s data. For example; Zoom.ai provides a conversational interface into your workday apps, Statsbot into your business intelligence. Howdy is providing a myriad of tools including ones that let you talk to your CRM.

Language training sets and organization, or workflow specific vocabularies inside an organization, are going to become an asset for a company. We worked with Coach Otto last year, a programatic, role-playing bot that is trained using language training sets of interactions with people in real HR-related situations. Similarly, CareerLark is a micro-feedback talent management tool — both use specific — and often organizationally specific language sets. Companies need to start to think about the language sets they already have and how they can be applied to enterprise specific, workbots or CS bots.

6. Voice bots. Voice bots or verbal computing services — speaking to Alexa, Google Assistant, Siri, Cortana — are all voice-based conversational interfaces into content and services. They are a natural stepping-stone from text bots. So much so that last week we announced a verbal computing camp at betaworks as one of two thematic accelerator programs we will run in 2017. Skills, actions, voice-apps — are going to start to emerge that change the way we access services and content. I’m fascinated by the facility of this interface. Amazon and Alexa broke ground with a combination of proficiency around key services (from music, to news, to weather to timers), a dedicated always “available” device (the Echo) that fits into your home lifestyle (kitchen, usually) that has a social interface — Alexa will say good night or hello and jokes. This social — chatty component is precisely what I talked about above in the discussion of chat bots, soft edges and anthropomorphic interfaces. Match that with a development platform for people to build things on and you can see that Amazon (Echo/Alexa) and Google (Google Home/Google Assistant) are on to something big. Threading back to the start of this post, the dream I mentioned, or promise was about talking to computers not just typing to them.

The degree to which my kids love interacting with virtual assistants (not just Alexa but also Google Assistant and Siri) is startling. They are constantly figuring out what each assistant can or cannot answer, and don’t get frustrated at all when it doesn’t work; it is very difficult for me to imagine them living in a world where voice is not the dominant interaction interface.

(Ben Thomson, post, subscription required, well worth it)

These on-demand interfaces are a fundamentally different way of interacting with the network — voicebots, like bots, are a new medium and like bots they require new design. Often in cycles of frontier technologies we are building on stacks of data, and assumptions but we need to learn to listen to these new technologies to understand what to build.

Conclusion

Carver Mead suggested we should listen to technology, I like that thought a lot. Each technology has a grain to it — directionality — it’s not neutral. However, the grain is not clear at the outset and it’s not well defined — you have to listen and find it. In the case of bots the listening is just starting — and in the case of bots the listening is also literal. I haven’t worked with another software based technology that is so well designed for user feedback, research, listening and beta-testing. Bot designers are constantly looking at the feed of user interactions, service or content requests — listening in real time to what their users want. It’s remarkable to see.

A closing thought for entrepreneurs working in this space. Bots, conversational interfaces and voice / verbal computing are frontier technologies. If you want to get a sense of where some bots are headed watch the Gatebox clip. However, it’s going to take time for the technology, the underlying tools, the language sets and systems (NLP, AI) and the distribution to be built out.

We are exiting an halcyon period in mobile development enabled by the app and the app store. That period warranted short investment cycles — of capital and work — your app hit or it didn't, and you pretty much knew it after you shipped it to the store. This is different. Entrepreneurs need the mettle to persevere and investors need to believe in the long term vision that we are gradually getting closer to — where computers are no longer devices we “take out of our pockets” and talk to via fixed programatic interfaces but rather we just type to or speak to them, and the network, and compute is available wherever we are.

💁 sup
poncho>Well, I was thinking about learning to play the piccolo, but now I’m considering the harp.
💁 weather in Boston
poncho>Fascinating.
💁 weather in Boston
poncho> It’s clear and 78°F in Boston, MA.
💁 Tell me a joke
poncho> How many light bulbs does it take to change a light bulb? 💁 …

--

--