In early 2014, I single-handedly created Yahoo Mail’s Autocompose feature in under two weeks for a highly succesful Valentine’s Day marketing campaign. I then patented a machine-learning iteration of Autocompose—nearly two years before the launch of Google’s Smart Reply. (And now ML-generated text is everywhere.)
I joined Yahoo’s marketing department as a Director in the autumn of 2013, having just put in a year and change at Google Creative Lab as a Creative Technologist. I’d started feeling out this job opportunity earlier that summer as ex-Googler Marisa Mayer celebrated her first anniversary as Yahoo’s CEO. Yahoo was sporting enormous renaissance potential and after joining I often encountered other former Google folks eager and excited to help restore the big Y to its prior glory. There were also “boomerangs”—former Yahoo employees drawn back into the fold by the company’s potential for rebirth. Optimism was in the air.
Yahoo Mail was in trouble
I wasn’t on a product team, but as I understood it—as an outsider—making headway was an interesting challenge: The longterm effects of infrastructure and strategic decisions authored in previous eras were bubbling to the fore long after those authors themselves had departed. In the early days of 2014 one pain point in particular was Yahoo Mail. Though the product itself was evolving rapidly under expert direction, some of that legacy infrastructure had caused service outages over the past few weeks, greatly frustrating users. As a result Mail took a hammering in the press and on social media.
As the Mail team worked tirelessly to upgrade and replace the broken bits, we in marketing were tasked with brainstorming a “charm offensive” for Valentine’s Day that could help users “rediscover their love” for Yahoo Mail. If you’ve never worked in marketing that premise may sound ridiculous. And if you were one of the affected users it might even be infuriating. But as a marketing team we were fighting two fronts: First, for the sake of the product itself we needed to change the conversation around Yahoo Mail. Secondly—and this was a more subtle, personal cause of mine—the Mail team was working tirelessly to not only shore up those loose ends but to launch new features. It’s difficult to give it your all, day after day, if you feel demoralized. I felt some empathy for those engineers and wanted some of that “rediscovered love” to reflect back onto them; to help re-energize them for their work ahead.
Our team was assigned the Charm brief on Thursday, January 23rd—just three weeks before Valentine’s Day—and because we needed our campaign to launch ahead of the actual holiday our real timeline was down to a mere two weeks. In those two weeks we’d have to come up with ideas, pitch them, receive approval on one, and race it through production. All this on a budget that could be described as “modest” at best.
It’s funny how you are the tools you bring to the table. As we pitched Valentine-themed ideas amongst ourselves I found that the video director thought we ought to make a clever video campaign. The poster designers sketched on poster campaigns. There were passionate arguments for humorous celebrity endorsements, a romantic dinner package giveaway, free movie tickets, and so on. But all of these things seemed like distractions from the product, rather than a means of showcasing it. And the execution of these ideas would surely exceed our budget. The clock was ticking.
I had an idea that felt a bit different from the rest. It required no additional budget, could be done in-house quickly, needed no media buy, and even had the potential to make money. To explain, let’s back up a bit.
iQuit: My past personal project
In the summer of 2006 I was 24 years old, living in Brooklyn, and had just been accepted into Yale’s graduate graphic design program. I was both excited and anxious as I researched financial aid, New Haven housing, and student health insurance. I became completely distracted from the scattered freelance design projects I was wrapping up and these last sources of income that I should have been thankful for suddenly felt like burdens on my remaining time in New York. I knew I was on the precipice of a major life change and I badly wanted to shed whatever remaining duties I had immediately and enjoy this last summer of my pre-grad school life. Self-indulgent, perhaps, but I was young and relatively unburdened by real responsibility.
I didn’t leave my clients hanging but I did procrastinate by packaging my restlessness into a one-off eruption of Web sarcasm. It was the height of the glossy “iSomething” craze as Steve Jobs rocketed Apple back into cultural relevance with iMac, iTunes, iWork, iLife, and of course—the crown jewel—iPod. (iPhone and iPad were still secret skunkworks projects waiting to someday be unveiled.)
I named my personal project iQuit.
iQuit generates a formal letter of resignation from its library of pointed, yet ambiguous, sentences. Perfect for quitting almost any sticky situation. Your job. Your relationship. Your friends. Get out of anything instantly with iQuit.
It wasn’t just a visual joke, it was a functional Web app. It would truly generate a unique resignation letter on the fly and email it to your intended recipients. iQuit had real users! (Or at least it had curiosity seekers who tried it out in jest.) I lost count of how many times friends “quit” me for a laugh.
The faux testimonials well illustrated my dark sense of humor:
Proud Parent. My wife and I have always preferred our older son to his younger brother, but infanticide is so expensive. Thanks to iQuit I’ve cut the excess baggage and have a family I’m truly proud of!
Delighted Designer. I used to do corporate advertising until I realized having a soul is nice. I fired up iQuit and BAM! Now I can starve with pride.
Smirking Single. I knew I wanted out of my relationship, but couldn’t find the words. iQuit made it quick and easy, just like me!
The randomly-generated resignation letters carried an absurdly conservative tone:
I was rather pleased with the result. Even complete strangers were enjoying my joke: It became a hit online, and garnered a second round of attention when I last updated it a few years later. In retrospect my summer of 2006 might have been better spent making code-friends around the Harvard campus, but hindsight is always 20/20, right?
Birth of Autocompose
So in early 2014 Yahoo Mail needed a way to speak for itself ahead of Valentine’s Day. “Charming” was the goal, but I’d settle for “engaging.” I knew I wanted Mail itself to be the center of attention, not a celebrity spokesperson and not a slew of charts and graphs about increasing stability or number of active users. I wanted this to be an emotional connection.
What’s Mail good for? Writing letters. What kind of letter might one write on Valentine’s Day? A love letter. And that’s when it hit me—writing love letters is hard:
“Dear Jane, you are so…neat.” This slide of my handwritten scrawl prefaced my pitch to the team.
What if Yahoo Mail could write it for you? I quickly made some visual mockups of what this feature might look like incorporated into Yahoo Mail. But that didn’t feel like enough to sell the idea. I needed a demo. Fortunately in addition to graphic design I’ve got a knack for programming so I went to work writing a bookmarklet that could inject this functionality right into a browser window running live Yahoo Mail. This meant I could go beyond merely driving a demo during a presentation, I could also share the bookmarklet to each of the team members or stakeholders and they could easily test drive it themselves. Within a day I had coded a working prototype and written enough copy for it to craft thousands of compositions. (I was wearing all the hats.)
I called it Autocompose:
With each click of the Autocompose button a unique, randomly-generated love letter appeared in your compose window, ready for you to edit or send straight away.
It was quirky. It was absurd. But it was engaging. Each time you clicked the Autocompose button it generated a new random love letter right in your Mail’s compose window. What was it going to say next? Do I dare click the Send button?
My little demo was not far off from the final product. It needed a little visual polish and the code needed vetting for production use, but the real point of concern for me was the text itself.
Unique love letters for every user
Autocompose relies entirely upon the tone of the generated composition and to me this seemed like a copywriter’s dream project. I wanted the number of possible permutations to exceed the number of active users so that in theory a unique love letter could be generated for each one of them. But for the output to feel cohesive a copywriter can’t just jot off a list of individual lines and call it a day—they need to see their lines in context together as they write them; to gut-check random permutations as they draft out new ideas. To facilitate that I threw together an “Autocompose Drafter” complete with its own rudimentary programming language for creating categories—like “Subject lines”, “Openers”, “Body sentences”, etc.— and Mad Libs-style word replacements. Our copywriter could enter her lines as a list on the left half of the screen, then click the Autocompose button repeatedly to see randomly-generated love letters appear on the right.
Autocompose wasn’t the artifact of a corporation being tone deaf to the nuances of romantic expressions of lust and admiration. Autocompose was us having a laugh — and letting our audience share in that laugh. I didn’t intend for Mail users to seriously automate their love letters. I intended for them to discover this temporary feature — funny for its absurdity — and to feel compelled to experiment with it for a while, then to share it with their friends.
And it worked.
Changing the conversation
Autocompose won the pitch bakeoff and when we launched ahead of Valentine’s Day we changed the online conversation around Yahoo Mail. Online sentiment can be measured in broad terms like negative, neutral, and positive and is based on the content of social posts from Twitter, Facebook, and so on. “Neutral” is more valuable than one might assume because it includes posts that merely link to something which is — when void of negative sentiment — a form of endorsement. Sentiment for Yahoo Mail had been decidedly negative. After launching Autocompose sentiment shot up to 90% “neutral to positive” across all social media platforms. We charmed.
Because the campaign took place in the product itself any “links to the campaign” meant links directly to Yahoo Mail. And as users were clicking away on the Autocompose button we were serving up ads in the sidebar. Mail was literally making money off of its own marketing campaign. Autocompose was iQuit, but put to better use.
Machine learning and style transfer
From a marketing perspective that’s the end of the story. But I’ve never been content to stick to one discipline. I find it much more exciting when marketing, product, and experimental research can intersect and play off one another. Anything less is a waste of potential.
Yahoo had recently acquired a product called Summly that ingested scores of news articles on a given item, then used machine learning to automatically generate a concise summary of that news item. I was impressed with the visual simplicity of the new app they were developing within Yahoo—later released as Yahoo News Digest—and this made me curious to know more about what Summly was and how it worked. The history of machine learning—from Markov chains to Word2Vec, to artificial neural networks—became part of my daily curiosity.
I was filled with questions. What if Mail could automatically summarize each conversation thread of your inbox? What if you could use your previous email exchanges with someone in particular to autocompose a reply to their next message? What if you could extract the style from one particular message author and transfer that style to another composition? (Today, in machine learning that is literally called “style transfer” and it makes for some intriguing visual results.) What if Autocompose could better predict what sort of reply you’d want to use based on the context in which you’d selected previous autocompositions?
Strange things could be possible. Give me a happy birthday email in the unraveling investigative style of Rachel Maddow. Setup an out-of-office response in the combative rhyming rhythms of O’Shea Jackson. Apologize to my friends for not making it out this weekend because I have the flu and I’m too dazed to write a proper apology myself.
My first artificial intelligence patent
It was a fun thought experiment and I felt the idea had legs. I put together a deck describing the concept—how it could function and how it might be integrated into Mail. Yahoo’s patent team felt my ideas met their criteria for pursuing a filing and before long I was taking phone calls with patent lawyers and fielding their helpful interrogations. By mid-August we had the blueprint for a new “smart” version of Autocompose filed with the US Patent and Trademark office as “Expressing statements in messages with personalized expression style.” (I’m listed as the sole inventor but it’s important to note that the intellectual property was owned by Yahoo and I have no financial stake in its use, defense, or licensing.)
It’s not often you get a patent off the back of a marketing campaign—let alone a successful one that snowballed from concept to launch in just two weeks. I was convinced this was only the beginning of what could be a string of offbeat wins for Yahoo. I aimed to leverage Autocompose’s success to form and lead a new “Marketing Innovation Group” that would exist somewhere between traditional marketing, product, and experimental research. With a small group devoted to this kind of thinking, how many fun and impactful interventions could we bang out in a quarter? In a year?
On cross-discipline creativity
Real creativity is having novel ideas of value. This may sound obvious, but to create novel ideas of value you have to actually be engaged with the culture and tools you’re hoping to invent with. And because it takes about a thousand well-intentioned bad ideas before you can arrive at a good one, I make a habit of having as many bad ideas as possible as often as possible. Play is of the utmost importance here. You have to work really hard to create good “bad ideas” and making playful experiments is the best way to generate that back catalog of important learnings. If your team isn’t “wasting time” with play they aren’t learning. The role of a creative manager is to case-by-case assess what aspects of that play are producing value and to know when time and budget dictate that play must finally be put aside in favor of execution. (I cannot articulate this better than Monty Python’s John Cleese did in his 1991 lecture on Creativity in Management.)
You don’t have to become an expert in the field you’re conceptualizing for—and really how could you without dropping your current practices and delving exclusively into that field for years of intense immersion?—but you do have to be excited about the process of investigation. I love learning as much as I can about something new and using that excitement of discovery to guide what aspects of this new thing I can use for marketing or business strategy. And it is always incredibly satisfying when I have the opportunity to sit down with experts in a field and feel empowered to ask the dumbest questions imaginable; receiving thoughtful, digestible responses in return.
Increasingly this seems to be what corporations hire me to do. Whether it’s data animation, quantum computation, virtual reality, machine learning, CRISPR, or something else entirely—my value comes from my excitement to learn, to experiment, and to eventually distill that play into something more immediately useful. We are the tools we bring to the table. As a Creative Director I’m more than some graphic design degree, or years of coding experience, or weekend writing, or business strategy… I’m bringing all of these interests to the table, and more. My favorite people to work with are just as varied, just as playful, and always more talented than I am.
Google Smart Reply
How I was lured away from Yahoo back into Google Creative Lab is a story for another time. But two years later there I was, once again working from Google’s Manhattan office and half-watching the keynote livestream of their annual developer’s conference, Google I/O, when something caught my attention. The Google employee on the main stage was touting this new feature called “Smart Reply” that had been integrated into their chat program and would soon rollout to Gmail as well. As he described how it worked I smiled to myself.
Smart Reply sounded a lot like Autocompose.