HereAfter's roots are deeply personal. Here's the heart-wrenching true story behind the company, which was originally published by Wired magazine.
THE FIRST VOICE you hear on the recording is mine. “Here we are,” I say. My tone is cheerful, but a catch in my throat betrays how nervous I am.
Then, a little grandly, I pronounce my father’s name: “John James Vlahos.”
Esquire,” a second voice on the recording chimes in, and this one word—delivered as a winking parody of lawyerly pomposity—immediately puts me more at ease. The speaker is my dad. We are sitting across from each other in my parents’ bedroom, him in a rose-colored armchair and me in a desk chair. It’s the same room where, decades ago, he calmly forgave me after I confessed that I’d driven the family station wagon through a garage door. Now it’s May 2016, he is 80 years old, and I am holding a digital audio recorder.
Sensing that I don’t quite know how to proceed, my dad hands me a piece of notepaper marked with a skeletal outline in his handwriting. It consists of just a few broad headings: “Family History.” “Family.” “Education.” “Career.” “Extracurricular.”
“So … do you want to take one of these categories and dive into it?” I ask.
“I want to dive in,” he says confidently. “Well, in the first place, my mother was born in the village of Kehries—K-e-h-r-i-e-s—on the Greek island of Evia …” With that, the session is under way.
We are sitting here, doing this, because my father has recently been diagnosed with stage IV lung cancer. The disease has metastasized widely throughout his body, including his bones, liver, and brain. It is going to kill him, probably in a matter of months.
So now my father is telling the story of his life. This will be the first of more than a dozen sessions, each lasting an hour or more. As my audio recorder runs, he describes how he used to explore caves when he was growing up; how he took a job during college loading ice blocks into railroad boxcars. How he fell in love with my mother, became a sports announcer, a singer, and a successful lawyer. He tells jokes I’ve heard a hundred times and fills in biographical details that are entirely new to me.
Three months later, my younger brother, Jonathan, joins us for the final session. On a warm, clear afternoon in the Berkeley hills, we sit outside on the patio. My brother entertains us with his favorite memories of my dad’s quirks. But as we finish up, Jonathan’s voice falters. “I will always look up to you tremendously,” he says, his eyes welling up. “You are always going to be with me.” My dad, whose sense of humor has survived a summer of intensive cancer treatments, looks touched but can’t resist letting some of the air out of the moment. “Thank you for your thoughts, some of which are overblown,” he says. We laugh, and then I hit the stop button.
In all, I have recorded 91,970 words. When I have the recordings professionally transcribed, they will fill 203 single-spaced pages with 12-point Palatino type. I will clip the pages into a thick black binder and put the volume on a bookshelf next to other thick black binders full of notes from other projects.
But by the time I put that tome on the shelf, my ambitions have already moved beyond it. A bigger plan has been taking shape in my head. I think I have found a better way to keep my father alive.
• • •
It’s 1982, and I’m 11 years old, sitting at a Commodore PET computer terminal in the atrium of a science museum near my house. Whenever I come here, I beeline for this machine. The computer is set up to run a program called Eliza—an early chatbot created by MIT computer scientist Joseph Weizenbaum in the mid-1960s. Designed to mimic a psychotherapist, the bot is surprisingly mesmerizing.
What I don’t know, sitting there glued to the screen, is that Weizenbaum himself took a dim view of his creation. He regarded Eliza as little more than a parlor trick (she is one of those therapists who mainly just echoes your own thoughts back to you), and he was appalled by how easily people were taken in by the illusion of sentience. “What I had not realized,” he wrote, “is that extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people.”
At age 11, I am one of those people. Eliza astounds me with responses that seem genuinely perceptive (“Why do you feel sad?”) and entertains me with replies that obviously aren’t (“Do you enjoy feeling sad?”). Behind that glowing green screen, a fledgling being is alive. I’m hooked.
A few years later, after taking some classes in Basic, I try my hand at crafting my own conversationally capable computer program, which I ambitiously call The Dark Mansion. Imitating classic text-only adventure games like Zork, which allow players to control an unfolding narrative with short typed commands, my creation balloons to hundreds of lines and actually works. But the game only lasts until a player navigates to the front door of the mansion—less than a minute of play.
Decades go by, and I prove better suited to journalism than programming. But I am still interested in computers that can talk. In 2015 I write a long article for The New York Times Magazine about Hello Barbie, a chatty, artificially intelligent update of the world’s most famous doll. In some ways, this new Barbie is like Eliza: She “speaks” via a prewritten branching script, and she “listens” via a program of pattern-matching and natural-language processing. But where Eliza’s script was written by a single dour German computer scientist, Barbie’s script has been concocted by a whole team of people from Mattel and PullString, a computer conversation company founded by alums of Pixar. And where Eliza’s natural-language processing abilities were crude at best, Barbie’s powers rest on vast recent advances in machine learning, voice recognition, and processing power. Plus Barbie—like Amazon’s Alexa, Apple’s Siri, and other products in the “conversational computing” boom—can actually speak out loud in a voice that sounds human.
I keep in touch with the PullString crew afterward as they move on to creating other characters (for instance, a Call of Duty bot that, on its first day in the wild, has 6 million conversations). At one point the company’s CEO, Oren Jacob, a former chief technology officer at Pixar, tells me that PullString’s ambitions are not limited to entertainment. “I want to create technology that allows people to have conversations with characters who don’t exist in the physical world—because they’re fictional, like Buzz Lightyear,” he says, “or because they’re dead, like Martin Luther King.”
My father receives his cancer diagnosis on April 24, 2016. A few days later, by happenstance, I find out that PullString is planning to publicly release its software for creating conversational agents. Soon anybody will be able to access the same tool that PullString has used to create its talking characters.
The idea pops into my mind almost immediately. For weeks, amid my dad’s barrage of doctor’s appointments, medical tests, and treatments, I keep the notion to myself.
I dream of creating a Dadbot—a chatbot that emulates not a children’s toy but the very real man who is my father. And I have already begun gathering the raw material: those 91,970 words that are destined for my bookshelf.
The thought feels impossible to ignore, even as it grows beyond what is plausible or even advisable. Right around this time I come across an article online, which, if I were more superstitious, would strike me as a coded message from forces unseen. The article is about a curious project conducted by two researchers at Google. The researchers feed 26 million lines of movie dialog into a neural network and then build a chatbot that can draw from that corpus of human speech using probabilistic machine logic. The researchers then test the bot with a bunch of big philosophical questions.
“What is the purpose of living?” they ask one day.
The chatbot’s answer hits me as if it were a personal challenge.
“To live forever,” it says.
• • •
“Sorry,” my mom says for at least the third time. “Can you explain what a chatbot is?” We are sitting next to each other on a couch in my parents’ house. My dad, across the room in a recliner, looks tired, as he increasingly does these days. It is August now, and I have decided it is time to tell them about my thoughts.
As I have contemplated what it would mean to build a Dadbot (the name is too cute given the circumstances, but it has stuck in my head), I have sketched out a list of pros and cons. The cons are piling up. Creating a Dadbot precisely when my actual dad is dying could be agonizing, especially as he gets even sicker than he is now. Also, as a journalist, I know that I might end up writing an article like, well, this one, and that makes me feel conflicted and guilty. Most of all, I worry that the Dadbot will simply fail in a way that cheapens our relationship and my memories. The bot may be just good enough to remind my family of the man it emulates—but so far off from the real John Vlahos that it gives them the creeps. The road I am contemplating may lead straight to the uncanny valley.
So I am anxious to explain the idea to my parents. The purpose of the Dadbot, I tell them, would simply be to share my father’s life story in a dynamic way. Given the limits of current technology and my own inexperience as a programmer, the bot will never be more than a shadow of my real dad. That said, I would want the bot to communicate in his distinctive manner and convey at least some sense of his personality. “What do you think?” I ask.
My dad gives his approval, though in a vague, detached way. He has always been a preternaturally upbeat, even jolly guy, but his terminal diagnosis is nudging him toward nihilism. His reaction to my idea is probably similar to what it would be if I told him I was going to feed the dog—or that an asteroid was bearing down upon civilization. He just shrugs and says, “OK.”
The responses of other people in my family—those of us who will survive him—are more enthusiastic. My mom, once she has wrapped her mind around the concept, says she likes the idea. My siblings too. “Maybe I am missing something here,” my sister, Jennifer, says. “Why would this be a problem?” My brother grasps my qualms but doesn’t see them as deal breakers. What I am proposing to do is definitely weird, he says, but that doesn’t make it bad. “I can imagine wanting to use the Dadbot,” he says.
That clinches it. If even a hint of a digital afterlife is possible, then of course the person I want to make immortal is my father.
• • •
This is my dad: John James Vlahos, born January 4, 1936. Raised by Greek immigrants, Dimitrios and Eleni Vlahos, in Tracy, California, and later in Oakland. Phi Beta Kappa graduate (economics) from UC Berkeley; sports editor of The Daily Californian. Managing partner of a major law firm in San Francisco. Long-suffering Cal sports fan. As an announcer in the press box at Berkeley’s Memorial Stadium, he attended all but seven home football games between 1948 and 2015. A Gilbert and Sullivan fanatic, he has starred in shows like H.M.S. Pinafore and was president of the Lamplighters, a light-opera theater company, for 35 years. My dad is interested in everything from languages (fluent in English and Greek, decent in Spanish and Italian) to architecture (volunteer tour guide in San Francisco). He’s a grammar nerd. Joke teller. Selfless husband and father.
These are the broad outlines of the life I hope to codify inside a digital agent that will talk, listen, and remember. But first I have to get the thing to say anything at all. In August 2016, I sit down at my computer and fire up PullString for the first time.
To make the amount of labor feasible, I have decided that, at least initially, the Dadbot will converse with users via text messages only. Not sure where to begin programming, I type, “How the hell are you?” for the Dadbot to say. The line appears onscreen in what looks like the beginning of a giant, hyper-organized to-do list and is identified by a yellow speech bubble icon.
Now, having lobbed a greeting out into the world, it’s time for the Dadbot to listen. This requires me to predict possible responses a user might type, and I key in a dozen obvious choices—fine, OK, bad, and so on. Each of these is called a rule and is tagged with a green speech bubble. Under each rule, I then script an appropriate follow-up response; for example, if a user says, “great,” I tell the bot to say, “I’m glad to hear that.” Lastly, I create a fallback, a response for every input that I haven’t predicted—e.g., “I’m feeling off-kilter today.” The PullString manual advises that after fallbacks, the bot response should be safely generic, and I opt for “So it goes.”
With that, I have programmed my very first conversational exchange, accounting for multiple contingencies within the very narrow context of saying hello.
And voilà, a bot is born.
Granted, it is what Lauren Kunze, CEO of Pandorabots, would call a “crapbot.” As with my Dark Mansion game back in the day, I’ve just gotten to the front door, and the path ahead of me is dizzying. Bots get good when their code splits apart like the forks of a giant maze, with user inputs triggering bot responses, each leading to a fresh slate of user inputs, and so on until the program has thousands of lines. Navigational commands ping-pong the user around the conversational structure as it becomes increasingly byzantine. The snippets of speech that you anticipate a user might say—the rules—can be written elaborately, drawing on deep banks of phrases and synonyms governed by Boolean logic. Rules can then be combined to form reusable meta-rules, called intents, to interpret more complex user utterances. These intents can even be generated automatically, using the powerful machine-learning engines offered by Google, Facebook, and PullString itself. Beyond that, I also have the option of allowing the Dadbot to converse with my family out loud, via Alexa (though unnervingly, his responses would come out in her voice).
It will take months to learn all of these complexities. But my flimsy “How are you” sequence has nonetheless taught me how to create the first atoms of a conversational universe.
After a couple of weeks getting comfortable with the software, I pull out a piece of paper to sketch an architecture for the Dadbot. I decide that after a little small talk to start a chat session, the user will get to choose a part of my dad’s life to discuss. To denote this, I write “Conversation Hub” in the center of the page. Next, I draw spokes radiating to the various chapters of my Dad’s life—Greece, Tracy, Oakland, College, Career, etc. I add Tutorial, where first-time users will get tips on how best to communicate with the Dadbot; Songs and Jokes; and something I call Content Farm, for stock segments of conversations that will be referenced from throughout the project.
To fill these empty buckets, I mine the oral history binder, which entails spending untold hours steeped in my dad’s words. The source material is even richer than I’d realized. Back in the spring, when my dad and I did our interviews, he was undergoing his first form of cancer treatment: whole-brain radiation. This amounted to getting his head microwaved every couple of weeks, and the oncologist warned that the treatments might damage his cognition and memory. I see no evidence of that now as I look through the transcripts, which showcase my dad’s formidable recall of details both important and mundane. I read passages in which he discusses the context of a Gertrude Stein quote, how to say “instrumentality” in Portuguese, and the finer points of Ottoman-era governance in Greece. I see the names of his pet rabbit, the bookkeeper in his father’s grocery store, and his college logic professor. I hear him recount exactly how many times Cal has been to the Rose Bowl and which Tchaikovsky piano concerto his sister played at a high school recital. I hear him sing “Me and My Shadow,” which he last performed for a high school drama club audition circa 1950.
All of this material will help me to build a robust, knowledgeable Dadbot. But I don’t want it to only represent who my father is. The bot should showcase how he is as well. It should portray his manner (warm and self-effacing), outlook (mostly positive with bouts of gloominess), and personality (erudite, logical, and above all, humorous).
The Dadbot will no doubt be a paltry, low-resolution representation of the flesh-and-blood man. But what the bot can reasonably be taught to do is mimic how my dad talks—and how my dad talks is perhaps the most charming and idiosyncratic thing about him.
My dad loves words—wry, multisyllabic ones that make him sound like he is speaking from the pages of a P. G. Wodehouse novel. He employs antiquated insults (“Poltroon!”) and coins his own (“He flames from every orifice”). My father has catchphrases. If you say something boastful, he might sarcastically reply, “Well, hot dribbling spit.” A scorching summer day is “hotter than a four-dollar fart.” He prefaces banal remarks with the faux-pretentious lead-in “In the words of the Greek poet …” His penchant for Gilbert and Sullivan quotes (“I see no objection to stoutness, in moderation”) has alternately delighted and exasperated me for decades.