Here’s what very few people know about one of the most important minds in AI: though he’s sometimes accused of being a robot himself, he’s actually very funny. It’s a bone-dry, snap-quick, intensely human humor that doesn’t at all resemble an AI entity. Few people also know that his father, a composer, died of heart disease at a relatively young age and passed his genetic predisposition for elevated cholesterol down to his son.
Kurzweil was also diagnosed with type 2 diabetes in the mid-1980s. The insulin treatment his doctor put him on made him a good deal fatter but no healthier. It seemed his congenital heart condition and treatment were locked in a race to kill him. So he decided to treat himself. Doing so meant constantly researching new ways to improve his health and carefully logging dozens of different biological reactions to drugs, diet, sleep, and exercise.
In 1999 he met his current physician and sometime collaborator Terry Grossman, on whom he made an immediate impression, as Grossman recounted for the Futurist magazine.
“I am not surprised when patients come to see me with a notebook of spreadsheets detailing various data extracted from their daily lives: blood pressure, weight, cholesterol, blood sugar levels, amount of exercise, etc., carefully tabulated for several years. But all previous data collections I had seen, even those organized into Excel and meticulously graphed, paled in comparison to Ray’s. His data collection was so thorough and meticulous that he could tell me what he ate for lunch on June 23, 1989 (as well as what he ate for lunch every other day for several years before that time). And not only what he ate, but the number of grams of each serving and calories consumed, as well as the number of calories he burned that day through exercise—every day for decades!”8
Ray Kurzweil has been experimenting on himself and carefully recording the results for almost thirty years now. When I ran into him at a San Francisco event in October 2012, he told me that he credits self-quantification with helping him overcome the threat of heart disease and diabetes. At the very least, it’s helped him outlive his father, who died at age fifty-eight.
“These numbers, you can change them . . . very radically. You’re not a prisoner to this characteristic. I’ve dramatically changed who I am.”
Improving health is, today, the most common use of self-quantification. But there are entire areas of life where a rigorous approach to data collection and analysis can lead to better outcomes; this includes analyzing communication, work, and buying patterns. Computer scientist Stephen Wolfram shares a lot in common with Ray Kurzweil. He too was an early pioneer of the use of personal computers to create and store records of virtually any signal, transaction, or change that could be recorded, even though he didn’t know exactly what he intended to do with that information when he started collecting it.
Wolfram is best known as the creator of the Wolfram Alpha search engine. It’s similar to Google except that the inputs—what you type in the search box—and the outputs are mathematical, sort of like a calculator that can read the Web. Unlike the Google engine, which responds to every question with a ranked list of pages, Alpha actually computes specific answers.
When you use Wolfram Alpha, a lot of queries come back with results like “need more information,” but when it works, it’s miraculous. Need to know the average life span for a human being in France? Alpha can give you the number in years (81.4) and a breakdown for the percentage of the population dying older or younger, as well as a graph for how the number has changed over time. For instance, there’s a dip during the two world wars, but a strange blip upward in the middle of both. Need to predict how long it’s going to take you to read 1,000 pages in a standard textbook? The answer, via Wolfram Alpha, is thirty hours, assuming this 1,000-page textbook is perfectly statistically average, contains 3.05 megabytes of information, 500,000 words, 45,000 lines, and 5.1 characters per word.
In March 2012 Wolfram surprised the world with a revelation on his blog that he had been Wolfram Alpha–izing his own life for close to thirty years.9 He had a log of not only every phone call he had made in that time but every keystroke since 2002 (more than 100 million), and every e-mail he had sent since 1989 (more than 200,000). He knows, roughly, how many steps he’s taken in the last two years, according to GPS and room-by-room motion sensor data. He has all of his medical test results and his personal genome, which he describes as “not yet very useful.” If it’s a Thursday night and it’s 10 P.M., he knows there’s a 50 percent probability that he’ll be on the phone. When I had the opportunity to speak with him I wanted to know what he’s learned.
“One thing I’ve found out is that I’m far more habitual than I ever imagined,” he told me. For instance, he now knows that if he ignores his in-box for five days, it will take him two weeks to get up to date. He has a formula to estimate the likelihood that he will procrastinate on a given project, the average time of procrastination, and the best way to avoid it.
“I’m one of these people that for certain types of tasks, there’s no point in starting early; I won’t finish it until just in time anyway. I have to know how long it’s actually going to take to finish so that I can get it done in an efficient way. If I start it too early, the task expands into the available space,” he says.
Wolfram hopes that one day he’ll be able to use this painfully extracted self-knowledge to figure out when he’s at his creative peak. This is not an insignificant matter, as his life is devoted to solving problems that are supposed to be impossible. “What I’ve had to do is figure out an efficient scheme to create on demand. I’ve gotten better at it. I know that if I think about something for a certain period of time, a window will appear, a period of a few hours when I’ll either have a good idea or I won’t.”
Wolfram’s experience speaks to one of the more important, near-term applications of QS. In the decade ahead, a lot more people will be tracking themselves to guard against something like the inside view.
But Wolfram and Kurzweil have created what Ben Franklin couldn’t: a reference class of personal data that meets the criteria of “large enough to be significant.” And they did so at a cost of time and effort that—while still high—was lower for them than it would be for someone else. As a result, they are perhaps better armed against the inside view than the rest of us.
If you’re not exactly eager to tabulate every step you take and every gram of sodium you ingest, some recent research suggests that you can greatly improve your health by simply watching just one signal: how you react to stressful events.
In 2012 University of Pennsylvania researcher David Almeida and some colleagues published a paper showing that the most important predictor of a future chronic health condition (aside from smoking, drinking, and engaging in conspicuously unhealthy behavior) was overreacting to routine, psychologically taxing incidents. When they interviewed subjects about how little stressors such as car breakdowns, angry e-mails, small disappointments, the little annoyances of modern life, affected them emotionally, they found that “for every one unit increase in affective reactivity [people reporting a big emotional change resulting from the stressful event], there was a 10% increase in the risk of reporting a chronic health condition 10 years later.”10
The researchers didn’t find that people who were exposed to more stressful experiences were more likely to develop a chronic health condition. Rather, the increase was isolated to people who reported feeling very different emotionally on a day that they encountered a stressor than on a day when they did not.11
This is a classic inside-view problem. Very few people keep track of how they react to little stressors. The costs of keeping such a record, in terms of inconvenience, are too high. Yet hidden in those reactions may be powerful clues to our future health. If it were easy and cheap to keep that data around, and if we were able to make sense of it quickly, we would surely keep a log of how stressed we felt at any given moment.
When I asked Kahneman via Skype at the Singularity Summit 2012
what he thought of the self-quant trend, he was guardedly optimistic about the potential applications of quantification techniques for physicians. Adopting the outside view will never be intuitive, he said. “But at least in principle there is an opportunity for people to discover regularities in their own lives. There will be an opportunity to look at the outcome of similar cases . . . A physician could have intuitions about a patient, but supplementing that intuition with instantly available statistics will likely result in fewer mistakes” (emphasis added).
Though Stephen Wolfram is the seminal self-quantifier, he believes that personal data collection has to become both a great deal easier and more immediately rewarding before it becomes mainstream. The technology itself is no longer the problem. What’s lacking is expertise. We imagine computerized data to be neat and clean. But getting our data into a usable form isn’t as simple as just pressing the Return key.
“There’s all sorts of plumbing that has to be done,” says Wolfram, meaning that service providers need to make these services easier to use for everyone, not just computer geniuses. “If everyone used one vendor’s equipment, that would be a start. But that won’t happen. It will stay a complex, multiproduct, multivendor environment. This data [is] in computers; [it’s] in pedometers; [it’s] in lots of different places. First step is to make it possible to upload to somewhere. Maybe upload it to a cloud that’s shared but some people are too paranoid to do that. I don’t think there’s anything technically difficult about this. It just requires all sorts of work.”
In the past couple of years, some enterprising start-ups have sprung up to relieve the amount of work involved in keeping track of signals, physical states, and so on. The United Kingdom’s Tictrac is a platform that allows you to take your data from different devices and sites and create a snapshot of yourself in the present. I spoke with founder Martin Blinder while the company was still beta testing. It’s since opened to the public and has been steadily gaining users. He knows that Tictrac will only succeed if it can offer personal analytics in a way that’s intuitive and user-friendly. It needs to be able to take your data and present you with a future prediction you couldn’t have reached yourself, and do it in a way that’s perfectly understandable at any given moment.
“We want to offer a breakdown of the food you’ve eaten in the last month, type of food and calories, and whatnot. But we also want you to be able to see analytics across different data sets, so you can pull in your calendar so you can see that you spend about twelve hours a week on business lunches—and then cross that with weight and find a correlation between the two,” he told me.
The average Tictrac user has as many as ten thousand different points of data that users bring to the practice of lifestyle management. Those points include everything from Facebook posts to e-mails to GPS or fitness logs of physical activity from such devices as Fitbit, Nike+, and Blipcare. Some of his more exceptional users have twenty thousand data points. And the exceptional is quickly becoming the average. As a measure for how much personal data there is today versus ten years ago, only two digital data streams that Tictrac users build into the graphs were around prior to 2003: e-mail and calendar.
Perhaps the most encouraging aspect of the Tictrac platform is that the number of data sources and possible insights is limited only by how much data can easily be collected through an API. With enough streaming data, it’s possible to see how all of your life areas interact, how overbooking appointments affects your exercise levels, how your communications with one person changes your drinking and sleeping or monthly expenses, even how what you eat influences your electricity usage. These sorts of data streams will grow as we integrate more sensing and broadcasting capability into more objects and our environment.
Devices like the Q Sensor from Affectiva, which looks like a strange wristwatch, can measure your level of interest and engagement in a given activity based on electrodermal conductance (i.e., how much electricity your skin is emitting), and in 2012 a company called Cogito, founded by two MIT grads working under a Defense Advanced Research Projects Agency (DARPA) grant, created a platform that can detect—and predict—your mood levels based on tone and cadence of speech. It was intended to help mental health professionals working with returning vets better anticipate the rise of depression. One day, a future version of Cogito could make its way into an iPhone app capable of helping you anticipate and plan around your future feelings. It would telegraph your emotional future the way horoscopes are supposed to but would be based on data points accumulated through actual life, as opposed to unproven notions about the effects of planets. There are countless other examples of self-diagnostic apps waiting to be developed. Research from University of Maryland scholar Lisa M. Viser has shown that it’s possible to detect dementia in keystroke patterns, or simply based on changes in the way someone types over a period of weeks or months. (You can also detect whether someone’s been drinking at lunch.)12 Bream Brush is a smart toothbrush that dialogues with the user’s smartphone via an app to keep a log of brushing time. The data can be shared with the user’s dentist, insurance provider, et cetera.13
Let’s assume one of these start-ups, or one not yet conceived, makes it to mainstream adoption. Once that occurs, the personal costs for self-quantification will have collapsed in just a few decades. In the 1980s, when Ray Kurzweil decided to flout the advice of his doctor, take himself off insulin, and begin keeping a detailed log of every meal he ate and what was in it, few other people would have had the patience, know-how, or inclination to attempt anything similar. The behavior at the time seemed positively bizarre. When Kurzweil first began his self-quantification experiments, the costs in terms of time and effort were a bit lower for him than they would be for anybody else, except Stephen Wolfram. Today, they’re joined by enterprising people such as Sacha Chua, and the numbers are growing.
We are one app away from becoming Ray Kurzweil.
Here’s what that app might look like to you in practice. You would give the program access to your biophysical signals, gleaned from your activity levels, mood analyzers, implants if you have any, e-mail and voice mail, et cetera. The program in turn would give you a rapidly evolving window into your future health. On any given day, you might receive a notification with the following warning: “Dear Patrick, as a result of that stress event you had a couple of weeks ago, the dizzy spell you complained of last night, and the fact that you’ve recently increased your daily alcohol consumption from two glasses of Merlot to four, your probability for stroke in the next year has just increased to 10 percent.”
Naturally, if you received this message, you would act to avert this stroke before it happened, rendering the prediction incorrect, but still invaluable.
Yet more cloud processing and an abundance of carefully collected personal data aren’t the magic ingredients that are going to bring the above scenario to life. Even with the right technology and a seamless interface or analytics engine to take the difficult work out of making usable predictions from your data, the most important component of your changing health picture is other people’s health data. Here’s the trade-off, the point where our outmoded ideas of privacy begin to get in the way of progress and better health.
The Network Is Your Doctor
In June 2012 a group of researchers from MIT and Columbia created a system that can predict future illness. They call the system the Hierarchical Association Rule Model or HARM (a bizarre but at least memorable acronym for a medical algorithm). It can’t tell you what’s wrong with you right now; instead, the algorithm determines what you’re likely to get next on the basis of a current diagnosis combined with certain demographic features such as race, age, and so on. To build it, they used clinical trial data from around 42,000 patient encounters and 2,300 patients, all at least forty years old. These were people who had signed up to test new medicines so their medical records were more thoroughly filled out than is the norm. The subjects were also encouraged to r
eport back on what they were feeling and experiencing, as such data could have an effect on the drug’s marketability. What were the results? It turns out stroke is more predictable than many statisticians had believed, even though the proximate causes for stroke are still difficult to determine.14
The system works by finding correlations among thousands of patients sharing their history, not by looking for causes. This is a big departure from the way medicine is traditionally practiced and taught. It also only works when thousands of people elect to share their most personal data.
Future breakthroughs in the application of highly personalized data to medicine will depend tremendously on a willingness to share that sort of information. The question is how to do it in a way that doesn’t come back to haunt you. From a researcher’s perspective, the problem becomes one of how to build privacy controls that allow your users to share but that allow your model, algorithm, system, or Web site to access the most valuable information.
The fledgling Consent to Research project headed by John Wilbanks is a great example of an organization that fully understands the importance of sharing health data for future medical practice but also understands the risks people take in exposing themselves. Wilbanks points out that more than one in ten people in the United States have a rare disease (defined as a disease that fewer than two hundred thousand people are diagnosed with per year), have a family member with a rare disease, or have a first-degree friend with a rare disease. There are a lot of illnesses that very few people have, but the sheer number of them affects us all.
Wilbanks’s sister is one such person. “Our best guess is that she has some kind of psoriatic arthritis. We don’t know what kind,” he told me and a few other folks at a Singularity Summit event in San Francisco. “The insurance industry is already quite good at denying her care based on the actuarial tables. So we pay for PET scans out of our own pocket.”
The Naked Future Page 6