Computer's speed reading advantage

In May last year, a supercomputer in San Jose, California, read 100,000 research papers in 2 hours. It found completely new biology hidden in the data. Called KnIT, the computer is one of a handful of systems pushing back the frontiers of knowledge without human help.

KnIT didn't read the papers like a scientist – that would have taken a lifetime. Instead, it scanned for information on a protein called p53, and a class of enzymes that can interact with it, called kinases. Also known as "the guardian of the genome", p53 suppresses tumours in humans. KnIT trawled the literature searching for links that imply undiscovered p53 kinases, which could provide routes to new cancer drugs.

Having analysed papers up until 2003, KnIT identified seven of the nine kinases discovered over the subsequent 10 years. More importantly, it also found what appeared to be two p53 kinases unknown to science. Initial lab tests confirmed the findings, although the team wants to repeat the experiment to be sure.

KnIT is a collaboration between IBM and Baylor College of Medicine in Houston, Texas. It is the latest step into a weird world where autonomous machines make discoveries that are beyond scientists, simply by rifling more thoroughly through what we already know, and faster than any human can.

The full article is short and worth reading.

As human history progresses, the body of previous research and knowledge in that field expands, and at some point humans may not have the time in their lives to learn it all (I'm setting immortality aside for now, though that is one potential solution). Computers, on the other hand, can read much more quickly than humans, and it would not surprise me if we start to see more and more of these computer-generated discoveries. The value of Big Data is still being debated, but this breakthrough suggests one path to unlocking it is shedding the limitations of human intelligence.

Unintended consequences

In 2003, fearing that overworked medical residents were committing errors due to fatigue, the Accreditation Council for Graduate Medical Education put limits on how many consecutive hours residents could work on a shift.

Now, ten years later, it's not clear the change has had the desired effect. 

One study, led by Sanjay Desai at Johns Hopkins, randomly assigned first-year residents to either a 2003- or 2011-compliant schedule. While those in the 2011 group slept more, they experienced a marked increase in handoffs, and were less satisfied with their education. Equally worrisome, both trainees and nurses perceived a decrease in the quality of care—to such an extent that one of the 2011-compliant schedules was terminated early because of concerns that patient safety was compromised. And another study, comparing first-year residents before and after the 2011 changes, found a statistically significant increase in self-reported medical error.

While these studies suggest the complex nature of patient safety—that manipulating one variable, like hours worked, inevitably affects another, like the number of handoffs—there is another tradeoff, more philosophical than quantifiable. It has less to do with the variables within the system and how we tinker with them, and more to do with what we overlook as we focus relentlessly on what we can count.

Caveat: this essay by Lisa Rosenbaum in the New Yorker is a bit short on data for my liking, the above study feeling like just one insufficient data point. 

But the meta point about unintended consequences and complexity is worth noting. The increase in handoffs of patients, the decrease in time any one doctor spends with a patient, these all have consequences that work against the quality of healthcare, even as I believe more well-rested residents are a good thing, many of my doctor friends having been put through grueling rotations.

Peer effects and social policy

When illness in one person is treated or prevented, others to whom that person is connected also benefit.


This leads to a problem. Taking network effects seriously means that we should value socially connected people more. From a policy perspective—if not a moral perspective—the connected should get more healthcare attention.

More from Nicholas Christakis here (PDF). As Christakis notes, a healthcare system that replaces the current one that grants inexplicit privileges to the wealthy with one that favors the networked might be more just, but the notion makes him uncomfortable.

Without even debating the ethics of such a system, I don't think measuring a person's peer effect multiplier is anywhere near precise enough today. I have nightmarish visions of an angry mob of people waving their Klout scores at the ER waiting room attendant.

Baumol's Cost Disease

It may not seem like an honor to have a term like "cost disease" named after you, but William Baumol's new book The Cost Disease is one of the more concise, enlightening economics books I've read recently.

Baumol's thesis is that certain service sectors, most notably healtchare and education, are doomed to outpace inflation because they are so dependent on labor. 

Is there hope for education costs coming down? Will Harvard and other universities with massive endowments decide to subsidize higher education? Unlikely. But disruption tends to not come from incumbents, so we wouldn't look there anyway.

Perhaps education cost disruption comes from something like online education. As Alex Tabarrok writes, online education gives teachers massive leverage.

In 2009, I gave a TED talk on the economics of growth. Since then my 15 minute talk has been watched nearly 700,000 times. That is far fewer views than the most-watched TED talk, Ken Robinson’s 2006 talk on how schools kill creativity, which has been watched some 26 million times. Nonetheless, the 15 minutes of teaching I did at TED dominates my entire teaching career: 700,000 views at 15 minutes each is equivalent to 175,000 student-hours of teaching, more than I have taught in my entire offline career.[1] Moreover, the ratio is likely to grow because my online views are increasing at a faster rate than my offline students.
Teaching students 30 at a time is expensive and becoming relatively more expensive. Teaching is becoming relatively more expensive for the same reason that butlers have become relatively more expensive–butler productivity increased more slowly than productivity in other fields, so wages for butlers rose even as their output stagnated; as a result, the opportunity cost of butlers increased. The productivity of teaching, measured in, say, kilobytes transmitted from teacher to student per unit of time, hasn’t increased much. As a result, the opportunity cost of teaching has increased, an example of what's known as Baumol’s cost disease. Teaching has remained economic only because the value of each kilobyte transmitted has increased due to discoveries in (some) other fields. Online education, however, dramatically increases the productivity of teaching. As my experience with TED indicates, it’s now possible for a single professor to teach more students in an afternoon than was previously possible in a lifetime.

I don't think online universities will ever adequately replace attending a university in person for certain things (socialization, live human feedback, and the signaling value of a degree from an actual university have tangible value), but I've taken several online courses and for certain subject matters they are more than adequate at transmitting knowledge.

Baumol argues that we shouldn't panic as much about the rising costs in healthcare and education because we're saving a lot of money in areas which aren't as dependent on labor, but that doesn't mean we shouldn't take a hard look at how to keep the costs in both of those areas down.

In healthcare, consumers are so removed from the actual cost side of the equation that it doesn't function much like an efficient marketplace at all. I see the doctor, I pay my copay of $15, and then when the bill comes out I have no idea whether I was given a good deal or not, I just hope my insurance covers as much of it as possible.

As for the value of what I receive from the healthcare industry, it's extremely difficult to gauge. Early in life, it tends to be very binary what I want. Cure my sinus infection. Fix my broken leg. Reconstruct my ACL. At the end of my life, the value equation shifts dramatically; still difficult to value, but in a completely different way. How do you quantify the value of an additional month of life? An additional year? Three years? And can you assign the proper amount of credit to the physician for

One reason Baumol's Cost Disease is more prevalent than it would otherwise be is that wages tend to be sticky. This was hammered home recently in the story of Hostess, which, in the face of declining sales, asked their worker unions to accept pay cuts, which the unions refused. That the executives had awarded themselves pay raises or that the root of the issue is really that no one eats Ding Dongs and Twinkies anymore doesn't negate the point that wages rarely go down in the real world, as they might in a truly efficient marketplace. I've actually never been at a company where any employees were asked to take a pay cut. Generally companies just freeze the salary of low performers, and that's enough of a signal that folks move on.

I find new-fangled labor marketplaces like TaskRabbit and Zaarly intriguing mostly as economic experiments in true wage elasticity. Companies don't generally try to ask people to work for lower wages, that's not a good signal in the recruiting marketplace. Rather, they'll approach it by trying to squeeze more work out of people at the same salary, which is a more subtle approach.

Low end disruption n the tech labor marketplace can happen, though it's most likely if initiated by the laborers themselves. In practice we call these people interns.

The healthcare market

Yichuan Wang argues that conservatives who believe that replacing ObamaCare with a market for individual health insurance is not as simple a free market solution as that for many other goods and services in his post Healthcare and Cars are Not Isomorphic.

Wang notes that as healthcare is a service, not a good like a car, "one productive hospital cannot simply 'export' its health care services across the country.

Some of the commenters on the post argue that healthcare is easily exportable, just like any other service (McDonald's and WalMart are cited as examples, among others). However, a study by Georgetown University that looked at George, Maine, and Wyoming, three states that allowed sale of insurance across state lines found that nothing happened. The primary deterrent was "the localized nature of how health care is delivered."

Respondents universally reported the enormous difficulty that out-of-state insurers face in building a network of local providers, and insurers identified doing so as a significant barrier to market entry that far surpasses concerns about a state’s regulatory environment or benefit mandates. State officials and insurers also noted that across state lines legislation ignores the primary cause of high prices—the cost of delivering care—and fails to account for often dramatic differences in the cost of care between states and regions.

Practical barriers and administrative obstacles also hinder success. Many state regulators are reluctant to relinquish some or all authority to enforce state standards by taking the risk of allowing another state to establish and enforce consumer protections that affect their residents. Respondents in five states reported difficulties in implementation because other states have little incentive to establish across state lines partnerships. In addition, officials and insurers in all six states noted the complexity of health insurance as a practical barrier to across state lines proposals and that establishing the rules under which an interstate health insurance compact would operate would likely demand more time and resources than states are willing to commit.


While it is certainly the case that many consumers and small businesses lack meaningful choices among insurers and struggle to find affordable coverage, our findings suggest that across state lines legislation does not appear to be the “silver bullet” that proponents are searching for.

More importantly, though, Wang notes one critical difficulty with healthcare: determining the quality of service.

But this is an incomplete justification. Haircuts are also services, but few people think barbershops require regulation. Yet, they differ in one key respect. While it is easy for me  to determine the quality of my haircut, it is much more difficult for me to determine what counts as “good” health care. A large part of this is linked to the uncertainty inherent in any kind of biological process. How does one determine if one is receiving sufficient care? Of course, if there is any gross negligence, it can be detected. Yet if the care is just slightly worse than it should be, there's no real way for the consumer to know. The problem is compounded by the fact that most people rarely get sick. If you go to get a haircut every month, you can quickly determine which barbershop is the best. Unfortunately for neoclassical economists but fortunately for society as a whole, people need catastrophic care substantially less often than they need haircuts, thereby impeding the market from finding the most efficient solution.

I think of this every time I need to find a doctor for one reason or another. Determining who to choose from the hundreds of doctors listed on the website of my insurance provider might as well be like picking a blind date out of a phone book. I have so much more information to help me pick the best tablet computer or the best running back to pick up off of waivers this week in fantasy football than I do to choose a doctor to care for a critical health issue that it's like a plot scenario out of an absurdist tragedy.