Sie sind auf Seite 1von 6


part 3




88  Scientific American, September 2018

sad0918Domi4p.indd 88 7/18/18 4:22 PM

September 2018, 89

sad0918Domi3p.indd 89 7/17/18 5:59 PM

H humanness in unprecedented ways?
umans are the only animals that build machines. by doing so,
we expand our capabilities beyond our biological limits. Tools
turn our hands into more versatile appendages. Cars let us
travel faster, and airplanes give us wings. Computers endow us
with bigger brains and memory capacity, and smartphones
orchestrate daily life. Now we are creating technology that can
evolve on its own by encoding into it an ability to learn through
data and effort. Will it ultimately supplant us? Or will it augment our abilities, enhancing our

Machine learning started in the 1950s with the

work of pioneering scientists such as Frank Rosenblatt,
who built an electronic neuron that learned to recog-
nize digits, and Arthur Samuel, whose checkers pro-
gram learned by playing against itself until it could
beat some humans. But it is only in the past decade
that the field has truly taken off, giving us self-driving
translate languages with uncanny accuracy. Machine
learning also draws on psychology. Like humans,
these analogy-based algorithms solve new problems
by finding similar ones in memory. This ability al-
lows for the automation of customer support, as well
as e-commerce sites that recommend products based
on your tastes.
cars, virtual assistants that understand our commands Machines may also learn by automating the scien-
IN BRIEF (up to a point) and countless other applications. tific method. To induce a new hypothesis, symbolic
The pursuit of artificial Every year we invent thousands of new algo- learners invert the process of deduction: If I know
intelligence can be seen rithms, which are sequences of instructions telling a that Socrates is human, what else do I need to infer
as part of human evolu- computer what to do. The hallmark of learning ma- that he is mortal? Knowing that humans are mortal
tion. The next stage of chines, however, is that instead of programming would suffice, and this hypothesis can then be tested
automation will require them in detail, we give them general goals such as by checking if other humans in the data are also
the creation of a so-called “learn to play checkers.” Then, like humans, they im- mortal. Eve, a biologist robot at the University of
master algorithm. It prove with experience. These learning algorithms Manchester in England, has used this approach to
would integrate the five
tend to fall into five main categories, each inspired by discover a potential new malaria drug. Starting with
main ways that machines
currently learn into a sin- a different scientific field. Unsurprisingly, one way data about the disease and basic knowledge of mo-
gle, unified paradigm. that machines learn is by mimicking natural selec- lecular biology, Eve formulated hypotheses about
Technology i s simply tion, through evolutionary algorithms. In the Cre- what drug compounds might work, designed experi-
an extension of human ative Machines Lab at Columbia University, primi- ments to test them, carried out the experiments in a
capabilities. Machines tive robots try to crawl or fly, and the specifications of robotic lab, revised or discarded the hypotheses, and
do not have free will, those that perform best are periodically mixed and repeated until it was satisfied.
only goals that we give mutated to 3-D print the next generation. Starting Finally, learning can rely purely on mathematical
to them. It is the misuse
with randomly assembled bots that can barely move, principles, the most important of which is Bayes’s
of the technology by
people that we should this process eventually produces creatures such as theorem. The theorem says that we should assign
be worried about, not robot spiders and dragonflies after thousands or tens initial probabilities to hypotheses based on our
a robot takeover. of thousands of generations. knowledge, then let the hypotheses that are consis-
A more plausible n ear- But evolution is slow. Deep learning, currently tent with the data become more probable and those
term scenario for AI the most popular machine-learning paradigm, takes that are not become less so. It then makes predic-
is the proliferation of inspiration from the brain. We start with a highly tions by letting all the hypotheses vote, with the
“digital doubles”—virtual simplified mathematical model of how a neuron more probable ones carrying more weight. Bayesian
models of ourselves that
works and then build a network from thousands or learning machines can do some medical diagnoses
will interact with each
other in countless simu- millions of these units and let it learn by gradually more accurately than human doctors. They are also
lations to help us make strengthening the connections between neurons that at the heart of many spam filters and of the system
faster, more informed fire together when looking at data. These neural net- that Google uses to choose which ads to show you.
choices in our daily lives. works can recognize faces, understand speech and Each of these five kinds of machine learning has

90  Scientific American, September 2018

sad0918Domi3p.indd 90 7/17/18 6:00 PM

its strengths and weaknesses. Deep learning, for ex- guishes these algorithms from ordinary ones is that
ample, is good for perceptual problems such as vi- they have a lot of flexibility in figuring out how to
sion and speech recognition but not for cognitive reach the goals we set for them rather than needing
ones such as acquiring commonsense knowledge to execute a predefined series of steps. Even as they
and reasoning. With symbolic learning, the reverse get better at the task with experience, the goals re-
is true. Evolutionary algorithms are capable of solv- main unchanged. Solutions that do not make prog-
ing harder problems than neural networks, but it ress toward the goal are automatically discarded.
can take a very long time to solve them. Analogical Plus, humans get to check that what the machines
methods can learn from just a small number of in- produce does indeed satisfy our objectives. We are
stances but are liable to get confused when given too also able to verify that the machines do not violate
much information about each. Bayesian learning is any of the constraints we put on them, such as “obey
most useful for dealing with small amounts of data the rules of the road.”
but can be prohibitively expensive with big data. When we envision an AI, though, we tend to proj-
These vexing trade-offs are why machine-learn- ect onto it human qualities such as volition and con-
ing researchers are working toward combining the sciousness. Most of us are also more familiar with
best elements of all the paradigms. In the same way humanlike AIs, such as home robots, than with the
that a master key opens all locks, our goal is to create myriad other types that do their work behind the
a so-called master algorithm—one that can learn ev- scenes. Hollywood compounds this perception by
erything that can be extracted from data, deriving all depicting robots and AIs as humans in disguise—an
possible knowledge from it. understandable tactic that makes for a more com-
The challenge on us now is similar to the one pelling story. Artificial intelligence is just the ability
faced by physicists: quantum mechanics is effective to solve hard problems—a task that does not require
at describing the universe at the smallest scales and free will. It is no more likely to turn against us than
general relativity at the largest scales, but the two your hand is to slap you. Like any other technology,
are incompatible and need to be reconciled. And in AIs will always be extensions of us. The more power-
the same way that James Clerk Maxwell first unified ful we can make them, the better.
light, electricity and magnetism before the Standard What, then, might our AI-enabled future look
Model of particle physics could be developed, differ- like? Intelligent machines will indeed supplant many
ent research groups, including mine at the Universi- jobs, but the effects on society will likely be similar
ty of Washington, have proposed ways to unify two to previous forms of automation. Two hundred years
or more of the machine-learning paradigms. Be- ago the majority of Americans were farmers. Yet to-
cause scientific progress is not linear and instead day machines have replaced almost all of them with-
happens in fits and starts, it is difficult to predict out causing massive unemployment. Doomsayers
when the full unification of the master algorithm argue that this time is different because machines
might be complete. Regardless, achieving this goal are replacing our brains, not just our brawn, leaving
will not usher in a new, dominant race of machines. nothing for humans to do. But the day that AIs can
Rather, it will accelerate human progress. carry out all the tasks we can is still very distant, if
it ever comes. For the foreseeable future, AIs and
MACHINE TAKEOVER? humans will be good at different things. Machine
ONCE WE ATTAIN THE MASTER ALGORITHM and feed it the learning’s primary effect will be to greatly lower the
vast quantities of data each of us produce, artificial- cost of intelligence. This democratization will in-
intelligence systems will potentially be able to learn crease the variety of economically feasible uses of
very accurate and detailed models of individual peo- that intelligence, generating new jobs and trans-
ple: our tastes and habits, strengths and weaknesses, forming old ones to accomplish more with the same
memories and aspirations, beliefs and personalities, amount of human labor.
the people and things we care about, and how we Then there is the “singularity” scenario, popular-
will respond in any given situation. That models of ized by futurist Ray Kurzweil. It is one of ever accel-
us could essentially predict the choices we will make erating technological progress: machines learn to Pedro Domingos is
is both exciting and disquieting. make better machines, which in turn make even bet- a professor of computer
Many worry that machines with these capabili- ter ones, and so on. But we know that this cannot science at the University
ties will use their newfound knowledge to take all continue forever because the laws of physics place of Washington and author
our jobs, enslave us or even exterminate us. But that strict limits on how powerful even a quantum com- of The Master Algorithm
(Basic Books, 2015).
is unlikely to happen because they have no will of puter can be, and in some aspects, we are not far
A fellow of the Association
their own. Essentially all AI algorithms are driven by from hitting them. The progress of AI, like the prog- for the Advancement
goals that we program, such as “find the shortest ress of everything else, will eventually plateau. of Artificial Intelligence,
route from the hotel to the airport.” What distin- Another vision popular among futurists is that he lives near Seattle.

September 2018, 91

sad0918Domi3p.indd 91 7/17/18 6:00 PM

This sea star uses
evolutionary algo-
rithms to learn how
to simulate itself.
These algorithms
are one type of
machine learning
that could be unified
with others into a
“master algorithm,”
a singularly power-
ful human tool.

computer models of us will become so good that they Within a decade each one of us will probably have
will be practically indistinguishable from the real a “digital double,” an AI companion that will be even
thing. In this scenario, we could upload ourselves tomore indispensable than our smartphones are today.
the cloud and live on forever as pieces of software, Your digital double will not need to physically move
free of the pesky constraints of the physical world. around with you; most likely it will live somewhere
One problem with this scenario is that it may not be in the cloud, just as much of your data already does.
biologically feasible. To upload yourself, you would We can see its beginnings in virtual assistants such
presumably need an accurate model of each of your as Siri, Alexa and Google Assistant. At the heart of
neurons, complete with the memories they store. It your digital double will be a model of you, learned
would have to be captured so reliably that the model’s
from all the data you have ever generated in your in-
predictions would not rapidly diverge from the be- teractions with the digital world, from desktop com-
havior of the real neurons—a tall order indeed. But puters and Web sites to wearable devices and sensors
even if this were a realistic option, would you really
in the environment such as smart speakers, thermo-
upload yourself if you had the chance? How could you stats, cell-phone towers and video cameras.
know for sure that your model was not missing some The better our learning algorithms become and the
essential part of you—or that it was conscious at all?
more personal data we feed them, the more accurate
What if a thief stole your identity in the most absolute
our digital doubles will get. Once we have the master
and complete sense of the word? I believe that peoplealgorithm and then couple it with continuous capture
will opt to hang on to their squishy, carbon-based of your sensorimotor stream via an augmented reality
selves—the “wetware,” as computer scientists joking- headset and other personal sensors, your double will
ly call it—for as long as they can and then call it quits.
grow to know you better than your best friend.
The model and data will be maintained by a “data
CHERCHEZ L’HUMAIN bank,” not unlike a traditional bank that stores and
AI—MACHINE LEARNING IN PARTICULAR—is really just the invests your money. Many existing companies would
continuation of human evolution. In The Extended surely like to provide that service for you. Google co-
Phenotype, Richard Dawkins shows how common it founder Sergey Brin has said that Google wants to be
is for animals’ genes to control the environment be- “the third half of your brain,” but you probably would
yond their bodies, from cuckoo eggs to beaver dams. not want part of your brain to subsist by showing

(Dawkins serves on Scientific American’s board of you ads. You might be better served by a new kind of
advisers.) Technology is the extended phenotype of company with fewer conflicts of interest or by a data
humans, and what we are building today is another union you form with like-minded people.
layer of our technological exoskeleton. I think the After all, the central worry about AI is not that it
most likely scenario for how humans will use AI is will spontaneously turn evil but that the humans
more fascinating than the usual speculations. who control it will misuse it (cherchez l’humain, as

92 Scientific American, September 2018

sad0918Domi3p.indd 92 7/17/18 6:00 PM

the French might say—“look to the human”). So your
Your digital
double will take
data bank’s first duty will be to ensure that your
model is never used against your interests. Both you
and the data bank must be vigilant about monitor-

your place in
ing AI crime because this technology will empower
bad actors as much as anyone. We will need AI police
(the Turing police, as William Gibson called it in his
1984 book Neuromancer) to catch the AI criminals.

all kinds of virtual

If you have the misfortune of living under an au-
thoritarian regime, this scenario could usher in un-
precedented dangers because it will allow the govern-
ment to monitor and restrain you like never before.

Given the speed at which machine learning is pro-
gressing and the predictive policing systems already
in use, the Minority Report scenario—where people
are preemptively arrested when they are about to
commit a crime—no longer seems far-fetched. Then
there are the implications of inequality as the world probable lives in cyberspace so that the single one you
adapts to the speed of life with digital doubles before live in the physical world is likely to be the best ver-
all of us are able to afford one. sion. Whether your simulated lives are some­how “real”
Our first duty, as individuals, will be not to be- and your cyberselves have a kind of consciousness (as
come complacent and trust our digital doubles be- portrayed in the plots of some Black Mirror e pisodes,
yond their years. It is easy to forget that AIs are like for instance) are interesting philosophical questions.
autistic savants and will remain so for the foresee- Some people worry that this means that we are
able future. From the outside, AIs may seem objec- handing over control of our lives to computers. But
tive, even perfect, but inside they are as flawed as we it actually gives us more control, not less, because it
are or more, just in different ways. For example, AIs allows us to make choices we could not before. Your
lack common sense and can easily make errors that model will also learn from the results of each virtual
a human never would, such as mistaking a person experience (Did you enjoy the date? Do you like your
crossing the street for a windblown plastic bag. They new job?) so that over time, it will become better at
are also liable to take our instructions too literally, suggesting the things you would choose for yourself.
giving us precisely what we asked for instead of In fact, we are already accustomed to most of our
what we actually wanted. (So think twice before tell- decision making taking place without our conscious
ing your self-driving car to get you to the airport on intervention because that is what our brains do now.
time at all costs.) Your digital double will be like a greatly expanded
Practically speaking, your digital double will be subconscious, with one key difference: Whereas your
similar enough to you to take your place in all kinds subconscious lives alone inside your skull, your dig-
of virtual interactions. Its job will not be to live your ital double will continuously interact with those of
life for you but rather to make all the choices you do other people and organizations. Everyone’s doubles
not have the time, patience or knowledge for. It will will keep trying to learn models of one another, and
read every book on Amazon and recommend the few they will form a society of models, living at comput-
that you are most likely to want to read yourself. If er speeds, branching out in all directions, figuring
you need a car, it will research the options and hag- out what we would do if we were there. Our ma-
gle with the car dealer’s bots. If you are job hunting, chines will be our scouts, blazing a trail into the fu-
it will interview itself for all the positions that fit ture for us as individuals and as a species. Where
your needs and then schedule live interviews for you will they lead us? And where will we choose to go? 
for the most promising ones. If you get a cancer di-
agnosis, it will try all potential treatments and rec-
ommend the most effective ones. (It will be your eth- MORE TO EXPLORE
ical duty to use your digital double for the greater The Master Algorithm: How the Quest for the Ultimate Learning Machine Will Remake Our World.
good by letting it take part in medical research, too.) Pedro Domingos. Basic Books, 2015.
And if you are seeking a romantic partner, your dou- The Digital Mind: How Science Is Redefining Humanity. A
 rlindo Oliveira. MIT Press, 2017.
ble will go on millions of virtual dates with all eligi- FROM OUR ARCHIVES
ble doubles. The pairs that hit it off in cyberspace Self-Taught Robots.Diana Kwon; March 2018.
can then go on a date in real life.
s c i e n t if i c a m e r i c a n . c o m /m a g a zi n e /s a
Essentially your double will live out countless

September 2018, 93

sad0918Domi3p.indd 93 7/17/18 6:00 PM