Jim Beam Black® Kentucky Straight Bourbon Whiskey, 43% Alc./Vol. ©2018 James B. Beam Distilling Co.

, Clermont, KY.
 D E C 2 0 18
| intel inside

A.I. is already changing everything.

Let’s rethink how it’s made—now.

THE GENIUS NEUROSCIENTIST WHO WANTS TO TRANSFORM A.I. P. 96

HOW TO TEACH A NEURAL NET SOME COMMON SENSE P. 74

FEI-FEI LI’S QUEST TO MAKE MACHINES BETTER FOR HUMANITY P. 90

INSIDE THE DIY SMART-TECH MOVEMENT P. 82

 A CELE BR AT I O N O F
THE IDEAS , I NNOVAT I O NS ,
AND ICO NS T H AT MAD E
THE WO R L D W I R ED

JEFF BEZOS

FEI-FEI LI NICK THOMPSON AND JACK DORSEY

EDWARD SNOWDEN QUESTLOVE

SAN F R ANCI S CO
2 01 8

JONY IVE AND ANNA WINTOUR

SATYA NADELLA AND JENNY LAY-FLURRIE J O I I T O AND REID H O FFMA N

4 DAYS
72 + EVENTS
2 ,000 P EO P L E

DION WEISLER CASH CASH

CASH CASH TINKER HATFIELD

SEE WHAT YOU MISSED

WI R ED.CO M / 25

EVAN SHARP WITH MICHAEL CALORE AND ARIELLE PARDES

SUPPORTING
SPONSORS
 PROMOTION

WIRED.
NOW ON
YOUR
SMART TV.
WIRED’s Video App Now Available
on Your Smart TV
 26.12
LAUNCH

“ D O E S A TH EO RY THAT E XPL AIN S

EVERYTHING RUN THE RISK OF

EXPLAINING NOTHING?” P A G E 96

DEC 2018 DAN WINTERS 0 0 5

 MOVADO CONNECT TIME DESIGNED LIKE NEVER BEFORE.
POWERED WITH WEAR OS BY GOOGLE ™.
MACY’S
Wear OS by Google is a trademark of Google LLC. MOVADO.COM
 26.12
FEATURES

90
The Human in
the Machine
AI pioneer Fei-Fei
Li reboots the field
she helped invent.
BY JESSI HEMPEL

74
The
Miseducation
of AI
How to teach a
neural network some
common sense.
BY CLIVE THOMPSON

82
Self-Driven
Inside the DIY smart-
tech movement.
BY TOM SIMONITE

96
The Man
Who Explained
Everything
Karl Friston’s unified
theory of schizophrenia,
the universe, and AI.
BY SHAUN RAVIV

DEC 2018 CHRISTIE HEMM KLOK 0 0 7

 26.12
GADGET LAB

Stocking
stumper:
Hanayama’s
DOT puzzle

35
Wish List
48 awesome
holiday gift ideas.

0 0 8 JOSEPH SHIN DEC 2018

 PROMOTION

#MUSTWIRED.COM/STORE
VISIT US IN NEW YORK CITY
BROOKFIELD PLACE
 CONTENTS

A photodetector-
filled artificial
eye in Michael
McAlpine’s lab

22 28

Set to Stun Look Sharp

The new nonlethal arsenal A laser-powered autofocus for
star searchers

22 Chartgeist
Facebook, ebikes, and Bumblebee 30 Wired Guide:
The Future of Shopping
How tomorrow’s retailers will

24 get you to Buy Now

32 Farm to Cable
Rebooting the spirit of the
DIY internet
BY CLIVE THOMPSON

Disinfo War
SIX BY SIX
warming planet Big Tech slays the monsters it
BY VIRGINIA HEFFERNAN creates 108 Stories by WIRED readers
BY ZEYNEP TUFEKCI

20 It’s Alive
3D-printing bionic eyes, ears, 26 3 Smart Things ... O N T H E COV E R

and spinal cords … about our inner sense Art for WIRED by Axis of Strength

0 1 0 ACKERMAN + GRUBER DEC 2018

 RELEASE NOTES

Chiang), we now have

helpful algorithms
that’ll make us dinner
reservations, or as
Simonite explores on
page 82, can diag-
nose an ailing plant.
A development he’s
less taken with?
Smart replies in e-
mail. “They’re pretty
useful, I’ll admit, but
I think it illustrates
“You have to be OK a broader problem:
with looking at the When you let these
same item for hours things take over for
in a dark studio,” you, human subtle-
says photographer ties get lost—like
Joseph Shin, who the fact that I’m a
shot some 50 prod- grumpy Englishman.”
ucts over a five-day
period for our annual
Wish List, his third
time on the job. And
this work is not with-
out its technical chal-
lenges: He spent 40
minutes just getting
the reflections right
on a motorcycle hel-
met—four diferent
shots that were com- “How do you point out
bined into the final terrible behavior on
image. “Some people social media if point-
hate it, but I like it. ing it out brings new
It’s not as glamor- recruits?” That’s the
ous or as exciting as question University
shooting somebody of North Carolina
interesting, but for technosociologist
From left: me this is kind of my Zeynep Tufekci
Kelsey Lannin, own portrait.” explores in her new
Pia Ceres, column (page 24).
Halie Chavez, There should be
Rebecca ways to fight the
WIRED’S EDITORIAL FELLOWS Heilweil, and
internet’s worst trolls
Briana Flin
without legitimiz-
ing them, but social
media giants have
been reluctant to
address the problem
head-on. “The way
wice a year, wired welcomes a group of five editorial fel- things used to work

T
lows to do everything from gear and photography research When senior writer doesn’t work any-
Tom Simonite began more,” she says. “The
to fact-checking and video producing. The current crop, covering artificial new world doesn’t
who depart at year’s end, have their fingerprints all over intelligence around have its institu-
this issue. Reporting fellow Rebecca Heilweil writes about Get More W I R E D 2006, it was still the tions yet.” Tufekci
We launched a pay- stuff of science fic- is no stranger to
nonlethal weapons and 3D-printed body parts in the Alpha wall on W I R E D .com, tion and research social media’s com-
section, while products fellow Pia Ceres ofers suggestions but if you’re already labs. Though we plicated power. Her
for holiday gifts in Wish List (page 35). Research fellow a print subscriber, remain a long way book Twitter and
don’t worry. You can from conversing Tear Gas: The Power
Kelsey Lannin fact-checked Tom Simonite’s story about the DIY AI authenticate your with replicants (or and Fragility of Net-
movement (page 82), and photography fellow Halie Chavez shot subscription “digients,” to cite worked Protest looks
three of the photos on this page. Finally, check out video fellow and read all of our Simonite’s favorite at how the internet
stories ad-free. Get fictional AI, from has changed popular
Briana Flin’s work on the debut of our video series, Tech Efects, the hookup here: The Lifecycle of Soft- protest, from Occupy
available on wired.com and wired’s new streaming app. W I R E D .com/register. ware Objects, by Ted to the Arab Spring.

0 1 2 HALIE CHAVEZ DEC 2018

 @WIRED / [email protected]

Re: “The Great daughter has never

Upwelling”: As the showed interest in
world became con- my magazines, but
nected, regular something about
people acquired the 25th anniver-
superpowers. sary issue captured
“I remember the her attention.
moment Kevin Kelly Tonight, as part of
writes about, going her bedtime read-
from no URLs to ing, she picked
URLs in real life. It WIRED . As she was
was like our secret paging through the
club was leaking into magazine asking me
THE NEXT 25 the physical world.” questions, I noticed
Craig Mod the amazing num-
(@craigmod) ber of women who
via Twitter are featured. She
definitely noticed
“The trouble is, too. She asked me
the more people about many of them,
we connect to, the and I told her their
shorter and more stories. My 5-year-
terse our engage- old is growing up
ment is with them. in a world where
So are we really con- she sees women,
necting, or do we people of color,
just think we are?” LBGTQ, and people
Roman Koziol with disabilities
(@RomKoziol) featured in print,
via Twitter film, and TV. She
will never know any
different. There’s
a lot of ugliness
in the world today,
but it’s the strength
of those who are
blazing new trails
that is creating the
world my daughter
and son will live in.
I thank you for that.
My daughter gives
me hope. She went
Re: “Nobody Left to sleep with the
“During the past year I really techno-utopians to their origi- Behind”: With magazine. I know
enjoyed WIRED . I found the intro- nal hype program, maybe to fund mindful tech and that she will be
spective pessimism refresh- their latest life extensions?” design, companies dreaming of the lim-
ing and the articles informative. Frank Ashley via mail@WIRED .com can better serve itless possibilities
But the latest issue, WIRED @25, customers—and for her in the years
reminded me of why I never “In a lucid cover article WIRED employees— to come—because
liked the ‘militant optimism’ of admits they never got the with disabilities. of the path that
WIRED and how it always seemed future right. Every innovator “This gave me the the women before
like marketing and propaganda. and startup should read this.” hope that I desper- her paved.”
Hopefully we aren’t experiencing Philippe Méda (@icopilots) ately needed today. Pablo (@IamPablo)
the recommitment of these old via Twitter My 5-year-old via Twitter

0 1 4 DEC 2018
160+ HOURS OF TRAINING
ALL TO HELP YOU.
Our Small Business Technology Advisors have
the experience and training to recommend tech
solutions for your unique business needs.

SPEAK WITH AN ADVISOR TODAY:

877-BUY-DELL
D E L L .C O M /S M A L L B U S I N E S S

CLICK CALL CHAT

TECH. ADVICE. PARTNERSHIP.

MORGAN PAIGA
Small Business Technology Advisor

LATITUDE 7290
with an 8th Gen Intel ® Core™ i3 processor Connect securely wherever work takes you on
Starting at $1,049 America’s largest, most reliable 4G LTE network.

*Offers subject to change. Taxes, shipping, and other fees apply. Dell reserves the right to cancel orders arising from pricing or other errors. Intel, the Intel Logo, Intel Inside, Intel Core, and Core Inside are trademarks of Intel Corporation or its subsidiaries
in the U.S. and/or other countries. Microsoft and Windows are trademarks of Microsoft Corporation in the U.S. and/or other countries. Screens simulated, subject to change. Windows Store apps sold separately. App availability and experience may vary by
market. This product is certiied for use on the Verizon network. Dell, EMC, and other trademarks are trademarks of Dell Inc. or its subsidiaries. © 2018 Dell Inc. All rights reserved. 268900
 PURE VISIONARIES

When it comes to people and businesses that have big ideas for the future of mobility, you’ll ind them
in Michigan, aka PlanetM. With its progressive mobility legislation, Michigan was the irst state to legalize
self-driving vehicle testing and use on public roads. We also lead the nation in mobility-related patents
and connected and automated vehicle projects. If you’ve got a mobility idea that you want to bring to life,
come to the place that’s helping make the impossible possible. Come to PlanetM. Learn more at planetm.com
ARGUMENT

FOLLOW THE BEAVER

TRAPPERS, MEDIA THEORY,
AND OUR WARMING PLANET
BY VIRGINIA HEFFERNAN
ALPHA

HAVING GNAWED THEIR way across

the Bering Land Bridge with their
iron-glazed teeth, beavers by the
tens of millions straight-up built
North America. They worked like
rodent Romans, subjugating the
deciduous forests with formi-
dable infrastructure: canals,
lodges, dams that can last cen-
turies, and deep still-water pools
used to float building materials.
By clear-cutting trees and block-
ing streams, the nocturnal, semi-
aquatic creatures also damaged
the environment in some of the
same ways humans do. Much
later, beavers unexpectedly
became the toast of a rarefied
academic circle at the University
of Toronto, where they inspired,
of all things, media theory.
In The Fur Trade in Canada,
Harold Innis, a political econo-
mist known for originality and
intellectual derring-do, chroni-
cled a fierce four-way battle for
domination of Canada from the
17th to the 20th centu-
ries. The combatants were

DEC 2018 KATI SZILÁGYI 0 1 7

 ALPHA

beavers, indigenous trappers, est company in North America) is the message.”) 2. Psychologi-
European colonizers, and the aimed to weaken tribal bonds. cal and social states are created
merciless environment. These Europeans learned all they could primarily by media. 3. New media
observations led to Innis’ later from the better trappers, then technologies thoroughly change
books, Empire and Communica- encouraged them to depend societies and institutions.
tions and The Bias of Communi- on imported goods, including As once they bent the water-
cations, both about how certain brandy. This eventually broke ways of Canada to their will and
media (paper, radio, television) up the native communities and left humans scrambling to keep
contain implicit jingoistic values. gave the colonizers what’s known up, now the rodents have a fresh
It’s axiomatic: Humans follow in communications theory as an ambition: to establish dominion
beavers. When humans showed “information monopoly.” over still colder latitudes.
up in the pre-Columbian Amer- Humans may soon follow the
icas, various tribes built their beavers and push north again,
cultures around beaver dams,
where they harvested meat, fur,
THE BEAVERS INSPIRED A seeking not pelts but asylum
from extreme heat and drought,
and glands, including the musky MEDIA THEORY THAT ORAL, floods, and poverty. As if hurri-
secretion of the castor anal sac,
which is still used in perfume.
PRINT, AND DIGITAL MEDIA canes in the US and Revelation-
caliber fires as far north as the
Hundreds of years passed. ARE ALWAYS BIASED. Arctic last year weren’t signals
Europeans of the 17th century enough, the UN’s Intergovern-
became almost erotically fixated mental Panel on Climate Change
on a certain kind of supple men’s warned in October that, absent
high hat made of beaver, and they Innis discovered a dynamic rigorous intervention, Earth
skinned their continent’s supply tension between enduring media in 22 years will be almost 2.7
to near-extinction. So the English (artifacts like inscribed stone degrees Fahrenheit hotter than
established, in 1670, “the Gover- meant to memorialize traditions) it was in preindustrial days.
nor and Company of Adventurers and imperial media (artifacts like In a hot, parched, salty, and
of England trading into Hudson’s pamphlets meant to monopo- melting world, Canada can look
Bay” and sent the stouthearted lize trade). He identified a bias, like a life raft. But climate refu-
human subspecies known as toward time or space, implicit in gees should be warned: In the
trappers to chase the rodents the materials a group used. Hard, coming decades, even southern
up the Canadian waterways. heavy stuf could be passed down Canada might not be entirely
North of the Saint Lawrence through generations, where light, habitable. To get safely out of
River—and especially in the sub- ephemeral stuf was best used for the heat, you might even have
lime Precambrian shield, the far and wide proselytizing. to get all the way to the tundra
exposed section of billion-year- Because the parties to the fur of the far, far north.
old metamorphic crust that trade mimicked, and pushed, And you would hardly be trail-
runs from Michigan to Green- one another forward—beavers blazing. Beavers, ever adaptable
land—the beavers, with their imitated the damming styles of and enterprising, got to the tun-
lush pelts that fetched the high- humans, humans dressed as bea- dra first—and their now flour-
est prices from European mil- vers, animal and human cultures ishing Arctic empire can be seen
liners, turned haute couture. fought and fused—their ways of from space. As so many times
Because the indigenous groups communicating evolved rapidly. before, they pushed past the
had the advantage of experi- Innis’ germinal work inspired northern edge of their traditional
ence, trappers from Hudson’s the so-called Toronto School, habitat, out of their comfort
Bay Company (today the old- which helped shape the McLu- zone, exacerbating and repair-
han Centre for Culture and Tech- ing and fleeing climate change all
nology, named for Innis’ most at once, past Alaska’s boreal for-
famous protégé, Marshall McLu- est into the Arctic, ambling ever
han. The school then developed a Virginia Heffernan upward, their luxurious pelts
media theory with these tenets: (@page88) is a WIRED thickening and thickening. 
contributor. She wrote
1. Oral, print, and digital media about digital hyper-
are always biased. (“The medium inflation in issue 26.11.

0 1 8 DEC 2018
 T H E A R T O F E S S E N C E

With its ultra-thin wallpaper design, the LG SIGNATURE OLED TV is technological innovation at its greatest.
World-renowned for its superior picture quality and picture-on-wall design, the TV offers a perfect black canvas that plays host to vibrant, accurate color.

Find your LG SIGNATURE TV at www.LGSIGNATURE.com

Included TV-to-AV Box cable is required for TV operation.
 ALPHA

OBSESSED ELECTRONICS OFTEN DON’T mesh Engineer Michael McAlpine shows

off his new bionic peepers.
wel l w it h f lesh a nd blood.

IT’S ALIVE! Cochlear implants can irritate

the scalp; pacemaker wires dis-
lodge; VR headsets weigh heav-

3D-PRINTING ily on the face. That’s why, for the

past six years, Michael McAlpine

BODY PARTS
has been Frankensteining alter-
goopy primordial ingredients
natives. A mechanical engineer
into a mold. The result is then
at the University of Minnesota,
given a few weeks to rest in a
he creates prototypes of bionic
nutrient-packed bath, which
body parts with nice, soft com-
allows the cells to grow around
ponents—some of them alive.
and within any core electronics.
The key to his electro-organic
Before they’re human-ready,
organs is his custom-made 3D
though, these replacement bits
printer, which McAlpine loads
first need to work well in rats and
w ith silicones, metals, and
other animals, McAlpine says.
human cells sourced from the
While tests show that his imi-
university’s med school. (They
tation ear can successfully per-
come in a gel-like culture so they
ceive music—a recording of “Für
stay happy and functional, he
Elise”—he has yet to connect the
says, while they’re handled.)
prosthetic’s radio receivers to a
His 3D-printed “ear,” made by
living thing’s nervous system.
enveloping a coil antenna in liv-
Same goes for his latest creation,
ing matter, requires electrically
an eye filled with a web of photo-
conductive silver nanoparti-
detectors that can translate light
cles and cartilage-forming cells,
into electrical signals—a first
while his “spinal cord” calls for
step to artificial vision.
neuron-forming cells and a
Other researchers are excited
translucent column of silicone.
by recent advances in lab-grown
Whatever the desired organ, the
human organs, but McAlpine
computer-guided nozzles take up
doesn’t think that should be the
to an hour to extrude McAlpine’s
only goal. “I don’t know that you
necessarily need to replace biol-
ogy with more biology,” he says.
He imagines enhancing his body
parts with extrasensory capabili-
ties, pointing to the medical tech
company Second Sight, which
thinks its retinal implant might
one day allow blind people to see
infrared wavelengths most of us
can’t. “You could give impaired
people superhuman abilities,”
he says. “In the future, you’ll
give the average person these
abilities as well.” In McAlpine’s
worst sci-fi nightmares, robots
will be stronger and smarter than
humans—so let’s start building
bio-augmented cyborg defend-
ers now. —Rebecca Heilweil

0 2 0 ACKERMAN + GRUBER DEC 2018

 T H E A R T O F E S S E N C E

The LG SIGNATURE refrigerator's beautiful design goes beyond its stunning exterior to a variety of advanced features.
And with innovations like InstaView™ Door-in-Door®, you can see inside with just two simple knocks.

Find your LG SIGNATURE refrigerator at www.LGSIGNATURE.com

 ALPHA

1. Laser-Induced
Plasma Effect
Still in the lab, this
blaster employs When the ground
two lasers. The first panels detect incom-
pulses on and off ing wheels, they emit
to dislodge atmo- high-voltage pulses
spheric electrons and that disrupt the vehi-
spin up plasma, while cle’s engine—but
the second beams won’t electrocute
straight into the the passengers, giv-
resulting ball of ion- ing guards more time
izing gas to release to investigate suspi-
an ear-splitting burst cious visitors.
of sound energy.
4. Variable
2. Carbon Nanotube Kinetic System
Thermophone Once a Darpa sci-
Instead of a tradi- ence project, Pep-
tional loudspeaker’s perBall’s AR15-esque
array of cones, coils, gun can fire 180
and magnets, this rounds of micro-
lightweight projector pulverized burn-
pushes heat currents ing irritant (or stink
through cylinders of bombs or inky liq-
pure, finely twisted uid) before reloading.
carbon. Rapidly First military deploy-
warming and cooling ment? Afghanistan.
these tubes creates
noise, a technique 5. Maritime Vessel
that, the Defense Stopping Occlusion
Department hopes, Technology

SET TO STUN could one day imbue

tiny drones with the
power to scream
When the eely hag-
fish senses dan-
ger, it spews out a

THE NEW NONLETHAL ARSENAL “Drop your weapons,

enemy scum!”

3. Pre-Emplaced
cloud of milky slime
that blinds and gags
potential predators.
The DOD’s version
WE MAY NE VER KNOW whether Cuba attacked American diplomats with Electric Vehicle of that secretes its
microwave weapons—but we do know similar devices exist. The US Stopper own mucous gunk—
At Tinker Air Force one idea is to affix
Department of Defense’s Joint Non-Lethal Weapons Directorate, along Base in Oklahoma, the contraption to
with a host of private arms companies, has spent decades testing every- the DOD is experi- autonomous water-
thing from long-range wireless Taser bullets to sonic guns that can menting with elec- craft—fine-tuned to
trified speed bumps. ensnare the propel-
disable a car engine from 150 feet away. The one requirement: These lers of enemy ships
weapons must emit less than 10,000 joules, the amount of energy it and submarines. Now
takes to kill a person. Bombs incite wars, the thinking goes—but North the agency is looking
at whether spider’s
Korea miiight forgive the “accidental discharge” of a directed-energy silk could fortify
laser pulser (also, as it happens, in the works). —Rebecca Heilweil a next-gen recipe.

Reasons to Not
Reasons to See Bumblebee Delete Facebook Reasons to Buy an Ebike
CHARTGEIST
BY J O N J. E I L E N B E R G

It’s a Transformers movie Chance to “engage” with brands Up to 3X the weight of a normal bike

John Cena as “Agent Burns” Rebel against the masses by actually Up to 25 mph, uphill
watching Facebook Watch

The “plot” Uncle Ed’s shares of Russian trolls Yet another device to charge
posing as “America Freedom Warriors”

Opportunity to share your data with “I’m not riding, I’m exercising”
It’s not directed by Michael Bay
interested parties worldwide

0 2 2 JEREMY PERRODEAU DEC 2018

 T H E A R T O F E S S E N C E

With the minimal design of LG SIGNATURE laundry machines, technology is refined to its simplest, most exquisite form.
Combined with TWINWash™ technology which allows you to wash two loads at once, it's the perfect marriage of convenience and beauty.

Find your LG SIGNATURE laundry machine at www.LGSIGNATURE.com

 Jones is a kind of real-world
Voldemort. Speak his name to
condemn his conspiracy theories
and you draw more attention to
his hateful ideas. It’s like fight-
ing fire with oxygen tanks instead
of fire extinguishers. The tools
breathe more life into the flames.
This is attention-gaming, and
Jones excels at it. At the hear-
ings, Jones sat behind Facebook
COO Sheryl Sandberg and Twit-
ter CEO Jack Dorsey as they testi-
fied, streaming the action from his
phone. He heckled Marco Rubio
as the senator talked to report-
ers. His stunts blew up online and
got him into The New York Times.
Jones has been spreading his
rage-fueled disinformation for
a while, but I have rarely writ-
ten about him publicly. With the
exception of three tweets among
tens of thousands I’ve posted, I
haven’t referred to Jones by name
on Twitter. He was “you know
who” to me. This was a deliberate
decision; I knew that he counted
on his critics to amplify his mes-
sage. I didn’t want to broaden the
reach of his curse.
ALPHA
So why am I naming him now?
That fuss Jones made at the Capi-
tol was a last gasp. He’d just been
ZEYNEP TUFEKCI banned from YouTube, Facebook,
Apple, and Spotify. Soon after, he

DISINFO WARS
was also banned from Twitter.
He’d been deplatformed.
Now that his Voldemort-like

BIG TECH SLAYS THE powers have vanished, it’s not

just possible to discuss Jones—
it’s necessary. His deplatform-

MONSTERS IT CREATES ing is easy to celebrate. Though

some may wish that good speech
is the best way to drive out bad
WHEN ALEX JONES crashed the congressional hearings looking into speech, the harms he perpe-
big tech platforms back in September, Lord Voldemort kept com- trated can’t be dealt with in the
ing to my mind. Even if you haven’t read the Harry Potter books, marketplace of ideas. There is
you probably know that almost no one in the wizarding world will no reasoned debate or enlight-
speak this archvillain’s name aloud; he is referred to only as “he who ened compromise with the idea
must not be named” or “you know who.” In the final book, Volde- that parents of children gunned
mort puts a curse on the name, so that merely uttering it acts like down at Sandy Hook Elementary
a beacon for the wizard’s crew of Death Eaters. Zeynep Tufekci School in Connecticut were just
Eager to communicate something crucial about the evil lord’s (@zeynep) is a actors in a false-flag operation
W I R E D contributor
latest plot to his friends, Harry at one point blurts out Voldemort’s and a professor at later used to promote gun con-
name. What follows are many, many scary pages. UNC Chapel Hill. trol. Nor is there anything to say

0 2 4 ZELOOT DEC 2018

LIFT HERE

LIFT HERE MACY'S & MACYS.COM

 PROMOTIONS + SPECIAL OFFERS + EVENTS

EXCITEMENT AWAITS IN CHEE HOO! BY MAUI JIM

Experience breathtaking views, bask in uncompromised protection, and revel in the
sleek styling of this classic style.

Powered by PolarizedPlus2® lens technology, all Maui Jim sunglasses immerse your
world in the richest color, clarity, and detail. Shown in the HCL® Bronze lens (H765-57) and
available in three additional frame/lens color combinations, Chee Hoo! offers classic lines,
with a unique mix of materials; lightweight acetate with double-injected metal and rubber
temples. Go from the boardroom to the beach and everywhere in between with the ultimate in
performance and style—Maui Jim. The view’s better from here.

Visit www.mauijim.com

INSIDER

Join our growing

community of
Insiders for access
to the latest ideas,
intelligence, events,
and promotions.
W I R E D I N S I D E R .CO M

VISIT US ONLINE AT WIREDINSIDER.COM + FOLLOW @WIREDINSIDER ON TWITTER + LIKE WIREDINSIDER ON FACEBOOK

 ALPHA

about his claim that KKK mem- due process or accountability,

bers are “just Jewish actors” a frustrated public is left with
pretending to be Nazis. (Lots of appealing to a few powerful ref-
actors in his world.) Yet while erees—and crossing our fingers.
I’m happy that Jones has lost his This is complicated stuff.
megaphone, I’m troubled both We’re dealing with three ideas
by the system that let him have that are structurally in tension:
it and the way it was taken away. that hate speech, harassment,
Simply put, the influential digi- false accusations, and baseless
tal platforms are built to gener- conspiracies (like antivaccina-
ate more Voldemorts, while also tion claims) cause real harm; that
amassing worrisome amounts of free speech is a crucial value; and
centralized power. that it’s necessary to deal with
The platforms are in the busi- algorithmic amplification and
ness of harvesting attention, attention-gamers.
and Jones and his kind are good Legislators, courts, users,
at delivering it. Jones’ support- and the platforms themselves
ers lapped up his content and have to be involved. There are
stoked outrage, leading to even some precedents we could use
more views. On YouTube, Alex from older technologies. Some
Jones’ channel was so heavily rec- updated version of the fairness
ommended that watching regular doctrine, which required radio
political content often led to an and television stations to devote
autoplay of his red-faced rants. time to issues of public impor-
But that wasn’t the end of it. A tance and seek out a multiplic-
network of hateful or conspira- ity of views, could be revived for
torial content suppliers provide the digital age. We could come up
these platforms with enormous with a kind of Fair Credit Report-
amounts of “engaging” content ing Act that gives users a right to
to attract users. Their recommen- challenge a platform banishment.
dation and sorting algorithms, There could be antitrust actions
designed to maximize engage- against centralized platforms
ment and the amount of time (along with user protections), or
onsite, spread them farther. upstarts could ofer alternatives
But if the unaccountable man- (with better business models).
ner in which the tech platforms As with most social problems, we
can amplify harmful content has have to accept that there is no sin-
led to a crisis, so has the facility gle, perfect solution, no avoiding
with which they can eject it. Jones trade-ofs, and also that inaction
delivered eyeballs for many years. is a decision too.
Then the platforms succumbed At a WIRED event in October,
to pressure and banned him, all Jack Dorsey said people don’t
within the span of a few weeks. view Twitter as a service. “They
The tech platforms have arbi- see what looks like a public
trary power to decide what to square,” he said, “and they have
amplify, and thus what to bury, the same expectation as they have
and they have the power to ban- of a public square, and that is
ish as they wish. There is noth- what we have to get right.”
ing aside from backlash to stop There’s lots of work to be
them from deplatforming, say, done. But getting it right is too
tech critics or politicians who important to be left to Dorsey
call for shutting tax loopholes for (and Mark Zuckerberg and
massive corporations. Without Susan Wojcicki) alone.

0 2 6 DEC 2018
BE MERRY AND LITE
Great Taste. Only 96 Calories.

MILLER LITE. HOLD TRUE.

 ALPHA

THE TWINKLING OF a starry night The device shoots four 22-watt

sky is romantic, sure. But for laser beams—each nearly a foot
astronomer Dominika Wyleza- in diameter and 4,000 times as
lek, “it’s a nightmare.” That’s powerful as a laser pointer—
because Wylezalek studies gal- into the sky, where they radiate
axies billions of light-years away, at a wavelength of 590 nanome-
and all that finicky glimmering— ters, exciting sodium atoms in
she’d call it “turbulence”—gets the upper atmosphere. When the
in the way. So she and her col- atoms get whipped into a frenzy,
leagues at the European South- they emit photons that shine like
ern Observatory fire up this stars. Those artificial stars form
Star Wars–worthy laser canon reference points for the tele-

LOOK SHARP
in northern Chile to measure dis- scope, helping its internal mir-
tortion in the atmosphere and ror correct for pesky twinkling.
improve the astronomical image. It’s like the autofocus on a cam-

AUTOFOCUS
era. The result: vastly sharper
photos. “We’re trying to push the
telescope to not see any distor-

FOR FAR-
tion from the atmosphere at all,”
Wylezalek says. For the past two
years, this laser blaster has been
shedding new light on the darkest

FLUNG STARS reaches of space. In 2019 it will

be used to study super-massive
black holes in far-of galaxies and
dense star clusters within the
Milky Way, revealing formerly
hazy secrets of stellar evolution
in hi-def. —Laura Mallonee

0 2 8 ENRICO SACCHETTI DEC 2018

 Fast is
a Gig-speed network
that gets business done.

Beyond Fast is
technology that makes
business boom.
There was a time when megabits per second was
all that really mattered. Which is why Comcast built
the nation’s largest Gig-speed network to deliver
unrelenting speed to more businesses in more places.
Now we’re going beyond.
Beyond connecting businesses. To helping you
create experiences that anticipate what customers
and employees want.
Beyond network complexity. To zero-touch,
one-box solutions that simplify your network and
amplify your performance.
Beyond the best for your money. To the best solutions
for your business at an even greater value.
We’re taking our relentless speed and adding smart
technologies and advanced applications that enable
your business to do what you never imagined—at
speeds you never thought possible.
Take your business beyond.
ComcastBusiness.com

Restrictions apply. Not available in all areas.

Maximum download 940 Mbps when hardwired via
Ethernet. Actual speeds vary and are not guaranteed.
©2018 Comcast. All rights reserved.
 ALPHA

When the first Ama-

zon Go store opened
to the public in Seat-
WIRED GUIDE tle this year, visitors
said it felt like shop-
lifting. That was the

THE
point: walk in, grab
what you need, and
go, without ever tak-
ing out your wallet.

FUTURE The shop’s balletic

system of computer
vision, weight sen-
sors, and deep learn-

OF ing is similar to the

tech being used in
self-driving cars. But

SHOPPING
this application ren-
ders checkout lines
obsolete. Amazon
plans to open another
3,000 cashier-free Download the Amazon Go app, Scan the QR code at an entry turnstile
which is linked to your Amazon account. to identify yourself. Shopping in a group?
MALLS, HOW QUAINT. The brick- stores by 2021, The app generates a unique QR code for Scan the phone’s QR code once per person
reports Bloomberg. each shopping trip. to link multiple members to one account.
and-mortar business model,
already battered by ecom-
merce, now has to fend off
m-com merce (mobi le, of
4 Ways to Game Dynamic Pricing
course), even v-commerce— Online prices are a 1. Wait It Out 3. Stay Anonymous
voice—thanks to the prolif- moving target—you Price-tracking services moni- If you’re an infrequent online
eration of smart speakers. have to know when tor millions of Amazon products shopper, log out of any store
to strike. Yuri Levin, and send alerts via email or Twit- accounts before browsing. Peo-
Data-tracking retailers are a Queen’s University ter when prices drop. The site ple who rarely make online pur-
plotting new ways to topple analytics professor CamelCamelCamel has a tool that chases tend to pay more for
Amazon, which accounts for and dynamic pricing charts an item’s price over time, their purchases than anony-
expert, reveals how to revealing patterns that help you mous shoppers—potential new
49 percent of all US ecom- score online deals on know which month, week, day, or customers—whom retailers are
merce sales (trouncing its everything from air- hour to pull the trigger. eager to woo.
closest competitor, eBay, by fares to appliances.
a factor of seven). Meanwhile, 2. Toss Your Cookies 4. Relocate Yourself
physical stores and their Online retailers keep tabs on Online prices can difer based on
your clicks, and some might your location, often according
flighty cousin, the pop-up, bump up the price if, say, you’ve to your proximity to the nearest
are being transformed into been lingering on that winter store or warehouse. Try plugging
AI-enabled spaces, where a break flight to Miami. Browse in in a friend’s or family member’s zip
Incognito mode and clear your code and have the item shipped to
robocart can optimize your cookies regularly. their address to save.
shopping route and show-
rooms are tricked out with
touchscreen mirrors. Soon,
Buy Me Maybe
even walls will be unnec- TOTAL RETAIL SALES:
e s s a r y : V i r t u a l b r o w s- More than a third of shoppers use $3.7 Trillion
ing is on the rise, thanks smartphones to browse, research,
or compare prices before they buy.
to VR adopters like Lowe’s
and Ikea. Welcome to the
hyper connected, friction- PURCHASES MADE AFTER
RESEARCHING PRODUCTS
f ree f ut u re of shoppin g. OR COMPARING PRICES
BY SMARTPHONE:
$1.3 TRILLION (36%)

PURCHASES MADE ONLINE:

$506.7 BILLION (14%)

By Caitlin Harrington
PURCHASES MADE BY
(@caitharr), a WIRED
SMARTPHONE:
researcher. Additional report- $117.9 BILLION (3%)
ing by Rebecca Heilweil.

DEC 2018
 More than 150 companies are
developing register-free retail,
according to CB Insights.
Zippin: In August this cashier-
less San Francisco concept store
became the first Amazon Go
emulator to open in the States.
BingoBox: This startup operates
more than 300 human-free RFID-
powered convenience stores
throughout China, with plans to
reach 2,000 locations by 2019.
Kroger: The supermarket giant’s
Scan, Bag, Go model lets shop-
pers scan the barcodes on their
groceries and pay straight from
their smartphones. The tech is
already in nearly 400 stores.
Standard Cognition: This San
Francisco startup has partnered
Smart cameras create a 3D model of each Computer-vision and AI algorithms rec- When you’re finished, just walk out. with Japanese drugstore supplier
patron. Tech called articulated motion ognize products, and weight sensors in A camera array clocks your departure.
analysis maps your body movement and the shelves detect when merch is removed. Minutes later, Amazon charges your Paltac to open 3,000 checkout-
distinguishes you from other shoppers. Pick up an item to add it to your virtual cart. account and sends a receipt via the app. free shops by 2020.

Ask a “Conversational How are Watson’s suggestions of its ingredients. We incorpo-

different from those based on rate health and nutrition apps.

26
Commerce” Expert previous purchases? We educate the bot on the micro
“The assistant uses conversational and macro nutrients required
Watson Assistant knows what interaction to understand what throughout the pregnancy cycle.”
you want before you do. IBM’s AI- your goal is. If you’re looking at
powered chatbot, already in use outerwear, maybe you’re shop- What else are these shopping
by thousands of companies, asks ping for a ski vacation. With gro- chatbots capable of?
online customers about their pref- ceries, maybe your goal is weight “It’s not all just about chat.
erences and mines additional data loss or athletic performance. Humans learn through videos
(like shopping history and the local Watson considers context.” and pictures. A clothing shopper
weather) to recommend prod- might say, ‘That’s about what I’d
ucts. The bot can be embedded in How are the recommendations like, but the thing Selena Gomez
stores’ mobile apps, chat apps like personalized? was wearing the other night was
Facebook Messenger, and desk- “Using the grocery example, we closer.’ We’re working on an idea
top sites. In the near future, says don’t just load Watson with data- where we show them that image
IBM’s consumer-AI specialist, Ted bases of grocery products. We and they can pick from similar
Percentage of Westerheide, everyone will have train the AI with recipes. Rather options. We’ve also begun using
smart-speaker owners their own virtual personal shopper. than simply identifying an item linguistic and tonal analysis to
who have made a We asked him how these chatbots as gluten free, for example, Wat- detect customer emotions, such
purchase by voice can shape our buying habits. son knows the molecular makeup as anger, happiness, and disgust.”

The Latest Shopping Tech, From AI to VR

Stock Bots Virtual Showrooms Face Time Fashion Forward Cinderella Scanners Swipe and Shop
Meet your new Hardware giant Gourmet confectioner Fashion retailer Far- Fleet Feet debuted Now users can tap
mechanical store Lowe’s debuted its Lolli & Pops recently fetch unveiled touch- scanners by Volumen- stickers on Instagram
clerks. Walmart part- “Holoroom How To” installed facial- screen mirrors and tal that generate a 3D Stories to display mer-
nered with Bossa last year, which guides recognition cameras clothing racks that model of your feet in chandise details and
Nova Robotics to headset-clad DIYers in stores to flag regu- sense when an item is five seconds. An AI shopping links. Insta-
deploy inventory- through home-im- lars and compile cus- removed—then beam algorithm extracts gram reportedly plans
tracking droids in provement projects tomized shopping a virtual version to 10 metrics to recom- to launch a stand-
50 stores this year. in VR. recommendations. your smartphone. mend your ideal shoe. alone shopping app.

JAMIE CULLEN
0 3 1
 ALPHA

was able to open in Grove, Oklahoma (popu-

lation 7,060) in 2016. Area schools are hand-
ing out Chromebooks, doctors are exploring
CLIVE THOMPSON telemedicine, and people no longer need to
hoof it to a library for faster connections.

FARM TO CABLE
In one sense, this is merely a story about
how to end the rural-urban digital divide:
Don’t rely on big corporations, and instead

REBOOTING THE SPIRIT help locals band together with the kind of
government grants or low-cost loans that
helped bring electrification in the 1930s.

OF THE INTERNET (And indeed, the federal government has

been ofering loans to the co-ops.)
In a deeper sense, the gumption of these
co-ops is super inspirational. It’s more inter-
net than the internet these days. In Silicon
BACK IN THE early 1930s, farmers couldn’t get wired. The big-city Valley, Big Tech has turned cyberspace into
electric utilities claimed that delivering power to customers spread a dreary strip mall of centralized services
out in rural areas wasn’t profitable. So eventually the locals rolled up and corporate choke points. It’s gotten big;
their sleeves and did it themselves. They formed electric co-ops and it’s gotten boring. Meanwhile, the folks in
strung their own damn wires, aided by cheap federal loans. Today Oklahoma are building networks with the
there are nearly 900 rural co-ops still providing their communities self-governing fervor of early local ISPs or
with electricity. A DIY success story! ¶ Now history repeats itself—with the old-school blogosphere or even, hell,
broadband. Thirty-nine percent of rural Americans had no access to Usenet communities.
home broadband in 2016 (compared with 4 percent of folks in urban That spirit is worth emulating in the rest
areas), because big telcos say it’s too expensive to build afordable of the online world. Many folks are annoyed
fiber-optic broadband in the countryside. Residents have to make do at Big Tech for tolerating abuse, for spying,
with dialup or Wi-Fi from a library. ¶ So co-ops are solving the problem for sneakily triggering compulsive use. What
again. In rural Oklahoma, for example, the Northeast Oklahoma Elec- if, instead of kvetching and waiting for tech
tric Cooperative recently laid 2,497 miles of fiber-optic cable—a feat monopolies to reform, we set up more user-
that required blasting through some bedrock—to launch its broadband run co-ops to operate upstart services we
Bolt Fiber Optic Services. Today Bolt serves almost 9,000 members, actually want? Imagine co-op social net-
ofering gigabit connections for less than you’d pay for comparable works that wouldn’t need to algorithmi-
service in a city. ¶ As for the local impact? “It’s been huge,” Ricky Hig- cally lure users into endless feed-scrolling
nite, Bolt’s director of IT, tells me. The rollout of broadband meant “engagement” to keep the ad dollars sluic-
that an aerospace factory with the potential of 100 good-paying jobs ing. (They also wouldn’t have to chase meta-
static growth to please VCs.) “Co-ops are
owned by the members, so it’s very bottom
up,” notes Jim Matheson, head of the National
Rural Electric Cooperative Association.
Yeah, I know, this is nuttily idealistic. Alter-
native social platforms like Diaspora haven’t
exactly thrived. But then again, maybe the
goal isn’t to be huge but rather, as with the
DIY co-ops, to serve tightly focused commu-
nities. Even little efforts could effectively
spook incumbents into reform. The academic
Rob Seamans has found that “the threat of
entry is enough.” When a farmers’ co-op
plans to roll out broadband, the big compa-
nies suddenly decide it’s time to upgrade.
The DIY spirit is out there, and it’s blast-
ing through the bedrock of Oklahoma.
Write to [email protected].

0 3 2 ZOHAR LAZAR DEC 2018

Goodnight, normal trading hours.
Say hello to 24-hour trading, five days a week. TD Ameritrade is the first retail
brokerage to offer around-the-clock trading on select securities. The future
of trading has arrived.
Get up to $600 when you open and fund an account.

Visit tdameritrade.com/trade24-5 to learn more.

Extended hours trading is subject to unique rules and risks, including lower liquidity and higher volatility. Extended hours trading not available on
market holidays. See tdameritrade.com/600offer for offer details and restrictions. This is not an offer or solicitation in any jurisdiction where we are
not authorized to do business. TD Ameritrade, Inc., member FINRA/SIPC. © 2018 TD Ameritrade.
 GO BIG
THIS HOLIDAY SEASON.
Introducing the limited-edition one-liter bottle of Patrón Silver, filled with the same ultra-premium
tequila found in our traditional bottle, only 33% more. Featuring a vibrant design inspired by Mexico’s
rich tradition and history of glass craftsmanship, it’s the perfect gift for anyone on your list.

THE FOURTH IN OUR SERIES OF ONE-LITER

COLLECTIBLE HOLIDAY BOTTLES.

The perfect way to enjoy Patrón is responsibly. Handcrafted and imported exclusively from Mexico by The Patrón Spirits Company, Las Vegas, NV. 40% abv.
 Joseph Shin

December 2018
Issue 26.12

Contributors: Michael Calore, Pia Ceres, Alexander Davies, Lauren Goode,

Brendan Nystedt, Arielle Pardes, Adrienne So, and Jeffrey Van Camp

Page 035

TAL LEMING, T Y P E S U P P L Y
 Ride the Lightning

Vanmoof banishes bulky batteries

by hiding a 504-watt-hour cell
inside this ebike’s 42-pound frame.
The front-wheel motor has a city-
shrinking 93-mile range, can pump
out 500 hill-killing watts, and will
hit 20 mph. Even if some knuckle-
head conquers the built-in lock,
a GPS tracker will help you (well,
the cops) recover your steed.
Vanmoof Electrified S2 | $3,398

Galaxy Notes

Perhaps the best gift is a gentle

reminder of how tiny we are in the
universe. Each postcard in this
set of 50 features an image of the
cosmos, from ancient star maps to
The Night Sky NASA archival photos, like Apollo
16’s shot of the far side of the
moon. Pin them up for inspiration,
or buy some stamps (remember
those?) and mail them to distant
SEE SOMETHING YOU LIKE? stargazers. The Night Sky | $19
Buy all of this cool gear now at
wired.com/wishlist2018
DEC 2018
0 3 6
 Pyro Brainiac

How do real campers tame fire?

With Bluetooth! Ignite your fuel
(charcoal or dead trees) in the
FirePit’s mesh body, then dial the
flames up or down in the app,
which controls 51 teeny air jets in
the burn chamber. In addition to
stoking the fire, the jets keep the
smoke in check. The recharge-
able power pack is good for up to
24 hours of computer-controlled
burn. BioLite FirePit | $200

Magic Hour

Cameras and watches share

mechanical DNA, which Leica is
celebrating with its new line of
timepieces. Assembled at the
company’s Leitz-Park campus, this
41-mm manual watch ticks for 60
hours between windings and fea-
tures a jet-setter-friendly GMT
complication. The brushed steel L2
even sports Leica’s iconic red dot
on its crown. Leica L2 | $16,500

Blipblox

DJ Playdate

Blipblox might be designed for

DJs ages 3 and up, but this is no
rinky-dink xylophone. The easy-
to-use synthesizer has pro-ready
features like MIDI support, mul-
tiple modulation filters, and an
integrated drum machine. So
your tot can mix beats before
bedtime, then you can “borrow”
it for your own midnight set.
Blipblox | $189

DEC 2018
0 3 8
Play just got serious
in the all-new Toyota Avalon.
Reboot the thrill of driving in Avalon Touring. Featuring available
Sport+ Mode with Adaptive Variable Suspension (AVS), 19-in. sport
wheels and daring new styling, Avalon pulls a complete 180 on the
driving experience itself. Let’s Go Places.
 Wizard of iOS

Skip the trip to Ollivanders and

pick up this coding wand instead.
The Kano app for tablets and
computers gives Hogwarts hope-
fuls step-by-step instructions to
build JavaScript routines. A wave
of the wand casts an enchant- Apple Watch Series 4
ing onscreen spell to levitate a
feather, play a ghostly harp, or
float a butterbeer. Harry Potter
Kano Coding Kit | $100

Harry Potter
Kano Coding Kit

Beat Master

The latest Apple Watch is an

upgrade over the last version in
almost every way, with a display
that’s 30 percent larger and a new
sensor array that can perform
an electrocardiogram to spot an
irregular ticker. Just remember,
you’ll need an iPhone to use the
thing. Apple Watch Series 4 |
$399 and up

Green Platform

With shapes optimized by compu-

tational fluid-dynamics software,
Firewire’s surfboards would be
a blast to ride even if they didn’t
meet the eco-friendly standards
set by Sustainable Surf’s Eco-
board program. This versatile keel-
fin fish is meant to tackle midsize
waves and keep the ocean a little
bit cleaner along the way. Firewire
Surfboards Go Fish | $765

DEC 2018
0 4 0
To see your home in a new light,
switch your switch.

Introducing NOON.
8LIçVWXWQEVXPMKLXMRKW]WXIQXLEXÙWEGXYEPP]WQEVX
3RP]2332EYXSQEXMGEPP]HIXIGXW]SYVI\MWXMRKFYPFWXSGVIEXIFIEYXMJYP
GSSVHMREXIHPMKLXMRKWGIRIW2SQSVIW[MXGLèMTTMRKSVHMQQIVWPMHMRK
7IXXLIVMKLXPMKLXJSVER]EGXMZMX][MXLEW[MTISJEW[MXGL

;MXLXLI2332ETTGVIEXIGYWXSQWGIRIWWIXEWGLIHYPIERHQSVI
2332PSSOWEJXIV]SY[MXLPIEVRIH:EGEXMSR1SHILERHWJVII2MKLX0MKLX
ERHZSMGIGSRXVSP0SZI]SYVW[MXGLIWEWQYGLEW]SYPSZI]SYVPMKLXW

0IEVRQSVIEX RSSRLSQIGSQ
 Pinch Hit

Tireless, indestructible—every kid

wants to be part robot. Now your
ofspring can live the cyber dream.
This programmable arm has grip-
pers with treads to help grasp
objects. Your budding engineer
can teach it up to 100 movements
as they learn coding, automation,
and the basics of its Arduino brain.
HamiltonBuhl STEAM Robo-
Arm Kit for Arduino | $219

Flap Your Jaws

This flat-pack sculpture kit is

a crafty take on the blood-
thirsty terror of Amity Island.
Bend the 16 vinyl flaps, secure
them with the included screw-
pegs, and a 14-inch great white
slowly emerges. Gently wiggle
the assembled elasmobranch to
admire the serene swing of its
tail. Shark Kinetic Toy Kit | $25

Shark Kinetic Toy Kit

Tile Pro

tion in the app or make it chirp

audibly so the precious is easier
to find. The newly updated Tile
Pro has an extended range of
300 feet and runs on replaceable Xbox Design Lab
coin cell batteries that last Controller
at least a year. Tile Pro | $35

DEC 2018
0 4 2
 insurance and you could save.

geico.com | 1-800-947-AUTO | Local Office

Some discounts, coverages, payment plans and features are not available in all states, in all GEICO companies, or in all situations. Boat and PWC coverages are underwritten by GEICO Marine Insurance Company. Homeowners, renters
and condo coverages are written through non-affiliated insurance companies and are secured through the GEICO Insurance Agency, Inc. Motorcycle and ATV coverages are underwritten by GEICO Indemnity Company. GEICO is a registered
service mark of Government Employees Insurance Company, Washington, DC 20076; a Berkshire Hathaway Inc. subsidiary. © 2018 GEICO
 Roland Aerophone Go
Too Saxy

Learn a new wind or reed instru-

ment, minus all the skronking.
Finger the keys of Roland’s digi-
tal horn to draw out the simulated
tones of 11 instruments, includ-
ing flute, clarinet, and four types
of sax. Unlock 50 more sounds—
from bagpipes to strings—in the
app. Practicing late at night? Plug
in headphones to jam in private.
Roland Aerophone Go | $500

Amazon Echo Sub

Low Talker

Amazon’s Echo speakers are fun

to chat with, but they don’t really
get down. The Echo Sub is here
to help and will pump a lot more
bass into Alexa’s DJ sets. The
6-inch downward-firing speaker,
powered by a 100-watt onboard
amp, wirelessly pairs with an Echo
or Echo Plus speaker—or two of
the same Echo devices for stereo
sound. Amazon Echo Sub | $130

Machine Learning

Gaming obsessives will love Evan

Amos’ historical book, which
offers vividly detailed exploded
views of every canonical console
from the Magnavox Odyssey to
the Nintendo Wii. It’s the next-
best thing to taking apart vintage
machines yourself, and it looks
much nicer on your coffee table.
The Game Console | $25

Kitty Critter

Your cat thinks it’s so smart. Let’s

see if it can outwit a semi-sentient
toy. The sensors in this “mouse”
detect your cat’s movements so it
can react accordingly, flicking its
tail or scurrying to and fro. With
three levels of robotic excitability,
an app-controlled mode, and 10
swappable tail types, your clawed
menace will never get bored.
Petronics Mousr | $149

DEC 2018
0 4 4
This magazine was
made with the help
of G Suite.
Wired uses Docs
to edit stories from
contributors across
the globe.
Share
Share
e
Shar
Wired used Slides
to coordinate the
design and content
of this spread.
 See Captains

A collaboration between eyewear

designer Salt Optics and outdoor
clothing outfit Aether Apparel,
these shades are handmade in
Japan, with polarized, photochro-
mic lenses, light titanium frames,
and wind-blocking silicone side
shields. Like James Bond, they’re
meant to tackle adventure and
look sharp peering over a martini.
Salt + Aether Voyage | $650

Chat Window

A Facebook device with a camera

on it? We hear you—but the Por-
tal+ makes video chat so delight-
ful, you might be won over. The
AI-powered camera software
zooms in to tightly frame whoev-
er’s speaking. And it comes with a
privacy cap you can snap over the
lens when you don’t want Mark

AGV X3000 Ago Helmet

Crown Jewel

The new Legends helmet series

from AGV replicates the designs
of brain buckets worn by Moto
GP superstars like Giacomo
Agostini and Renzo Pasolini. The
Agostini helmet’s old-school style
is matched with modern safety
and comfort features, such as
sturdy double D-ring fasteners
and machine-washable padding.
AGV X3000 Ago Helmet | $700

DEC 2018
0 5 3
 Vinyl Fantasy

If you want to enjoy the past

while living in the present, give
this record player a spin. Its
audiophile-grade turntable sits
atop a case housing a 25-watt-
per-channel amp and a trio of
speakers. When it’s time to go dig-
ital, there’s a compartment for hid-
ing a Chromecast, Airport Express,
or Sonos Connect. Symbol Audio
Modern Record Player | $3,295

Holes in One

This funky camera is an unex-

Lomography pected treat for the photographer
Sprocket Rocket who has everything. It shoots
wide panoramic images across
multiple frames of 35-mm film,
exposing the whole surface, even
around the sprocket holes. Toss in
some rolls of Lomochrome Purple
film and they’ll be snapping artsy
panos for weeks. Lomography
Sprocket Rocket | $70

DEC 2018
0 5 4
 Shadow Play

Most table lamps fade into your

decor, but Blu Dot’s light leaves
an impression. Instead of a sim-
ple shade, it uses perforated,
powder-coated steel plates on
either side to diffuse and soften
the shine. It comes in four eye-
catching colors, and the cloth
cord includes a mood-altering
dimmer switch. Blu Dot Filter
Table Lamp | $299

Long View

Surprise the inquisitive photog-

rapher in your life with a clever
macro device for their DSLR.
Designed to capture close-ups
of tiny subjects, this 24-mm, f/14
lens has a waterproof barrel and
built-in LEDs that lets them snap
portraits of insects or ambush fish
in a stream, unlocking an exciting
new world for exploration. Laowa
Macroprobe | $1,500 and up

Master & Dynamic MW07

Achtung, Elon

Sure, Audi’s first fully electric car

comes to the US without the drag-
cutting, range-boosting cameras
that replace side-view mirrors in Audi E-tron

match, with high-performance

beryllium drivers and wings that
help fasten them to your ears to
form a bass-optimizing seal. Bat-
tery life is solid too—good for 3.5
hours of rocking between visits
to the charging case. Master &
Dynamic MW07 | $299

DEC 2018
0 5 6
GIVE THE GIFT OF PRINCIPLES THIS HOLIDAY

“EVERYONE WITH GOALS AND DREAMS

C A N L E A R N F R O M R AY ’ S A P P R O A C H ”
MIKE BLOOMBERG

“PRINCIPLES IS THE BOOK I WISH I HAD AS A YOUNG ENTREPRENEUR”

MARK CUBAN

“ PRINCIPLES BY RAY DALIO IS A MASTERPIECE—IT’S A MUST-READ!”

MARC BENIOFF

“P R I N C I P L E S I S H E L P F U L I N S H O W I N G P E O P L E H O W T O
O P E R AT E B Y C L E A R LY A R T I C U L AT E D A N D S H A R E D
P R I N C I P L E S T H AT L E A D T O FA S T E R P R O G R E S S . ”
JACK DORSEY

“ B E A U T I F U L LY W R I T T E N A N D F I L L E D W I T H S U C H W I S D O M ”
ARIANNA HUFFINGTON

“ R AY H A S A G R E AT B O O K O U T Y O U A L L N E E D T O
GET CALLED PRINCIPLES. IT’S A MAJOR KEY!!”
SEAN (P.DIDDY) COMBS

“PRINCIPLES HAD A PROFOUND POSITIVE

I M PA C T O N M Y L E A D E R S H I P S T Y L E ”
REED HASTINGS

“RAY DALIO HAS PROVIDED ME WITH INVALUABLE GUIDANCE”

BILL GATES

“IT HAS ALREADY CHANGED HOW I THINK ABOUT

MAKING DECISIONS IN MY LIFE AND IN MY BUSINESS.”
TIM FERRISS

“I F O U N D I T T O B E T R U LY E X T R A O R D I N A R Y. ”
TONY ROBBINS

“REMARKA B LY E N G A G I N G ”
ANDREW ROSS SORKIN

“RAY DALIO HAS CREATED THE STRONGEST CULTURE

I ’ V E E V E R S E E N I N A N O R G A N I Z AT I O N ”
ADAM GRANT

“HALF BUSINESS AND HALF LIFE ADVICE, RAY CAPTURES A SERIES

OF PRAGMATIC RULES FOR LIVING AND WORKING EFFECTIVELY.”
KEVIN SYSTROM

#1 NEW YORK TIMES BESTSELLER

#1 AMAZON BUSINESS BOOK OF THE YEAR
O VE R 1 . 5 M I L L I O N C O P I E S S O L D

WATCH PRINCIPLES FOR SUCCESS & RAY’S TED TALK ON PRINCIPLES.COM

 Fitbit Charge 3

A Step Up

The Charge is Fitbit’s best-selling

activity tracker, and for good
reason: It offers step counting,
workout tracking, and heart-rate
measuring, all without the weight,
expense, or complexity of most
smartwatches. The third gener-
ation has a more accurate heart-
rate-sensing module and lasts
up to seven days on a single, er,
charge. Fitbit Charge 3 | $150

Sole Mates

Rothy’s 3D-knitted footwear

uses a proprietary yarn manufac-
tured from recycled plastic
water bottles. Like the compa-
ny’s other shoes, the Sneaker is
a machine-washable, sockless
slip-on and comes in nine vibrant
colors (garnet!). It’s comfortable
enough to become the shoes
your feet request every morning.
Rothy’s The Sneaker | $125

Stage Right

Along with doodling album covers

and dreaming up groovy names,
tripping over your cables is part
of being in a band. Tidy up your
practice space with Boss’ wire-
less amplifier. Plug the included
transmitter into your ax and get
everything from a glassy shimmer
to a raw crunch from the 30-watt
amp. It even wakes up automati-
cally when you pick up your guitar.
Boss Katana-Air | $400

DEC 2018
0 5 8
 ENJOY RESPONSIBLY
©2018. DEWAR’S BLENDED SCOTCH WHISKY 40% ALC. BY VOL. IMPORTED BY JOHN DEWAR & SONS COMPANY, CORAL GABLES, FL.
 Sweet Relief

This beautiful tome by W I R E D

alumni Betsy Mason and Greg
Miller charts the fascinating his- Nimble 5-Day Fast
tory of cartography. The scores Portable Charger
of maps in the book range from
the whimsical (origins of meats
All Over the Map supplied to Parisian butchers) to
the political (North Dakota frack-
ing sites) to the fantastical (Jerry
Gretzinger’s imaginary world).
All Over the Map | $50

Reduce, Reuse, Recharge

Battery packs are handy, but the

environmental impact of electron-
ics is out of hand. Let your giftee
rest easy—this juice box is made
from eco-friendly bioplastics and
recyclable aluminum. Support for
USB-C and the new Power Deliv-
ery standard means they can top
of any device without ticking of
Mother Nature. Nimble 5-Day
Fast Portable Charger | $60

Clean Cut

The century-old Japanese firm

Yoshiharu Cutlery earned its
stripes handcrafting chisels. Now
it makes elegant knives like this
pocketable specimen. Its take
on a Kiridashi utility knife pairs a
carbon-steel blade with a brass
body. Keep it on your desk to slice
open packages, or simply display
it as a sharp conversation piece.
Penato Knife | $12

All Together Now

That fidget spinner is so last year.

This three-dimensional matrix puz-
zle is a far more attractive way to
keep your fingers busy, and its rel-
atively low difficulty is perfectly
suited for novice puzzlers. Once
you’ve cracked the code, the DOT
becomes a clever challenge for all
those chatterboxes who stop by
your desk to sabotage your pro-
ductivity. DOT Pocket Puzzle | $14

DEC 2018
0 6 0
*Re quires th e option al M B - D18 multi-p owe r b at te r y p a ck an d EN - EL18 a/b/c b at te r y. N ikon is a re gis te re d tra d e m ark of N iko n Corp o ration . ©2018 N ikon In c .

CAPTURE THE WIN

E V E RY T I M E

Delly Carr

With the exceptionally flexible full-frame Nikon D850 in his hands, award-winning sports
photographer Delly Carr creates striking shots like never before. The D850’s superb combination
of resolution and speed — 45.7MP FX full-frame CMOS sensor and 9 fps* continuous
shooting — makes it our best all-around DSLR ever. Also featuring full-frame 4K UHD video,
the D850 helps Delly capture the sports world’s most thrilling moments in dramatic detail.
See how the D850 captures sports action in every detail at nikonusa.com/d850.
 Bag in Black

This tool bag from Mythbusters

host Adam Savage is based on
Neil Armstrong’s McDivitt tote.
But while the astronaut’s lunar
pack was pure white, Savage’s
newest satchel is deep-space
black. Each bag is sewn from
repurposed sailcloth, so it’s dura-
ble enough to take your lunch,
headphones, and bike lock to
the moon and back. Savage
Industries EDC Two | $175

Blue Notes

UE’s Boom series has topped our

list of favorite Bluetooth speakers
for a few years running. The shiny
new Boom 3 inherits that portable
party crown. It’s the same size as a
tallboy of Budweiser, but it sounds
like a magnum of Veuve Clicquot
—which is to say, outstanding.
The waterproof design (it floats!)
comes in a variety of beach-ready
colors. UE Boom 3 | $150

No Sweat

Step out of yoga class feeling

fresh in this odor-minimizing
sweatshirt made from a moisture-
wicking blend of recycled polyes-
ter, hemp, and Tencel, a material
derived from wood pulp. The addi-
tion of fleece lining lends a hint of
luxury, and stitching details ele-
vate the silhouette from schlubby
Prana Cozy Up Sweatshirt
sweatshirt to Savasana chic.
Prana Cozy Up Sweatshirt | $79

DEC 2018
0 6 2
 Hot Seat

You won’t find many knobs or

levers on the underside of Herman
Miller’s new Cosm chair, but rather
a four-point suspension system
Dyson Cyclone V10 Absolute that constantly tilts and adapts
to keep you balanced. The flexi-
ble Y-shaped frame and one-piece
mesh seat offer responsive sup-
port as you fidget during those
endless conference calls. Herman
Miller Cosm | $895 and up

Whirl Jam

Our go-to stick vac gets a reboot

with Dyson’s latest sucker, the V10.
New to the mix is a torque-drive
head with a motor inside the rotat-
ing brush, designed to give car-
pets and hardwood floors a more
thorough scrubbing. A larger bin
and an increased run time of up to
an hour can handle even the gnar-
liest holiday-party messes. Dyson
Cyclone V10 Absolute | $700

DEC 2018
0 6 4
 PROMOTION

Hey
Handsome!
We know that
caring for your
hair, skin, and
beard isn’t
easy, but you
always look
and smell your
best. If you’re
a grooming
junkie, and want
to sample the
product first,
then sign up
for our Try It
Program. Just fill
out a grooming
profile to qualify.*

Sign up now > GQ.com/tryit

*Spaces and samples are limited. Products are targeted to skin and hair types and preferences where applicable.
 Fill-Up Glass

Your rarest whiskey deserves a

most worthy vessel. Molten crystal
is formed into this curvy, faceted,
1¼-pound chalice, then a patented
process cools the varying thick-
nesses at different rates to keep
the glass from cracking. Prefer an
old-fashioned? The ridged bot-
tom grips your ingredients for
superior muddling. Norlan Rauk
Heavy Tumbler | $50

Mission Control

Programming your smart home

from your smartphone is easy, but
what does your family do when
you walk off with the all-in-one
door lock, music player, and TV
remote in your pocket? Brilliant’s
two-switch touchscreen panel lets
everyone turn on the lights, crank
up the Sonos, switch off the ther-
mostat, and bark at Alexa. Bril-
liant Two Switch Panel | $349

Loving Cup

This multitasking travel mug has

a French press inside that brews
coffee or steeps loose-leaf tea.
Prepare your concoction right in
the cup, or pour in your elixir of
choice. Drinks stay hot (or cold)
for hours, thanks to double-wall
vacuum insulation. And at just
1 pound, the Espro won’t weigh
down your adventures. Espro
Ultralight Press | $40

Carry All

If you’re not yet sold on the tablet

PC future, try the Go, Microsoft’s
Microsoft Surface Go With Type Cover
10-inch touchscreen computer
with a kickstand. Pair it with the
Type Cover ($100 and up) and
the Surface Pen ($100) for easier
Espro Ultralight Press swiping, tapping, and doodling,
and you’ll see how well the tiny
Go fits—and works—in places
other Windows machines can’t.
Microsoft Surface Go | $399

DEC 2018
0 6 6
 PROMOTION

AN INV ITE - O N LY OAS I S S T E PS AWAY F R O M T EC H WES T.

JA N UA R Y 8 & 9, 2 0 1 9
LAS VEGAS

E M A I L W I R E D I N S I D E R @ W I R E D.C O M
F O R M O R E I N F O R M AT I O N

ADDITIONAL SUPPORT FROM

 Device Drivers

Why wait for tech support? This

iFixit Pro Tech Toolkit handy kit from the nerds at iFixit
contains all the tools you need
to replace or repair parts deep
inside your devices. With an
87-piece suite of bits, tweezers,
drivers, and spudgers, you can
MacGyver pretty much anything
from a current-gen iPhone to
your vintage NES. iFixit Pro Tech
Toolkit | $60

Mint Condition

Crack open this bottle of schnapps

for an aperitif, or simply freshen
up your bar cart with its bright
aqua-green hue. The Wisconsin
craft distillery Death’s Door Spir-
its uses a special blend of pepper-
mint extract, almond, rosewater,
and absinthe to create the minty
flavor. Enjoy it neat, mixed with
hot cocoa, or as the highlight of a
Mint Mule. Wondermint | $24

Tweet Storm

The flashy metal cones on KEF’s

newest speakers can’t camou-
flage their smartest design detail:
The concentric driver nestles the
tweeter at the center of the mid-
range speaker, creating a single
acoustic point of origin for the high
and mid frequencies that keeps
your Floyd from getting flabby.
Hook them up to your stereo and
they’ll intensify every track on your
playlist. KEF R3 | $2,000 a pair

DEC 2018
0 6 8
 M I R R O R L E S S
R E I N V E N T E D
* For compatibility details with the new optional adapter, please visit nikonusa.com/mirrorless
Nikon is a registered trademark of Nikon Corporation. ©2018 Nikon Inc.

Finally, a full-frame mirrorless camera with the optical excellence you’ve come
to expect from Nikon. With precision optics, reliability and intuitive design,
the new Z 7 is the first mirrorless that feels and acts like a Nikon camera.
Its revolutionary new mount lets in more light than ever before, opening
the door to groundbreaking new lens designs and boundless imaging
possibilities. Take spectacular stills and 4K UHD video with the new NIKKOR
Z lenses, or use approximately 360 NIKKOR® lenses with the new adapter.*
Behold a future full of creative possibility with mirrorless, reinvented by Nikon.

4 5 . 7 M P | 6 4 –2 5 , 6 0 0 I S O | 49 3 P D A F S YS T E M | U P TO 9 F P S |
4K UHD FULL-FRAME VIDEO | QUAD VGA ELECTRONIC VIEWFINDER

N I KO N U S A .CO M /M I R RO R L E S S
 Oloid in Brass

Sure Shot

Phones make convenient shoot-

ers, but even the best of them
can’t beat a real camera for qual-
ity. This point-and-shoot packs
DSLR-grade imaging into a trim
body. With its 28-mm-equivalent
lens and gorgeous film simulation
Fujifilm XF10 modes, the pocketable XF10 is
well suited for candid street pho-
tography, landscapes, and por-
traiture. Fujifilm XF10 | $500

Heady Metal

This precision-machined, 2-inch-

tall brass sculpture by artist Pat
Kim rolls across a flat surface
with a gentle push. Based on the
dual-circle oloid shape discovered
by inventor and mathematician
Paul Schatz in 1929, this update
is the perfect desk-topping gift
for the contemplative (or pro-
crastinating) creative in your life.
Oloid in Brass | $120

Samsung 65-Inch Q8FN 4K TV

Light Box

This set is primed for movie night:

It can accurately display the pano-
ply of colors supported by the film
industry’s DCI-P3 projection stan-
dard. And when you’re streaming
4K encoded for HDR playback, its
backlight control handles extreme
lights and darks—Black Panther’s
nighttime car chase will thrill,
even in bright rooms. Samsung
65-Inch Q8FN 4K TV | $2,300

Sound Design

It’s easy to buy an Eames chair;

the design duo made scads of
them, and they’re still in produc-
tion. But their 1946 table radio
was never manufactured—until
now. The revived radio gets a
21st-century update with the
addition of Bluetooth connectiv-
ity. Fewer than 1,000 have been
made; it’s a true Eames original.
Eames Radio by Vitra | $999

DEC 2018
0 7 0
 145
92

WILLIAM, heart attack survivor.

THIS IS WHAT HIGH BLOOD PRESSURE LOOKS LIKE.

You might not see or feel its symptoms, but the results – a heart attack or stroke – are far from invisible or silent.
If you’ve come off your treatment plan, get back on it, or talk with your doctor to create a new exercise,
diet and medication plan that works better for you.

Go to LowerYourHBP.org before it’s too late.

 FEATURES | 2 6 .1 2

JAMES NOLAN GANDY 0 7 3

 THE
MISEDUCATION OF
ARTIFICIAL
INTELLIGENCE

By CLIVE THOMPSON

Illustration by RON LICATA

PHOTOGRAPHS BY BETH HOLZER

0 7 5
We’ve spent years F I V E Y E A R S A G O , the coders at DeepMind,
a London-based artificial intelligence company,
feeding neural nets watched excitedly as an AI taught itself to play a

vast amounts classic arcade game. They’d used the hot tech-
nique of the day, deep learning, on a seemingly

of data, teaching whimsical task: mastering Breakout,1 the Atari

game in which you bounce a ball at a wall of

them to think like bricks, trying to make each one vanish.

Deep learning is self-education for machines;
human brains. you feed an AI huge amounts of data, and even-

They’re crazy-smart, tually it begins to discern patterns all by itself.

In this case, the data was the activity on the

but they have screen—blocky pixels representing the bricks,

the ball, and the player’s paddle. The DeepMind

absolutely AI, a so-called neural network made up of layered

algorithms, wasn’t programmed with any knowl-
NO COMMON SENSE. edge about how Breakout works, its rules, its

What if we’ve been goals, or even how to play it. The coders just let
the neural net examine the results of each action,

doing it all wrong? each bounce of the ball. Where would it lead?
To some very impressive skills, it turns out.
During the first few games, the AI flailed around.
But after playing a few hundred times, it had
begun accurately bouncing the ball. By the 600th
game, the neural net was using a more expert
move employed by human Breakout players,
chipping through an entire column of bricks
and setting the ball bouncing merrily along the
top of the wall.
“That was a big surprise for us,” Demis Hass-
abis, CEO of DeepMind, said at the time. “The
strategy completely emerged from the under-
lying system.” The AI had shown itself capable
of what seemed to be an unusually subtle piece
of humanlike thinking, a grasping of the inher-
ent concepts behind Breakout. Because neural
nets loosely mirror the structure of the human
brain, the theory was that they should mimic,
in some respects, our own style of cognition.
This moment seemed to serve as proof that the
theory was right.
Then, last year, computer scientists at Vicar-
ious, an AI firm in San Francisco, offered an
interesting reality check. They took an AI like
the one used by DeepMind and trained it on

C L I V E T H O M P S O N (@pomeranian99) is a
columnist for WIRED.
Breakout. It played great. But then they slightly tweaked
the layout of the game. They lifted the paddle up higher in
one iteration; in another, they added an unbreakable area in
the center of the blocks.
A human player would be able to quickly adapt to these
changes; the neural net couldn’t. The seemingly supersmart
AI could play only the exact style of Breakout it had spent hun-

G
dreds of games mastering. It couldn’t handle something new.
“We humans are not just pattern recognizers,” Dileep
George, a computer scientist who cofounded Vicarious, A R Y M A R C U S I S a pensive, bespecta-
tells me. “We’re also building models about the things we cled 48-year-old professor of psychology and
see. And these are causal models—we understand about neuroscience at New York University, and he’s
cause and efect.” Humans engage in reasoning, making logi- probably the most famous apostate of orthodox
cal inferences about the world around us; we have a store deep learning.
of common-sense knowledge that helps us figure out new Marcus first got interested in artificial intel-
situations. When we see a game of Breakout that’s a little ligence in the 1980s and ’90s, when neural nets
diferent from the one we just played, we realize it’s likely were still in their experimental phase, and he’s
to have mostly the same rules and goals. The neural net, on been making the same argument ever since. “It’s
the other hand, hadn’t understood anything about Break- not like I came to this party late and want to
out. All it could do was follow the pattern. When the pattern pee on it,” Marcus told me when I met him at
changed, it was helpless. his apartment near NYU. (We are also personal
Deep learning is the reigning monarch of AI. In the six years friends.) “As soon as deep learning erupted, I said
since it exploded into the mainstream, it has become the ‘This is the wrong direction, guys!’ ” 1. Steve Jobs was
working at Atari when
dominant way to help machines sense and perceive the world Back then, the strategy behind deep learn- he was commissioned
around them. It powers Alexa’s speech recognition, Waymo’s ing was the same as it is today. Say you wanted to create 1976’s
Breakout, a job no
self-driving cars, and Google’s on-the-fly translations. Uber is a machine to teach itself to recognize daisies. other engineer wanted.
He roped his friend
in some respects a giant optimization problem, using machine First you’d code some algorithmic “neurons,” Steve Wozniak, then
learning to figure out where riders will need cars. Baidu, the connecting them in layers like a sandwich (when at Hewlett-Packard,
into helping him.
Chinese tech giant, has more than 2,000 engineers cranking you use several layers, the sandwich gets thicker
away on neural net AI. For years, it seemed as though deep or deep—hence “deep” learning). You’d show an 2. A member of the
Asteraceae family, dai-
learning would only keep getting better, leading inexorably image of a daisy to the first layer, and its neu- sies are distinguished
by a yellow disc sur-
to a machine with the fluid, supple intelligence of a person. rons would fire or not fire based on whether rounded by ray flowers.
But some heretics argue that deep learning is hitting a the image resembled the examples of daisies This plant family also
includes dahlias, mari-
wall. They say that, on its own, it’ll never produce gener- it had seen before. The signal would move on golds, dandelions,
alized intelligence, because truly humanlike intelligence to the next layer, where the process would be lettuce, and artichokes.

isn’t just pattern recognition. We need to start figuring out repeated. Eventually, the layers would winnow 3. In 1975 the psycholo-
gist Jean Piaget and
how to imbue AI with everyday common sense, the stuf of down to one final verdict. the linguist Noam
human smarts. If we don’t, they warn, we’ll keep bumping up At first, the neural net is just guessing blindly; Chomsky met in France
for what would prove
against the limits of deep learning, like visual-recognition it starts life a blank slate, more or less. The key is to be a historic debate.
systems that can be easily fooled by changing a few inputs, to establish a useful feedback loop. Every time Grossly simplified,
Piaget argued that
making a deep-learning model think a turtle is a gun. But the AI misses a daisy,2 that set of neural connec- human brains are blank-
slate self-learning
if we succeed, they say, we’ll witness an explosion of safer, tions weakens the links that led to an incorrect machines, and Chomsky
more useful devices—health care robots that navigate a clut- guess; if it’s successful, it strengthens them. that they are endowed
with some prepro-
tered home, fraud detection systems that don’t trip on false Given enough time and enough daisies, the neu- grammed smarts.
positives, medical breakthroughs powered by machines that ral net gets more accurate. It learns to intuit
ponder cause and efect in disease. some pattern of daisy-ness that lets it detect
But what does true reasoning look like in a machine? And the daisy (and not the sunflower or aster) each
if deep learning can’t get us there, what can? time. As the years went on, this core idea—start
with a naive network and train by repetition—
was improved upon and seemed useful nearly
anywhere it was applied.
But Marcus was never convinced. For him,
the problem is the blank slate: It assumes that
humans build their intelligence purely by
observing the world around them, and that
machines can too. But Marcus doesn’t think
that’s how humans work. He walks the intel-

0 7 7
 lectual path laid down by Noam Chomsky,3 who
argued that humans are born wired to learn,
programmed to master language and interpret
the physical world.
For all their supposed braininess, he notes,
neural nets don’t appear to work the way
human brains do. For starters, they’re much

O
too data-hungry. In most cases, each neural net
requires thousands or millions of examples to
learn from. Worse, each time you want a neural REN ETZIONI is a smiling bear of a guy. A computer
net to recognize a new type of item, you have to scientist who runs the Allen Institute for Artificial Intelli-
start from scratch. A neural net trained to rec- gence in Seattle, he greets me in his bright oice wearing
ognize only canaries isn’t of any use in recog- jeans and a salmon-colored shirt, ushering me in past a
nizing, say, birdsong or human speech. whiteboard scrawled with musings about machine intelli-
“We don’t need massive amounts of data to gence. (“DEFINE SUCCESS,” “WHAT’S THE TASK?”) Out-
learn,” Marcus says. His kids didn’t need to see side, in the sun-drenched main room of the institute, young
a million cars before they could recognize one. AI researchers pad around sylphlike, headphones attached,
Better yet, they can generalize; when they see quietly pecking at keyboards.
a tractor for the first time, they understand Etzioni and his team are working on the common-sense
that it’s sort of like a car. They can also engage problem. He defines it in the context of two legendary AI
in counterfactuals. Google Translate can map moments—the trouncing of the chess grandmaster Garry
the French equivalent of the English sentence Kasparov 5 by IBM’s Deep Blue in 1997 and the equally
“The glass was pushed, so it fell of the table.” shocking defeat of the world’s top Go player by DeepMind’s
But it doesn’t know what the words mean, so AlphaGo last year. (Google bought DeepMind in 2014.) “With
it couldn’t tell you what would happen if the Deep Blue we had a program that would make a superhu-
glass weren’t pushed. Humans, Marcus notes, man chess move—while the room was on fire,” Etzioni jokes.
grasp not just the patterns of grammar but the “Right? Completely lacking context. Fast-forward 20 years,
logic behind it. You could give a young child a we’ve got a computer that can make a superhuman Go move—
fake verb like pilk, and she’d likely be able to while the room is on fire.” Humans, of course, do not have
reason that the past tense would be pilked. She this limitation. His team plays weekly games of bughouse
hasn’t seen that word before, of course. She chess,6 and if a fire broke out the humans would pull the
hasn’t been “trained” on it. She has just intu- alarm and run for the doors.
ited some logic about how language works and Humans, in other words, possess a base of knowledge about
can apply it to a new situation. the world (fire burns things) mixed with the ability to reason
“These deep-learning systems don’t know about it (you should try to move away from an out-of-control
how to integrate abstract knowledge,” says Mar- fire). For AI to truly think like people, we need to teach it the
cus, who founded a company that created AI to stuf that everyone knows, like physics (balls tossed in the air
4. Pronounced learn with less data (and sold the company to will fall) or the relative sizes of things (an elephant can’t fit
“archive,” this online Uber in 2016). in a bathtub). Until AI possesses these basic concepts, Etzi-
database of prepub-
lication papers in Earlier this year, Marcus published a white oni figures, it won’t be able to reason.
physics, math, and paper on arXiv,4 arguing that, without some With an infusion of hundreds of millions of dollars from
computer science
has been hosted by new approaches, deep learning might never Paul Allen,7 Etzioni and his team are trying to develop a
Cornell University
since 2001. get past its current limitations. What it needs layer of common-sense reasoning to work with the exist-
is a boost—rules that supplement or are built ing style of neural net. (The Allen Institute is a nonprofit,
5. In 1996, Kasparov—
then the best chess in to help it reason about the world. so everything they discover will be published, for anyone
player in the world— to use.) The first problem they face is answering the ques-
beat Deep Blue. During
a rematch a year later, tion, What is common sense?
Kasparov surrendered
after 19 moves. He Etzioni describes it as all the knowledge about the world
later told a reporter: that we take for granted but rarely state out loud. He and his
“I’m a human being.
When I see something colleagues have created a set of benchmark questions that
that is well beyond a truly reasoning AI ought to be able to answer: If I put my
my understanding,
I’m afraid.” socks in a drawer, will they be there tomorrow? If I stomp on
6. Also known as dou- someone’s toe, will they be mad?
bles chess, bughouse One way to get this knowledge is to extract it from people.
chess is played on
two boards with four Etzioni’s lab is paying crowdsourced humans on Amazon
players, usually in Mechanical Turk to help craft common-sense statements. The
extremely fast rounds
of three minutes or less. team then uses various machine-learning techniques—some

0 7 8
7. Microsoft cofounder old-school statistical analyses, some deep-learn- sentence “Oren threw the ball” and infer that the ball must
and philanthropist Paul
Allen donated billions ing neural nets—to draw lessons from those be smaller than Oren.
to science, climate, and statements. If they do it right, Etzioni believes Another challenge is visual reasoning. Aniruddha Kembhavi,
health research, as well
as to Seattle causes. they can produce reusable Lego bricks of com- another of Etzioni’s AI scientists, shows me a virtual robot
He died of complica-
tions from cancer on puter reasoning: One set that understands writ- wandering around an onscreen house. Other Allen Institute
October 15 at age 65. ten words, one that grasps physics, and so on. scientists built the Sims-like house, filling it with everyday
8. Don’t put tomatoes Yejin Choi, one of Etzioni’s leading common- items and realistic physics—kitchen cupboards full of dishes,
in the refrigerator. sense scientists, has led several of these crowd- couches that can be pushed around. Then they designed the
A 2016 study found
that tomatoes kept sourced eforts. In one project, she wanted to robot, which looks like a dark gray garbage canister with arms,
below 54 degrees lose
the ability to produce develop an AI that would understand the intent and told it to hunt down certain items. After thousands of
their natural flavor. or emotion implied by a person’s actions or state- tasks, the neural net gains a basic grounding in real-life facts.
ments. She started by examining thousands of “What this agent has learned is, when you ask it ‘Do I have
online stories, blogs, and idiom entries in Wik- tomatoes?’ it doesn’t go and open all the cabinets. It prefers to
tionary and extracting “phrasal events,” such as open the fridge,” 8 Kembhavi says. “Or if you say ‘Find me my
the sentence “Jef punches Roger’s lights out.” keys,’ it doesn’t try to pick up the television. It just looks behind
Then she’d anonymize each phrase—“Person the television. It has learned that TVs aren’t usually picked up.”
X punches Person Y’s lights out”—and ask the Etzioni and his colleagues hope that these various compo-
Turkers to describe the intent of Person X: Why nents—Choi’s language reasoning, the visual thinking, other
did they do that? When she had gathered 25,000 work they’re doing on getting an AI to grasp textbook science
of these marked-up sentences, she used them information—can all eventually be combined. But how long
to train a machine-learning system to analyze will it take, and what will the final products look like? They
sentences it had never seen before and infer the don’t know. The common-sense systems they’re building
emotion or intent of the subject. still make mistakes, sometimes more than half the time. Choi
At best, the new system worked only half estimates she’ll need around a million crowdsourced human
the time. But when it did, it evinced some very statements as she trains her various language-parsing AIs.
humanlike perception: Given a sentence like Building common sense, it would seem, is uncommonly hard.
“Oren cooked Thanksgiving dinner,” it predicted

T
that Oren was trying to impress his family. “We
can also reason about others’ reactions, even if
they’re not mentioned,” Choi notes. “So X’s fam- HERE ARE OTHER pathways to making machines that
ily probably feel impressed and loved.” Another reason, and they’re even more labor-intensive. For exam-
system her team built used Turkers to mark ple, you could simply sit down and write out, by hand, all
up the psychological states of people in sto- the rules that tell a machine how the world works. This is
ries; the resulting system could also draw some how Doug Lenat’s Cyc project works. For 34 years, Lenat has
sharp inferences when given a new situation. It employed a team of engineers and philosophers to code 25
was told, for instance, about a music instructor million rules of general common sense, like “water is wet”
getting angry at his band’s lousy performance or “most people know the first names of their friends.” This
and that “the instructor was furious and threw lets Cyc deduce things: “Your shirt is wet, so you were prob-
his chair.” The AI predicted that the musicians ably in the rain.” The advantage is that Lenat has exquisite
would “feel fear afterwards,” even though the control over what goes into Cyc’s database; that isn’t true of
story doesn’t explicitly say so. crowdsourced knowledge.
Choi, Etzioni, and their colleagues aren’t aban- Brute-force, handcrafted AI has become unfashionable
doning deep learning. Indeed, they regard it as in the world of deep learning. That’s partly because it can
a very useful tool. But they don’t think there is a be “brittle”: Without the right rules about the world, the AI
shortcut to the laborious task of coaxing people can get flummoxed. This is why scripted chatbots are so frus-
to explicitly state the weird, invisible, implied trating; if they haven’t been explicitly told how to answer a
knowledge we all possess. Deep learning is gar- question, they have no way to reason it out. Cyc is enormously
bage in, garbage out. Merely feeding a neural more capable than a chatbot and has been licensed for use in
net tons of news articles isn’t enough, because health care systems, financial services, and military projects.
it wouldn’t pick up on the unstated knowledge, But the work is achingly slow, and it’s expensive. Lenat says
the obvious stuf that writers didn’t bother to it has cost around $200 million to develop Cyc.
mention. As Choi puts it, “People don’t say ‘My But a bit of hand coding could be how you replicate some
house is bigger than me.’ ” To help tackle this of the built-in knowledge that, according to the Chomsky-
problem, she had the Turkers analyze the physi- ite view, human brains possess. That’s what Dileep George
cal relationships implied by 1,100 common verbs, and the Vicarious researchers did with Breakout. To create
such as “X threw Y.” That, in turn, allowed for an AI that wouldn’t get stumped by changes to the layout of
a simple statistical model that could take the the game, they abandoned deep learning and built a system
 9. Russian blues have
extremely soft, thick
coats that make them
appear larger than
they actually are.

10. Captcha stands

for “Completely Auto-
mated Public Turing
test to tell Computers
and Humans Apart.”
It originated at Carne-
gie Mellon University
in 2000; Yahoo was
the first big company
that included hard-coded basic assumptions. Without too to make its use com- would be able to do this only for numbers it had
much trouble, George tells me, their AI learned “that there are monplace. seen before; train it up to 100 and it would know
objects, and there are interactions between objects, and that 11. LeCun created an that 99 is “fizz” and 100 is “buzz.” But it wouldn’t
the motion of one object can be causally explained between early AI chip at Bell know what to do with 105. In contrast, the hybrid
Labs, the storied
the object and something else.” research arm of AT&T, DeepMind system seemed to understand the rule
in 1992. Bell Labs
As it played Breakout, the system developed the ability to developed many essen- and went past 100 with no problem. Edward Gre-
weigh diferent courses of action and their likely outcomes. tial innovations: the fenstette, one of the DeepMind coders who built
transistor, the silicon
This worked in reverse too. If the AI wanted to break a block in solar cell, communica- the hybrid, says, “You can train systems that will
the far left corner of the screen, it reasoned to put the paddle in tions satellites, cell generalize in a way that deep-learning networks
phone networks,
the far right corner. Crucially, this meant that when Vicarious fiber-optic cable sys- simply couldn’t on their own.”
tems, and the Unix
changed the layout of the game—adding new bricks or rais- operating system.

Y
ing the paddle—the system compensated. It appeared to have
12. HAL was originally
extracted some general understanding about Breakout itself. supposed to be voiced
Granted, there are trade-ofs in this type of AI engineer- by Martin Balsam, A N N L E C U N , 11 a deep-learning pioneer
an actor with a thick
ing. It’s arguably more painstaking to craft and takes careful Bronx accent. After and the current head of Facebook’s AI research
recording, however,
planning to figure out precisely what foreordained logic to director Stanley wing, agrees with many of the new critiques of
feed into the system. It’s also hard to strike the right balance Kubrick decided Bal- the field. He acknowledges that it requires too
sam sounded “too
of speed and accuracy when designing a new system. George colloquially American.” much training data, that it can’t reason, that it
says he looks for the minimum set of data “to put into the He was replaced doesn’t have common sense. “I’ve been basi-
by Canadian actor
model so it can learn quickly.” The fewer assumptions you Douglas Rain. cally saying this over and over again for the
need, the more eiciently the machine will make decisions. past four years,” he reminds me. But he remains
Once you’ve trained a deep-learning model to recognize cats, steadfast that deep learning, properly crafted,
you can show it a Russian blue 9 it has never seen and it ren- can provide the answer. He disagrees with the
ders the verdict—it’s a cat!—almost instantaneously. Having Chomskyite vision of human intelligence. He
processed millions of photos, it knows not only what makes a thinks human brains develop the ability to reason
cat a cat but also the fastest way to identify one. In contrast, solely through interaction, not built-in rules. “If
Vicarious’ style of AI is slower, because it’s actively making you think about how animals and babies learn,
logical inferences as it goes. there’s a lot of things that are learned in the
When the Vicarious AI works well, it can learn from much first few minutes, hours, days of life that seem
less data. George’s team created an AI to bust captchas,10 those to be done so fast that it looks like they are hard-
“I’m not a robot” obstacles online, by recognizing characters wired,” he notes. “But in fact they don’t need to
in spite of their distorted, warped appearance. Much as with be hardwired, because they can be learned so
the Breakout system, they endowed their AI with some abili- quickly.” In this view, to figure out the physics
ties up front, such as knowledge that helps it discern the likely of the world, a baby just moves its head around,
edges of characters. With that bootstrapping in place, they data-crunches the incoming imagery, and con-
only needed to train the AI on 260 images before it learned cludes that, hey, depth of field is a thing.
to break captchas with 90.4 percent accuracy. In contrast, Still, LeCun admits it’s not yet clear which
a neural net needed to be trained on more than 2.3 million routes will help deep learning get past its humps.
images before it could break a captcha. It might be “adversarial” neural nets, a relatively
Others are building common-sense-like structure into neu- new technique in which one neural net tries to
ral nets in diferent ways. Two researchers at DeepMind, for fool another neural net with fake data—forc-
instance, recently created a hybrid system—part deep learn- ing the second one to develop extremely subtle
ing, part more traditional techniques—known as inductive internal representations of pictures, sounds,
logic programming. The goal was to produce something that and other inputs. The advantage here is that you
could do mathematical reasoning. don’t have the “data hungriness” problem. You
They trained it on the children’s game fizz-buzz, in which don’t need to collect millions of data points on
you count upward from 1, saying “fizz” if a number is divisi- which to train the neural nets, because they’re
ble by 3 and “buzz” if it is divisible by 5. A regular neural net learning by studying each other. (Apocalyptic
 side note: A similar method is being used to create those make AI safer; the idea that humanity shouldn’t
profoundly troubling “deepfake” videos in which someone be wholesale slaughtered is, of course, arguably
appears to be saying or doing something they are not.) a piece of common-sense knowledge itself. (Part
I met LeCun at the oices of Facebook’s AI lab in New York. of the Allen Institute’s mandate is to make AI
Mark Zuckerberg recruited him in 2013, with the promise that safer by making it more reasonable.)
the lab’s goal would be to push the limits of ambitious AI, not Etzioni notes that the dystopic sci-fi visions
just produce minor tweaks for Facebook’s products. Like an of AI are less risky than near-term economic
academic lab, LeCun and his researchers publish their work displacement.14 The better AI gets at common
for others to access. sense, the more rapidly it’ll take over jobs that
LeCun, who retains the rich accent of his native France and currently are too hard for mere pattern-matching
has a Bride of Frankenstein shock of white in his thick mass of deep learning: drivers, cashiers, managers, ana-
dark hair, stood at a whiteboard energetically sketching out lysts of all stripes, and even (alas) journalists.
theories of possible deep-learning advances. On the facing But truly reasoning AI could wreak havoc even
wall was a set of gorgeous paintings from Stanley Kubrick’s beyond the economy. Imagine how good polit-
2001: A Space Odyssey—the main spaceship floating in deep ical disinformation bots would be if they could
space, the wheel-like ship orbiting Earth. “Oh, yes,” LeCun use common-sense knowledge to appear indis-
said, when I pointed them out; they were reprints of artwork tinguishably human on Twitter or Facebook or
Kubrick commissioned for the movie. in mass phone calls.
It was weirdly unsettling to discuss humanlike AI with those Marcus agrees that reasoning AI will have dan-
images around, because of course HAL 9000, the humanlike AI gers. But the upsides, he says, would be huge. AI
in 2001, 12 turns out to be a highly eicient murderer. And this that could reason and perceive like humans yet
pointed to a deeper philosophical question that floats over move at the speed of computers could revolu-
the whole debate: Is smarter AI even a good idea? Vicarious’ tionize science, teasing out causal connections
system cracked captcha, but the whole point of captcha is to at a pace impossible for us alone. It could follow
prevent bots from impersonating humans. Some AI think- if-then chains and ponder counterfactuals, run-
ers worry that the ability to talk to humans and understand ning mental experiments the way humans do,
their psychology could make a rogue AI incredibly dangerous. except with massive robotic knowledge. “We
Nick Bostrom 13 at the University of Oxford has sounded the might finally be able to cure mental illness, for
alarm about the dangers of creating a “superintelligence,” example,” Marcus adds. “AI might be able to
an AI that self-improves and rapidly outstrips humanity, understand these complex biological cascades
able to outthink and outflank us in every way. (One way he of proteins that are involved in building brains
suggests it might amass control is by manipulating people— and having them work correctly or not.”
something for which possessing a “theory of mind” would be Sitting beneath the images from 2001, LeCun
quite useful.) Elon Musk is suiciently convinced of this dan- makes a bit of a heretical point himself. Sure,
ger that he has funded OpenAI, an organization dedicated to making artificial intelligence more humanlike
the notion of safe AI. helps AI to navigate our world. But directly rep-
This future doesn’t keep Etzioni up at night. He’s not wor- licating human styles of thought? It’s not clear
ried about AI becoming maliciously superintelligent. “We’re that’d be useful. We already have humans who
13. In 2003, Bostrom worried about something taking over the world,” he scofs, can think like humans; maybe the value of smart
published the now-
famous paper-clip “that can’t even on its own decide to play chess again.” It’s machines is that they are quite alien from us.
warning about not clear how an AI would develop a desire to do so or what “They will tend to be more useful if they have
superintelligence: “A
well-meaning team that desire would look like in software. Deep learning can capabilities we don’t have,” he tells me. “Then
of programmers [could]
make a big mistake conquer chess, but it has no inborn will to play. they’ll become an amplifier for intelligence. So
in designing its goal What does concern him is that current AI is woefully inept. to some extent you want them to have a nonhu-
system. This could
result … in a super- So while we might not be creating HAL with a self-preserving man form of intelligence ... You want them to be
intelligence whose top intelligence, an “inept AI attached to deadly weapons can eas- more rational than humans.” In other words,
goal is the manufac-
turing of paper clips, ily kill,” he says. This is partly why Etzioni is so determined maybe it’s worth keeping artificial intelligence
with the consequence
that it starts transform- to give AI some common sense. Ultimately, he argues, it will a little bit artificial.
ing first all of Earth
and then increasing
portions of space
into paper-clip manu-
facturing facilities.”

14. A 2016 White House

report estimated
that up to 47 percent
of jobs would be threat-
ened by AI in the next
few decades, with
low-skilled jobs the
most endangered.

0 8 1
 Artificial intelligence tech has trickled
down to the public, and a new wave
of DIY TINKERERS is harnessing its power.

0 8 2
BY TOM SIMONITE
 In late winter of 1975, a scrap
of paper started appearing
on bulletin boards around
the San Francisco Peninsula.
“Are you building your own
computer?” it asked. “Or
some other digital black-
magic box? If so, you might
like to come to a gathering.”

The invite drew 32 people to a Menlo Park, Cal- kits for free, along with some of their research. Hackers and
ifornia, garage for the first meeting of the Home- hobbyists are now playing with nearly the same technology
brew Computer Club, a community of hobbyists that’s driving Silicon Valley’s wildest dreams. “High school
intrigued by the potential of a newly afordable students can now do things that the best researchers in the
component called the microprocessor. One was world could not have done a few years ago,” says Andrew Ng,
a young engineer named Steve Wozniak, who an AI researcher and entrepreneur who has led big projects
later brought a friend named Steve Jobs into at Google and China’s Baidu.
the club. “It was a demonstration that individu- People like Ng have big hopes for the amateur AI explosion:
als could make technological progress and that They want it to spread the technology’s potential far from Sili-
it doesn’t all have to happen at big companies con Valley, physically and culturally, to see what happens when
and universities,” says Len Shustek, a retired tech outsiders “train” neural networks according to their own
entrepreneur who was also in the garage that priorities and ways of seeing the world. Ng likes to imagine
first night. “Now the same thing is happening that one day a person in India might use what they learn in
for artificial intelligence.” online videos about AI to make their local water safer to drink.
Since 2012, computers have become dramati- Of course, not every DIY neural network will be quite so
cally better at understanding speech and images, G-rated. Late last year, a Reddit account posted a pornographic
thanks to a once obscure technology called arti- video that seemed to star Wonder Woman’s Gal Gadot. The
1. Darpa launched a
ficial neural networks. True mastery of this AI clip circulated around Reddit’s seamier corners and beyond program in 2016 to
technique requires powerful computers, years to adult-video sites. But attentive viewers noticed that Gad- invent ways to smoke
out video forgeries.
of research experience, and a yen for deep math. ot’s face occasionally flickered or slipped on her head like a One technique scans
for an unnatural lack
If you have all those things, congratulations: loose mask. The poster explained that the clip was fake, cre- of blinking. The con-
Chances are you’re already a well-remunerated ated by training a neural network to generate images of Gad- cern goes way beyond
porn: Deepfake vid-
employee of Amazon, Facebook, Google, or the ot’s face that matched the expressions of the video’s original eos could exacerbate
other select few giants vying to shape the world star. They then released the code and methodology online so the kind of disinfor-
mation that’s already
with their massively complicated AI strategies. anyone could make similar “deepfake” 1 clips of their own. corroding politics.
Yet the battle for AI supremacy also has lit- So the age of homebrew AI may not be all sweetness and
tered the ground with tools and spare parts light. Nor will it be all darkness and porn. Mostly, its expres-
that anyone can pick up. To draw in top-flight sions will be marvelous in their specificity. Meet some of the
scientists and app developers, tech giants have pioneers showing what happens when the masses can teach
released some of their in-house AI-building tool- computers new tricks.

Senior writer T O M S I M O N I T E (@tsimonite) covers intelligent machines for WIRED.

0 8 4
 Self-taught coder Robbie Barrat
tapped into the power of AI for art’s
sake—generating hip hop rhymes
and designing off-kilter fashion.

I’mma Let This Neural Network Finish My Lyrics

WHEN ROBBIE BARRAT was in middle school in That foul-mouthed AI system proved to be Barrat’s ticket
rural West Virginia, he started scavenging old of the farm. His grades weren’t good enough to get into the
computers from a local recycling center, ripping schools where he’d hoped to study math or computer sci-
them apart, and putting them back together ence. But the project helped him land an internship with a
again. Then he taught himself to code on his fam- self-driving-car project in the heart of Silicon Valley. From
ily’s farm. He took up AI in high school after he there he moved to Stanford University, where he now works
got into an argument with friends over whether in a biomedical lab, trying to develop neural networks that
computers could be creative. Barrat’s retort can identify molecules with medicinal potential. But training
was to teach a neural net to rap by training it neural networks to make art is still his passion.
on the lyrics of Kanye West. (Sample couplet: These days, in his spare time, Barrat uses video clips and
“I’mma need a fix, girl you was celebrating / photos from fashion shows to produce AI-generated images
Mayonnaise colored Benz I get my engine rev- of models wearing new outfits. The results are smeary, glitch-
ving.”) At school, Barrat’s friends loved it, but ridden, and weird—ever thought you’d like pants with a bag
some adults were shocked. “The teacher got a wrapped around the lower leg, or a sweater with a giant pouch
little bit upset because the neural network was hanging from one side?—but Barrat is working with a designer
extremely profane,” he says. to translate them into real clothes. He can’t wait to try them on.

PETER PRATO
 Now a freshman in computer science
at the University of Georgia, Shaza
Mehdi trained a neural network to
identify plant diseases on sight.

Diagnosis of Botanical Ailments? There’s an App for That

THE ROSEBUSHES in Shaza Mehdi’s front yard she leaned on strangers in discussion forums. “I was really
are beautiful but prone to sickness. One day last annoying about it,” she recalls cheerfully.
year, Mehdi, a fan of Star Trek, asked herself why Mehdi was particularly inspired by a YouTube video starring
her phone couldn’t function like a tricorder to a Stanford researcher who built a neural network that rivals
diagnose the plants’ alictions. “How would a board-certified dermatologists at identifying skin cancers.
computer be able to know?” wondered the high An online tutorial told her how to implement the researcher’s
school senior from Lawrenceville, Georgia. Soon trick herself. Step one was to download software trained to
she, together with a friend named Nile Ravenell, recognize everyday objects such as toilets and teapots. Step
2. Only 29 percent of was tinkering with neural networks between two was to retune its visual sense by feeding it roughly 10,000
rural high schools and
34 percent of urban going to class, getting her nails done, and hang- labeled images of ailing plants that Mehdi had diligently col-
ones offer computer
science courses.
ing out at the Wale House near her school. lected from the web, identified by disease.
About 45 percent of Mehdi didn’t know how to code, and the adults Late in 2017 she finally put her app, which she had chris-
suburban high schools
do. Mehdi taught in her life could provide encouragement but not tened plantMD, to the test. Mehdi looked nervously at a
herself through expertise; her school didn’t ofer introductory sickly looking grapevine with light green patches and brown
tutorials like “Tensor-
Flow for Poets.” classes in computer science. 2 Lying on her bed spots on its leaves. A pockmarked leaf snapped into focus on
at night with the family dog, Teddy, and her the phone’s screen. A few tense heartbeats later, the phrase
underpowered Dell laptop, Mehdi taught her- “grapevine anthracnose” blinked into view above it. A quick
self the programming language Python and the web search confirmed the diagnosis: a clear case of the fun-
basics of neural networks from YouTube vid- gal infection also known as bird’s-eye rot. “I was incredibly
eos and online tutorials. As she ran into bugs, relieved,” Mehdi recalls. The tricorder had worked.

IRINA ROZOVSKY
The Camera Is Ready to Check in Your Clothes 3. It’s only going to get
tougher: According to
the UN, Japan’s popula-
tion is expected to
DRY CLEANING is a tough business in the small, them. Online, he read about machine learning, stretching decline by roughly 20
percent by 2050, while
aging cities of Japan. Daisuke Tahara’s family his English and programming skills to the limit. In the store, the working-age pop-
owns eight dry cleaners in Tagawa, a shrink- he took 40,000 photos of suits, shirts, skirts, and other gar- ulation could plummet
by nearly a third.
ing southern-prefecture town of about 50,000, ments, and used them to train his code.
where it can be tough to find good employees. 3 In July, Tahara started testing his system in one of his
So Tahara began thinking about having com- stores. Customers lay their clothes on a table with a cam-
puters augment his workforce. era mounted overhead. His software gives them a look, then
First Tahara, 38, tried to modernize his busi- displays its verdict (two shirts, one jacket) for confirmation
ness with a better computer system to log and on a tablet computer. Employees typically have to help the
track orders. But most of his employees had customer the first time. After that, they can use it alone.
little experience with technology, and they Tahara says his workers were at first suspicious of his
struggled to adapt. “They forgot easily,” Tahara creation but have come around after finding it makes their
says. So the self-taught coder began research- jobs easier. He doesn’t plan to use the project as an excuse to
ing how software could check in a customer’s eliminate jobs, but he hopes it will help him expand. “I want
garments automatically, just by looking at to open a store with only the system and no staf,” he says.

Daisuke Tahara studied machine

learning online and has used it to
automate some of the work at his
family’s dry-cleaning business.

TARO KARIBE
0 8 7
 Waymo in Miniature

IN A WAREHOUSE in Oakland, California, a small,

nerdy crowd watches Will Roscoe tap a phone
with his thumb. At his feet, an RC car with its
plastic skin ripped of starts to drive around a
racetrack marked in yellow and white tape on
the scratched concrete floor—with no further
input from Roscoe. The Frankenvehicle, which
has a camera and a bunch of electronics zip-tied
to its top, is called a Donkey Car. Roscoe is no AI
expert, but his creation uses neural network soft-
ware similar to what Waymo’s street-legal auton-
omous minivans rely on to perceive the world.
A civil engineer by training, Roscoe was
inspired to create the Donkey Car by a politi-
cal defeat. In 2016 he ran for a seat on the board
of the Bay Area subway system, BART. Roscoe
pledged to expand capacity by replacing trains
with self-driving electric buses, but he finished
third. Building his own pint-size autonomous
vehicle seemed a good way to show voters that
the technology wasn’t pure fancy. “I wanted to
demonstrate it can work at a small scale,” he says.
As it turned out, his timing was perfect—a
robotics hobbyist group dedicated to hacking RC
cars was about to hold its first meeting in nearby
Berkeley. There he met a fellow tinkerer, Adam
Conway, who ofered to build the vehicle. Ros-
coe, a self-taught coder, crafted its self-driving
autopilot using TensorFlow, software created
by Google and later released as open source.
He also borrowed some neural network code
from an attendee of the RC car meetup. Ros-
coe’s final design learns to drive by watching a
human steer the vehicle during demonstration
runs. He named his creation Donkey Car after
what he considers its spirit animal—safe for
kids, not conventionally elegant, and prone to
fits of disobedience.
Roscoe and Conway put all their software and
hardware designs online for others to use. Don-
key Cars now race in Hong Kong, Paris, and Mel-
bourne, Australia. At the Oakland warehouse in
January, nine home-built autonomous vehicles
vied to complete the fastest lap around the track;
among the contenders was a Donkey Car built
by a trio of nervous high schoolers. The vehicles
are also beginning to venture beyond 4 the race-
track. Two hobbyists near Los Angeles modified
theirs to spot and scoop up trash on the beach.
4. If a rogue Donkey
In Oakland, Roscoe’s car had leaves lodged in Car is coming your way,
its suspension. “I’ve been trying to take it out how do you stop it?
Roscoe says: “Just put
on sidewalks,” he says. “I even have a leash.” your foot in front of it.”

PETER PRATO
0 8 8
Will Roscoe trails his self-driving
vehicle in Oakland, California. He
cobbled it together as a proof of con-
cept for autonomous mass transit.
 SOMETIME AROUND 1 AM ON A WARM NIGHT
last June, Fei-Fei Li was sitting in her pajamas
in a Washington, DC, hotel room, practicing a
speech she would give in a few hours. Before
going to bed, Li cut a full paragraph from her
notes to be sure she could reach her most
important points in the short time allotted.
When she woke up, the 5 3 expert in artificial
intelligence put on boots and a black and navy
knit dress, a departure from her frequent uni-
form of a T-shirt and jeans. Then she took an
Uber to the Rayburn House Oice Building, just
south of the US Capitol.
Before entering the chambers of the US House
Committee on Science, Space, and Technology,
she lifted her phone to snap a photo of the over-
size wooden doors. (“As a scientist, I feel spe-
cial about the committee,” she said.) Then she
stepped inside the cavernous room and walked
THE to the witness table.
HUMAN IN THE The hearing that morning, titled “Artificial
Intelligence—With Great Power Comes Great
MACHINE Responsibility,” included Timothy Persons,
ARTIFICIAL chief scientist of the Government Account-
INTELLIGENCE has ability Oice, and Greg Brockman, cofounder
a problem: and chief technology officer of the nonprofit
The biases of its OpenAI. But only Li, the sole woman at the table,
creators are getting could lay claim to a groundbreaking accomplish-
hard-coded ment in the field of AI. As the researcher who
into its future. built ImageNet, a database that helps comput-
FEI-FEI LI has a plan ers recognize images, she’s one of a tiny group
to fix that—by of scientists—a group perhaps small enough to
rebooting the field she fit around a kitchen table—who are responsible
helped invent. for AI’s recent remarkable advances.
That June, Li was serving as the chief AI sci-
entist at Google Cloud and was on leave from
her position as director of the Stanford Artifi-
cial Intelligence Lab. But she was appearing in
front of the committee because she was also the
cofounder of a nonprofit focused on recruiting
women and people of color to become builders
of artificial intelligence.
It was no surprise that the legislators sought
her expertise that day. What was surprising
was the content of her talk: the grave dangers
brought on by the field she so loved.
The time between an invention and its impact
can be short. With the help of artificial intelli-
gence tools like ImageNet, a computer can be
By JESSI HEMPEL taught to learn a specific task and then act far
Photograph by CHRISTIE HEMM KLOK faster than a person ever could. As this technol-

0 9 1
 ogy becomes more sophisticated, it’s being deputized to fil- father took her to a gas station and asked her
ter, sort, and analyze data and make decisions of global and to tell the mechanic to fix his car. She spoke lit-
social consequence. Though these tools have been around, tle English, but through gestures Li figured out
in some way or another, for more than 60 years, in the past how to explain the problem. Within two years,
decade we’ve started using them for tasks that change the Li had learned enough of the language to serve
trajectory of human lives: Today artificial intelligence helps as a translator, interpreter, and advocate for her
determine which treatments get used on people with illnesses, mother and father, who had learned only the
who qualifies for life insurance, how much prison time1 a per- most basic English. “I had to become the mouth
son serves, which job applicants get interviews. and ears of my parents,” she says.
Those powers, of course, can be dangerous.2 Amazon had to She was also doing very well in school. Her
ditch AI recruiting software that learned to penalize résumés father, who loved to scour garage sales, found
that included the word “women.” And who can forget Google’s her a scientific calculator, which she used in
2015 fiasco, when its photo identification software mislabeled math class until a teacher, sizing up her mistaken
black people as gorillas, or Microsoft’s AI-powered social chat- calculations, figured out that it had a broken
bot that started tweeting racial slurs. But those are problems function key. Li credits another high school math
that can be explained and therefore reversed. In the pretty instructor, Bob Sabella, for helping her navigate
near future, Li believes, we will hit a moment when it will be her academic life and her new American identity.
impossible to course-correct. That’s because the technology Parsippany High School didn’t have an advanced
is being adopted so fast, and far and wide. calculus class, so he concocted an ad hoc ver-
Li was testifying in the Rayburn building that morning sion and taught Li during lunch breaks. Sabella
because she is adamant her field needs a recalibration. Prom- and his wife also included her in their family,
inent, powerful, and mostly male tech leaders have been bringing her on a Disney vacation and lending
warning about a future in which artificial-intelligence-driven her $20,000 to open a dry-cleaning 3 business for
technology becomes an existential threat to humans. But Li her parents to run. In 1995, she earned a schol-
thinks those fears are given too much weight and attention. arship to study at Princeton. While there, she
She is focused on a less melodramatic but more consequential traveled home nearly every weekend to help
question: how AI will afect the way people work and live. It’s run the family business.
bound to alter the human experience—and not necessarily for At college, Li’s interests were expansive. She
the better. “We have time,” Li says, “but we have to act now.” majored in physics and studied computer science
If we make fundamental changes to how AI is engineered— and engineering. In 2000, she began her doctor-
and who engineers it—the technology, Li argues, will be a ate at Caltech in Pasadena, working at the inter-
transformative force for good. If not, we are leaving a lot of section of neuroscience and computer science.
humanity out of the equation. Her ability to see and foster connections
At the hearing, Li was the last to speak. With no evidence between seemingly dissimilar fields is what
of the nerves that drove her late-night practice, she began. led Li to think up ImageNet. Her computer-vision
“There’s nothing artificial about AI.” Her voice picked up peers were working on models to help computers
momentum. “It’s inspired by people, it’s created by people, perceive and decode images, but those models
and—most importantly—it impacts people. It is a power- were limited in scope: A researcher might write
ful tool we are only just beginning to understand, and that one algorithm to identify dogs and another to
is a profound responsibility.” Around her, faces brightened. identify cats. Li began to wonder if the prob-
The woman who kept attendance agreed audibly, with an lem wasn’t the model but the data. She thought
“mm-hmm.” that, if a child learns to see by experiencing the
visual world—by observing countless objects
and scenes in her early years—maybe a com-

FEI-FEI Li grew up in Chengdu, an

industrial city in southern China. She was a lonely, brainy kid,
puter can learn in a similar way, by analyzing
a wide variety of images and the relationships
between them. The realization was a big one for
as well as an avid reader. Her family was always a bit unusual: Li. “It was a way to organize the whole visual
In a culture that didn’t prize pets, her father brought her a concept of the world,” she says.
puppy. Her mother, who had come from an intellectual fam- But she had trouble convincing her colleagues
ily, encouraged her to read Jane Eyre. (“Emily is my favor- that it was rational to undertake the gargantuan
ite Brontë,” Li says. “Wuthering Heights.”) When Li was 12, task of tagging every possible picture of every
her father emigrated to Parsippany, New Jersey, and she object in one gigantic database. What’s more, Li
and her mother didn’t see him for several years. They joined had decided that for the idea to work, the labels
him when she was 16. On her second day in America, Li’s would need to range from the general (“mam-

J E S S I H E M P E L wrote about Uber CEO Dara Khosrowshahi in issue 26.05.

 mal”) to the highly specific (“star-nosed mole”). When Li, of her peers were women. At that moment she
who had moved back to Princeton to take a job as an assis- felt this acutely; she saw how the disparity was
tant professor in 2007, talked up her idea for ImageNet, she increasingly going to be a problem. Most scien-
had a hard time getting faculty members to help out. Finally, tists building AI algorithms were men, and often
a professor who specialized in network architecture agreed men of a similar background. They had a partic-
to join her as a collaborator. ular worldview that bled into the projects they
Her next challenge was to get the giant thing built. That pursued and even the dangers they envisioned.
meant a lot of people would have to spend a lot of hours doing Many of AI’s creators had been boys with sci-fi
the tedious work of tagging photos. Li tried paying Princeton dreams, thinking up scenarios from The Termi-
students $10 an hour, but progress was slow going. Then a nator and Blade Runner. There’s nothing wrong
student asked her if she’d heard of Amazon Mechanical Turk.4 with worrying about such things, Li thought.
1. Tools that predict Suddenly she could corral many workers, at a fraction of the But those ideas betrayed a narrow view of the
a criminal’s propensity
for future crime have cost. But expanding a workforce from a handful of Princeton possible dangers of AI.
been used by courts to students to tens of thousands of invisible Turkers had its Deep learning systems are, as Li says, “bias
justify sentencing and
probation decisions. own challenges. Li had to factor in the workers’ likely biases. in, bias out.” Li recognized that while the algo-
Researchers found “Online workers, their goal is to make money the easiest way, rithms that drive artificial intelligence may
that one such tool was
no more effective at right?” she says. “If you ask them to select panda bears from appear to be neutral, the data and applications
predicting recidivism
than an online poll 100 images, what stops them from just clicking everything?” that shape the outcomes of those algorithms are
of people who were So she embedded and tracked certain images—such as pic- not. What mattered were the people building it
given a few details
about an inmate. tures of golden retrievers that had already been correctly and why they were building it. Without a diverse
identified as dogs—to serve as a control group. If the Turks group of engineers, Li pointed out that day on
2. Back in the ’80s,
not long after Li was labeled these images properly, they were working honestly. Capitol Hill, we could have biased algorithms
born, a British medical
school was using a In 2009, Li’s team felt that the massive set—3.2 million making unfair loan application decisions, or
computer program to images—was comprehensive enough to use, and they pub- training a neural network only on white faces—
automate admissions.
A few years into the lished a paper on it, along with the database. (It later grew to creating a model that would perform poorly on
process, two professors 15 million.) At first the project got little attention. But then black ones. “I think if we wake up 20 years from
figured out that the
model was filtering out the team had an idea: They reached out to the organizers of now and we see the lack of diversity in our tech
women and people
with non-European- a computer-vision competition taking place the following and leaders and practitioners, that would be my
sounding names. year in Europe and asked them to allow competitors to use doomsday scenario,” she said.
3. Li later sold the dry- the ImageNet database to train their algorithms. This became It was critical, Li came to believe, to focus the
cleaning business and the ImageNet Large Scale Visual Recognition Challenge. development of AI on helping the human expe-
paid back the Sabellas.
“My first and last Around the same time, Li joined Stanford as an assistant rience. One of her projects at Stanford was a
startup exit!” she says.
professor. She was, by then, married to Silvio Savarese, a partnership with the medical school to bring AI
4. The marketplace, roboticist. But he had a job at the University of Michigan, and to the ICU in an efort to cut down on problems
operated by Amazon,
allows “requesters” the distance was tough. “We knew Silicon Valley would be eas- like hospital-acquired infections. It involved
to set a price for ier for us to solve our two-body problem,” Li says. (Savarese developing a camera system that could monitor
people to complete
“human intelligence joined Stanford’s faculty in 2013.) “Also, Stanford is special a hand-washing station and alert hospital work-
tasks” online. One
study found that 2,676 because it’s one of the birthplaces of AI.” ers if they forgot to scrub properly. This type of
workers who performed In 2012, University of Toronto researcher Geofrey Hinton interdisciplinary collaboration was unusual. “No
3.8 million tasks
earned a median wage entered the ImageNet competition, using the database to train one else from computer science reached out to
of about $2 an hour. a type of AI known as a deep neural network. It turned out to me,” says Arnold Milstein, a professor of medi-
be far more accurate than anything that had come before— cine who directs Stanford’s Clinical Excellence
and he won. Li hadn’t planned to go see Hinton get his award; Research Center.
she was on maternity leave, and the ceremony was happening That work gave Li hope for how AI could
in Florence, Italy. But she recognized that history was being evolve. It could be built to complement peo-
made. So she bought a last-minute ticket and crammed herself ple’s skills rather than simply replace them. If
into a middle seat for an overnight flight. Hinton’s ImageNet- engineers would engage with people in other
powered neural network changed everything. By 2017, the final disciplines (even people in the real world!), they
year of the competition, the error rate for computers identify- could make tools that expand human capac-
ing objects in images had been reduced to less than 3 percent, ity, like automating time-consuming tasks
from 15 percent in 2012. Computers, at least by one measure, to allow ICU nurses to spend more time with
had become better at seeing than humans. patients, rather than building AI, say, to auto-
ImageNet enabled deep learning to go big—it’s at the root of mate someone’s shopping experience and elim-
recent advances in self-driving cars, facial recognition, phone inate a cashier’s job.
cameras that can identify objects (and tell you if they’re for sale). Considering that AI was developing at warp
Not long after Hinton accepted his prize, while Li was still speed, Li figured her team needed to change the
on maternity leave, she started to think a lot about how few roster—as fast as possible.

0 9 3
 LI HAS always been drawn to math, so she
recognizes that getting women and people of color into
Initially the experience was enlivening. She
met with companies that had real-world uses for
computer science requires a colossal efort. According to her science. She led the rollout of public-facing
the National Science Foundation, in 2000, women earned AI tools that let anyone create machine learn-
28 percent of bachelor’s degrees in computer science. In ing algorithms without writing a single line of
2015 that figure was 18 percent. Even in her own lab, Li code. She opened a new lab in China and helped
struggles to recruit underrepresented people of color and to shape AI tools to improve health care. She
women. Though historically more diverse than your typical spoke at the World Economic Forum in Davos,
AI lab, it remains predominantly male, she says. “We still do rubbing elbows with heads of state and pop stars.
not have enough women, and especially underrepresented But working in a private company came with
minorities, even in the pipeline coming into the lab,” she says. new and uncomfortable pressures. Last spring,
“Students go to an AI conference and they see 90 percent 5 Li was caught up in Google’s very public drub-
people of the same gender. And they don’t see African Amer- bing over its Project Maven contract with the
icans nearly as much as white boys.” Defense Department. The program uses AI to
Olga Russakovsky had almost written of the field when Li interpret video images that could be used to
became her adviser. Russakovsky was already an accomplished target drone strikes; according to Google, it
computer scientist—with an undergraduate degree in math was “low-res object identification using AI”
and a master’s in computer science, both from Stanford—but and “saving lives was the overarching intent.”
her dissertation work was dragging. She felt disconnected Many employees, however, objected to the use
from her peers as the only woman in her lab. Things changed of their work in military drones. About 4,000 of
when Li arrived at Stanford. Li helped Russakovsky learn some them signed a petition demanding “a clear pol-
skills required for successful research, “but also she helped icy stating that neither Google nor its contrac-
build up my self-confidence,” says Russakovsky, who is now tors will ever build warfare technology.” Several
an assistant professor in computer science at Princeton. workers resigned in protest.
Four years ago, as Russakovsky was finishing up her PhD, Though Li hadn’t been involved directly with
she asked Li to help her create a summer camp to get girls the deal, the division that she worked for was
interested in AI. Li agreed at once, and they pulled volunteers charged with administering Maven. And she
together and posted a call for high school sophomores. Within became a public face of the controversy when
a month, they had 200 applications for 24 spots. Two years emails she wrote that looked as if they were try-
later they expanded the program, launching the nonprofit ing to help the company avoid embarrassment
AI4All 6 to bring underrepresented youth—including girls, were leaked to The New York Times. Publicly
people of color, and people from economically disadvantaged this seemed confusing, as she was well known
backgrounds—to the campuses of Stanford and UC Berkeley. in the field as someone who embodied ethics.
AI4All is on the verge of growing out of its tiny shared oice In truth, before the public outcry she had con-
at the Kapor Center in downtown Oakland, California. It now sidered the technology to be “fairly innocu-
has camps at six college campuses. (Last year there were 900 ous”; she hadn’t considered that it could cause
applications for 20 spots at the newly launched Carnegie an employee revolt.
Mellon camp.) One AI4All student worked on detecting eye But Li does recognize why the issue blew up:
diseases using computer vision. Another used AI to write a “It wasn’t exactly what the thing is. It’s about
program ranking the urgency of 911 calls; her grandmother the moment—the collective sense of urgency
had died because an ambulance didn’t reach her in time. Con- for our responsibility, the emerging power of
firmation, it would seem, that personal perspective makes a AI, the dialog that Silicon Valley needs to be
diference for the future of AI tools. in. Maven just became kind of a convergence
point,” she says. “Don’t be evil” was no longer
a strong enough stance.

AFTER THREE years running

the AI Lab at Stanford, Li took a leave in 2016 to join Goo-
The controversy subsided when Google
announced it wouldn’t renew the Maven con-
tract. A group of Google scientists and exec-
gle as chief scientist for AI of Google Cloud, the company’s utives—including Li—also wrote (public)
enterprise computing business. Li wanted to understand guidelines pledging that Google would focus
how industry worked and to see if access to customers anx- its AI research on technology designed for social
ious to deploy new tools would shift the scope of her own good, would avoid implementing bias into its
cross-disciplinary research. Companies like Facebook, Goo- tools, and would avoid technology that could
gle, and Microsoft were throwing money into AI in search end up facilitating harm to people. Li had been
of ways to harness the technology for their businesses. And preparing to head back to Stanford, but she felt
companies often have more and better data than universi- it was critical to see the guidelines through.
ties. For an AI researcher, data is fuel. “I think it’s important to recognize that every

0 9 4
 organization has to have a set of principles and respon- debacle. But she said the reason for her return
sible review processes. You know how Benjamin Franklin to Stanford was that she didn’t want to forfeit
said, when the Constitution was rolled out, it might not be her academic position. She also sounded tired.
perfect but it’s the best we’ve got for now,” she says. “Peo- After a tumultuous summer at Google, the eth-
ple will still have opinions, and diferent sides can continue ics guidelines she helped write were “the light
the dialog.” But when the guidelines were published, she at the end of the tunnel,” she says.
says, it was one of her happiest days of the year: “It was so And she was eager to start a new project at
important for me personally to be involved, to contribute.” Stanford. This fall, she and John Etchemendy,
the former Stanford provost, announced the
creation of an academic center that will fuse the

IN JUNE, I visited Li at her home, a

modest split-level in a cul-de-sac on the Stanford campus.
study of AI and humanity, blending hard science,
design research, and interdisciplinary studies.
“As a new science, AI never had a field-wide efort
It was just after 8 in the evening, and while we talked her to engage humanists and social scientists,” she
husband put their young son and daughter through their says. Those skill sets have long been viewed as
bedtime routines upstairs. Her parents were home for inconsequential to the field of AI, but Li is ada-
the night in the in-law unit7 downstairs. The dining room mant that they are key to its future.
had been turned into a playroom, so we sat in her living Li is fundamentally optimistic. At the hearing
room. Family photos rested on every surface, including a in June, she told the legislators, “I think deeply
broken 1930s-era telephone sitting on a shelf. “Immigrant about the jobs that are currently dangerous
parents!” she said when I ask her about it. Her father still and harmful for humans, from fighting fires to
likes to go to yard sales. search and rescue to natural disaster recovery.”
As we talked, text messages started pinging on Li’s phone. She believes that we should not only avoid put-
Her parents were asking her to translate a doctor’s instruc- ting people in harm’s way when possible, but
tions for her mother’s medication. Li can be in a meeting at that these are often the very kind of jobs where
the Googleplex or speaking at the World Economic Forum technology can be a great help.
or sitting in the green room before a congressional hearing There are limits, of course, to how much a
and her parents will text her for a quick assist. She responds single program at a single institution—even a
without breaking her train of thought. 5. In 2017 women made prominent one—can shift an entire field. But
up just 12 percent of
For much of Li’s life, she has been focused on two seem- AI researchers who
Li is adamant she has to do what she can to
ingly diferent things at the same time. She is a scientist who were published at major train researchers to think like ethicists, who
machine-learning
has thought deeply about art. She is an American who is Chi- conferences, according are guided by principle over profit, informed
nese. She is as obsessed with robots as she is with humans. to research by Element by a varied array of backgrounds.
AI, in collaboration
Late in July, Li called me while she was packing for a fam- with WIRED . On the phone, I ask Li if she imagines there
ily trip and helping her daughter wash her hands. “Did you 6. About half of the
could have been a way to develop AI diferently,
see the announcement of Shannon Vallor?” she asks. Vallor students in the 2018 without, perhaps, the problems we’ve seen so far.
AI4All summer camp
is a philosopher at Santa Clara University whose research at Princeton were “I think it’s hard to imagine,” she says. “Scientific
focuses on the philosophy and ethics of emerging science of African American, advances and innovation come really through
Latinx, or Native
and technologies, and she had just signed on to work for American descent. generations of tedious work, trial and error. It
Eighty-seven percent
Google Cloud as a consulting ethicist. Li had campaigned of students received
took a while for us to recognize such bias. I only
hard for this; she’d even quoted Vallor in her testimony in some financial aid. woke up six years ago and realized ‘Oh my God,
Washington, saying: “There are no independent machine 7. Roughly 4.5 million we’re entering a crisis.’ ”
values. Machine values are human values.” The appoint- parents lived with On Capitol Hill, Li said, “As a scientist, I’m hum-
their grown children
ment wasn’t without precedent. Other companies have also in the US in 2010, bled by how nascent the science of AI is. It is the
according to the
started to put guardrails on how their AI software can be Census Bureau, a
science of only 60 years. Compared to classic
used, and who can use it. Microsoft established an internal 15 percent increase sciences that are making human life better every
from 2007.
ethics board in 2016. The company says it has turned down day—physics, chemistry, biology—there’s a long,
business with potential customers owing to ethical concerns long way to go for AI to realize its potential to help
brought forward by the board. It’s also begun placing lim- people.” She added, “With proper guidance AI will
its on how its AI tech can be used, such as forbidding some make life better. But without it, the technology
applications in facial recognition. stands to widen the wealth divide even further,
But to speak on behalf of ethics from inside a corpora- make tech even more exclusive, and reinforce
tion is, to some extent, to acknowledge that, while you can biases we’ve spent generations trying to over-
guard the henhouse, you are indeed a fox. When we talked in come.” This is the time, Li would have us believe,
July, Li already knew she was leaving Google. Her two-year between an invention and its impact.
sabbatical was coming to an end. There was plenty of spec-
ulation about her stepping down after the Project Maven Additional reporting by Gregory Barber.
THE MAN
WHO
EXPLAINED

Photographs by Kate Peters

0 9 6
 whose brains need work. The National Hospi-
tal for Neurology and Neurosurgery—where a
modern-day royal might well seek treatment—
dominates one corner of Queen Square, and the
world-renowned neuroscience research facili-
ties of University College London round out its
perimeter. During a week of perfect weather last
July, dozens of neurological patients and their
families passed silent time on wooden benches
at the outer edges of the grass.
On a typical Monday, Karl Friston arrives on

Karl Friston’s FREE ENERGY PRINCIPLE Queen Square at 12:25 pm and smokes a cigarette
in the garden by the statue of Queen Charlotte.
might be the most all-encompassing A slightly bent, solitary figure with thick gray

idea since Charles Darwin’s theory hair, Friston is the scientific director of Univer-
sity College London’s storied Functional Imaging
of natural selection. Some believe it’s Laboratory, known to everyone who works there

the missing link to TRUE ARTIFICIAL as the FIL. After finishing his cigarette, Friston
walks to the western side of the square, enters
INTELLIGENCE. But to understand it, a brick and limestone building, and heads to a

you need to peer inside the mind seminar room on the fourth floor, where any-
where from two to two dozen people might be
of FRISTON himself. by Shaun Raviv facing a blank white wall waiting for him. Fris-
ton likes to arrive five minutes late, so everyone
else is already there.
His greeting to the group is liable to be his
first substantial utterance of the day, as Friston
prefers not to speak with other human beings
before noon. (At home, he will have conversed

THING
with his wife and three sons via an agreed-upon
series of smiles and grunts.) He also rarely meets
people one-on-one. Instead, he prefers to hold
open meetings like this one, where students,
postdocs, and members of the public who desire
Friston’s expertise—a category of person that
has become almost comically broad in recent
years—can seek his knowledge. “He believes
WHEN KING GEORGE III of England began to show signs of that if one person has an idea or a question or
acute mania toward the end of his reign, rumors about the project going on, the best way to learn about it
royal madness multiplied quickly in the public mind. One leg- is for the whole group to come together, hear
end had it that George tried to shake hands with a tree, believ- the person, and then everybody gets a chance to
ing it to be the King of Prussia. Another described how he was ask questions and discuss. And so one person’s
whisked away to a house on Queen Square, in the Bloomsbury learning becomes everybody’s learning,” says
district of London, to receive treatment among his subjects. David Benrimoh, a psychiatry resident at McGill
The tale goes on that George’s wife, Queen Charlotte, hired University who studied under Friston for a year.
out the cellar of a local pub to stock provisions for the king’s “It’s very unique. As many things are with Karl.”
meals while he stayed under his doctor’s care. At the start of each Monday meeting, every-
More than two centuries later, this story about Queen Square one goes around and states their questions at the
is still popular in London guidebooks. And whether or not it’s outset. Friston walks in slow, deliberate circles
true, the neighborhood has evolved over the years as if to con- as he listens, his glasses perched at the end of his
form to it. A metal statue of Charlotte stands over the north- nose, so that he is always lowering his head to
ern end of the square; the corner pub is called the Queen’s see the person who is speaking. He then spends
Larder; and the square’s quiet rectangular garden is now all the next few hours answering the questions in
but surrounded by people who work on brains and people turn. “A Victorian gentleman, with Victorian
manners and tastes,” as one friend describes
S H A U N R A V I V (@ShaunRaviv) is a writer living in Friston, he responds to even the most confused
Atlanta, Georgia. questions with courtesy and rapid reformulation.

0 9 8
 The Q&A sessions—which I started calling “Ask Karl” meet- it, including researchers whose work depends
ings—are remarkable feats of endurance, memory, breadth on it, told me they didn’t fully comprehend it.
of knowledge, and creative thinking. They often end when it But often those same people hastened to add
is time for Friston to retreat to the minuscule metal balcony that the free energy principle, at its heart, tells a
hanging of his oice for another smoke. simple story and solves a basic puzzle. The second
Friston first became a heroic figure in academia for devis- law of thermodynamics tells us that the universe
ing many of the most important tools that have made human tends toward entropy, toward dissolution; but
brains legible to science. In 1990 he invented statistical para- living things fiercely resist it. We wake up every
metric mapping, a computational technique that helps—as morning nearly the same person we were the day
one neuroscientist put it—“squash and squish” brain images before, with clear separations between our cells
into a consistent shape so that researchers can do apples-to- and organs, and between us and the world with-
apples comparisons of activity within diferent crania. Out out. How? Friston’s free energy principle says
of statistical parametric mapping came a corollary called that all life, at every scale of organization—from
voxel-based morphometry, an imaging technique that was single cells to the human brain, with its billions of
used in one famous study to show that the rear side of the hip- neurons—is driven by the same universal imper-
pocampus of London taxi drivers grew as they learned “the ative, which can be reduced to a mathematical
knowledge.” 1 A study published in Science in 2011 used yet a function. To be alive, he says, is to act in ways
third brain-imaging-analysis software invented by Friston— that reduce the gulf between your expectations
dynamic causal modeling—to determine if people with severe and your sensory inputs. Or, in Fristonian terms,
brain damage were minimally conscious or simply vegetative. it is to minimize free energy.
When Friston was inducted into the Royal Society of Fel- To get a sense of the potential implications of
lows in 2006, the academy described his impact on studies of this theory, all you have to do is look at the array
the brain as “revolutionary” and said that more than 90 per- of people who darken the FIL’s doorstep on Mon-
cent of papers published in brain imaging used his methods. day mornings. Some are here because they want
Two years ago, the Allen Institute for Artificial Intelligence, a to use the free energy principle to unify theo-
research outfit led by AI pioneer Oren Etzioni, calculated that ries of the mind, provide a new foundation for
Friston is the world’s most frequently cited neuroscientist. biology, and explain life as we know it. Others
He has an h-index—a metric used to measure the impact of hope the free energy principle will finally ground
1. To earn a London taxi
a researcher’s publications—nearly twice the size of Albert psychiatry in a functional understanding of the
license, drivers must Einstein’s. Last year Clarivate Analytics, which over more brain. And still others come because they want
memorize 320 routes
and many landmarks than two decades has successfully predicted 46 Nobel Prize to use Friston’s ideas to break through the road-
within 6 miles of Charing winners in the sciences, ranked Friston among the three blocks in artificial intelligence research. But they
Cross. The grueling pro-
cess includes a written most likely winners in the physiology or medicine category. all have one reason in common for being here,
test as well as a series
of one-on-one meetings
What’s remarkable, however, is that few of the researchers which is that the only person who truly under-
with an examiner. who make the pilgrimage to see Friston these days have come stands Karl Friston’s free energy principle may
2. The account is called to talk about brain imaging at all. Over a 10-day period this be Karl Friston himself.
@FarlKriston. Sample summer, Friston advised an astrophysicist, several philoso-
tweet: “Life is an inevi-
table & emergent prop- phers, a computer engineer working on a more personable FRISTON ISN'T JUST one of the most influ-
erty of any (ergodic)
random dynamical sys-
competitor to the Amazon Echo, the head of artificial intel- ential scholars in his field; he’s also among the
tem that possesses a ligence for one of the world’s largest insurance companies, most prolific in any discipline. He is 59 years old,
Markov blanket. Don’t
leave with out it!” a neuroscientist seeking to build better hearing aids, and a works every night and weekend, and has pub-
psychiatrist with a startup that applies machine learning to lished more than 1,000 academic papers since
3. A 2018 article in
Nature analyzed the help treat depression. And most of them had come because the turn of the millennium. In 2017 alone, he was
phenomenon of “hyper-
prolific” scholars,
they were desperate to understand something else entirely. a lead or coauthor of 85 publications3—which
which the authors For the past decade or so, Friston has devoted much of his amounts to approximately one every four days.
defined as anyone with
more than 72 publica- time and efort to developing an idea he calls the free energy But if you ask him, this output isn’t just the fruit
tions in a year. principle. (Friston refers to his neuroimaging research as a of an ambitious work ethic; it’s a mark of his
day job, the way a jazz musician might refer to his shift at the tendency toward a kind of rigorous escapism.
local public library.) With this idea, Friston believes he has Friston draws a carefully regulated boundary
identified nothing less than the organizing principle of all around his inner life, guarding against intru-
life, and all intelligence as well. “If you are alive,” he sets out sions, many of which seem to consist of “wor-
to answer, “what sorts of behaviors must you show?” rying about other people.” He prefers being
First the bad news: The free energy principle is madden- onstage, with other people at a comfortable
ingly diicult to understand. So diicult, in fact, that entire distance, to being in private conversations. He
rooms of very, very smart people have tried and failed to grasp does not have a mobile phone. He always wears
it. A Twitter account 2 with 3,000 followers exists simply to navy-blue suits, which he buys two at a time
mock its opacity, and nearly every person I spoke with about at a closeout shop. He finds disruptions to his
weekly routine on Queen Square “rather nerve-racking” and
so tends to avoid other human beings at, say, international
conferences. He does not enjoy advocating for his own ideas.
At the same time, Friston is exceptionally lucid and forth-
coming about what drives him as a scholar. He finds it incred-
ibly soothing—not unlike disappearing for a smoke—to lose
himself in a diicult problem that takes weeks to resolve. And
he has written eloquently about his own obsession, dating
back to childhood, with finding ways to integrate, unify, and
make simple the apparent noise of the world.
Friston traces his path to the free energy principle back to
a hot summer day when he was 8 years old. He and his family
were living in the walled English city of Chester, near Liver-
pool, and his mother had told him to go play in the garden. He
turned over an old log and spotted several wood lice—small
bugs with armadillo-shaped exoskeletons—moving about, he
initially assumed, in a frantic search for shelter and darkness.
After staring at them for half an hour, he deduced that they
were not actually seeking the shade. “That was an illusion,”
Friston says. “A fantasy that I brought to the table.”
He realized that the movement of the wood lice had no
larger purpose, at least not in the sense that a human has
a purpose when getting in a car to run an errand. The crea-
tures’ movement was random; they simply moved faster in
the warmth 4 of the sun. Friston calls this his first scientific
insight, a moment when “all these contrived, anthropomor-
phized explanations of purpose and survival and the like all
seemed to just peel away,” he says. “And the thing you were
observing just was. In the sense that it could be no other way.”
Friston’s father was a civil engineer who worked on bridges
all around England, and his family moved around with him.
In just his first decade, the young Friston attended six difer-
ent schools. His teachers often didn’t know what to do with
him, and he drew most of his fragile self-esteem from soli-
tary problem solving. At age 10 he designed a self-righting
robot that could, in theory, traverse uneven ground while
carrying a glass of water, using self-correcting feedback described how he enjoyed electronics design
actuators and mercury levels. At school, a psychologist was and being alone in nature, so the computer sug-
brought in to ask him how he came up with it. “You’re very gested he become a television antenna installer.
intelligent, Karl,” Friston’s mother reassured him, not for That didn’t seem right, so he visited a school
the last time. “Don’t let them tell you you’re not.” He says career counselor and said he’d like to study the
he didn’t believe her. brain in the context of mathematics and physics.
When Friston was in his mid-teens, he had another wood- The counselor told Friston he should become a
lice moment. He had just come up to his bedroom from watch- psychiatrist, which meant, to Friston’s horror,
ing TV and noticed the cherry trees in bloom outside the that he had to study medicine.
window. He suddenly became possessed by a thought that Both Friston and the counselor had confused
has never let go of him since. “There must be a way of under- psychiatry with psychology, which is what he
standing everything by starting from nothing,” he thought. probably ought to have pursued as a future
“If I’m only allowed to start of with one point in the entire researcher. But it turned out to be a fortunate
universe, can I derive everything else I need from that?” He error, as it put Friston on a path toward study-
stayed there on his bed for hours, making his first attempt. ing both the mind and body,5 and toward one of
“I failed completely, obviously,” he says. the most formative experiences of his life—one
Toward the end of secondary school, Friston and his class- that got Friston out of his own head.
mates were the subjects of an early experiment in computer- After completing his medical studies, Friston
assisted advising. They were asked a series of questions, and moved to Oxford and spent two years as a res-
their answers were punched into cards and run through a ident trainee at a Victorian-era hospital called
machine to extrapolate the perfect career choice. Friston had Littlemore. Founded under the 1845 Lunacy Act,

1 0 0
 4. Young Friston was
probably right. Many
species of wood
lice will dry out in direct
sunlight, and some
respond to a rise in
temperature with kine-
sis, an increased level
of random movement.

5. Friston found time

for other pursuits as
well. At age 19, he spent
an entire school vaca-
The Markov blanket in Karl Friston’s tion trying to squeeze
office—“keeping your internal states all of physics on one
warm since 1856.” page. He failed but
did manage to fit all of
quantum mechanics.
Littlemore had originally been instituted to help transfer all was “as rampant and incorrigible a pedophile as 6. The names of
“pauper lunatics” from workhouses to hospitals. By the mid- you could ever imagine,” Friston says. Friston’s patients at
Littlemore have been
1980s, when Friston arrived, it was one of the last of the old And then there was Robert, an articulate changed in this story.
asylums on the outskirts of England’s cities. young man who might have been a university
Friston was assigned a group of 32 chronic schizophrenic student had he not sufered severe schizophre-
patients, the worst-of residents of Littlemore, for whom treat- nia. Robert ruminated obsessively about, of all
ment mostly meant containment. For Friston, who recalls his things, angel shit; he pondered whether the
former patients with evident nostalgia, it was an introduction stuf was a blessing or a curse and whether it
to the way that connections in the brain were easily broken. was ever visible to the eye, and he seemed per-
“It was a beautiful place to work,” he says. “This little com- plexed that these questions had not occurred
munity of intense and florid psychopathology.” to others. To Friston, the very concept of angel
Twice a week he led 90-minute group therapy sessions in shit was a miracle. It spoke to the ability of peo-
which the patients explored their ailments together, remi- ple with schizophrenia to assemble concepts
niscent of the Ask Karl meetings today. The group included that someone with a more regularly function-
colorful characters who still inspire Friston’s thinking more ing brain couldn’t easily access. “It’s extremely
than 30 years later. There was Hillary,6 who looked like she diicult to come up with something like angel
could play the senior cook on Downton Abbey but who, before shit,” Friston says with something like admira-
coming to Littlemore, had decapitated her neighbor with a tion. “I couldn’t do it.”
kitchen knife, convinced he had become an evil, human-sized After Littlemore, Friston spent much of the
crow. There was Ernest, who had a penchant for pastel Marks early 1990s using a relatively new technology—
& Spencer cardigans and matching plimsoll shoes, and who PET scans—to try to understand what was going
on inside the brains of people with schizophrenia. He invented The blanket is a gag gift from Friston’s son, a
statistical parametric mapping along the way. Unusually for plush, polyester inside joke about an idea that
the time, Friston was adamant that the technique should has become central to the free energy princi-
be freely shared rather than patented and commercialized, ple. Markov is the eponym of a concept called
which largely explains how it became so widespread. Fris- a Markov blanket, which in machine learning
ton would fly across the world—to the National Institutes of is essentially a shield that separates one set of
Health in Bethesda, Maryland, for example—to give it to other variables from others in a layered, hierarchical
researchers. “It was me, literally, with a quarter of biometric system. The psychologist Christopher Frith—
tape, getting on an airplane, taking it over there, download- who has an h-index on par with Friston’s—once
ing it, spending a day getting it to work, teaching somebody described a Markov blanket as “a cognitive ver-
how to use it, then going home for a rest,” Friston says. “This sion of a cell membrane, shielding states inside
is how open source software worked in those days.” the blanket from states outside.”
Friston came to Queen Square in 1994, and for a few years his In Friston’s mind, the universe is made up of
oice at the FIL sat just a few doors down from the Gatsby Com- Markov blankets inside of Markov blankets. Each
putational Neuroscience Unit. The Gatsby—where research- of us has a Markov blanket that keeps us apart
ers study theories of perception and learning in both living from what is not us. And within us are blankets
and machine systems—was then run by its founder, the cog- separating organs, which contain blankets sep-
nitive psychologist and computer scientist Geofrey Hinton. arating cells, which contain blankets separating
While the FIL was establishing itself as one of the premier their organelles. The blankets define how biolog-
labs for neuroimaging, the Gatsby was becoming a training ical things exist over time and behave distinctly
ground for neuroscientists interested in applying mathemat- from one another. Without them, we’re just hot
ical models to the nervous system. gas dissipating into the ether.
Friston, like many others, became enthralled by Hinton’s “That’s the Markov blanket you’ve read about.
“childlike enthusiasm” for the most unchildlike of statistical This is it. You can touch it,” Friston said dryly
models, and the two men became friends. 7 Over time, Hinton when I first saw the throw in his oice. I couldn’t
convinced Friston that the best way to think of the brain was help myself; I did briefly reach out to feel it under
as a Bayesian probability machine. The idea, which goes back my fingers. Ever since I first read about Markov
to the 19th century and the work of Hermann von Helmholtz, blankets, I’d seen them everywhere. Markov
is that brains compute and perceive in a probabilistic manner,
constantly making predictions and adjusting beliefs based on Friston’s office. A friend describes
what the senses contribute. According to the most popular him as “a Victorian gentleman, with
Victorian manners and tastes.”
modern Bayesian account, the brain is an “inference engine”
that seeks to minimize “prediction error.”
In 2001, Hinton left London for the University of Toronto,
where he became one of the most important figures in artifi-
cial intelligence, laying the groundwork8 for much of today’s
research in deep learning. Before Hinton left, however, Friston
visited his friend at the Gatsby one last time. Hinton described
a new technique he’d devised to allow computer programs to
emulate human decisionmaking more eiciently—a process
for integrating the input of many diferent probabilistic mod-
els, now known in machine learning as a “product of experts.”
The meeting left Friston’s head spinning. Inspired by Hin-
ton’s ideas, and in a spirit of intellectual reciprocity, Friston
sent Hinton a set of notes about an idea he had for connect-
ing several seemingly “unrelated anatomical, physiological,
and psychophysical attributes of the brain.” Friston published
those notes in 2005—the first of many dozens of papers he
would go on to write about the free energy principle.

EVEN FRISTON HAS A HARD TIME deciding where to start

when he describes the free energy principle. He often sends
people to its Wikipedia page. But for my part, it seems apt to
begin with the blanket draped over the futon in Friston’s oice.
It’s a white fleece throw, custom-printed with a black-
and-white portrait of a stern, bearded Russian mathemati-
cian named Andrei Andreyevich Markov, who died in 1922.
 blankets around a leaf and a tree and a mosquito. In London, solving. When I get into the car to run an errand,
I saw them around the postdocs at the FIL, around the black- I am minimizing free energy by confirming my
clad protesters at an antifascist rally, and around the people hypothesis—my fantasy—through action.
living in boats in the canals. Invisible cloaks around everyone, For Friston, folding action and movement into
and underneath each one a diferent living system that min- the equation is immensely important. Even per-
imizes its own free energy. ception itself, he says, is “enslaved by action”: To
The concept of free energy itself comes from physics, which gather information, the eye darts, the diaphragm
means it’s diicult to explain precisely without wading into draws air into the nose, the fingers generate fric-
mathematical formulas. In a sense that’s what makes it pow- tion against a surface. And all of this fine motor
erful: It isn’t a merely rhetorical concept. It’s a measurable movement exists on a continuum with bigger
quantity that can be modeled, using much the same math plans, explorations, 10 and actions. “We sam-
that Friston has used to interpret brain images to such world- ple the world,” Friston writes, “to ensure our
changing efect. But if you translate the concept from math into predictions become a self-fulfilling prophecy.”
English, here’s roughly what you get: Free energy is the difer- So what happens when our prophecies are not
ence between the states you expect to be in and the states your self-fulfilling? What does it look like for a system
sensors tell you that you are in. Or, to put it another way, when to be overwhelmed by surprise? The free energy
you are minimizing free energy, you are minimizing surprise. principle, it turns out, isn’t just a unified theory of
According to Friston, any biological system 9 that resists a action, perception, and planning; it’s also a the-
tendency to disorder and dissolution will adhere to the free ory of mental illness. When the brain assigns too
energy principle—whether it’s a protozoan or a pro basket- little or too much weight to evidence pouring in
7. At the time, Hinton ball team. A single-celled organism has the same imperative from the senses, trouble occurs. Someone with
was living in a partic-
ularly noisy building in to reduce surprise that a brain does. schizophrenia, for example, may fail to update
Camden. The neigh-
bors’ water pipes were
The only diference is that, as self-organizing biological sys- their model of the world to account for sensory
so loud that he built tems go, the human brain is inordinately complex: It soaks in input from the eyes. Where one person might see
a soundproof box in a
basement bedroom out information from billions of sense receptors, and it needs to a friendly neighbor, Hillary might see a giant, evil
of rubber and ¾-inch organize that information eiciently into an accurate model crow. “If you think about psychiatric conditions,
drywall where he and
his wife could sleep. of the world. “It’s literally a fantastic organ in the sense that and indeed most neurological conditions, they
8. In 2012, Hinton won
it generates hypotheses or fantasies that are appropriate for are just broken beliefs or false inference—hallu-
the ImageNet Chal- trying to explain these myriad patterns, this flux of sensory cinations and delusions,” Friston says.
lenge, a competition to
identify objects in a information that it is in receipt of,” Friston says. In seeking to Over the past few years, Friston and a few
15-million-image data- predict what the next wave of sensations is going to tell it— other scientists have used the free energy prin-
base built by Fei-Fei Li
(see story on page 90). and the next, and the next—the brain is constantly making ciple to help explain anxiety, depression, and
ImageNet helped pro-
pel neural networks—
inferences and updating its beliefs based on what the senses psychosis, along with certain symptoms of
and Hinton—to the relay back, and trying to minimize prediction-error signals. autism, Parkinson’s disease, and psychopathy.
forefront of AI.
So far, as you might have noticed, this sounds a lot like In many cases, scientists already know—thanks
9. In 2013, Friston ran a the Bayesian idea of the brain as an “inference engine” that to Friston’s neuroimaging methods—which
model that simulated
a primordial soup full of Hinton told Friston about in the 1990s. And indeed, Friston regions of the brain tend to malfunction in dif-
floating molecules. He
programmed it to obey
regards the Bayesian model as a foundation of the free energy ferent disorders and which signals tend to be
both basic physics and principle (“free energy” is even a rough synonym for “pre- disrupted. But that alone isn’t enough to go
the free energy princi-
ple. The model gener- diction error”). But the limitation of the Bayesian model, for on. “It’s not sufficient to understand which
ated results that looked Friston, is that it only accounts for the interaction between synapses, which brain connections, are work-
like organized life.
beliefs and perceptions; it has nothing to say about the body ing improperly,” he says. “You need to have a
10. Friston’s term for
this kind of exploration
or action. It can’t get you out of your chair. calculus that talks about beliefs.”
is “epistemic foraging.” This isn’t enough for Friston, who uses the term “active So: The free energy principle ofers a unifying
He is notorious among
his colleagues for inference” to describe the way organisms minimize surprise explanation for how the mind works and a unify-
his coinages, known while moving about the world. When the brain makes a pre- ing explanation for how the mind malfunctions.
as Fristonese.
diction that isn’t immediately borne out by what the senses It stands to reason, then, that it might also put us
relay back, Friston believes, it can minimize free energy in one on a path toward building a mind from scratch.
of two ways: It can revise its prediction—absorb the surprise,
concede the error, update its model of the world—or it can A FEW YEARS AGO, a team of British research-
act to make the prediction true. If I infer that I am touching ers decided to revisit the facts of King George
my nose with my left index finger, but my proprioceptors tell III’s madness with a new analytic tool. They
me my arm is hanging at my side, I can minimize my brain’s loaded some 500 letters written by the king
raging prediction-error signals by raising that arm up and into a machine-learning engine and laboriously
pressing a digit to the middle of my face. trained the system to recognize various textual
And in fact, this is how the free energy principle accounts features: word repetition, sentence length, syn-
for everything we do: perception, action, planning, problem tactical complexity, and the like. By the end of

1 0 3
 the training process, the system was able to predict whether left the monster tended to move to the right.
a royal missive had been written during a period of mania or After a while it became clear that, even in the toy
during a period of sanity. environment of the game, the reward-maximizing
This kind of pattern-matching technology—which is roughly agent was “demonstrably less robust”; the free
similar to the techniques that have taught machines to recog- energy agent had learned its environment better.
nize faces, images of cats, and speech patterns—has driven “It outperformed the reinforcement-learning
huge advances in computing over the past several years. But agent because it was exploring,” Moran says. In
it requires a lot of up-front data and human supervision, and another simulation that pitted the free-energy-
it can be brittle. Another approach to AI, called reinforce- minimizing agent against real human players, the
ment learning, has shown incredible success at winning story was similar. The Fristonian agent started
games: Go, chess, Atari’s Breakout. Reinforcement learning slowly, actively exploring options—epistemi-
doesn’t require humans to label lots of training data; it just cally foraging, Friston would say—before quickly
requires telling a neural network to seek a certain reward, attaining humanlike performance.
often victory in a game. The neural network learns by play- Moran told me that active inference is starting
ing the game over and over, optimizing for whatever moves to spread into more mainstream deep-learning
might get it to the final screen, the way a dog might learn to research, albeit slowly. Some of Friston’s students
perform certain tasks for a treat. have gone on to work at DeepMind and Google
But reinforcement learning, too, has pretty major limita- Brain, and one of them founded Huawei’s Artificial
tions. In the real world, most situations are not organized Intelligence Theory lab. “It’s moving out of Queen
around a single, narrowly defined goal. (Sometimes you have Square,” Moran says. But it’s still not nearly as
to stop playing Breakout to go to the bathroom, put out a fire, common as reinforcement learning, which even
or talk to your boss.) And most environments aren’t as stable undergraduates learn. “You don’t teach under-
and rule-bound as a game is. The conceit behind neural net- graduates the free energy principle—yet.”
works is that they are supposed to think the way we do; but The first time I asked Friston about the con-
reinforcement learning doesn’t really get us there. nection between the free energy principle and
To Friston and his enthusiasts, this failure makes com- artificial intelligence, he predicted that within
plete sense. After all, according to the free energy principle, five to 10 years, most machine learning would
the fundamental drive of human thought isn’t to seek some incorporate free energy minimization. The sec-
arbitrary external reward. It’s to minimize prediction error. ond time, his response was droll. “Think about
Clearly, neural networks ought to do the same. It helps that why it’s called active inference,” he said. His
the Bayesian formulas behind the free energy principle— straight, sparkly white teeth showed through his
the ones that are so diicult to translate into English—are smile as he waited for me to follow his wordplay.
already written in the native language of machine learning. “Well, it’s AI,” Friston said. “So is active infer-
Julie Pitt, head of machine-learning infrastructure at Netflix, ence the new AI? Yes, it’s the acronym.” Not for
discovered Friston and the free energy principle in 2014, and it the first time, a Fristonian joke had passed me by.
transformed her thinking. (Pitt’s Twitter bio reads, “I infer my
own actions by way of Active Inference.”) Outside of her work WHILE I WAS IN LONDON, Friston gave a
at Netflix, she’s been exploring applications of the principle in talk at a quantitative trading firm. About 60
a side project called Order of Magnitude Labs. Pitt says that baby-faced stock traders were in attendance,
the beauty of the free energy model is that it allows an artifi- rounding out the end of their workday. Friston
cial agent to act in any environment, even one that’s new and described how the free energy principle could
unknown. Under the old reinforcement-learning model, you’d model curiosity in artificial agents. About 15
have to keep stipulating new rules and sub-rewards to get your minutes in, he asked his listeners to raise a
agent to cope with a complex world. But a free energy agent hand if they understood what he was saying.
always generates its own intrinsic reward: the minimization He counted only three hands, so he reversed the
of surprise. And that reward, Pitt says, includes an imperative question: “Can you put your hand up if that was
to go out and explore. complete nonsense and you don’t know what I
In late 2017, a group led by Rosalyn Moran, a neuroscientist was talking about?” This time, a lot of people
and engineer at King’s College London, pitted two AI players raised their hands, and I got the feeling that
against one another in a version of the 3D shooter game Doom. the rest were being polite. With 45 minutes
The goal was to compare an agent driven by active inference left, Friston turned to the organizer of the talk
to one driven by reward-maximization. and looked at him as if to say, What the hell?
The reward-based agent’s goal was to kill a monster inside The manager stammered a bit before saying,
the game, but the free-energy-driven agent only had to min- “Everybody here’s smart.” Friston graciously
imize surprise. The Fristonian agent started of slowly. But agreed and finished his presentation.
eventually it started to behave as if it had a model of the game, The next morning, I asked Friston if he thought
seeming to realize, for instance, that when the agent moved the talk went well, considering that few of those

1 0 4
 bright young minds seemed to understand him. “There is going Guide to the Galaxy; it’s the town where “a girl
to be a substantial proportion of the audience who—it’s just sitting on her own in a small café” suddenly dis-
not for them,” he said. “Sometimes they get upset because covers the secret to making the world “a good
they’ve heard that it’s important and they don’t understand and happy place.” But fate intervenes. “Before
it. They think they have to think it’s rubbish and they leave. she could get to a phone to tell anyone about it,
You get used to that.” a terrible stupid catastrophe occurred, and the
In 2010, Peter Freed, a psychiatrist at Columbia Univer- idea was lost forever.”
sity, gathered together 15 brain researchers to discuss one of It’s unclear whether the free energy princi-
Friston’s papers. Freed described what happened in the jour- ple is the secret to making the world a good and
nal Neuropsychoanalysis: “There was a lot of mathematical happy place, as some of its believers almost seem
knowledge in the room: three statisticians, two physicists, to think it might be. Friston himself tended to
a physical chemist, a nuclear physicist, and a large group of take a more measured tone as our talks went on,
neuroimagers—but apparently we didn’t have what it took. suggesting only that active inference and its cor-
I met with a Princeton physicist, a Stanford neurophysiol- ollaries were quite promising. Several times he
ogist, a Cold Springs Harbor neurobiologist to discuss the conceded that he might just be “talking rubbish.”
paper. Again blanks, one and all: too many equations, too During the last group meeting I attended at the
many assumptions, too many moving parts, too global a the- FIL, he told those in attendance that the free
ory, no opportunity for questions—and so people gave up.” energy principle is an “as if” concept—it does
But for all the people who are exasperated by Friston’s not require that biological things minimize free
impenetrability, there are nearly as many who feel he has energy in order to exist; it is merely suicient
unlocked something huge, an idea every bit as expansive as as an explanation for biotic self-organization.
11. In Foundation, Darwin’s theory of natural selection. When the Canadian Friston’s mother died a few years ago, but
published in 1951, one
of Asimov’s characters
philosopher Maxwell Ramstead first read Friston’s work lately he has been thinking back to her frequent
defines psychohistory in 2014, he had already been trying to find ways to connect reassurances during his childhood: You’re very
as “that branch of
mathematics which complex living systems that exist at diferent scales—from intelligent, Karl. “I never quite believed her,”
deals with the reactions cells to brains to individuals to cultures. In 2016 he met Fris- he says. “And yet now I have found myself sud-
of human conglomer-
ates to fixed social ton, who told him that the same math that applies to cellular denly being seduced by her argument. Now I do
and economic stimuli.”
diferentiation—the process by which generic cells become believe I’m actually quite bright.” But this new-
12. On a recent Saturday, more specialized—can also be applied to cultural dynamics. found self-esteem, he says, has also led him to
a man came to the door
asking if Friston’s wife “This was a life-changing conversation for me,” Ramstead examine his own egocentricity.
was home. When Fris- says. “I almost had a nosebleed.” Friston says his work has two primary moti-
ton said yes, the man
said, “Good, because I “This is absolutely novel in history,” Ramstead told me as vations. Sure, it would be nice to see the free
got a dead cat here.”
He wanted it stuffed.
we sat on a bench in Queen Square, surrounded by patients energy principle lead to true artificial conscious-
and staf from the surrounding hospitals. Before Friston came ness someday, he says, but that’s not one of his
along, “We were kind of condemned to forever wander in this top priorities. Rather, his first big desire is to
multidisciplinary space without a common currency,” he con- advance schizophrenia research, to help repair
tinued. “The free energy principle gives you that currency.” the brains of patients like the ones he knew at the
In 2017, Ramstead and Friston coauthored a paper, with old asylum. And his second main motivation, he
Paul Badcock of the University of Melbourne, in which they says, is “much more selfish.” It goes back to that
described all life in terms of Markov blankets. Just as a cell evening in his bedroom, as a teenager, looking
is a Markov-blanketed system that minimizes free energy in at the cherry blossoms, wondering, “Can I sort
order to exist, so are tribes and religions and species. it all out in the simplest way possible?”
After the publication of Ramstead’s paper, Micah Allen, a “And that is a very self-indulgent thing. It has
cognitive neuroscientist then at the FIL, wrote that the free no altruistic clinical compassion behind it. It
energy principle had evolved into a real-life version of Isaac is just the selfish desire to try and understand
Asimov’s psychohistory,11 a fictional system that reduced all of things as completely and as rigorously and as
psychology, history, and physics down to a statistical science. simply as possible,” he says. “I often reflect on the
And it’s true that the free energy principle does seem to have jokes that people make about me—sometimes
expanded to the point of being, if not a theory of everything, maliciously, sometimes very amusingly—that I
then nearly so. (Friston told me that cancer and tumors might can’t communicate. And I think: I didn’t write it
be instances of false inference, when cells become deluded.) for you. I wrote it for me.”
As Allen asked: Does a theory that explains everything run Friston told me he occasionally misses the
the risk of explaining nothing? last train home to Rickmansworth, lost in one of
On the last day of my trip, I visited Friston in the town of those problems that he drills into for weeks. So
Rickmansworth, where he lives in a house filled with taxider- he’ll sleep in his oice, curled on the futon under
mied animals 12 that his wife prepares as a hobby. As it happens, his Markov blanket, safe and securely separated
Rickmansworth appears on the first page of The Hitchhiker’s from the external world.
Statement Required by 39 U.S.C. 3685 showing the Ownership, Management and Circulation of WIRED, published monthly
COLOPHON
(12 issues) for October 1, 2018. Publication No. 011-724. Annual subscription price $24.00. DEEP LEARNINGS THAT
HELPED GET THIS ISSUE OUT:
1. Location of known oice of Publication is One World Trade Center, New York, NY 10007.
How to unclog the bathroom sink, YouTube
2. Location of the Headquarters or General Business Oices of the Publisher is One World Trade Center, New York, NY 10007. edition; fumbling through the third section
of Scott Joplin’s “Maple Leaf Rag”; IPAs and
empty stomachs don’t mix; reacquainting
3. The names and addresses of the Publisher, Editor and Managing Editor are: Publisher, Kim Kelleher, One World Trade Cen- myself with the intricate SFUSD assign-
ter, New York, NY 10007. Editor, Nicholas Thompson, 520 Third Street, 3rd Floor, San Francisco, CA 94107. Managing Editor, ment system; realizing our phones now
Erica Jewell, 520 Third Street, 3rd Floor, San Francisco, CA 94107.
know everything about us and being OK
with that; overcoming my primal fear of the
ocean and swimming with sharks; Wired SF
4. The owner is: Advance Magazine Publishers Inc., published through its Condé Nast division, One World Trade Center, New food > Wired NY food; replacing the battery
York, NY 10007. Stockholder: Directly or indirectly through intermediate corporations to the ultimate corporate parent, in a MacBook for less than a third of what
Apple charges; the scary prescience of late-
Advance Publications Inc., 950 Fingerboard Road, Staten Island, NY 10305.
20th-century media theory; deadlines are
just unreasonable expectations we place
5. Known bondholders, mortgagees and other security holders owning or holding 1 percent or more of total amount of bonds, on ourselves and each other; Earth is but
mortgages or other securities are: None. a mote of dust suspended in a sunbeam;
regular worship of the Godhead at WOTF.

6. Extent and nature of circulation WIRED is a registered trademark of

Advance Magazine Publishers Inc. Copy-
Average no. copies Single issue
right ©2018 Condé Nast. All rights
reserved. Printed in the USA. Volume 26,
each issue during nearest to No. 12. WIRED (ISSN 1059–1028) is pub-
preceding 12 months filing date lished monthly by Condé Nast, which is a
division of Advance Magazine Publishers
Inc. Editorial office: 520 Third Street, Ste.
a.*Total No. Copies 889,018 870,344
305, San Francisco, CA 94107-1815. Prin-
b.*Paid Circulation cipal office: Condé Nast, 1 World Trade
(1) Mailed Outside-County Paid 620,846 618,845 Center, New York, NY 10007. Robert A.
Subscriptions Stated on Sauerberg, Jr., President and Chief Execu-
tive Officer; David E. Geithner, Chief Finan-
PS Form 3541 cial Officer; Pamela Drucker Mann, Chief
(2) Mailed In-County Paid 0 0 Revenue & Marketing Officer. Periodicals
Subscriptions Stated on postage paid at New York, NY, and at addi-
PS Form 3541
tional mailing offices. Canada Post Pub-
lications Mail Agreement No. 40644503.
(3) Paid Distribution Outside the 28,921 25,190 Canadian Goods and Services Tax Regis-
Mails Including Sales Through tration No. 123242885 RT0001.
Dealers and Carriers, Street
POSTMASTER: Send all UAA to CFS (see
Vendors, Counter Sales, and
DMM 707.4.12.5); NONPOSTAL AND MIL-
Other Paid Distribution ITARY FACILITIES: Send address correc-
Outside USPS® tions to WIRED , PO Box 37617, Boone, IA
(4) Paid Distribution by Other 0 0 50037-0662. For subscriptions, address
changes, adjustments, or back issue
Classes of Mail Through inquiries: Please write to WIRED , PO Box
the USPS 37617, Boone, IA 50037-0662, call (800)
c.*Total Paid Distribution 649,767 644,035 769 4733, or email subscriptions@ WIRED
d.*Free or Nominal Rate Distribution .com. Please give both new and old
addresses as printed on most recent
(1) Free or Nominal Rate 158,517 153,544 label. First copy of new subscription will
Outside-County Copies be mailed within eight weeks after receipt
included on PS Form 3541 of order. Address all editorial, business,
and production correspondence to WIRED
(2) Free or Nominal Rate 0 0
Magazine, 1 World Trade Center, New
In-County Copies York, NY 10007. For permissions and
included on PS Form 3541 reprint requests, please call (212) 630
(3) Free or Nominal Rate Copies 0 0 5656 or fax requests to (212) 630 5883.
Visit us online at www.WIRED .com. To sub-
Mailed at Other Classes
scribe to other Condé Nast magazines on
Through the USPS the web, visit www.condenet.com. Occa-
(4) Free or Nominal Rate 21,002 19,503 sionally, we make our subscriber list avail-
Distribution Outside the Mail able to carefully screened companies
that ofer products and services that we
e. Total Free or Nominal Rate 179,519 173,047 believe would interest our readers. If you
Distribution do not want to receive these ofers and/or
f. Total Distribution 829,286 817,082 information, please advise us at PO Box
37617, Boone, IA 50037-0662, or call
g. Copies not Distributed 59,733 53,262
(800) 769 4733.
h. Total 889,018 870,344
i. Percent Paid 78.35% 78.82% W I R E D is not responsible for the return

j. Paid Electronic Copies 53,843 51,131 or loss of, or for damage or any other
injury to, unsolicited manuscripts,
k. Total Paid Print Copies (line 15c) 703,610 695,166
unsolicited artwork (including, but not
+Paid Electronic Copies limited to, drawings, photographs, and
l. Total Print Distribution (line 15f) 883,129 868,213 transparencies), or any other unsolici-
+Paid Electronic Copies ted materials. Those submitting
manuscripts, photographs, artwork,
m. Percent Paid (Both Print & 79.67% 80.07% or other materials for consideration
Electronic Copies) should not send originals, unless spe-
cifically requested to do so by W I R E D
in writing. Manuscripts, photographs,
7. I certify that all information furnished
artwork, and other materials sub-
on this form is true and complete. mitted must be accompanied by a
(Signed) David Geithner, Chief Financial Oicer self-addressed, stamped envelope.
 #WIREDBACKPAGE

SIX BY SIX: STORIES BY WIRED READERS

Each month, we publish a six-word story—and it could be written by you. Submit your six words on
Twitter, Facebook, or Instagram, along with #WIREDBACKPAGE. We’ll pick one story to illustrate
here. Your next assignment: In six words, write a courtroom drama set in the year 2058.

IMAGINE LIFE UNDER AN AI PRESIDENT:

FIRST
TERM
STRUGGLES,
SECOND
TERM
PERFECTION.
@RAMSAYTAPLIN, VIA TWITTER

HONORABLE MENTIONS: RAN ON A VERY PROGRESSIVE PLATFORM. (@ABGL ASSMAN, VIA INSTAGRAM) // ANALYSIS COMPLETE: PRESIDENT REQUIRES HARD RESET.
(@MICHAEL_LOUIS, VIA INSTAGRAM) // BUILD A FIREWALL; HUMANS CODE IT. (DANIEL GABRIEL SECARA, VIA FACEBOOK) // THE PRESIDENT IS MADE IN CHINA. (@LASSELYHN,
VIA INSTAGRAM) // THE WORLD’S UTOPIAN DREAM: UNBIASED POLITICS. (AD NAVARRO, VIA FACEBOOK) // I WON’T NEED THE OVAL OFFICE. (@RODOLPHORR, VIA INSTAGRAM)

1 0 8 YEYEI GÓMEZ DEC 2018

 SRPB43 | www.SeikoUSA.com

Congratulations Wired

SHOP THE COLLECTION AT SEIKOUSA.COM