Rise of the machines Computers could achieve superhuman
levels of intelligence in this century. Could they pose a threat to humanity?
By The Week Staff | 8:00am ET
How smart are today's computers?
They can tackle increasingly complex tasks with an almost
human-like intelligence. Microsoft has developed an Xbox game console that can
assess a player's mood by analyzing his or her facial expressions, and in 2011,
IBM's Watson supercomputer won Jeopardy — a quiz show that often requires
contestants to interpret humorous plays on words. These developments have
brought us closer to the holy grail of computer science: artificial
intelligence, or a machine that's capable of thinking for itself, rather than
just respond to commands. But what happens if computers achieve
"superintelligence" — massively outperforming humans not just in
science and math but in artistic creativity and even social skills? Nick
Bostrom, director of the Future of Humanity Institute at the University of
Oxford, believes we could be sleepwalking into a future in which computers are
no longer obedient tools but a dominant species with no interest in the
survival of the human race. "Once unsafe superintelligence is
developed," Bostrom warned, "we can't put it back in the
bottle."
When will AI become a reality?
There's a 50 percent chance that we'll create a computer
with human-level intelligence by 2050 and a 90 percent chance we will do so by
2075, according to a survey of AI experts carried out by Bostrom. The key to AI
could be the human brain: If a machine can emulate the brain's neural networks,
it might be capable of its own sentient thought. With that in mind, tech giants
like Google are trying to develop their own "brains" — stacks of
coordinated servers running highly advanced software. Meanwhile, Facebook co-founder
Mark Zuckerberg has invested heavily in Vicarious, a San Francisco–based
company that aims to replicate the neocortex, the part of the brain that
governs vision and language and does math. Translate the neocortex into
computer code, and "you have a computer that thinks like a person,"
said Vicarious co-founder Scott Phoenix. "Except it doesn't have to eat or
sleep."
Why is that a threat?
No one knows what will happen when computers become
smarter than their creators. Computer power has doubled every 18 months since
1956, and some AI experts believe that in the next century, computers will
become smart enough to understand their own designs and improve upon them
exponentially. The resulting intelligence gap between machines and people,
Bostrom said, would be akin to the one between humans and insects. Computer
superintelligence could be a boon for the human race, curing diseases like
cancer and AIDS, solving problems that overwhelm humans, and performing work
that would create new wealth and provide more leisure time. But
superintelligence could also be a curse.
What could go wrong?
Computers are designed to solve problems as efficiently
as possible. The difficulty occurs when imperfect humans are factored into
their equations. "Suppose we have an AI whose only goal is to make as many
paper clips as possible," Bostrom said. That thinking machine might
rationally decide that wiping out humanity will help it achieve that goal —
because humans are the only ones who could switch the machine off, thereby
jeopardizing its paper-clip-making mission. In a hyperconnected world,
superintelligent computers would have many ways to kill humans. They could
knock out the internet-connected electricity grid, poison the water supply,
cause havoc at nuclear power plants, or seize command of the military's
remote-controlled drone aircraft or nuclear missiles. Inventor Elon Musk
recently warned that "we need to be super careful with AI,'' calling it
"potentially more dangerous than nukes.''
Is that bleak future inevitable?
Many computer scientists do not think so, and question
whether AI is truly achievable. We're a long way from understanding the
processes of our own incredibly complex brains — including the nature of
consciousness itself — let alone applying that knowledge to produce a sentient,
self-aware machine. And though today's most powerful computers can use
sophisticated algorithms to win chess games and quiz shows, we're still far
short of creating machines with a full set of human skills — ones that could
"write poetry and have a conception of right and wrong," said Ramez
Naam, a lecturer at the Silicon Valley–based Singularity University. That being
said, technology is advancing at lightning speed, and some machines are already
capable of making radical and spontaneous self-improvements. (See below.)
What safeguards are in place?
Not many thus far. Google, for one, has created an AI
ethics review board that supposedly will ensure that new technologies are
developed safely. Some computer scientists are calling for the machines to come
pre-programmed with ethical guidelines — though developers then would face
thorny decisions over what behavior is and isn't "moral." The
fundamental problem, said Danny Hillis, a pioneering supercomputer designer, is
that tech firms are designing ever-more intelligent computers without fully
understanding — or even giving much thought to — the implications of their
inventions. "We're at that point analogous to when single-celled organisms
were turning into multicelled organisms," he said. "We're amoeba, and
we can't figure out what the hell this thing is that we're creating."
When robots learn to lie
In 2009, Swiss researchers carried out a robotic
experiment that produced some unexpected results. Hundreds of robots were
placed in arenas and programmed to look for a "food source," in this
case a light-colored ring. The robots were able to communicate with one another
and were instructed to direct their fellow machines to the food by emitting a
blue light. But as the experiment went on, researchers noticed that the
machines were evolving to become more secretive and deceitful: When they found
food, the robots stopped shining their lights and instead began hoarding the
resources — even though nothing in their original programming commanded them to
do so. The implication is that the machines learned
"self-preservation," said Louis Del Monte, author of The Artificial
Intelligence Revolution. "Whether or not they're conscious is a moot
point."
Comments
Post a Comment