How much do we need humans in space? How much do we want them there? Astronauts embody the triumph of human imagination and engineering. Their efforts shed light on the possibilities and problems posed by travel beyond our nurturing Earth. Their presence on the moon or on other solar-system objects can imply that the countries or entities that sent them there possess ownership rights. Astronauts promote an understanding of the cosmos, and inspire young people toward careers in science.
When it comes to exploration, however, our robots can outperform astronauts at a far lower cost and without risk to human life. This assertion, once a prediction for the future, has become reality today, and robot explorers will continue to become ever more capable, while human bodies will not.
Fifty years ago, when the first geologist to reach the moon suddenly recognized strange orange soil (the likely remnant of previously unsuspected volcanic activity), no one claimed that an automated explorer could have accomplished this feat. Today, we have placed a semi-autonomous rover on Mars, one of a continuing suite of orbiters and landers, with cameras and other instruments that probe the Martian soil, capable of finding paths around obstacles as no previous rover could.
Since Apollo 17 left the moon in 1972, the astronauts have journeyed no farther than low Earth orbit. In this realm, astronauts’ greatest achievement by far came with their five repair missions to the Hubble Space Telescope, which first saved the giant instrument from uselessness and then extended its life by decades by providing upgraded cameras and other systems. (Astronauts could reach the Hubble only because the Space Shuttle, which launched it, could go no farther from Earth, which produces all sorts of interfering radiation and light.) Each of these missions cost about a billion dollars in today’s money. The cost of a telescope to replace the Hubble would likewise have been about a billion dollars; one estimate has set the cost of the five repair missions equal to that for constructing seven replacement telescopes.
Today, astrophysicists have managed to send all of their new spaceborne observatories to distances four times farther than the moon, where the James Webb Space Telescope now prepares to study a host of cosmic objects. Our robot explorers have visited all the sun’s planets (including that former planet Pluto), as well as two comets and an asteroid, securing immense amounts of data about them and their moons, most notably Jupiter’s Europa and Saturn’s Enceladus, where oceans that lie beneath an icy crust may harbor strange forms of life. Future missions from the United States, the European Space Agency, China, Japan, India, and Russia will only increase our robot emissaries’ abilities and the scientific importance of their discoveries. Each of these missions has cost far less than a single voyage that would send humans—which in any case remains an impossibility for the next few decades, for any destination save the moon and Mars.
In 2020, NASA revealed of accomplishments titled “20 Breakthroughs From 20 Years of Science Aboard the International Space Station.” Seventeen of those dealt with processes that robots could have performed, such as launching small satellites, the detection of cosmic particles, employing microgravity conditions for drug development and the study of flames, and 3-D printing in space. The remaining three dealt with muscle atrophy and bone loss, growing food, or identifying microbes in space—things that are important for humans in that environment, but hardly a rationale for sending them there.
The inside of a tokamak—the donut-shaped vessel designed to contain a nuclear fusion reaction—presents a special kind of chaos. Hydrogen atoms are smashed together at unfathomably high temperatures, creating a whirling, roiling plasma that’s hotter than the surface of the sun. Finding smart ways to control and confine that plasma will be key to unlocking the potential of nuclear fusion, which has been mooted as the clean energy source of the future for decades. At this point, the science underlying fusion seems sound, so what remains is an engineering challenge. “We need to be able to heat this matter up and hold it together for long enough for us to take energy out of it,” says Ambrogio Fasoli, director of the Swiss Plasma Center at École Polytechnique Fédérale de Lausanne.
That’s where DeepMind comes in. The artificial intelligence firm, backed by Google parent company Alphabet, has previously turned its hand to video games and protein folding, and has been working on a joint research project with the Swiss Plasma Center to develop an AI for controlling a nuclear fusion reaction.
In stars, which are also powered by fusion, the sheer gravitational mass is enough to pull hydrogen atoms together and overcome their opposing charges. On Earth, scientists instead use powerful magnetic coils to confine the nuclear fusion reaction, nudging it into the desired position and shaping it like a potter manipulating clay on a wheel. The coils have to be carefully controlled to prevent the plasma from touching the sides of the vessel: this can damage the walls and slow down the fusion reaction. (There’s little risk of an explosion as the fusion reaction cannot survive without magnetic confinement).
But every time researchers want to change the configuration of the plasma and try out different shapes that may yield more power or a cleaner plasma, it necessitates a huge amount of engineering and design work. Conventional systems are computer-controlled and based on models and careful simulations, but they are, Fasoli says, “complex and not always necessarily optimized.”
DeepMind has developed an AI that can control the plasma autonomously. A paper published in the journal Nature describes how researchers from the two groups taught a deep reinforcement learning system to control the 19 magnetic coils inside TCV, the variable-configuration tokamak at the Swiss Plasma Center, which is used to carry out research that will inform the design of bigger fusion reactors in future. “AI, and specifically reinforcement learning, is particularly well suited to the complex problems presented by controlling plasma in a tokamak,” says Martin Riedmiller, control team lead at DeepMind.
The neural network—a type of AI setup designed to mimic the architecture of the human brain—was initially trained in a simulation. It started by observing how changing the settings on each of the 19 coils affected the shape of the plasma inside the vessel. Then it was given different shapes to try to recreate in the plasma. These included a D-shaped cross-section close to what will be used inside ITER (formerly the International Thermonuclear Experimental Reactor), the large-scale experimental tokamak under construction in France, and a snowflake configuration that could help dissipate the intense heat of the reaction more evenly around the vessel.
DeepMind’s AI was able to autonomously figure out how to create these shapes by manipulating the magnetic coils in the right way—both in the simulation, and when the scientists ran the same experiments for real inside the TCV tokamak to validate the simulation. It represents a “significant step,” says Fasoli, one that could influence the design of future tokamaks or even speed up the path to viable fusion reactors. “It’s a very positive result,” says Yasmin Andrew, a fusion specialist at Imperial College London, who was not involved in the research. “It will be interesting to see if they can transfer the technology to a larger tokamak.”
Fusion offered a particular challenge to DeepMind’s scientists because the process is both complex and continuous. Unlike a turn-based game like Go, which the company has famously conquered with its AlphaGo AI, the state of a plasma constantly changes. And to make things even harder, it can’t be continuously measured. It is what AI researchers call an “under–observed system.”
“Sometimes algorithms which are good at these discrete problems struggle with such continuous problems,” says Jonas Buchli, a research scientist at DeepMind. “This was a really big step forward for our algorithm because we could show that this is doable. And we think this is definitely a very, very complex problem to be solved. It is a different kind of complexity than what you have in games.”
This incessant surveillance is antidemocratic, and it’s also a loser’s game. The price of accurate intel increases asymptotically; there’s no way to know everything about natural systems, forcing guesses and assumptions; and just when a complete picture is starting to coalesce, some new player intrudes and changes the situational dynamic. Then the AI breaks. The near-perfect intelligence veers into psychosis, labeling dogs as pineapples, treating innocents as wanted fugitives, and barreling eighteen-wheelers into kindergarten busses that it sees as highway overpasses.
The dangerous fragility inherent to optimization is why the human brain did not, itself, evolve to be an optimizer. The human brain is data-light: It draws hypotheses from a few data points. And it never strives for 100 percent accuracy. It’s content to muck along at the threshold of functionality. If it can survive by being right 1 percent of the time, that’s all the accuracy it needs.
The brain’s strategy of minimal viability is a notorious source of cognitive biases that can have damaging consequences: close-mindedness, conclusion jumping, recklessness, fatalism, panic. Which is why AI’s rigorously data-driven method can help illuminate our blindspots and debunk our prejudices. But in counterbalancing our brain’s computational shortcomings, we don’t want to stray into the greater problem of overcorrection. There can be enormous practical upside to a good enough mentality: It wards off perfectionism’s destructive mental effects, including stress, worry, intolerance, envy, dissatisfaction, exhaustion, and self-judgment. A less-neurotic brain has helped our species thrive in life’s punch and wobble, which demands workable plans that can be flexed, via feedback, on the fly.
These antifragile neural benefits can all be translated into AI. Instead of pursuing faster machine-learners that crunch ever-vaster piles of data, we can focus on making AI more tolerant of bad information, user variance, and environmental turmoil. That AI would exchange near-perfection for consistent adequacy, upping reliability and operational range while sacrificing nothing essential. It would suck less energy, haywire less randomly, and place less psychological burdens on its mortal users. It would, in short, possess more of the earthly virtue known as common sense.
Here’s three specs for how.
Building AI to Brave Ambiguity
Five hundred years ago, Niccolò Machiavelli, the guru of practicality, pointed out that worldly success requires a counterintuitive kind of courage: the heart to venture beyond what we know with certainty. Life, after all, is too fickle to permit total knowledge, and the more that we obsess over ideal answers, the more that we hamper ourselves with lost initiative. So, the smarter strategy is to concentrate on intel that can be rapidly acquired—and to advance boldly in the absence of the rest. Much of that absent knowledge will prove unnecessary, anyway; life will bend in a different direction than we anticipate, resolving our ignorance by rendering it irrelevant.
We can teach AI to operate this same way by flipping our current approach to ambiguity. Right now, when a Natural Language Processor encounters a word—suit—that could denote multiple things—an article of clothing or a legal action—it devotes itself to analyzing ever greater chunks of correlated information in an effort to pinpoint the word’s exact meaning.
Last month, Stanford researchers declared that a new era of artificial intelligence had arrived, one built atop colossal neural networks and oceans of data. They said a new research center at Stanford would build—and study—these “foundational models” of AI.
Critics of the idea surfaced quickly—including at the workshop organized to mark the launch of the new center. Some object to the limited capabilities and sometimes freakish behavior of these models; others warn of focusing too heavily on one way of making machines smarter.
“I think the term ‘foundation’ is horribly wrong,” Jitendra Malik, a professor at UC Berkeley who studies AI, told workshop attendees in a video discussion.
Malik acknowledged that one type of model identified by the Stanford researchers—large language models that can answer questions or generate text from a prompt—has great practical use. But he said evolutionary biology suggests that language builds on other aspects of intelligence like interaction with the physical world.
“These models are really castles in the air; they have no foundation whatsoever,” Malik said. “The language we have in these models is not grounded, there is this fakeness, there is no real understanding.” He declined an interview request.
A research paper coauthored by dozens of Stanford researchers describes “an emerging paradigm for building artificial intelligence systems” that it labeled “foundational models.” Ever-larger AI models have produced some impressive advances in AI in recent years, in areas such as perception and robotics as well as language.
Large language models are also foundational to big tech companies like Google and Facebook, which use them in areas like search, advertising, and content moderation. Building and training large language models can require millions of dollars worth of cloud computing power; so far, that’s limited their development and use to a handful of well-heeled tech companies.
But big models are problematic, too. Language models inherit bias and offensive text from the data they are trained on, and they have zero grasp of common sense or what is true or false. Given a prompt, a large language model may spit out unpleasant language or misinformation. There is also no guarantee that these large models will continue to produce advances in machine intelligence.
The Stanford proposal has divided the research community. “Calling them ‘foundation models’ completely messes up the discourse,” says Subbarao Kambhampati, a professor at Arizona State University. There is no clear path from these models to more general forms of AI, Kambhampati says.
Thomas Dietterich, a professor at Oregon State University and former president of the Association for the Advancement of Artificial Intelligence, says he has “huge respect” for the researchers behind the new Stanford center, and he believes they are genuinely concerned about the problems these models raise.
But Dietterich wonders if the idea of foundational models isn’t partly about getting funding for the resources needed to build and work on them. “I was surprised that they gave these models a fancy name and created a center,” he says. “That does smack of flag planting, which could have several benefits on the fundraising side.”
Stanford has also proposed the creation of a National AI Cloud to make industry-scale computing resources available to academics working on AI research projects.
Emily M. Bender, a professor in the linguistics department at the University of Washington, says she worries that the idea of foundational models reflects a bias toward investing in the data-centric approach to AI favored by industry.
Bender says it is especially important to study the risks posed by big AI models. She coauthored a paper, published in March, that drew attention to problems with large language models and contributed to the departure of two Google researchers. But she says scrutiny should come from multiple disciplines.
“There are all of these other adjacent, really important fields that are just starved for funding,” she says. “Before we throw money into the cloud, I would like to see money going into other disciplines.”
In recent years, researchers have used artificial intelligence to improve translation between programming languages or automatically fix problems. The AI system DrRepair, for example, has been shown to solve most issues that spawn error messages. But some researchers dream of the day when AI can write programs based on simple descriptions from non-experts.
On Tuesday, Microsoft and OpenAI shared plans to bring GPT-3, one of the world’s most advanced models for generating text, to programming based on natural language descriptions. This is the first commercial application of GPT-3 undertaken since Microsoft invested $1 billion in OpenAI last year and gained exclusive licensing rights to GPT-3.
“If you can describe what you want to do in natural language, GPT-3 will generate a list of the most relevant formulas for you to choose from,” said Microsoft CEO Satya Nadella in a keynote address at the company’s Build developer conference. “The code writes itself.”
Microsoft VP Charles Lamanna told WIRED the sophistication offered by GPT-3 can help people tackle complex challenges and empower people with little coding experience. GPT-3 will translate natural language into PowerFx, a fairly simple programming language similar to Excel commands that Microsoft introduced in March.
This is the latest demonstration of applying AI to coding. Last year at Microsoft’s Build, OpenAI CEO Sam Altman demoed a language model fine-tuned with code from GitHub that automatically generates lines of Python code. As WIRED detailed last month, startups like SourceAI are also using GPT-3 to generate code. IBM last month showed how its Project CodeNet, with 14 million code samples from more than 50 programming languages, could reduce the time needed to update a program with millions of lines of Java code for an automotive company from one year to one month.
Microsoft’s new feature is based on a neural network architecture known as Transformer, used by big tech companies including Baidu, Google, Microsoft, Nvidia, and Salesforce to create large language models using text training data scraped from the web. These language models continually grow larger. The largest version of Google’s BERT, a language model released in 2018, had 340 million parameters, a building block of neural networks. GPT-3, which was released one year ago, has 175 billion parameters.
Such efforts have a long way to go, however. In one recent test, the best model succeeded only 14 percent of the time on introductory programming challenges compiled by a group of AI researchers.