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 doughnut-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 in Switzerland.
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 the 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 re-create 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.”
Tech conglomerate Tencent caused a stir last year with the announcement that it would comply with China’s directive to incorporate facial recognition technology into its games in the country. The move was in line with China’s strict gaming regulation policies, which impose limits on how much time minors can spend playing video games—an effort to curb addictive behavior, since gaming is labeled by the state as “spiritual opium.”
The state’s use of biometric data to police its population is, of course, invasive, and especially undermines the privacy of underage users—but Tencent is not the only video game company to track its players, nor is this recent case an altogether new phenomenon. All over the world, video games, one of the most widely adopted digital media forms, are installing networks of surveillance and control.
In basic terms, video games are systems that translate physical inputs—such as hand movement or gesture—into various electric or electronic machine-readable outputs. The user, by acting in ways that comply with the rules of the game and the specifications of the hardware, is parsed as data by the video game. Writing almost a decade ago, the sociologists Jennifer R. Whitson and Bart Simon argued that games are increasingly understood as systems that easily allow the reduction of human action into knowable and predictable formats.
Video games, then, are a natural medium for tracking, and researchers have long argued that large data sets about players’ in-game activities are a rich resource in understanding player psychology and cognition. In one study from 2012, Nick Yee, Nicolas Ducheneaut, and Les Nelson scraped player activity data logged on the World of Warcraft Armory website—essentially a database that records all the things a player’s character has done in the game (how many of a certain monster I’ve killed, how many times I’ve died, how many fish I’ve caught, and so on).
The researchers used this data to infer personality characteristics (in combination with data yielded through a survey). The paper suggests, for example, that there is a correlation between the survey respondents classified as more conscientious in their game-playing approach and the tendency to spend more time doing repetitive and dull in-game tasks, such as fishing. Conversely, those whose characters more often fell to death from high places were less conscientious, according to their survey responses.
Correlation between personality and quantitative gameplay data is certainly not unproblematic. The relationship between personality and identity and video game activity is complex and idiosyncratic; for instance, research suggests that gamer identity intersects with gender, racial, and sexual identity. Additionally, there has been general pushback against claims of Big Data’s production of new knowledge rooted in correlation. Despite this, games companies increasingly realize the value of big data sets to gain insight into what a player likes, how they play, what they play, what they’ll likely spend money on (in freemium games), how and when to offer the right content, and how to solicit the right kinds of player feelings.
While there are no numbers on how many video game companies are surveilling their players in-game (although, as a recent article suggests, large publishers and developers like Epic, EA, and Activision explicitly state they capture user data in their license agreements), a new industry of firms selling middleware “data analytics” tools, often used by game developers, has sprung up. These data analytics tools promise to make users more amenable to continued consumption through the use of data analysis at scale. Such analytics, once available only to the largest video game studios—which could hire data scientists to capture, clean, and analyze the data, and software engineers to develop in-house analytics tools—are now commonplace across the entire industry, pitched as “accessible” tools that provide a competitive edge in a crowded marketplace by companies like Unity, GameAnalytics, or Amazon Web Services. (Although, as a recent study shows, the extent to which these tools are truly “accessible” is questionable, requiring technical expertise and time to implement.) As demand for data-driven insight has grown, so have the range of different services—dozens of tools in the past several years alone, providing game developers with different forms of insight. One tool—essentially Uber for playtesting—allows companies to outsource quality assurance testing, and provides data-driven insight into the results. Another supposedly uses AI to understand player value and maximize retention (and spending, with a focus on high-spenders).
Delphi taps the fruits of recent advances in AI and language. Feeding very large amounts of text to algorithms that use mathematically simulated neural networks has yielded surprising advances.
In June 2020, researchers at OpenAI, a company working on cutting-edge AI tools, demonstrated a program called GPT-3 that can predict, summarize, and auto-generate text with what often seems like remarkable skill, although it will also spit out biased and hateful language learned from text it has read.
The researchers behind Delphi also asked ethical questions of GPT-3. They found its answers agreed with those of the crowd workers just over 50 percent of the time—little better than a coin flip.
Improving the performance of a system like Delphi will require different AI approaches, potentially including some that allow a machine to explain its reasoning and indicate when it is conflicted.
The idea of giving machines a moral code stretches back decades both in academic research and science fiction. Isaac Asimov’s famous Three Laws of Robotics popularized the idea that machines might follow human ethics, although the short stories that explored the idea highlighted contradictions in such simplistic reasoning.
Choi says Delphi should not be taken as providing a definitive answer to any ethical questions. A more sophisticated version might flag uncertainty, because of divergent opinions in its training data. “Life is full of gray areas,” she says. “No two human beings will completely agree, and there’s no way an AI program can match people’s judgments.”
Other machine learning systems have displayed their own moral blind spots. In 2016, Microsoft released a chatbot called Tay designed to learn from online conversations. The program was quickly sabotaged and taught to say offensive and hateful things.
Efforts to explore ethical perspectives related to AI have also revealed the complexity of such a task. A project launched in 2018 by researchers at MIT and elsewhere sought to explore the public’s view of ethical conundrums that might be faced by self-driving cars. They asked people to decide, for example, whether it would be better for a vehicle to hit an elderly person, a child, or a robber. The project revealed differing opinions across different countries and social groups. Those from the US and Western Europe were more likely than respondents elsewhere to spare the child over an older person.
Some of those building AI tools are keen to engage with the ethical challenges. “I think people are right to point out the flaws and failures of the model,” says Nick Frost, CEO of Cohere, a startup that has developed a large language model that is accessible to others via an API. “They are informative of broader, wider problems.”
Cohere devised ways to guide the output of its algorithms, which are now being tested by some businesses. It curates the content that is fed to the algorithm and trains the algorithm to learn to catch instances of bias or hateful language.
Frost says the debate around Delphi reflects a broader question that the tech industry is wrestling with—how to build technology responsibly. Too often, he says, when it comes to content moderation, misinformation, and algorithmic bias, companies try to wash their hands of the problem by arguing that all technology can be used for good and bad.
When it comes to ethics, “there’s no ground truth, and sometimes tech companies abdicate responsibility because there’s no ground truth,” Frost says. “The better approach is to try.”
There’s an old joke that physicists like to tell: Everything has already been discovered and reported in a Russian journal in the 1960s, we just don’t know about it. Though hyperbolic, the joke accurately captures the current state of affairs. The volume of knowledge is vast and growing quickly: The number of scientific articles posted on arXiv (the largest and most popular preprint server) in 2021 is expected to reach 190,000—and that’s just a subset of the scientific literature produced this year.
It’s clear that we do not really know what we know, because nobody can read the entire literature even in their own narrow field (which includes, in addition to journal articles, PhD theses, lab notes, slides, white papers, technical notes, and reports). Indeed, it’s entirely possible that in this mountain of papers, answers to many questions lie hidden, important discoveries have been overlooked or forgotten, and connections remain concealed.
Artificial intelligence is one potential solution. Algorithms can already analyze text without human supervision to find relations between words that help uncover knowledge. But far more can be achieved if we move away from writing traditional scientific articles whose style and structure has hardly changed in the past hundred years.
Text mining comes with a number of limitations, including access to the full text of papers and legal concerns. But most importantly, AI does not really understand concepts and the relationships between them, and is sensitive to biases in the data set, like the selection of papers it analyzes. It is hard for AI—and, in fact, even for a nonexpert human reader—to understand scientific papers in part because the use of jargon varies from one discipline to another and the same term might be used with completely different meanings in different fields. The increasing interdisciplinarity of research means that it is often difficult to define a topic precisely using a combination of keywords in order to discover all the relevant papers. Making connections and (re)discovering similar concepts is hard even for the brightest minds.
As long as this is the case, AI cannot be trusted and humans will need to double-check everything an AI outputs after text-mining, a tedious task that defies the very purpose of using AI. To solve this problem we need to make science papers not only machine-readable but machine-understandable, by (re)writing them in a special type of programming language. In other words: Teach science to machines in the language they understand.
Writing scientific knowledge in a programming-like language will be dry, but it will be sustainable, because new concepts will be directly added to the library of science that machines understand. Plus, as machines are taught more scientific facts, they will be able to help scientists streamline their logical arguments; spot errors, inconsistencies, plagiarism, and duplications; and highlight connections. AI with an understanding of physical laws is more powerful than AI trained on data alone, so science-savvy machines will be able to help future discoveries. Machines with a great knowledge of science could assist rather than replace human scientists.
Mathematicians have already started this process of translation. They are teaching mathematics to computers by writing theorems and proofs in languages like Lean. Lean is a proof assistant and programming language in which one can introduce mathematical concepts in the form of objects. Using the known objects, Lean can reason whether a statement is true or false, hence helping mathematicians verify proofs and identify places where their logic is insufficiently rigorous. The more mathematics Lean knows, the more it can do. The Xena Project at Imperial College London is aiming to input the entire undergraduate mathematics curriculum in Lean. One day, proof assistants may help mathematicians do research by checking their reasoning and searching the vast mathematics knowledge they possess.