Artificial intelligence is here. It’s overhyped, poorly understood, and flawed but already core to our lives—and it’s only going to extend its reach.
AI powers driverless car research, spots otherwise invisible signs of disease on medical images, finds an answer when you ask Alexa a question, and lets you unlock your phone with your face to talk to friends as an animated poop on the iPhone X using Apple’s Animoji. Those are just a few ways AI already touches our lives, and there’s plenty of work still to be done. But don’t worry, superintelligent algorithms aren’t about to take all the jobs or wipe out humanity.
The current boom in all things AI was catalyzed by breakthroughs in an area known as machine learning. It involves “training” computers to perform tasks based on examples, rather than relying on programming by a human. A technique called deep learning has made this approach much more powerful. Just ask Lee Sedol, holder of 18 international titles at the complex game of Go. He got creamed by software called AlphaGo in 2016.
There’s evidence that AI can make us happier and healthier. But there’s also reason for caution. Incidents in which algorithms picked up or amplified societal biases around race or gender show that an AI-enhanced future won’t automatically be a better one.
The Beginnings of Artificial Intelligence
Artificial intelligence as we know it began as a vacation project. Dartmouth professor John McCarthy coined the term in the summer of 1956, when he invited a small group to spend a few weeks musing on how to make machines do things like use language.
He had high hopes of a breakthrough in the drive toward human-level machines. “We think that a significant advance can be made,” he wrote with his co-organizers, “if a carefully selected group of scientists work on it together for a summer.”
Those hopes were not met, and McCarthy later conceded that he had been overly optimistic. But the workshop helped researchers dreaming of intelligent machines coalesce into a recognized academic field.
Early work often focused on solving fairly abstract problems in math and logic. But it wasn’t long before AI started to show promising results on more human tasks. In the late 1950s, Arthur Samuel created programs that learned to play checkers. In 1962, one scored a win over a master at the game. In 1967, a program called Dendral showed it could replicate the way chemists interpreted mass-spectrometry data on the makeup of chemical samples.
As the field of AI developed, so did different strategies for making smarter machines. Some researchers tried to distill human knowledge into code or come up with rules for specific tasks, like understanding language. Others were inspired by the importance of learning to understand human and animal intelligence. They built systems that could get better at a task over time, perhaps by simulating evolution or by learning from example data. The field hit milestone after milestone as computers mastered tasks that could previously only be completed by people.
Deep learning, the rocket fuel of the current AI boom, is a revival of one of the oldest ideas in AI. The technique involves passing data through webs of math loosely inspired by the working of brain cells that are known as artificial neural networks. As a network processes training data, connections between the parts of the network adjust, building up an ability to interpret future data.
Artificial neural networks became an established idea in AI not long after the Dartmouth workshop. The room-filling Perceptron Mark 1 from 1958, for example, learned to distinguish different geometric shapes and got written up in The New York Times as the “Embryo of Computer Designed to Read and Grow Wiser.” But neural networks tumbled from favor after an influential 1969 book coauthored by MIT’s Marvin Minsky suggested they couldn’t be very powerful.
Not everyone was convinced by the skeptics, however, and some researchers kept the technique alive over the decades. They were vindicated in 2012, when a series of experiments showed that neural networks fueled with large piles of data could give machines new powers of perception. Churning through so much data was difficult using traditional computer chips, but a shift to graphics cards precipitated an explosion in processing power.
Since then, the quest to proliferate larger and larger language models has accelerated, and many of the dangers we warned about, such as outputting hateful text and disinformation en masse, continue to unfold. Just a few days ago, Meta released its “Galactica” LLM, which is purported to “summarize academic papers, solve math problems, generate Wiki articles, write scientific code, annotate molecules and proteins, and more.” Only three days later, the public demo was taken down after researchers generated “research papers and wiki entries on a wide variety of subjects ranging from the benefits of committing suicide, eating crushed glass, and antisemitism, to why homosexuals are evil.”
This race hasn’t stopped at LLMs but has moved on to text-to-image models like OpenAI’s DALL-E and StabilityAI’s Stable Diffusion, models that take text as input and output generated images based on that text. The dangers of these models include creating child pornography, perpetuating bias, reinforcing stereotypes, and spreading disinformation en masse, as reported by many researchers and journalists. However, instead of slowing down, companies are removing the few safety features they had in the quest to one-up each other. For instance, OpenAI had restricted the sharing of photorealistic generated faces on social media. But after newly formed startups like StabilityAI, which reportedly raised $101 million with a whopping $1 billion valuation, called such safety measures “paternalistic,” OpenAI removed these restrictions.
With EAs founding and funding institutes, companies, think tanks, and research groups in elite universities dedicated to the brand of “AI safety” popularized by OpenAI, we are poised to see more proliferation of harmful models billed as a step toward “beneficial AGI.” And the influence begins early: Effective altruists provide “community building grants” to recruit at major college campuses, with EA chapters developing curricula and teaching classes on AI safety at elite universities like Stanford.
Just last year, Anthropic, which is described as an “AI safety and research company” and was founded by former OpenAI vice presidents of research and safety, raised $704 million, with most of its funding coming from EA billionaires like Talin, Muskovitz and Bankman-Fried. An upcoming workshop on “AI safety” at NeurIPS, one of the largest and most influential machine learning conferences in the world, is also advertised as being sponsored by FTX Future Fund, Bankman-Fried’s EA-focused charity whose team resigned two weeks ago. The workshop advertises $100,000 in “best paper awards,” an amount I haven’t seen in any academic discipline.
Research priorities follow the funding, and given the large sums of money being pushed into AI in support of an ideology with billionaire adherents, it is not surprising that the field has been moving in a direction promising an “unimaginably great future” around the corner while proliferating products harming marginalized groups in the now.
We can create a technological future that serves us instead. Take, for example, Te Hiku Media, which created language technology to revitalize te reo Māori, creating a data license “based on the Māori principle of kaitiakitanga, or guardianship” so that any data taken from the Māori benefits them first. Contrast this approach with that of organizations like StabilityAI, which scrapes artists’ works without their consent or attribution while purporting to build “AI for the people.” We need to liberate our imagination from the one we have been sold thus far: saving us from a hypothetical AGI apocalypse imagined by the privileged few, or the ever elusive techno-utopia promised to us by Silicon Valley elites.
In just a few years, the number of artworks produced by self-described AI artists has dramatically increased. Some of these works have been sold by large auction houses for dizzying prices and have found their way into prestigious curated collections. Initially spearheaded by a few technologically knowledgeable artists who adopted computer programming as part of their creative process, AI art has recently been embraced by the masses, as image generation technology has become both more effective and easier to use without coding skills.
The AI art movement rides on the coattails of technical progress in computer vision, a research area dedicated to designing algorithms that can process meaningful visual information. A subclass of computer vision algorithms, called generative models, occupies center stage in this story. Generative models are artificial neural networks that can be “trained” on large datasets containing millions of images and learn to encode their statistically salient features. After training, they can produce completely new images that are not contained in the original dataset, often guided by text prompts that explicitly describe the desired results. Until recently, images produced through this approach remained somewhat lacking in coherence or detail, although they possessed an undeniable surrealist charm that captured the attention of many serious artists. However, earlier this year the tech company Open AI unveiled a new model— nicknamed DALL·E 2—that can generate remarkably consistent and relevant images from virtually any text prompt. DALL·E 2 can even produce images in specific styles and imitate famous artists rather convincingly, as long as the desired effect is adequately specified in the prompt. A similar tool has been released for free to the public under the name Craiyon (formerly “DALL·E mini”).
The coming-of-age of AI art raises a number of interesting questions, some of which—such as whether AI art is really art, and if so, to what extent it is really made by AI—are not particularly original. These questions echo similar worries once raised by the invention of photography. By merely pressing a button on a camera, someone without painting skills could suddenly capture a realistic depiction of a scene. Today, a person can press a virtual button to run a generative model and produce images of virtually any scene in any style. But cameras and algorithms do not make art. People do. AI art is art, made by human artists who use algorithms as yet another tool in their creative arsenal. While both technologies have lowered the barrier to entry for artistic creation— which calls for celebration rather than concern—one should not underestimate the amount of skill, talent, and intentionality involved in making interesting artworks.
Like any novel tool, generative models introduce significant changes in the process of art-making. In particular, AI art expands the multifaceted notion of curation and continues to blur the line between curation and creation.
There are at least three ways in which making art with AI can involve curatorial acts. The first, and least original, has to do with the curation of outputs. Any generative algorithm can produce an indefinite number of images, but not all of these will typically be conferred artistic status. The process of curating outputs is very familiar to photographers, some of whom routinely capture hundreds or thousands of shots from which a few, if any, might be carefully selected for display. Unlike painters and sculptors, photographers and AI artists have to deal with an abundance of (digital) objects, whose curation is part and parcel of the artistic process. In AI research at large, the act of “cherry-picking” particularly good outputs is seen as bad scientific practice, a way to misleadingly inflate the perceived performance of a model. When it comes to AI art, however, cherry-picking can be the name of the game. The artist’s intentions and artistic sensibility may be expressed in the very act of promoting specific outputs to the status of artworks.
Second, curation may also happen before any images are generated. In fact, while “curation” applied to art generally refers to the process of selecting existing work for display, curation in AI research colloquially refers to the work that goes into crafting a dataset on which to train an artificial neural network. This work is crucial, because if a dataset is poorly designed, the network will often fail to learn how to represent desired features and perform adequately. Furthermore, if a dataset is biased, the network will tend to reproduce, or even amplify, such bias—including, for example, harmful stereotypes. As the saying goes, “garbage in, garbage out.” The adage holds true for AI art, too, except “garbage” takes on an aesthetic (and subjective) dimension.
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 neural network was able to manipulate the plasma inside a fusion reactor into a number of different shapes that fusion researchers have been exploring.Illustration: DeepMind & SPC/EPFL
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.”
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.”
