Source: Neural Reality

Image source: Generated by Wujie AI

Cocoon

In your brain, neurons are arranged in networks of varying sizes. Every action and every thought of yours can change these networks: neurons are either included or excluded, and the connections between them are either strengthened or weakened. This process is constantly ongoing—while you are reading these words, it is changing, and its scale is beyond imagination. There are approximately 80 billion neurons in your brain, sharing 10 trillion or even more connections. Your skull is like a galaxy, and this galaxy is ever-changing.

Computer scientist Geoffrey Hinton, often referred to as the "godfather of artificial intelligence," handed me a cane. He said, "You'll need this here." Then, he walked along a path through the woods towards the shore of a lake. The path meandered through a shady clearing of green trees, past pairs of sheds, and then descended along stone steps to a small dock. As Hinton walked down, he warned, "It's very slippery here."

New knowledge will subtly integrate into your existing neural network. Sometimes they are fleeting: for example, meeting a stranger at a party, their name may only leave a brief impression in your memory network. But sometimes they may last a lifetime—if that stranger becomes your spouse. Because new knowledge blends with old knowledge, what you know will influence what you learn. If someone at a party talks to you about their trip to Amsterdam, the next day, when you visit a museum, your neural network may push you towards Vermeer (a 17th-century Dutch painter). Similarly, tiny changes often lead to huge transformations.

Hinton said, "We used to have bonfires here." We were on a rock protruding into Ontario's Georgian Bay, which extends westward to Lake Huron. The bay is dotted with islands. In 2013, at the age of 65, Hinton sold a company he co-founded with three others to Google for $44 million and then bought this island. Before that, he had been a computer science professor at the University of Toronto for 30 years, leading the way in this then-obscure field called neural networks. The inspiration for this field came from the way neurons are connected in the brain. Since artificial neural networks had only achieved relative success in tasks such as image classification and speech recognition, most researchers considered them at best mildly interesting and at worst a waste of time. Hinton recalled, "Our neural networks couldn't even compete with a child." In the 1980s, when he watched the movie "The Terminator," he was not troubled by the portrayal of the world-destroying artificial intelligence "Skynet" as a neural network; instead, he was pleased to see the potential of this technology being depicted.

The heat caused cracks in the rock to burst outward from the fire pit, and Hinton poked the spot where the fire was with the stick. He was tall and thin, with an English face, and as a thorough scientist, he always commented on what was happening in the material world: the lives of animals, the water flow in the bay, and the geology of the island. "I put a steel mesh under the wood so that air can come in, and the temperature is high enough to soften even the metal," he said in a strange tone, "This is real fire, something to be proud of!"

For decades, Hinton has been trying to cleverly build larger neural networks. He came up with new methods to train them and help them progress. He recruited graduate students and made them believe that neural networks were not a failed endeavor. He thought he was participating in a project that might not yield results until a century later, after he was dead. At the same time, he found himself a widower, raising two young children alone. During a particularly difficult period, the pressures of family life and research left him breathless, and he felt he had given it his all. He said, "I gave up on neural networks when I was 46." He did not anticipate that neural network technology would advance rapidly about a decade ago. As computers became faster, neural networks began transcribing speech, playing games, translating languages, and even driving cars using data from the internet. Around the time Hinton's company was acquired, artificial intelligence began to flourish, and systems like OpenAI's ChatGPT and Google's Bard emerged, many people believe they are changing the world in unpredictable ways.

Hinton set off along the coast, and I followed closely, the fractured rocks swaying under my feet. "Look!" he said, standing in front of a boulder the size of a person, blocking our way. "You can do this, throw the stick over first," he threw his stick to the other side of the boulder, "then there are fulcrums here and here, and there's something to grab onto here." I watched him climb over with ease and took the same steps tentatively.

Every time we learn, our neural networks change, but how exactly do they change? Many researchers like Hinton, who work with computers, are trying to explore the "learning algorithms" of neural networks—a program that acquires new knowledge by changing the statistical "weights" of artificial neurons' connections. In 1949, psychologist Donald Hebb proposed a simple rule for how people learn, often summarized as "neurons that fire together wire together." Once a group of neurons in the brain is synchronously activated, they are more likely to be activated again. This helps explain why we find it easier to do something the second time. But soon, people realized that computerized neural networks needed a different approach to solve complex problems. In the 1960s and 1970s, when Hinton was a young researcher, he drew neural networks in his notebook and imagined how new knowledge would reach their boundaries. How could a network of several hundred artificial neurons store a concept? And how would these networks correct themselves if the concept was flawed?

We circled the coast and arrived at Hinton's cabin, the only one on the island. It was built of enclosed glass and stood high on a wide, dark stone staircase. "Once we came here, and a huge water snake stuck its head out. It was a wonderful memory," Hinton said as we approached the house. His father was a renowned entomologist who named a little-known stage of metamorphosis. It was his father who instilled in him a unique love for cold-blooded animals. When he was young, he and his father kept many snakes, turtles, frogs, toads, and lizards in the garage. Now, when Hinton is on the island (he often goes there in the warm months), he often looks for snakes and brings them home so he can observe them in a terrarium. He has spent his life thinking about how to think from the bottom up, so he is good at observing non-human minds.

Earlier this year, Hinton left Google (he had been working at Google since his company was acquired). He is concerned that artificial intelligence may cause harm and has begun to talk about the "survival threat" this technology may pose to humanity in interviews. The more he uses ChatGPT (an artificial intelligence system trained on a large corpus of human writing), the more uneasy he becomes.

One day, someone from Fox News wrote to him, hoping to interview him about artificial intelligence. Hinton prefers to send sharp single-sentence replies via email, such as when he received a long report from a Canadian intelligence agency, he replied, "Snowden is my hero." So he tried to be witty and wrote, "Fox News is an oxy moron*." Then, he had a stroke of genius and asked ChatGPT to explain his joke. The system told him that his sentence implied that Fox News is fake news, and when he asked ChatGPT to pay attention to the space before "moron," the system explained that Fox News is addictive, like the drug OxyContin. Hinton was astonished. This level of understanding seems to represent a new era of artificial intelligence.

*Translator's note:

"Oxymoron" in English means a figure of speech in which contradictory terms appear in conjunction, but it is a compound word derived from Greek. In this case, Hinton intentionally separated the word with a space, using the Greek origins: "oxy" from the Greek "oxys," meaning sharp, keen, or acidic; and "moron" meaning idiot. ChatGPT seemed to "understand" Hinton's witticism, as it also used "oxy" in OxyContin, where "oxy" comes from its component oxycodone, whose etymology is also from "oxys" (meaning "acid").

We have many reasons to fear the advent of artificial intelligence. For example, the fear of human workers being replaced by computers is a common concern. However, Hinton, along with many prominent tech experts including OpenAI CEO Sam Altman, has issued a warning that artificial intelligence systems may begin to think for themselves, and even attempt to replace or eliminate human civilization. It is shocking that one of the most outstanding researchers in artificial intelligence has expressed such alarming views.

Standing in his own kitchen (he has been tormented by back pain for most of his life, and it eventually became so severe that he gave up sitting. Since 2005, he has not sat for more than an hour at a time), he told me, "People say that artificial intelligence is just glorified 'autocomplete.' Let's analyze this. Suppose you want to be a master at predicting the next word: if you want to be a real master, you have to understand what people are saying, there's no other way. So, training something to be really good at predicting the next word is actually forcing it to understand. Yes, it is indeed 'autocomplete,' but you haven't thought about what it means to have a really good 'autocomplete.'" Hinton believes that "large language models," such as GPT that supports OpenAI's chatbot, can understand the meaning of words and ideas.

Those who are skeptical that we overestimate artificial intelligence point out that there is still a gap between human thinking and neural networks. First, the way neural networks learn is different from us: we accumulate experience, grasp the relationship between experience and reality, and acquire knowledge organically; while neural networks learn abstractly, they process a vast repository of information about the world, a world they do not truly inhabit. But Hinton believes that the intelligence displayed by artificial intelligence systems surpasses their artificial origins.

"When you eat something, you take it in, then break it down into smaller components," he told me, "so you could say that parts of my body are made up of parts of other animals. But that's misleading." He believes that by analyzing human writing, large language models like GPT are able to understand how the world works, thus creating a thinking system, and writing is just a small part of what this system can do. He continued, "It's like a caterpillar turning into a butterfly. In the chrysalis, you turn the caterpillar into a broth, and then you make a butterfly out of the broth."

He started rummaging in a small cabinet by the kitchen. "Aha!" he exclaimed as he placed something on the counter—a dead dragonfly. It was preserved very well. He explained, "I found this at the dock. It had just hatched on the rock and was drying its wings, so I caught it. Look underneath." The dragonfly he caught had just emerged from its larval form. The larva looked different, with its own eyes and legs. It had a hole in its back, and the dragonfly emerged from that hole.

"The larva of a dragonfly is a monster living in the water," Hinton said, "just like in the movie 'Alien,' the dragonfly breaks out of the monster's back. The larva turns into a broth at one stage, and then the dragonfly is born from the broth." In his metaphor, the larva represents the data used to train modern neural networks, and the dragonfly represents the agile artificial intelligence produced from it. Deep learning (a technology Hinton helped pioneer) has led to this transformation. I bent down to see more clearly. Hinton stood straight, as he almost always does, carefully maintaining his posture. "Very beautiful," he said softly, "now you understand. It started as one thing, and now it has become another."

The Branches of the Giant Tree

A few weeks ago, when Hinton invited me to visit his small island, I imagined various possible scenarios. Perhaps he was an introvert who wanted to be alone, or a tech tycoon with a god complex and futuristic ideas. In the days leading up to my arrival, he sent me a photo he had taken via email, showing a rattlesnake coiled in the island's grass. I didn't know whether to feel excited or scared.

In fact, as private islands go, Hinton's island is quite modest, totaling only two acres. Hinton himself is the opposite of a Silicon Valley tech tycoon. At 75, he has an English face reminiscent of a Joshua Reynolds painting, with white hair framing a broad forehead. His blue eyes are usually very steady, leaving his mouth to express emotions. He is a talkative person, especially when talking about himself, "‘Jeff’ is a deformation of 'ego fortissimo,'" he told me.

But he is not a conceited person, as his life has been overshadowed by a sad shadow. "I should probably tell you about my wives," he said to me when we first talked, "I've been married three times. One ended amicably, and the other two ended tragically." He remains on friendly terms with his first wife Joanne, but his second and third wives, Rosalind and Jackie, both died of cancer in 1994 and 2018, respectively. For the past four years, Hinton has been with retired sociologist Rosemary Gartner. She gently told me, "I think he's the kind of person who always needs a partner."

He is a romantic rationalist, with a balanced emotional understanding of science and emotions. In the cabin, a single large room occupies most of the ground floor, with a burgundy canoe inside. He and Jackie found this long-neglected canoe in the island's woods. Jackie, an art historian, and a group of female canoe makers rebuilt the canoe during her illness. Hinton said, "She completed the maiden voyage." Since then, no one has used it.

He placed the dragonfly carefully and then walked to a standing desk, where there was a laptop, a pile of Sudoku puzzles, and a notebook with computer passwords (he rarely takes notes because he has designed a memory system that can generate and recall very long passwords in his mind). He asked, "How about we make a family tree?" Using two fingers (he doesn't have a fixed typing method), he typed "Geoffrey Hinton family tree" and pressed enter. In 2013, Google acquired Hinton's startup, in part because the team discovered how to significantly improve image recognition using neural networks. Now, the screen was filled with an endless family tree.

Hinton comes from a special British scientific family: politically radical and creatively rich. In his family tree, his great-uncle Sebastian Hinton was the inventor of the jungle gym, and his aunt Joan Hinton was a physicist in the Manhattan Project. Before her, there was Lucy Everest, the first woman elected to the Royal Institute of Chemistry; Charles Howard Hinton, a mathematician who created the concept of the four-dimensional tesseract, a gateway to the fourth dimension (the same hypercube that appears in the movie "Interstellar"); James Hinton, a pioneering ear surgeon and advocate of polygamy (he reportedly said, "Christ is the savior of men, and I am the savior of women."). In the mid-19th century, Hinton's great-grandfather, the British mathematician George Boole, invented the binary logic system, now known as Boolean algebra, the foundation of all computing. Boole's wife, Mary Everest, was a mathematician and writer, and also the niece of George Everest, a surveyor. Mount Everest is named after George.

"Jeff was born to do science," Yann LeCun, Hinton's former student and collaborator, now head of AI at Meta, told me. However, Hinton's family is even more extraordinary. His father, Howard Everest Hinton, grew up in Mexico during the Mexican Revolution in the 1910s, working on a silver mine managed by his father. "He was very tough," Hinton said of his father. Family legend has it that at the age of 12, Howard threatened to shoot his boxing coach because the coach was too strict, and the coach took him seriously and had to flee the town. Howard's native language was Spanish, and he was ridiculed for his accent while attending Berkeley. "He hung out with a group of similarly discriminated Filipinos and became a radical at Berkeley." Howard had mature Marxist and Stalinist political views.

In school, Hinton favored science. But for ideological reasons, his father forbade him from studying biology. Howard believed that the possibility of genetic determinism contradicted the communist belief in the ultimate malleability of human nature. "I hated all kinds of beliefs," Hinton recalled. Howard, who taught at the University of Bristol, was like an "Indiana Jones" in the world of entomology: he smuggled oddities from around the world in his luggage back to England and edited an important journal in the field. Hinton's middle name is also Everest, which put immense pressure on him to make his own mark. He remembered his father telling him, "If you work twice as hard as me, when you're my age, you might only be half as successful as me."

At Cambridge University, Hinton tried different majors but was disheartened to find that he was never the smartest student in the class. He briefly left university to "read depressing novels" and worked odd jobs in London, then returned to try architecture for a day. Finally, he dabbled in physics, chemistry, physiology, and philosophy, and settled on a degree in experimental psychology. He often "haunted" the office of moral philosopher Bernard Williams, finding himself interested in computers and the mind. One day, Williams pointed out that our different thoughts must reflect different physical arrangements in our brains, which is completely different from the situation in computers, where software is independent of hardware. Hinton was struck by this observation.

He remembered a friend telling him in high school that memory might be stored in a "holographic" way in the brain. That is, although memory is distributed, it can be accessed as a whole through any part. What he encountered was "connectionism"—a method that combines neuroscience, mathematics, philosophy, and programming, aimed at exploring how neurons work together to "think." One of the goals of connectionism was to create a brain-like system in computers. This had made some progress at the time: in the 1950s, psychologist and connectionism pioneer Frank Rosenblatt created a machine called the "Perceptron," which used simple computer hardware to simulate a network of hundreds of artificial neurons. When connected to optical sensors, the device could track which patterns of light activated which artificial neurons, thus recognizing letters and shapes.

In the cabin, Hinton stood for a while, then paced back and forth behind the kitchen counter on the ground floor. He made some toast, grabbed an apple for each of us, and then used a footstool to prop up a small table for himself. The pressure from his family prevented him from finding brief satisfaction. "I always liked woodworking," he reminisced playfully as we ate, "at school, you could volunteer for woodworking in the evenings. I often thought, if I became an architect, would I be happier because I wouldn't have to force myself to do these things. With science, I always had to force myself. Because of family reasons, I had to succeed, I had to have a way out. There was happiness in it, but more anxiety. Now that I've succeeded, it makes me feel extremely relieved."

Hinton's laptop chimed. Since leaving Google, he had been receiving requests to comment on artificial intelligence. He walked over to check his email and then got lost in the forest of family trees, all of which seemed to have some issue or another.

"Look at this," he said.

I walked over to look at the screen. It was an "academic family tree," with Hinton at the top, followed by his students and his students' students. The "tree" was very wide, and he had to scroll horizontally to see the extent of his influence. "Oh my goodness," Hinton said as he studied the academic family tree, "she wasn't actually my student." He scrolled the mouse again, "He's brilliant but not good at being an advisor because he always thinks he can do better than others." Hinton is someone who carefully nurtures talent, and he seems to enjoy the feeling of being surpassed by his students. When evaluating job candidates, he often asks their advisors, "Are they better than you?" Recalling his father, who passed away in 1977, Hinton said, "He was extremely competitive. I often wonder if he would be happy to see me achieve this level of success. Because now I am more successful than he was."

According to Google Scholar data, Hinton is now the second most cited researcher among psychologists and the most cited among computer and cognitive scientists. If his start at Cambridge University was slow and unusual, it was due to his research in an emerging field. He closed his laptop and said, "Back then, there were very few people doing neural networks at good universities. You couldn't do it at MIT, you couldn't do it at Berkeley, and you couldn't do it at Stanford." Hinton became a hub for a nascent network, which had its advantages. Over the years, many top talents came to him.

Boltzmann Machines

"Today's weather is really nice," Hinton said the next morning, "we should go chop down a tree." He was wearing a dress shirt tucked into khaki pants, not looking much like a lumberjack. Nevertheless, he rubbed his hands together. On the island, he was always chopping down trees to create a more orderly and beautiful landscape.

The house is actually not yet finished, and few contractors are willing to come to such a remote place. The people hired by Hinton also made some unnecessary mistakes (connecting the drainage pipe to an uphill slope, leaving the floor unfinished), which still makes him angry. Almost every room has a small project that needs to be corrected, and when I visited, Hinton had made some small markings on the building materials to help the new contractors. These markings are usually written directly on the building materials. In the bathroom on the first floor, a note on the baseboard against the wall reads, "This baseboard should be used in the bathroom (maple decoration only in front of the shower)." In the closet of the guest room, masking tape extends along the shelf: "Do not prime for the shelf, prime for the shelf bracket."

Labeling things is also useful for the brain, helping it grasp reality. But what does labeling mean for artificial minds? When Hinton obtained his PhD in artificial intelligence at the University of Edinburgh, he pondered how to simulate "cognition" in the brain on a computer. At that time, in the 1970s, most artificial intelligence researchers were "symbolists." In their view, understanding ketchup might involve many concepts such as "food," "sauce," "condiment," "sweetness," "freshness," "red," "tomato," "Americans," "French fries," "mayonnaise," and "mustard." These concepts mixed together to form a new concept, "ketchup." There was a well-funded large-scale artificial intelligence project called Cyc, which aimed to build a massive knowledge base that scientists could input concepts, facts, rules, and inevitable exceptions into using a special language (e.g., birds can fly, but penguins and birds with damaged wings…).

However, Hinton was skeptical of this approach. It seemed too rigid, too focused on the reasoning abilities held by philosophers and linguists. He knew that in the natural world, many animals can exhibit intelligent behavior without having concepts that can be expressed in language. They simply learn how to be smart through experience. The source of intelligence is learning, not knowledge.

Human complex thinking seems to often occur through symbols and words. However, Hinton and his collaborators James L. McClelland and David Rumelhart believed that many behaviors occur at the sub-conceptual level. They wrote, "Note that if you learn a new fact about something, your expectations about other similar things often change."

For example, if you are told that gorillas like onions, you might guess that orangutans also like onions. This suggests that knowledge is likely "distributed" in the brain, composed of small modules shared between related ideas. The concepts of "gorilla" and "orangutan" would not have two separate networks of neurons, but rather bundles of neurons representing various specific or abstract "features" - fur, quadrupedal, primate, animal, intelligence, wildness, etc., might be activated in a certain way to represent "gorilla," and slightly differently to represent "orangutan." In addition to these features, we can also add "onion" and other features. A brain constructed in this way could become confused and make mistakes: mixing various features in the wrong arrangement would result in a creature that is neither a gorilla nor an orangutan. However, a brain with the right learning algorithm could adjust the weights between neurons, making logical combinations superior to illogical ones.

Hinton continued to explore these ideas, first as a postdoctoral researcher at the University of California, San Diego (and married Joanne, who was his supervisor in computer vision), then as a research fellow in applied psychology at Cambridge University, and later as a computer science professor at Carnegie Mellon University in Pittsburgh in 1982. There, he spent most of his research budget on a computer capable of running neural networks. Soon after, he married for the second time, to molecular biologist Rosalind Zalin. At Carnegie Mellon University, Hinton made a breakthrough. He and computer scientist and neuroscientist Terrence Sejnowski developed a neural network called the "Boltzmann Machine." The name of this system pays tribute to Ludwig Boltzmann, the 19th-century Austrian physicist who used mathematical methods to describe the behavior of gases on a large scale and the behavior of their constituent particles on a small scale. Hinton and Sejnowski combined these equations with "learning theory."

Hinton was reluctant to explain the Boltzmann Machine to me. "Let me tell you what it feels like," he said, "it's like taking a child for a walk. There's a mountain ahead, and you have to take the child to the top and then back down." He sighed as he looked at me (the metaphorical child). His concern was valid; I might be misled by a simplified explanation and then mislead others. "Trying to explain complex ideas that you don't understand is useless. First, you have to understand how something works. Otherwise, it's all just nonsense to you." Finally, he picked up a few sheets of paper and began drawing neural diagrams connected by arrows and writing equations, and I tried to understand these things (I had studied linear algebra on Khan Academy before visiting).

Hinton suggested that one way to understand the Boltzmann Machine is to imagine a set of facial features used to create a composite image of a criminal. Through this system, various facial features - thick eyebrows, blue eyes, crooked nose, thin lips, large ears, and so on - can be combined to generate a composite sketch similar to the ones used by the police. To make the composite image work, the features themselves must be properly designed. By changing the connection weights between artificial neurons, the Boltzmann Machine can not only learn to combine features, but also learn to design features. The Boltzmann Machine starts with features that are chaotic, like static on a TV screen, and then goes through two phases: "wakefulness" and "sleep" to refine these features. During "wakefulness," it adjusts these features to make them more like real faces. During "sleep," it imagines a face that doesn't exist, then modifies the features to make them fit poorly.

It tells itself what not to learn in its dreams. This system is very elegant: over time, it can gradually move away from mistakes and towards reality, without needing anyone to tell it right from wrong. It just needs to see the real and dream the unreal.

In 1983, Hinton and Sejnowski described the Boltzmann Machine in a paper. Yann LeCun told me, "I read that paper at the beginning of my graduate studies and said, 'I have to talk to these people, they are the only ones in the world who understand that we need a learning algorithm.'" In the mid-1980s, Yoshua Bengio, a pioneer in natural language processing and computer vision and the current scientific director of the Quebec Artificial Intelligence Institute Mila, trained a Boltzmann Machine to recognize spoken syllables as part of his master's thesis. "Jeff was one of the external examiners," Bengio recalled, "and he wrote, 'This won't work.'" However, Bengio's version of the Boltzmann Machine was more effective than Hinton expected, and it took Bengio several years to figure out why it was successful. This pattern would become all too familiar - over the next few decades, neural networks often performed better than expected, perhaps because neurons formed new architectures during the training process. Bengio recalled, "The experimental part came before the theory. We usually try new methods and see what the neural network can produce on its own."

Hinton said that partly because Rosalind disliked Ronald Reagan, they moved to the University of Toronto. They adopted a boy and a girl from Latin America and lived in a house in the city. Hinton said, "I'm the kind of selfless professor who is dedicated to his work."

Rosalind had struggled with infertility and had unpleasant experiences with cold and unfeeling doctors. Perhaps because of this, when she was later diagnosed with ovarian cancer, she chose homeopathy. "It doesn't make any sense," Hinton said, "you can't dilute something and make it stronger." He couldn't understand how a molecular biologist could endorse homeopathy. Nevertheless, Rosalind was determined to treat her cancer herself, even refusing surgery despite the tumor being as large as a grapefruit. Later, although she agreed to surgery, she refused chemotherapy and sought increasingly expensive homeopathic treatments, first in Canada and later in Switzerland. She developed secondary tumors and asked Hinton to sell their house to pay for the new homeopathic treatments. He recalled, "I drew the line there," Hinton said, his eyes narrowed in pain, "I told her, 'No, we're not selling the house. Because if you die, I have to take care of the kids, and it would be better for them if we have a house.'"

Translator's Note

The theory is based on "treating the same disease with the same preparation."

Rosalind returned to Canada and immediately checked into the hospital. She persisted for several months, but she refused to let her children see her until the day before she passed away because she didn't want them to see how sick she was. Throughout her illness, she firmly believed that she would soon get better. As Hinton described all of this, he still seemed deeply pained: angry, guilty, hurt, and confused. When Rosalind passed away, Hinton was 46 years old, with a 5-year-old son and a 3-year-old daughter. He said, "She hurt everyone because she refused to accept the reality of her impending death."

The sound of the waves filled the afternoon tranquility. The intense golden sunlight streamed through the floor-to-ceiling windows of the room, and the small spider webs outside became particularly clear in the light. Hinton stood for a while, collecting himself.

He said, "I think I need to go chop down a tree."

We walked out the front door and down the path to the shed. Hinton took a green chainsaw and some safety goggles from one of the sheds.

"Rosemary said I can't chop trees when no one's around, in case I cut off my arm or something," he said to me. "Have you ever sailed a boat?"

"No," I said.

"Then I won't chop off my right arm."

He put on a pair of protective shoes over his khaki pants.

"I don't want to give you the impression that I know what I'm doing," he said. "But the basic idea is, you make a lot of V-shaped cuts in the tree, and then it falls."

Hinton walked through the path to the tree he had chosen in his mind, checking for snakes in the bushes as he went. It was a lush cedar tree, about 20 feet tall. Hinton looked up at the direction the tree was leaning, then started the chainsaw and began cutting on the opposite side of the trunk from the lean. He took the saw off and made another cut, creating a V-shaped notch. Then he stopped and turned to explain to me, "Because the tree is leaning away from the notch, as you cut in, the V will open up, and the saw won't get stuck."

Hinton silently operated the chainsaw, occasionally stopping to wipe his brow. The sun was scorching, and mosquitoes swarmed from every dark corner. I glanced at the side of the shed, where ants and spiders were engaged in unknown, endless activities. At the end of the path, the water shimmered, a beautiful and peaceful place. But I understood why Hinton wanted to cut it down: a lovely round hill extended down into a gentle cave, and without the extra tree, light could flow into the cave. The tree was a mistake.

Finally, he made the second cut on the other side of the tree, angling towards the first cut. Then he moved back and forth, deepening the two cuts, making the tree teeter. Suddenly, gravity almost silently took over. The large tree fell, with astonishing grace, towards the bottom of the cave. Light poured in.

Backpropagation

Hinton fell in love with the Boltzmann Machine. He hoped that the Boltzmann Machine, or something similar, would be based on the learning of the real brain. "It should come true," he told me, "if I were God, I would make it come true." However, further experiments found that as the Boltzmann Machine grew, they often became overwhelmed by inherent randomness. "Jeff and I didn't see eye to eye on the Boltzmann Machine," Yann LeCun said, "Jeff thought it was the most beautiful algorithm. I thought it was ugly. It's stochastic, meaning it's partly based on randomness. In contrast, I thought the backpropagation algorithm was more elegant."

From the 1960s, some researchers explored the backpropagation algorithm. While Hinton was researching the Boltzmann Machine with Sejnowski, he also collaborated with Rumelhart and another computer scientist, Ronald Williams, on backpropagation. They suspected that this technique had untapped potential in learning, especially when combined with neural networks running across multiple layers.

One way to understand backpropagation is to imagine a Kafkaesque judicial system. The upper layers of the neural network can be thought of as a jury, constantly trying cases. In the dystopian world of backpropagation, just after the jury makes a decision, the judge can tell the jurors that their decision was wrong, and they will be punished until they mend their ways. The jurors find that three of them have a particularly strong influence in leading everyone down the wrong path. This shared responsibility is the first step of backpropagation.

Next, the three hot-headed jurors must determine how they themselves were misled. They consider the influences they were under - parents, teachers, experts, and so on - and find the people who misled them. In turn, these influencers must find their own influencers and share the blame among them. This leads to a round of mutual blame, with each layer of influencers demanding that their influencers take responsibility. This is a chain of reverse blame. Once it's known who misled whom and how much, the neural network will adjust itself proportionally, allowing each person to receive a little less "bad" influence and a little more "good" influence. The entire process is repeated with mathematical precision until all decisions (not just in this case, but in all cases) are as "correct" as possible.

In 1986, Hinton, Rumelhart, and Williams published a three-page paper in Nature, demonstrating how this system operates in neural networks. They pointed out that, like the Boltzmann Machine, backpropagation is not a "reasonable model for brain learning": unlike computers, the brain cannot rewind to review its past performance. But backpropagation can still achieve neural properties similar to the brain (neural specialization). In the real brain, neurons are sometimes arranged in structures designed to solve specific problems: for example, in the visual system, neurons in different "columns" can recognize the edges of what we see. Similar situations occur in the backpropagation network: higher-level neurons exert an evolutionary pressure on lower-level neurons. As a result, certain layers of a network responsible for deciphering handwriting, for example, may become specialized in recognizing lines, curves, or edges. Eventually, the entire system can develop "appropriate internal representations." The neural network will understand and utilize the knowledge it possesses.

Translator's Note

In recent years, some researchers have suggested that the brain also has the ability to execute the backpropagation algorithm, see the translated full text of "Backpropagation in the Brain" published in Nature Neuroscience, Neural Reality, click the link to read.

In the 1950s and 1960s, the "perceptron" and other connectionist research results had a huge impact. However, in the following years, people's enthusiasm for connectionism gradually waned. The backpropagation paper was one of the contributors to the revival of interest, and it gained widespread attention. However, due to practical and conceptual reasons, progress in building backpropagation networks was slow. In practical terms, this was due to the slow development of computers. "The progress of backpropagation basically depends on how much the computer can learn overnight," Hinton recalled, "and the answer is often not much." Conceptually, neural networks are mysterious, because they cannot be programmed using traditional methods, and you cannot edit the weights of connections between artificial neurons. Moreover, it's difficult to understand the meaning of the weights, as they continuously adjust and change through self-training.

The learning process of backpropagation also has many pitfalls. For example, in the process of "overfitting," the network may choose to memorize the training data instead of learning to generalize from the data. Avoiding various traps is not simple, as it depends entirely on the network itself. It's like cutting down a tree: researchers can make cuts here and there, but it's entirely up to the tree where it falls. Researchers can try techniques such as "ensembling" (combining weak networks into a strong network) or "early stopping" (allowing the network to learn, but not too much), and they can also use the Boltzmann Machine for "pre-training" to let it study some knowledge and then build a backpropagation network on top of it. This way, the system has to wait until it has mastered some basic knowledge before starting "supervised" training. After that, they let the neural network learn freely, hoping it will meet their expectations.

New neural network "architectures" also emerged. "Recurrent" and "convolutional" networks allowed the system to make progress based on various ways of working within itself. However, researchers seemed to have discovered a kind of alien technology and didn't know how to use neural networks. They turned the Rubik's Cube around, trying to find order in chaos. "I have always believed that neural networks are not nonsense," Hinton said, "it's not a matter of faith for me, it's obvious." Since the brain learns using neurons, it must be feasible to use neural networks for complex learning. He would work twice as hard and twice as long to prove this.

When networks are trained through backpropagation, they need to be told when they make mistakes and how much they are wrong. This requires a large amount of accurately labeled data, so the network can learn the difference between handwritten "7" and "1," or between a golden retriever and a red setter. However, it's difficult to find a large and accurately labeled dataset, and creating more datasets is also challenging. LeCun and his collaborators developed a huge handwritten digit database, which they later used to train a neural network to read postal code samples provided by the United States Postal Service. A computer scientist at Stanford University named Fei Fei Li led the development of a massive database called ImageNet. Creating this database required collecting over 14 million images and manually categorizing them into 20,000 classes.

As the scale of neural networks continued to grow, Hinton came up with a method: transferring knowledge from large networks to small networks (small enough to run on devices like smartphones). He explained in the kitchen, "This is called knowledge distillation. In school, the art teacher would show us some slides and say, 'This is Rubens, that's Van Gogh, this is William Blake.' But suppose the art teacher tells you, 'Well, this is Titian Vecellio, but it's a peculiar Titian because in some respects it's very much like Raphael, which is unusual for Titian.' This seems more helpful to you. They not only tell you the right answer, but also other seemingly correct answers." In "distillation learning," one neural network provides not only the correct answer to another neural network, but also a series of possible answers and their probabilities. This is a richer form of knowledge.

Several years after Rosalind's death, Hinton reconnected with art historian Jacqueline Ford (hereinafter referred to as Jacqueline). Hinton had briefly dated her before moving to the United States. She was cultured, passionate, curious, and beautiful. Hinton's sister said, "You're way out of your league compared to her." Nevertheless, Jacqueline gave up her job in the UK and moved to Toronto with Hinton. They got married on December 6, 1997, Hinton's 50th birthday. The following decades were the happiest time of his life, and his family was whole again. His children liked their new mother, and he and Jacqueline began to explore the islands of Georgian Bay. Reflecting on this time, he gazed at the canoe in the living room. He said, "We found this canoe in the woods, it was upside down, covered in canvas, completely rotten, everything was rotten. But Jacqueline was determined to save it, just like she saved me and the kids."

Hinton didn't like backpropagation. He told me, "How unsatisfying it is intellectually. Unlike the Boltzmann Machine, it's completely deterministic. Unfortunately, it's also more practical." Slowly, as actual progress was made, the power of backpropagation became undeniable. Hinton told me that in the early 1970s, the British government hired a mathematician named James Lighthill to determine if artificial intelligence research had any chance of success. Lighthill's conclusion was that it was not possible. "He was right," Hinton said, "on the assumption that everyone agreed on at the time: the speed of computers might be a thousand times faster, but not a billion times faster." Hinton did the math in his head and figured that if he started running a program on an extremely fast research computer in 1985 and continued running it until now, and then started running the same program on the fastest system currently used in artificial intelligence, it would take less than a second to catch up to the previous one.

In early 2000, as multi-layer neural networks equipped with powerful computers began training on larger datasets, Hinton, Bengio, and LeCun began discussing the potential of "deep learning". In 2012, Hinton, Alex Krizhevsky, and Ilya Sutskever introduced AlexNet, an 8-layer neural network that could finally recognize objects in ImageNet with an accuracy level comparable to humans. Hinton, Krizhevsky, and Sutskever founded a company and sold it to Google. With this wealth, he and Jacqueline bought a small island in Georgian Bay. "That was my one true indulgence," Hinton said.

Two years later, Jacqueline was diagnosed with pancreatic cancer. The doctors estimated she had one or two years to live. "She was very brave and very rational," Hinton said, "she didn't desperately deny or try to escape the situation. Her view was, 'I can feel sorry for myself, or I can say I don't have much time left, I should enjoy this time as much as possible and make sure everyone else is okay.'" Before deciding on a treatment method, she and Hinton carefully studied the statistics. Through chemotherapy, she extended her time by one to three years. In their cottage, when she could no longer climb the stairs, Hinton made a small basket out of rope so she could lower tea from the second floor to the first, and then he could heat it in the microwave. (He later realized, "I should have just moved the microwave upstairs.")

Later that evening, we leaned on Hinton's desk, and he showed me photos of Jacqueline on his laptop. In one photo from their wedding day, she and Hinton stood with the children in the neighbor's living room exchanging vows. Hinton looked radiant and relaxed. Jacqueline gently held one of his hands. In the last photo he showed me, she was paddling a wine-red canoe on the mottled water near the dock, gazing into the camera. "That was in the summer of 2017," Hinton said. Jacqueline passed away the following year. In June of that year, Hinton, Bengio, and LeCun received the Turing Award (equivalent to the Nobel Prize in computer science).

Hinton firmly believes that neural networks are indeed capable of having emotions. "I think emotions are counterfactual statements about 'something that would lead to some behavior'," he told me earlier that day. "For example, I want to punch someone in the nose. What I mean is: if I didn't have social inhibitions, if I didn't stop myself, I would actually punch him. So when I say 'I feel angry,' it's actually a shorthand for 'I want to perform an aggressive act.' Emotions are just a way of stating an intention to act."

He told me about a "frustrated AI" he had seen in 1973. At that time, a computer was connected to two television cameras and a simple mechanical arm. The system's task was to assemble some blocks on the table into a toy car. He said, "It was difficult, especially in 1973. If the blocks were separate, the visual system could recognize them, but if they were stacked together, the visual system couldn't recognize them. So how did it do it? It pulled the blocks back a bit, and then with a 'bang!' it dropped them on the table. Basically, it couldn't deal with what it was dealing with, so it violently changed it. If a person did that, you would say they were frustrated. The computer couldn't see the right blocks, so it smashed them." To have feelings is to desire something you can't have.

"I love this house, but sometimes it's a sad place," he said as we looked at photos. "Because she used to like being here, and now she's not."

The sun was setting, and Hinton turned on a small lamp on his desk. He closed the computer, pushed his glasses up his nose, straightened his shoulders, and brought his thoughts back to the present.

He said, "I want you to know about Roz (Rosalind) and Jacqueline, because they are important parts of my life. But in reality, this is also related to artificial intelligence. People usually have two attitudes towards artificial intelligence, denial and fatalism. Everyone's first reaction to artificial intelligence is 'we must stop it.' Just like everyone's first reaction to cancer is 'we should cut it out.' But the important thing is to realize that 'cutting it out' is just a fantasy."

He sighed. "We can't deny it, we have to face reality. We need to think: how can we make artificial intelligence treat humans less badly?"

The Future Outlook

How useful or dangerous is artificial intelligence really? No one knows, in part because neural networks are so strange. In the 20th century, many researchers wanted to create computers that simulated the brain. However, despite neural networks (such as OpenAI's GPT model) involving billions of artificial neurons (similar to the brain in this respect), they are actually very different from the biological brain. Today's artificial intelligence is based on cloud computing and is housed in data centers with industrial-scale power consumption. They are clueless in some respects, but gifted in others. They provide reasoning services for millions of users, but rely on prompts from users. They are lifeless. They have likely passed the Turing test [a standard formulated by computer pioneer Alan Turing, which has long been acclaimed, and which considers any computer that can convincingly mimic human conversation to be reasonably considered to be capable of thinking]. However, our intuition tells us that nothing residing in a browser tab can truly think like us. These systems force us to ask, is our way of thinking the only one?

In his last years at Google, Hinton focused on creating artificial intelligence that was closer to traditional minds using hardware that was closer to the brain. In today's artificial intelligence, the connection weights between artificial neurons are stored in digital form, much like the brain constantly records its own information. In the actual brain, the weights are embedded in the physical connections between neurons. Hinton wanted to use special computer chips to create an artificial version of this system.

He told me, "If this can be achieved, it will be very remarkable." The chips can learn by changing "conductance." Because this method integrates weights into the hardware, it cannot be copied from one machine to another*. Each artificial intelligence must learn on its own. "They have to go to school like students," he said, "but the advantage is that you go from using one megawatt of power to using 30 watts." As he spoke, he leaned forward, his eyes fixed on me: I caught a glimpse of the missionary Hinton.

*Translator's note

Traditional methods store weights in digital form, so they can be copied, see "digital intelligence" below.

Because the knowledge gained by each artificial intelligence is lost when it is dismantled, he called this method "mortal computing". "This method makes us give up immortality," he said, "but in literary works, you always give up becoming an immortal god for the sake of a beloved woman, don't you? In this case, we will get something more valuable than 'immortality', which is energy efficiency." In addition, energy efficiency also encourages "individuality": because the human brain can run solely on the energy provided by cereal, the world can support billions of different brains. Each brain can continue to learn, instead of being trained once and then thrown into the world.

As a scientific endeavor, this "mortal" artificial intelligence may bring us closer to replicating our own brains. But Hinton regretfully believes that digital intelligence seems to be more powerful. He said, in simulated intelligence, "if the brain dies, the knowledge dies with it." In contrast, in digital intelligence, "if a computer dies, the same connection strength can be used on another computer. And even if all the digital computers die, if you store the connection strength somewhere, you can make another digital computer and run the same weights on it. Ten thousand neural networks can learn ten thousand different things at the same time, and then share what they have learned together." He said that the combination of immortality and replicability suggests that "we have reason to worry that digital intelligence will replace biological intelligence".

How should we describe the mental life of a digital intelligence that has no physical body or individual identity? In recent months, some artificial intelligence researchers have begun calling GPT a "reasoning engine". This may be to avoid the word "thinking," and yet we have been trying to define "thinking". Bengio told me, "People criticize us for using these words: 'thinking,' 'knowing,' 'understanding,' 'deciding,' and so on. But even though we don't fully understand the meanings of these words, the analogies they provide are still very effective in helping us understand our actions. It helps us articulate and explore words like 'imagination,' 'attention,' 'planning,' 'intuition,' and so on. Much of what we do is about solving the 'intuition' problem of the mind."

Intuition can be understood as thoughts we can't explain: our brains generate these thoughts for us by unconsciously linking current experiences with past experiences. We tend to be rational rather than intuitive, but Hinton believes that our intuition is stronger than we think. He told me, "For years, the symbolist AI people said our true nature is a 'reasoning machine'. I think that's nonsense. Our true nature is an 'analogy machine', with a bit of reasoning built on top, and when analogies give the wrong answer, we notice and correct it."

Overall, current artificial intelligence technology is articulate and overly rational: it stumbles on intuitive things in the physical world. LeCun told me, "Any teenager can learn to drive a car with almost no supervision after 20 hours of practice. Any cat can jump onto a series of furniture and climb to the top of a shelf. Today, apart from self-driving cars, we don't have any artificial intelligence system that comes close to doing these things." And these systems are overdesigned, requiring "mapping the entire city, hundreds of engineers, and tens of thousands of hours of training." Solving these tricky problems of physical intuition "will be a huge challenge in the next decade". However, the principle is simple: if neurons can do it, then neural networks can do it.

Hinton believes that people's skepticism about artificial intelligence, while comforting, often stems from an unfounded trust in "human exceptionalism." Researchers complain that AI chatbots can "hallucinate," fabricating seemingly plausible answers to difficult questions. But Hinton questions this notion: "We should use the word 'confabulate'," he told me, "'Hallucination' only exists with sensory input—auditory hallucinations, visual hallucinations, olfactory hallucinations. But if it's just making stuff up, then it's just confabulation." He cited the case of John Dean, White House counsel to President Richard Nixon, who was interviewed about the Watergate scandal before knowing his conversations were being recorded. Dean made up details, intentionally getting things wrong, and confusing who said what. "But the gist of it was right," Hinton said, "He recalled the situation and imposed that recollection on some characters in his mind. He fabricated a little script. That's how human memory works. In our minds, confabulation and truth-telling are really not that different. Telling the truth is just fabricating the right story, because everything is a matter of your judgment, isn't it?" From this perspective, ChatGPT's confabulation ability is a flaw, but at the same time, it is a symbol of human intelligence.

People often ask Hinton if he regrets his work. He doesn't. [He recently sent a journalist a message: "Here's a song for you," along with a link to Edith Piaf's "Non, Je Ne Regrette Rien."]. He said that when he started his research, no one thought this technology would succeed; even when it did, no one thought it would succeed so quickly. Because he believes that artificial intelligence is real intelligence, he expects it to make contributions in many fields. However, he sometimes worries, for example, he is concerned that powerful people will abuse it. He believes that autonomous weapons (which the US military is actively developing) should be banned, but he warns that even benign automated systems could cause serious harm. "If you want a system to be effective, you have to give it the ability to create its own subgoals. The problem now is that there's a very common subgoal that applies to almost all goals: gaining more control. The question we need to study is: How do we prevent artificial intelligence from seizing control? No one knows the answer." (He pointed out that "control" is not necessarily in a physical sense: "Just like Trump can invade the Capitol with language.")

In this field, opinions on Hinton's views vary. LeCun told me, "I'm not afraid of artificial intelligence. I think it would be relatively easy if we design them so that their goals align with ours." He continued, "Some people think that if a system is intelligent, it will want to dominate everything. But the desire for domination has nothing to do with intelligence, it has to do with testosterone." I remembered the spiders I saw in the cabin, and how their webs covered Hinton's windows. They didn't want to dominate, yet their insect intelligence allowed them to expand their territory. Life systems without centralized brains, such as ant colonies, don't "want" to do anything, but they can still forage, navigate upstream, and besiege competitors. Both Hinton and LeCun may be right. The transformation of artificial intelligence is not yet complete, and we don't know what it will become.

"Why not just unplug them?" I asked Hinton a popular question about artificial intelligence. "Is this question absurd?"

He said, "This argument is indeed not sound: that we would be better off without something—but the cost of doing so is not worth it. Just like we would be better off without fossil fuels, but we would become more primitive, so it may not be worth the risk." He added, "It's not going to happen. Because that's how society is, there will be competition between different countries. If the United Nations really worked, maybe it could prevent such things from happening. Nevertheless, artificial intelligence is very useful. It has great potential in fields like medicine, and of course, it can give a country an advantage through autonomous weapons." So earlier this year, Hinton refused to sign a petition calling for a "pause of at least six months in AI research."

"So what should we do?" I asked.

"I'm at a loss too," he said, "If it were like global warming, people could say, listen, we either stop burning carbon, or we find an effective way to remove carbon dioxide from the atmosphere. We know what the solution looks like in our hearts. But in the field of artificial intelligence, it's not that simple."

Hinton was wearing a blue waterproof jacket, and we were about to go to the dock to meet Rosemary. "She's bringing supplies!" he said with a smile. As we walked out, I looked back at the cabin. In the large room, a burgundy canoe was gleaming in the sunlight. Chairs were arranged in a semicircle in front of the canoe, facing the water through the window. Some magazines were piled on a small table. It was a beautiful house. Human thought is not just rational, it exists in time, coexists with life and death, and constructs a world around itself. It gathers many meanings, like gravity assisting. I thought, perhaps artificial intelligence could also imagine such a place. But does it need such a place?

We walked along a forest path, through a shed, down steps to the dock, and then boarded Hinton's boat. Under the captivating blue sky, a gentle breeze brushed the water. Hinton stood at the helm, and I sat at the bow, watching the islands pass by slowly, thinking about the story of artificial intelligence. For some, this is a Copernican story, in which the intuition that "human thought is special" is being overturned by machines that can think. For others, it's a Promethean story, where we stole fire from the gods, risking being burned alive. Some think we are deluding ourselves, being fooled by the machines we've created and the companies hoping to profit from them. From a novel perspective, it could also be a story about human limitations. If we were gods, we might create a different kind of artificial intelligence. In reality, we can only manage the current version of artificial intelligence. Meanwhile, I couldn't help but think of this story from the perspective of Eden. By reconstructing knowledge systems in our minds, we seized the forbidden fruit. Now, we risk being expelled from Eden. But who would choose not to understand how "understanding" itself works?

At the dock, Hinton used the wind to accelerate, turn, and guide himself into a berth, as if he were born to do it. "I'm learning!" he proudly said. We got off the boat and waited for Rosemary to arrive next to a store. After a while, Hinton went in to buy light bulbs. I stood there, enjoying the warmth, and then saw a tall, bright-eyed woman with long white hair walking briskly towards me from the parking lot.

Rosemary shook hands with me, then glanced over my shoulder. Hinton was coming out of the store near the greenery, grinning.

"What kind of medicine are you selling in your gourd?" she asked.

Hinton held up a black and yellow banded snake, about a meter long, coiling and uncoiling like a spring. "I brought a gift!" he said heroically, "I found it in the bushes."

Rosemary laughed happily and turned to me, "That's typical of him."

"It's not happy," Hinton observed the snake.

"If you were caught, would you be happy?" Rosemary asked.

"I'm handling it carefully," Hinton said, "The neck is very fragile."

He transferred the snake from one hand to the other, then extended the first palm. The palm was sticky with the musk of the snake.

"Smell it," he said.

We took turns smelling it. The smell was strange: mineral, spicy, reptilian, and chemical, unmistakably biological.

"You've got it all over your shirt!" Rosemary said.

"But I had to catch it!" Hinton explained.

He put the snake down, and it slithered into the bushes. He watched it leave with satisfaction.

"Okay, the weather is really nice today. Shall we bravely go outdoors and explore?" he said.

Author: Joshua Rothman

Translator: Lemon | Proofreader: Fluffy Bunny Paper

Formatting: Fluffy Bunny Paper | Cover: Call Me Death From Above

Original article:

https://www.newyorker.com/magazine/2023/11/20/geoffrey-hinton-profile-ai

免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。