The AI Revolution: How Machines Learned to Think — and Why We Should Be Both Hopeful and Afraid

Machines that think like humans. Our dream to create machines in our own image goes back to antiquity. And now, for the first time, it’s no longer a dream.

But here’s the strange thing about this moment: I was researching an interview subject the way I’ve done a thousand times in a forty-year career — except this time I wasn’t typing search terms into a browser. I was having a conversation with a computer. It felt like something big had changed.

It had. And what follows is an attempt to understand how much.

The Dream Goes Back to Turing

The modern origins of artificial intelligence can be traced to World War II and the remarkable brain of Alan Turing. The British mathematician built a machine capable of deciphering coded messages from the Nazis. After the war, he was among the first to predict that computers might one day match the human brain.

In 1951, he gave a short lecture on BBC radio. “I think it is probable, for instance, that at the end of the century, it will be possible to program a machine to answer questions in such a way that it will be extremely difficult to guess whether the answers are being given by a man or by the machine.”

And so the Turing test was born: could anyone build a machine that could converse with a human indistinguishably from another person?

In 1956, a group of pioneering scientists spent a summer at Dartmouth College and coined a new academic field. They called it artificial intelligence.

For decades, that ambition ran far ahead of what computers could actually do. In 1978, NOVA released its first film on the subject, and Arthur C. Clarke — as always — was prescient. “It doesn’t really exist yet at any level, because our most complex computers are still morons. High-speed morons, but still morons. Nevertheless, we have the possibility of machines which can outpace their creators, and therefore become more intelligent than us.”

Rules Versus Patterns

Early AI was built on expert systems: purpose-built machines that followed rules. The idea was simple — put all our knowledge into a machine and give it instructions for what to do with it.

That approach hit its high-water mark in 1997, when IBM’s Deep Blue defeated world chess champion Garry Kasparov. It searched about 200 million positions a second, navigating a tree of possibilities to find the best move. Effective — but brittle. Deep Blue wasn’t strategising the way a human does.

The human brain, with more than 80 billion neurons, doesn’t learn by following rules. It learns by taking in a steady stream of data and looking for patterns.

“The way that learning actually works in the human brain is by updating the weights of the synaptic connections that are underlying this neural network,” explains Manolis Kellis, Professor of Computer Science at MIT. “We have trillions of parameters in our brain that we can adjust based on experience. I’m getting a reward — I will update the strength of the connections that led to this reward. I’m getting punished — I will diminish the strength of the connections that led to the punishment. So this is the original neural network. We did not invent it. We inherited it.”

The question was whether an artificial neural network could be built in the same image. Turing imagined it. But computers were nowhere near powerful enough — until the data sets arrived.

How a Machine Actually Learns

An artificial neural network uses interconnected nodes that communicate with each other. Each node receives inputs, processes them, and produces outputs passed on to other nodes. It learns by adjusting the strength of the connections between nodes based on the data it is exposed to.

Training lets it recognise patterns and learn from experience, the way a child does. A child learns fast because it is constantly predicting the future, seeing what happens, and updating the weights in its neural network accordingly. That same self-supervised learning paradigm can be applied to machines.

Some of the first networks were trained on vintage Atari games. Games reduce the complexity of the real world to a narrow set of possible actions. The AI is shown every pixel, frame by frame, and learns to associate pixels with actions. When it destroys blocks or shoots aliens, the connections that enabled that success are strengthened. When it fails, no reward. Eventually the reinforced connections overrule the weaker ones. The program has learned to win.

That is a very old trick. It’s how you train a dog. It’s how you teach a child.

AlphaGo and the Move Nobody Expected

In 2014, DeepMind — co-founded by Mustafa Suleyman, later acquired by Google — began work on AlphaGo, a network that could play the ancient board game Go.

“Go was thought to be a game where machines would never win,” Kellis says. “The number of choices for every move is enormous.”

But compute power was growing exponentially, and in some sense computation is a proxy for how intelligent the model is.

AlphaGo was trained two ways. First it was fed a large data set of expert games — supervised learning. Then it played against itself millions of times — reinforcement learning. Its skills gradually improved.

In March 2016, it faced Lee Sedol, one of the world’s top players, in Seoul. AlphaGo won. But more remarkably, it made a move so novel that Go experts thought it was a blunder. It turned out to be a stroke of brilliance. That emergent creative behaviour was a hint of what was coming: generative AI.

The Chatbot That Surprised Its Creators

Meanwhile, a company called OpenAI was building something that would become ChatGPT. It let users hold a dialogue with a machine that seemed uncannily human. An early version appeared in 2018; a later one became a global sensation in late 2022.

It surprised even the pioneers. “It surprised me a lot that they’re able to do things that we didn’t think they could do simply by learning to imitate how humans respond,” says Yoshua Bengio. “I thought this kind of ability would take many more years, or decades.”

ChatGPT is a large language model. LLMs consume vast amounts of text — books, articles, websites — and by recognising patterns across billions of words, they can guess the next word in a sentence. That is how they generate original answers.

“If you’re good at predicting the next word, it means you’re understanding something about the sentence,” Kellis explains. “What the style of the sentence is, what the feeling of the sentence is. And you can’t tell whether this was a human or a machine. That’s basically the definition of the Turing test.”

The Prosthetic Arm That Learned Him

Here the story became personal. As an arm amputee, the reporter had used a body-powered prosthetic for a decade — a harness and cable, technology more than a century old. AI, coupled with small electric motors, is finally dragging prosthetics into the twenty-first century.

At Coapt, a small company in Chicago, electrodes in a cuff read the faint electrical signals generated in the stump when the wearer imagines moving a phantom hand. The wearer trains a machine-learning model to connect those signal patterns to specific movements.

“We see eight different signals on the screen,” says CEO Blair Locke. “All eight sensor sites feed in together and let the algorithm sort out the data. What you’re experiencing is your ability to teach the system what hand-closed is to you. And that’s different than what it would be to me.”

The system is a Bayesian classification model. As the wearer trains it, it labels EMG patterns into movement classes — hand open, hand closed, wrist rotation. Then it compares live signals to that library and uses statistical probability to pick the best match.

“Is it me that’s learning, or the algorithm that’s learning? Or are we learning together?” the reporter asks.

“You’re learning together.”

Spotting Cancer Before the Human Eye Can

Computer scientist Regina Barzilay started working on AI in the 1990s, when rule-based systems like Deep Blue were giving way to neural networks. She used the techniques to decipher dead languages — a small language model, intellectually challenging but not life-changing.

Then her life changed. In 2014 she was diagnosed with breast cancer.

Working with radiologist Constance Lehman at Massachusetts General Hospital, she identified the Achilles heel of the diagnostic system: the human eye. “What is the likelihood of these patients to develop cancer within the next five years? We, with our human eyes, cannot really make these assertions, because the patterns are so subtle.”

The tool they built, Mirai, uses a convolutional neural network — the kind that combs through an image with virtual magnifying glasses, looking for edges, shapes and textures, then assembling the puzzle. Trained on more than 128,000 mammograms from seven sites in four countries, Mirai is between 75% and 84% accurate at predicting future cancer diagnoses.

Then a friend of Barzilay’s developed lung cancer. Working with oncologist Lecia Sequist, they trained a model called Sybil on thousands of CAT scans. Sybil correctly forecast cancer between 80% and 95% of the time — and in some cases identified a problem on a baseline scan where the radiologist could see nothing at all.

“What is it seeing?” the reporter asks.

“That’s the million-dollar question,” comes the reply. “And maybe not the million-dollar question. Does it really matter?”

The point isn’t that the machine can explain itself. The point is that early prediction could shift medicine from treating advanced disease toward preventing it.

Proteins, Drugs, and 200 Million Structures

Back at DeepMind, the games were always a means to an end. In 2021 the company released AlphaFold, pattern-recognition software designed to help researchers understand proteins — long chains of amino acids involved in nearly every function in the body. How a protein folds into a three-dimensional shape determines how it interacts with other molecules. If you know the shape of a disease-causing protein, you can sometimes design a drug to disable it.

DeepMind trained AlphaFold on thousands of known protein structures, then used that knowledge to predict 200 million structures — nearly all the proteins known to science.

At Insilico Medicine, theoretical chemist Petrina Kamya uses AlphaFold and its own deep-learning models to design molecules. The process is, as she puts it, a bit like Airbnb: you set your criteria and the algorithm filters out the properties that don’t fit. Identifying the best candidates takes 48 to 72 hours of computing time. Asked how long it would have taken her as a computational chemist, she is honest: “I would have thought of some of these, but not all of them.”

Human trials remain unavoidable. But cheaper preclinical discovery could mean cheaper drugs, and could make neglected diseases worth pursuing.

Fire, Fraud, and the Edge of the Network

AI is also doing financial analysis, detecting fraud, discovering novel materials, and helping build clean energy technology. In Sonoma County, California, CAL FIRE has installed more than a thousand remotely operated cameras on mountaintops. A neural network trained with UC San Diego scientists, part of a programme called ALERT California, scans the video for early signs of smoke.

“The successes are the fires that you don’t hear about in the news,” says Chief Philip SeLegue.

Meanwhile, at MIT’s Computer Science and Artificial Intelligence Lab, Daniela Rus is working on making robots safe to work beside humans — using the same EMG-reading electrodes found in a prosthetic arm. The ambition is partnership: better reach, lift, precision, strength, vision. Physical superpowers through machines.

The obstacle is size. Most neural networks run on supercomputers occupying entire buildings. A self-contained robot body can’t carry that. The answer, developed at MIT, is liquid networks — an autonomous vehicle running on just 19 neurons, capable of driving in environments it has never seen before, without a connection to the cloud.

The inspiration? A worm called C. elegans, whose entire brain contains 300 neurons. Not the architecture we built our networks on — a different one entirely.

The Nightmare on the Other Side

Here’s where the hope curdles.

In 2018, Jordan Peele combined a pitch-perfect Obama impression with AI software to make a convincing fake video. “This is a dangerous time,” the fake Obama said.

Fakes are as old as photography. Mussolini, Hitler and Stalin all ordered images doctored, erasing those who fell out of favour. But as Berkeley computer scientist Hany Farid points out, you could once count on one hand the number of people in the world who could do this convincingly. Now you need almost no skill at all.

Farid’s team generated a fake newscaster in a day and a half. “It’s the classic automation story. We have lowered barriers to entry to manipulate reality. And when you do that, more and more people will do it. Some good people will do it, but lots of bad people will do it.”

How it works: generative adversarial networks pit two algorithms against each other in a dizzying student-teacher loop — one generates, one judges, and they iterate thousands of times until the output passes. Newer diffusion models work differently, corrupting labelled images with noise over billions of steps until the image is unrecognisable, then learning to reverse the process from pure noise back to a coherent image based on a text prompt.

“If you had asked me, ‘Will this work?’ I would have said no,” Farid admits. “It just doesn’t seem like it should work. And that’s the magic of getting this much data and very powerful algorithms and very powerful computing.”

In May 2023, a fake image of an explosion at the Pentagon, shared via what appeared to be a verified Bloomberg account, moved the stock market. Two minutes was enough to wipe half a trillion dollars off valuations. From a single fake image.

“We’re going to live in a world where we don’t know what’s real,” Farid says. “There is distrust of governments, there is distrust of media, there is distrust of academics. And now throw on top of that video evidence. So-called video evidence. I think this is the very definition of throwing jet fuel onto a dumpster fire.”

The Extinction Statement

Yoshua Bengio, one of the pioneers of the field, didn’t spend much time thinking about dystopia while he was building it. “The more I read, the more I thought about it, the more concerned I got. If we are not honest with ourselves, we’re gonna fool ourselves. We’re gonna lose.”

In May 2023 he signed a one-sentence public warning: “Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks, such as pandemics and nuclear war.”

Asked whether we can simply pull the plug, his answer is uncomfortably simple: yes — before it gets so powerful that it prevents us from pulling the plug.

Near-term dangers are already here. Deepfakes and misinformation. Supercharged bias and hate content. Displaced jobs. Easier routes to bioweapons. And systems so complex that nobody — including their creators — fully understands how they reach their decisions.

“Nobody really understands how those systems reach their decisions,” Rus says. “So we have to be much more thoughtful about how we test and evaluate them before releasing them.”

Governments are beginning to move. But the commercial incentive to win the race is immense — a predicted thirteen trillion dollars of economic impact by 2030 — and the tech industry has historically fought regulation hard.

“There’s no question that guardrails will slow things down,” Suleyman concedes. “But the risks are uncertain and potentially enormous. So it makes sense for us to start having the conversation right now.”

The Honest Ending

The reporter eventually went home with a new myoelectric arm and tried to use it. It worked — sometimes. The hook opened when it shouldn’t. The software complained that the electrodes weren’t making good skin contact. He reset it, retrained it, and tried again.

Amputees told him it can take months to really learn to use an arm like this one. The choke point isn’t the artificial intelligence. It’s figuring out the best way to communicate human intention to it.

Which is, in a way, the whole story.

Machine learning is not yet smart enough to give us Luke Skywalker’s hand. It is not yet capable of creating the Terminator. Many of our hopes and fears for AI are still rooted in science fiction.

But we are walking down a road to the unknown, and the door is opening.

Be hopeful. Be careful. And trust — but always verify.