Artificial Intelligence

Bob: The concept of creating "thinking machines" goes back centuries—philosophers long debated what such machines might look like and what they would be capable of. However, it wasn't until the 1940s, with the advent of the first digital computers, that the hardware necessary for AI development became available. Then, in 1950, Alan Turing published his landmark paper Computing Machinery and Intelligence, where, as we know, he introduced the Imitation Game as a way to answer the question "Can machines think?"

Hanne: Right.

Bob: Turing's paper is widely considered the foundational text of the field of AI, and he is regarded as one of the field's pioneers. Also, it's worth noting that in 1950, computers were essentially just number crunchers and were nowhere near capable of passing his Imitation Game. However, Turing made it clear that he wasn't asking whether current computers could succeed in the game, but whether future ones might:

… we are not asking whether all digital computers would do well in the game nor whether the computers at present available would do well, but whether there are imaginable computers which would do well.

His ability to foresee the potential of digital computers was remarkable.

Hanne: Impressive!

Bob: Furthermore, Turing outlined how one should go about building a machine to succeed in his Imitation Game. He suggested that instead of trying to build a complex machine from scratch, one should begin with something much simpler and then train it through education and experience:

Instead of trying to produce a programme to simulate the adult mind, why not rather try to produce one which simulates the child's? If this were then subjected to an appropriate course of education one would obtain the adult brain.

This was essentially the first recorded description of what we now call "machine learning"—a branch of artificial intelligence that involves programming a system to a simple starting state and then training it with data and practice so that it improves over time. Machine learning is central to the field of artificial intelligence today, and Turing's ability in 1950 to foresee how it should be implemented further highlights his great vision.

Alice: And interestingly, around the same time, Claude Shannon also demonstrated one of the earliest examples of machine learning with his electromechanical mouse, Theseus. This mouse could navigate a maze by remembering dead ends and eventually find its way out. Then, when placed in the same maze again, it could solve it in one go. Maze-solving AI competitions inspired by this concept still exist today.

Bob: Right. Shannon also published a paper in 1950 about how to program a computer to play chess. Programming computers to play chess was one of the first things that researchers were interested in when digital computers first appeared. In fact, Alan Turing, working with his colleague David Champernowne, actually wrote the world's first chess program—Turochamp—around this time as well. However, it was never executed during Turing's lifetime because no computer powerful enough to run it was available back then.

(Hanne thinks.)

Hanne: Wait… so who was first in these early AI developments? Was it Turing or Shannon?

Bob: Well, it's difficult to say who was first, even if we look at the publication dates. For instance, Shannon's paper on programming a computer to play chess was published in March 1950, while Turing's Computing Machinery and Intelligence paper was published in October of that same year. However, Turing had been giving lectures on the ideas discussed in his paper as early as 1948. Similarly, Shannon publicly demonstrated his Theseus mouse in 1951 but would have started working on it years earlier.

Hanne: Oh.

Bob: So, around this time, Turing and Shannon were working in parallel, and we don't really talk about who was first. We usually just say that around 1950, plus or minus a couple of years, these early developments in AI began to emerge.

Hanne: I see. And Alan Turing and Claude Shannon were the leading figures in the field at the time.

Bob: Exactly.

Alice: The next major development in artificial intelligence came in 1956 with the Dartmouth Workshop. This is often regarded as the "birthplace of AI", and it is considered the moment when artificial intelligence officially emerged as a distinct field. The workshop was organised by four American researchers:

• John McCarthy
• Marvin Minsky
• Nathaniel Rochester
• Claude Shannon

Their goal was to bring together academics interested in the idea of thinking machines with the hope of collaborating and advancing this new area of study.

Bob: The Dartmouth Workshop was especially significant because it set in motion many of the first dedicated AI research programs in the United States. John McCarthy was the main figure involved in planning the workshop, and he was also the one who coined the term "artificial intelligence" to describe the field. Because of this, as well as his early contributions to the field, he is considered to be "the father of artificial intelligence".

Hanne: Oh!

Alice: Both John McCarthy and Marvin Minsky would later go on to win the Turing Award for their contributions to AI—more about the Turing Award later. As for Turing himself, he had passed away in 1954 and so could not attend the workshop, but his ideas would have deeply influenced the discussions and ambitions of those present.

Hanne: Hmm. How old was Turing when he died?

Alice: He was 41 years old.

Hanne: Oh, that's young.

Alice: It is. We'll say a bit about the circumstances surrounding his death later. For now, the main takeaway from this part of the discussion is that the 1956 Dartmouth Workshop is considered the founding moment of artificial intelligence as a field of study.

Hanne: Okay. Got it.

Alice: Great. Next up, a few words about neural networks—a domain of artificial intelligence that uses computational models inspired by the human brain to solve problems. For this bit, we'll need to quickly jump back to the 1940s when researchers began exploring the possibility of artificial brains. Basically, they wondered that if the brain works through a network of neurons, would it be possible to build something similar for machines?

Hanne: Interesting.

Alice: In 1943, American neuroscientist Warren McCulloch and young logician Walter Pitts came up with a mathematical model of a single neuron. We won't go into all the technical details, but essentially, their model produced an output when its inputs passed a certain threshold. At this point, their work was still fully theoretical—but it planted the idea that maybe, by connecting many such neurons, machines could recognise patterns the way brains do.

Hanne: Huh!

Alice: Building on this, in 1957, American psychologist Frank Rosenblatt took the next step. He connected many such artificial neurons into a full network, which he called the "perceptron". His perceptron could learn to recognise simple visual patterns—like distinguishing vertical lines from horizontal ones—by adjusting the strengths of its connections, or "weights", based on examples.

Bob: Right. But the way it learnt was limited. The perceptron was a single-layer network, which meant it could only adjust the weights connected to its final output. This worked for simple problems, but it struggled with more complex tasks. By the late 1960s, critics highlighted these limitations, and interest in neural networks faded.

Hanne: Oh.

Alice: However, interest didn't completely disappear. In 1974, American researcher Paul Werbos theorised that if a network had multiple layers instead of just a single layer, it would be capable of tackling more complex problems. He introduced a smarter training method called "backpropagation", which was a way for the model to learn from its mistakes by sending error information back through the network layer by layer.

Bob: You can think of a multi-layer neural network as a network of several layers of decision-makers, each passing information to the next. Backpropagation meant that if the final decision was wrong, the blame could be passed backward so every layer could learn from the mistake—like the whole crew learning from an error instead of just the captain. However, at the time, this notion of backpropagation was still theoretical, and it wasn't clear that it could work in practice.

Alice: But in 1986, a breakthrough came when British-Canadian computer scientist Geoffrey Hinton—widely regarded as "the godfather of AI"—together with American psychologist David Rumelhart and American computer scientist Ronald Williams, published their famous paper Learning representations by back-propagating errors. This paper provided the first convincing demonstrations that backpropagation could actually train multi-layered networks in practice. They showed it could solve problems the perceptron couldn't—such as recognising handwritten digits and distinguishing between different types of shapes.

Hanne: Cool!

Bob: Now, given these developments, you might think that from the 1980s onwards, neural networks took off and became central to the field of AI. But… that was not to be. Despite their encouraging results, many researchers remained sceptical; they felt neural networks would be too slow and unlikely to scale to harder problems. On top of that, it wasn't exactly clear how best to train them or what kind of results to expect.

Hanne: Oh… poor neural networks.

Bob: Still, Geoffrey Hinton and a handful of others remained convinced of their potential. Years later, neural networks would make a comeback and would only then become central to the field of artificial intelligence.

Hanne: Yay!

Alice: So, that was a glimpse into the domain of neural networks and some of its key developments. However, it's important to note that there were many other individuals who also made notable contributions. Furthermore, titles like "the father of AI" for John McCarthy or "the godfather of AI" for Geoffrey Hinton are catchy, but they tend to oversimplify things. While these figures were influential, progress in AI has always been a collective effort shaped by numerous researchers from around the world. So, we encourage you to further explore developments in the subfield of neural networks to learn more about all the contributions made.

Hanne: Okay, will do. Neural networks seem fascinating.

Alice: They certainly are!

Bob: Next, let's look at how the broader field of AI had been developing between 1950 and 2000. Basically, what you'll find is that AI went through cycles of excitement and disappointment. This was because people expected computers to become intelligent very quickly, but progress turned out to be far slower and harder than anticipated.

Hanne: Hmm… why was that?

Bob: Well, when digital computers first appeared, they could perform mathematical calculations much faster than humans. The natural assumption was: if machines can outperform humans in math, surely they'll soon outperform us in other areas—like language translation and speech recognition. As a result, researchers gave optimistic predictions about when these breakthroughs would happen. But… it turned out that tasks that were easy for humans to carry out proved to be extremely difficult for machines to conquer, especially when those tasks required additional context.

Alice: For instance, in the 1960s, during the Cold War, the U.S. invested heavily in machine translation systems to quickly translate Russian documents. However, these systems failed to deliver. They had tiny vocabularies, struggled with grammar, and produced laughable results. In fact, there's a famous anecdote about the English sentence: "The spirit is willing, but the flesh is weak," being translated into Russian and then back into English, returning, "The vodka is good, but the meat is rotten."

Hanne: Oh no.

Bob: Speech recognition also ran into similar problems. Early military projects wanted pilots to talk naturally to their systems so they wouldn't need to adjust the controls manually. However, the machines could only handle rigid, pre-set phrases, and if a pilot said something slightly different from the expected command, the system failed completely.

Alice: The struggle continued in image recognition as well. A professor at an American university in the late 1960s even assigned a summer project for students to develop an algorithm to get a computer to identify objects in an image. The initial assumption was that this would be easy. But once the students began, they discovered the sheer difficulty of the problem—things like lighting, shadows, and perspective constantly messed up their algorithms. In the end, none of the students succeeded, and decades later, image recognition would remain one of AI's biggest challenges.

Bob: And even in robotics, the pattern was the same. Researchers could build machines that moved well in carefully controlled lab settings, but once placed in the real world, with its unpredictability, clutter, and change, robots failed. Tasks that were simple for humans—like picking up an object, opening a door, or walking across uneven ground—turned out to be extremely difficult for machines to master.

Hanne: Huh! So, AI was failing to deliver on multiple fronts.

Bob: Exactly. These repeated setbacks led to many ambitious promises being unfulfilled, and by the 1970s, governments and funding agencies were growing frustrated. In Britain, the government commissioned the Lighthill Report (1973), which reviewed AI research and concluded that progress was disappointing outside of a few niche areas like game-playing. As a result, AI funding in Britain was cut almost entirely. In the U.S., many projects had also lost major support. These cutbacks, combined with waning media interest, resulted in the first AI winter—lasting from the mid-1970s into the early 1980s.

Hanne: An AI winter! What a thing!

Bob: Yes. But in the 1980s, excitement returned with things called "expert systems", which were programs designed to capture the decision rules of human specialists. Around the same time, Japan launched its Fifth Generation Computer Project (1982)—a hugely ambitious government-backed program to build machines capable of reasoning and natural language conversation. This sparked a worldwide wave of investment, as the U.S. and Europe didn't want to fall behind. And for a while, AI began to look commercially promising.

Hanne: Ohhh…

Bob: But in the end, these expert systems turned out to be expensive and hard to maintain. By the late 1980s, the hype collapsed, funding dried up again, and the second AI winter set in—lasting from the late 1980s into the mid-1990s.

Hanne: Two AI winters!

Bob: That's right. Though, it's worth noting that some historians argue that "AI winter" isn't the best term. Even in those down periods, research didn't completely die—and some sub-fields of AI actually grew! What really declined was funding, media hype, and public belief in AI.

Hanne: Ah, I see.

Bob: So, in a nutshell, between 1950 and 2000, AI gained a reputation for overpromising and underdelivering. This scepticism carried over into the early 2000s—but slowly, the field began to deliver on some of those long-standing promises.

Hanne: Right. AI spring was coming!

Bob: Well, it is not called AI spring. For modern breakthroughs the term "AI boom" is used.

(Hanne thinks.)

Hanne: Hmm… boom… bloom. Well, in spring, the flowers bloom, and then they—boom! So, technically, AI boom is just another way of saying AI spring.

Bob: Huh. I never really thought about it like that. Interesting perspective.

Hanne: Indeed.

Alice: Alrighty, Hanne. So, as we move away from the early struggles of artificial intelligence, let's share with you some of the key breakthroughs in the field. The ones that we'd like to discuss are:

• Deep Blue
• Watson
• AlexNet
• AlphaGo

Are you ready for them?

Hanne: Hanne is ready!

Alice: Awesome! Here we go! Like we mentioned, from the onset, researchers were fascinated by the idea of teaching computers to play chess. In the 1960s, chess engines could barely compete with beginners. By the 1970s, they reached strong amateur strength. By the 1980s, some were good enough to challenge professional players—though still far from world champion level. Basically, progress was steady but slow.

Bob: Right. But there were still those who doubted whether computers would ever advance far enough to match the best humans at chess. And for good reason. The number of possible games of chess is astronomically large. In his 1950 chess paper, Claude Shannon estimated that the number of possible games was at least \(10^{120}\)—a figure now known as "Shannon's number". However, this was only a rough estimate, and the actual number of the total possible games is much, much larger.

Hanne: Wow!

Bob: So, programming a computer to find the perfect solution from the very first move was out of the question—there were simply too many possibilities.

Hanne: No kidding.

Alice: But then came the 1990s with IBM's Deep Blue project. Deep Blue was a multimillion-dollar chess-playing machine built by a team of computer scientists and chess experts at IBM. Its target opponent was Garry Kasparov, the reigning world chess champion at the time. Kasparov, however, wasn't just another strong player—he had been the undisputed world chess champion since 1985 and was widely considered to be the greatest player in the history of the game! This made him the ideal candidate for a man versus machine match.

Bob: The first clash took place in 1996 in Philadelphia. Deep Blue could calculate 200 million positions per second and see deeper into lines of play than humans could. Kasparov lost the first game. He later described the strange feeling of facing a machine that showed no emotion, no nervousness, and no fatigue. For a moment, he wondered if the age of human dominance in chess was already over.

Hanne: No…

Bob: But Kasparov regrouped and went on to win the six-game match 4–2! Human dominance in chess was safe.

Hanne: Wonderful!

Bob: Still, the fact that a computer had beaten Garry Kasparov in even a single game was historic—it sent shockwaves through the chess world.

Hanne: Oh.

Alice: Then, in 1997, IBM returned with a stronger version of Deep Blue. This time, the six-game match was played in New York, with extensive media coverage and millions following it like a boxing fight.

Hanne: Of course—this is humanity on the line! Who won?

Alice: Deep Blue won game 1. Kasparov struck back in game 2. Games 3, 4, and 5 ended in draws, leaving the match tied 2½-2½ heading into the final game.

Hanne: Ahhhh!

Alice: And… Deep Blue prevailed, winning the final game and the match 3½–2½. This marked the first time in history a computer had beaten the reigning world chess champion in a full match.

Hanne: Oh no.

Alice: But the victory was clouded by controversy. Kasparov accused IBM of human intervention during critical positions and demanded a rematch. IBM refused, dismantled Deep Blue, and never released the code. This left lingering questions: Was it a pure machine victory, or chess played by a committee through a computer?

Hanne: Rematch!

Alice: Either way, the moment was historic. For decades, chess had been a symbol of human superiority over machines—and now Deep Blue had not just defeated the world champion but Garry Kasparov, considered at the time to be the greatest player in the history of the game.

Bob: From there, chess engines—as they are called—only grew stronger. By the 2000s, many were already stronger than the best grandmasters, and today, engines such as Stockfish play at a level far beyond any human. In fact, there are now separate competitions just for chess engines—since humans no longer pose a serious challenge.

Hanne: Rats! Humans shall conquer these machines once more. Mark my words.

Alice: Moving on, the next breakthrough came in 2011, when IBM's Watson computer competed on Jeopardy!—an American game show. In this show, contestants are given clues in the form of answers and must respond with the correct question. For example, if the clue is "This U.S. state is known as the Sunshine State," the correct response would be "What is Florida?"

Hanne: Okay.

Alice: Now, while this might sound straightforward, the show is filled with wordplay, puns, cultural references, and clues that require both broad knowledge and the ability to interpret language quickly and accurately. Human champions rely on memory, intuition, and reasoning—all under time pressure.

Bob: And that's what made Watson's challenge so ambitious. Up to that point, computers had proven they could excel at structured tasks like chess, where the rules are clear. But language was another matter—human speech and writing are full of double meanings, context, and subtle cues. Because of this, many felt that a computer would never be able to compete with humans on Jeopardy!.

Hanne: Right.

Bob: Still, engineers at IBM took up the challenge. They designed their Watson machine—named after Thomas J. Watson Sr., IBM's first CEO—to take on the best human players at Jeopardy!. Watson used natural language processing, information retrieval, and machine learning to weigh possible answers and decide when to buzz in. To build its knowledge base, the team fed it millions of documents—everything from encyclopedias to dictionaries to news articles. Eventually, after years of development, Watson was ready to compete.

Alice: The contest took place over three televised episodes in February 2011. Watson faced two of the greatest Jeopardy! players of all time: Ken Jennings, who had the longest winning streak at 74 games, and Brad Rutter, the all-time highest money winner. And… Watson didn't just win—it dominated, earning $77,147 to Jennings' $24,000 and Rutter's $21,600.

Hanne: Oh… oh no.

Alice: The event was staged as a one-time showcase, but Watson's victory made headlines around the world. Some saw it as thrilling proof of AI's growing capabilities, while others were uneasy, wondering what it meant for the future of jobs that involved processing information. Ken Jennings even joked in his final answer, writing: "I for one welcome our new computer overlords."

Hanne: Yikes!

Alice: But beyond the spectacle, the significance was clear: Watson demonstrated that AI could tackle the messy, ambiguous world of human language—a domain long thought to be out of reach for machines.

Hanne: Right.

Alice: The next breakthrough came in 2012, when a system called AlexNet won the ImageNet Large Scale Visual Recognition Challenge. This challenge was an annual competition launched in 2010 by American computer scientist Fei-Fei Li. It brought together researchers from around the world to test computer vision systems on a massive dataset of over a million images—everything from animals to vehicles to household objects. The task was simple for humans but extremely difficult for machines—look at a picture and say what's in it.

Hanne: Ah, good old image recognition.

Alice: For the first few years of the challenge, progress was steady but modest. The best systems relied on handcrafted features and traditional machine learning techniques. Accuracy was improving little by little, but the consensus in the field was that computers just weren't very good at recognising images the way humans were.

Bob: And while neural networks had been around in theory for decades, most researchers had written them off as impractical—they were too slow, too hard to train, and didn't scale well. But all that changed in 2012 with the development of AlexNet.

Alice: This is where Geoffrey Hinton, "the godfather of AI", comes back into the story. Recall we mentioned that he was one of the few who kept on believing in the potential of neural networks even though they had fallen out of favour since the 1980s.

Hanne: Yep.

Alice: Well, in 2012, he was a professor at a university in Canada. One of his PhD students, Ilya Sutskever, approached him with the idea of applying deep neural networks to the ImageNet challenge. Sutskever and Hinton explored the idea but soon realised that building a neural network to compete in the ImageNet challenge was going to be a monumental task. Sutskever then brought his friend Alex Krizhevsky—who was also one of Hinton's students—on board.

Bob: After further exploration, the team decided to try something unconventional. Instead of using traditional CPUs—central processing units—which could handle general computing tasks but were slow for heavy computations, they trained their neural network using GPUs—graphics processing units. GPUs were originally designed for video games and 3D graphics, but they could also perform massive numbers of mathematical operations in parallel, making them perfect for training a large network efficiently.

Alice: Alex Krizhevsky, the one most capable of building such a network, became the lead architect and wrote most of the code to train the massive network on GPUs. This is why the system, AlexNet, was named after him. However, Geoffrey Hinton and Illya Sutskever made major contributions as well.

Bob: And in the 2012 ImageNet Challenge, AlexNet crushed the competition! Its error rate was nearly half that of the next-best system—a gap so large that it immediately overturned the widespread scepticism about neural networks. This was a pivotal moment in AI, as it proved that with enough data and computational power, deep neural networks could achieve human-level performance on tasks once thought to be beyond machines. In fact, nearly all of the subsequent AI developments can be traced back to the AlexNet breakthrough.

Alice: Right. The AlexNet breakthrough in 2012 essentially kicked off the modern AI boom. This boom, however, was mostly limited to academic and research circles until 2022, when the global AI boom truly began.

Hanne: I see… the boom before the BOOM!

Alice: Precisely. Hinton, Krizhevsky, and Sutskever documented their work in a paper titled ImageNet Classification with Deep Convolutional Neural Networks and presented it at a conference later that year. Soon after, they formed a small startup called DNNresearch. Google acquired their company in 2013 and brought the trio into its growing artificial intelligence division.

Bob: Geoffrey Hinton remained at Google until 2023 and was a key figure in the development of Google Brain, a deep learning research project aimed at scaling neural networks and applying them to real-world Google products. However, after leaving Google in 2023, he expressed regret about some of his contributions to AI. Specifically, he voiced concerns about the potential misuse of artificial intelligence, especially in areas like surveillance, weaponisation, and the spread of misinformation. But for his contributions to AI, he was a co-recipient of the Turing Award in 2018 and, interestingly, the Nobel Prize in Physics in 2024.

Alice: As for Alex Krizhevsky, he left Google in 2017 after reportedly losing interest in his work and has since kept a low profile. Illya Sutskever, on the other hand, left Google in 2015 to cofound a company called OpenAI and remains more active in the public eye.

Hanne: So, basically, these three figures—Geoffrey Hinton, Alex Krizhevsky, and Illya Sutskever—developed the AlexNet system, which in turn played a major role in shaping modern AI. They then went on to work for Google but eventually left for different reasons.

Alice: Correct.

Bob: Finally, the last AI breakthrough that we'd like to delve into is AlphaGo—a system built to compete in the ancient Chinese board game, Go. This game dates back over 2,500 years and seems simple on the surface—players place black or white stones on a grid, aiming to surround more territory than their opponent. However, the strategy is incredibly deep, and the number of possible board positions in Go is far greater than in chess! Additionally, Go relies heavily on intuition, creativity, and long-term strategy. Experts believed it would be decades before a computer could beat a top human player—some even thought it might never happen.

Hanne: Yeah… heard that one before.

Bob: But then came DeepMind—a leading artificial intelligence research company founded in London in 2010 and acquired by Google in 2014. They took on the challenge of building a system to compete in Go, and they called it AlphaGo.

Hanne: Here we go…

Bob: In October of 2015, AlphaGo defeated the European champion Fan Hui 5-0, marking the first time a computer program had beaten a professional Go player.

Hanne: Well… that was quick.

Bob: However, many Go players felt that Fan Hui was past his prime and that AlphaGo's victory wasn't all that impressive.

Hanne: Oh.

Bob: So, AlphaGo's next challenge was to play against one of the strongest Go players at the time—Lee Sedol. Many expected Lee Sedol to win easily. Lee was also very confident going into the five-game match, curious about whether he would win 5-0 or 4-1. For many in the Go community, the idea that a machine could compete with the best human players in Go was absurd.

Hanne: Hmm…

Bob: The match took place in March of 2016 in Seoul and drew millions of viewers from around the world. The first game was a hard-fought battle, but AlphaGo prevailed. It was a shock to Lee Sedol and the Go community. But Lee, now more aware of the capabilities of AI, regrouped for the second game. However, AlphaGo won again! In fact, the second game became famous for AlphaGo's Move 37. It was an unconventional move, with some even thinking it was a mistake when it was played. Many commentators said that it was a move that no human player would ever make. AlphaGo agreed, estimating that there was a 1 in 10,000 chance that a human would have played that move. However, it became clear later in the game that Move 37 was brilliant.

Alice: Right. And it just so happened that before AlphaGo's Move 37, Lee Sedol had stepped away for a break. When he returned, he was taken aback by the move, and it took him a while to process it. He later commented that it was this move that made him realise AI could be creative.

Hanne: Interesting.

Bob: And the strategy that AlphaGo used in Move 37 has now become standard strategy in Go. In fact, several techniques introduced by AI have now become standard strategy in Go—reshaping the way the thousand-year-old game is played. … But getting back to the match with Lee Sedol, AlphaGo went on to win 4-1, completely shocking the Go community and marking the end of human dominance in Go.

Hanne: Noooo.

Bob: And just like with chess engines, more sophisticated Go-playing systems have since been developed to the point where they now far surpass human players.

Alice: Right. One of those more sophisticated systems that emerged after AlphaGo is called AlphaGo Zero.

Hanne: Wow! AlphaGo without any sugar. Amazing!

Alice: Ha! Not quite. We'll get to why it's called Zero in a moment. Basically, DeepMind introduced AlphaGo Zero about a year after AlphaGo's clash with Lee Sedol and had it play a 100-game match against the original AlphaGo.

Hanne: Okay. And what was the result?

Alice: 100-0. AlphaGo Zero won every single game!

Hanne: Yikes!

Alice: However, what's particularly fascinating about AlphaGo Zero is how it was trained. Usually, these game-playing AI systems are given the rules of the game, as well as a massive amount of data on past games and board configurations to analyse and learn strategy. After that, they improve further by playing many games against themselves.

Hanne: Right.

Alice: But with AlphaGo Zero, the researchers only gave it the basic rules of Go. It wasn't given any data from past games to analyse and learn from. This is why it was called Zero—it began with zero knowledge of Go strategy.

Hanne: Ohhh.

Alice: But with only this basic knowledge of the rules of Go, AlphaGo Zero played millions of games against itself, learning and refining its approach through trial and error. And remarkably, this approach resulted in a system that was much stronger than AlphaGo, which had a foundation of human knowledge to build on.

Hanne: Wow! That's kind of scary. The less human knowledge it had to start with, the better it did.

Alice: Yep. That was a big part of it. Though it's also worth noting that AlphaGo Zero benefited from more efficient learning algorithms and improved neural network architectures.

Hanne: I see.

Alice: The technique was then generalised to AlphaZero, which mastered Go, chess, and the Japanese game Shogi. At its core, it relied on neural networks working hand in hand with reinforcement learning, allowing the system to evaluate positions and improve entirely through self-play. This combination meant AlphaZero wasn't just another strong program—it was proof that a general approach, powered by neural networks, could master different complex games without needing human examples.

Hanne: Well… the field of AI really did begin to deliver on some of its longstanding promises.

Alice: It certainly did:

• Deep Blue in 1997
• Watson in 2011
• AlexNet in 2012
• AlphaGo in 2016

At this point, the holographic illustrations filled the room, reflecting the trio's discussion so far. From the first part on chatbots, Hanne saw depictions of everything from ELIZA and PARRY all the way through to the digital assistants like Siri and Alexa. There were also images related to their discussion on artificial intelligence—Turing and Shannon's early work, the Dartmouth Workshop, and the evolution of neural networks. Hanne even observed depictions of AI's early struggles unfolding in a snowy setting.

Next, she saw illustrations of the key AI breakthroughs. There was Deep Blue locked in a match with Garry Kasparov, Watson facing Ken Jennings and Brad Rutter on Jeopardy!, AlexNet dominating the ImageNet competition, and AlphaGo competing against Lee Sedol.

Interestingly, some of the illustrations blended fact with whimsical elements, adding a playful tone to the room. ELIZA, PARRY, Watson, and AlexNet were depicted with robot bodies that moved around the room as if alive. ELIZA went around asking questions; PARRY scampered nervously, hiding from the Mafia; Watson quizzed itself with Jeopardy! clues and answers; and AlexNet enthusiastically described every image it encountered. The other chatbots and AI systems were also given exaggerated traits while still being true to their nature.

Hanne spent a while interacting with the illustrations—speaking to some of them and playing games with others—fascinated by the unfolding evolution of artificial intelligence around her. It was an experience like no other.

Eventually, the illustrations settled across the room and no longer interacted with her. Hanne walked among them, reflecting on all that she had learnt and thinking about the connection between artificial intelligence, cryptography, and cybersecurity.

She thought about how these AI systems surpassed humans in games like chess and Go, and wondered whether they could also be designed to surpass humans in codebreaking. Recalling the Enigma and Lorenz machines, she felt that if there had been sufficiently advanced AI codebreaking systems during World War II, they might have found many more vulnerabilities in the cipher machines than the human codebreakers were able to uncover.

Her thoughts then transitioned to modern cryptosystems. Even though she hadn't yet explored them, she realised that they would have to be designed with both human and AI threats in mind. But she couldn't help but wonder how difficult this task might be—if AI could already outperform humans in different games, then surely, it should be able to find ways to breach any cryptosystem that humans create.

When she posed the question to Alice and Bob, they acknowledged that this was indeed a serious concern. However, they explained that AI has shown little promise in codebreaking and has not made any significant breakthroughs in compromising modern cryptosystems thus far. But at the same time, they admitted that they weren't exactly sure how much effort had been spent on developing AI codebreaking systems.

Alice and Bob then went on to give Hanne a little taste of modern cryptosystems. They explained that these systems essentially form the backbone of the internet and that their security relies on the fact that for many years humans have been unable to find sufficiently fast algorithms to solve certain mathematical problems. However, they pointed out that if sufficiently fast algorithms were developed to solve these problems, it would break modern-day cryptography! Things like e-commerce, online banking, private messaging, and social media would collapse, as credit card numbers, bank details, personal messages, and social media credentials would be easily accessible by anyone. Again, they reiterated that, as of now, AI has not been able to develop sufficiently fast algorithms or find other effective hacks to compromise these systems, but security researchers are actively working to mitigate this threat.

Finally, Alice and Bob encouraged Hanne to keep the threat of AI top of mind when they eventually move on to more modern cryptosystems and the broader field of cybersecurity. They pointed out that just as these AI systems were optimised to master games, they could one day be optimised to break cryptosystems and could wreak havoc across the digital world.

Hanne spent a few more moments reflecting on the discussion and admiring the whimsical illustrations around her before the trio resumed their discussion on AI and chatbots.