Preface: DeepMind.

Demis Hassabis was born in North London in 1976 to a Greek Cypriot father and a Chinese Singaporean mother. He described the family as “quite bohemian,” which I’m guessing is the British equivalent of “hipster.” According to Demis, his dad did a lot of different things, including working as a singer-songwriter with aspirations of being like Bob Dylan, working as a teacher, and running a toy shop with his mom. His mom moved to Britain in the early 1970s. She originally trained as a nurse for children with special needs, later worked as a manager at the department store chain John Lewis, ran the family toy shop, and also worked as a teacher.

While Demis became a scientist, his siblings followed their parents’ more artistic path: his sister is a composer and pianist, and his brother, George, is a professional poker player who studied creative writing.

Hassabis has noted that his parents were not interested in computers and did not particularly like them, which he believes forced him to be self-reliant in his technical pursuits. He credits his mother’s roots for his early exposure to technology: the family spent summers in Singapore, where he encountered “futuristic” gadgets and Nintendo games that were not yet available in the UK.

At age 13, Hassabis was the second-highest-rated player in the world in the Under-14 category, only 35 Elo points behind the top-ranked player, Judit Polgár, who is widely considered the greatest female chess player of all time. He reached an Elo rating of 2300 by age 13, which is the “Master” standard. For perspective, the average casual chess player typically has a rating between 400 and 800.

Hassabis often credits chess as the reason he became obsessed with artificial intelligence. He has said that playing at such a high level forced him to look “under the hood” of his own mind. He would ask himself questions like: “Why did I make that mistake?” or “How am I visualizing these moves five steps ahead?” He was so good at thinking that he became curious about why he was so good at it: What is it that I’m doing when I think? How does that make me better? Why am I different?

Geniuses always realize their difference from a young age. Contrary to popular belief, though, while it is confusing to watch a genius do his work from the outside, it is equally confusing being the genius himself and not knowing where this genius-ness is coming from. He knew he was good at thinking, exceptionally good, yet he didn’t know why he had this gift. What is this gift that he clearly has that other people don’t?

Most eight-year-olds have to beg their parents for a computer. Hassabis didn’t have to. He used the prize money he won from a chess tournament to buy his first machine, a ZX Spectrum 48K. That was 1984, when he was just eight years old.

1984 was the height of the home computing revolution in the UK. The ZX Spectrum, developed by Clive Sinclair, was the affordable, rubber-keyed machine that essentially taught an entire generation of British engineers how to code. Because he bought it himself, he didn’t have to follow anyone else’s rules. He spent his time typing in code from magazines and books, which is how he began moving from playing chess to thinking about how a machine could be programmed to “think” like him.

It’s fascinating to think that the same brain that was winning chess prizes in 1984 to buy a 48KB computer would, exactly 40 years later in 2024, be awarded the Nobel Prize in Chemistry for using AI to map the “building blocks of life.”

In the documentary The Thinking Game, which premiered at the 2024 Tribeca Film Festival, Demis Hassabis describes the moment he decided to give up on chess. During one of his final competitions, he started to wonder what could be done with all the brainpower in the room, surrounded as he was by some of the best players in the world. What if all that brainpower went into solving real problems, like cancer, instead of being spent on a board game? A deep sense of meaninglessness swallowed him, and he started to feel sick. Right there, on the spot, he decided to give up on chess.

When Hassabis stepped back from competitive chess at 13, he didn’t stop being competitive; he simply shifted his focus to programming, simulation games, and early AI experiments. He taught himself Assembly language, which allowed him to write code that ran directly on the computer’s processor, making his programs much faster and more efficient. He was part of a small group of teenagers in the late 1980s who “hacked around,” making their own games and software tools.

A few years after the Spectrum, he upgraded to a Commodore Amiga, which was significantly more powerful. He bought the Computer Chess Handbook by David Levy to learn about search algorithms and evaluation functions, and used these concepts to write his first “real” AI program: a version of the board game Othello (also known as Reversi).

Writing that Othello program was Demis’s first real “Frankenstein moment,” the first time he built something that could outthink a human. Since he couldn’t just download an AI library like developers do today, he had to build the “brain” from scratch using the fundamental logic he had learned from chess.

To make the computer “think,” Demis used a classic AI concept called Minimax. In simple terms, Minimax is a decision-making algorithm used in two-player games like chess, tic-tac-toe, or Othello to find the optimal move. The name comes from its core goal: minimizing the maximum possible loss. It assumes your opponent is just as smart as you and will always make the best move for themselves.

Imagine a tree where the “trunk” is the current board position, and every “branch” is a possible move. The computer branches out every move it could make, then every move the opponent could make in response, and so on. Even on that limited hardware, Demis tried to make the computer look four or five moves ahead. In a game like Othello, where the board can flip almost entirely in a single turn, this required massive computational efficiency.

You want to choose the move that leads to the highest score (the “Max”). Your opponent wants to choose the move that results in the lowest score for you (the “Min”). The algorithm works backwards from the future. It looks ahead several turns, assigns a score to those future boards (for example, +10 for a win, −10 for a loss), and then “bubbles” those scores back up to the present. If a future branch leads to a guaranteed loss because your opponent is smart enough to take it, the algorithm discards that path and picks a safer one.

Think of Minimax as a hyper-cautious chess player who assumes their opponent is a literal genius. It doesn’t just look for a way to win; it obsessively looks for ways it might lose. Imagine you have two choices, Move A and Move B. After Move A, your opponent can either let you win (+10) or trap you (−10). Since Minimax assumes the opponent isn’t a fool, it scores Move A as −10. After Move B, no matter what the opponent does, the game stays tied (0). Minimax will choose Move B. Even though Move A could lead to a win, the algorithm discards it because it sees a guaranteed way for the opponent to make you lose. It prefers a boring tie over a risky gamble.

This is a very interesting strategy, and one I always use. I always assume my opponent is just as smart as I am, or even smarter. In reality that’s rarely true, but thinking this way yields me the best move, though it does take more work and brainpower. On top of that, I’m self-aware enough to know I’m quite a lazy person. I always try to do the least amount of work that yields the maximum results. I’m not greedy. I’m a very cautious and safe strategist; contrary to popular belief, I try to avoid too many rounds of back-and-forth. In real life, when I strategize, I usually go for the move that leaves my opponent with no comeback. Minimax makes a lot of sense to me. I just never summarized the way I strategize like this.

The problem with Minimax is what’s called the “state space explosion.” Think of it as the curse of possibilities. It is the primary reason why computers can easily beat humans at tic-tac-toe but struggled for decades with a game like Go. It’s what happens when the number of possible configurations of a system grows so rapidly that even the world’s fastest supercomputer runs out of memory and time to look at them all.

In tic-tac-toe, there are only 255,168 possible games. A computer can calculate every single one in a fraction of a second. In chess, the number of possible games is larger than the number of atoms in the observable universe.

It reminds me of chaos theory, although the two are not quite the same thing. Chaotic systems are deterministic, which means that given the same initial conditions, the output is always the same. But when the initial conditions change even a little, the final results can be drastically different. The weather is the classic example: it is so sensitive to its starting conditions that to predict it, you would need to measure them exactly, with no error at all, and that is impossible. In this sense, Earth itself could be a computer, and the weather a calculation so large that only the planet itself can run it. A chess game is different. Its rules are perfectly known and nothing is mismeasured. The difficulty is purely the number of branches. But both end in the same place: the future is technically determined, and still practically impossible to compute.

Because the computer couldn’t see to the end of the game, young Demis couldn’t use “pure” Minimax. He had to use heuristics, or shortcuts. A computer can see the moves, but it doesn’t know which board position is “good” or “bad” unless you tell it. So Demis wrote an evaluation function, a set of rules that gave points for certain positions. He essentially translated his own strategic intuition into a scoring system.

In chess, this would mean telling the computer: “Even if the game isn’t over, controlling the center of the board is worth +5 points.” Controlling the center is the golden rule of chess, but it’s only the foundation. Once you have a foothold in the center, you use that leverage for more specific strategies. The most common way beginners lose is by leaving their King in the middle of the board, so castling matters: it moves the King to a corner protected by a wall of three pawns and brings a Rook toward the center. A piece is only as good as the squares it controls. Knights are short-range, so to be effective they need to be deep in enemy territory. An “outpost” is a square, usually on the 5th or 6th rank, that is protected by one of your pawns and cannot be attacked by an enemy pawn; a Knight sitting there is a nightmare for your opponent. Rooks are like snipers: they hate being stuck behind their own pawns, so you want them on “open files,” columns with no pawns, where they can slide all the way into the enemy’s back row.

Grandmaster Philidor famously said, “Pawns are the soul of chess.” They aren’t just fodder; they are the terrain of the battlefield. A diagonal chain of pawns protecting each other forms a wall that blocks enemy Bishops. An isolated pawn, with no friendly pawns on the neighboring columns, is hard to defend because no other pawn can help it. Doubled pawns, two pawns of the same color on the same column, block each other and are usually a liability. Every one of these ideas can be turned into a number, given a weight, and ranked against the others. That is what an evaluation function is.

Othello has its own version of these rules. The most important is owning the corners: a disc in a corner can never be flipped, so it anchors everything around it. Demis gave his program points for corners and for the other positions he knew from experience to be strong.

Demis sat his younger brother George down in front of the Amiga and had him play against the code. When the program started beating George, Demis realized he had created a digital version of his own strategic mind. It was a profound moment: intelligence wasn’t just a “human” thing. It was a process that could be coded. This is exactly what DeepMind would eventually disrupt. In those days, Demis had to hard-code the “points” himself. With AlphaGo, the AI taught itself what a good position looked like. ☀️