New English Edition
Artificial Intelligence in Horticulture
Kim Kyung-jin, Attorney at Law
Across five chapters and ten sections, this book examines computer vision for crop diagnosis, harvesting robots and autonomous field systems, smart greenhouses and digital twins, precision irrigation and supply-chain quality control, high-throughput phenotyping, and predictive breeding.
New English Edition
Artificial Intelligence in Food Crop Agriculture
Kim Kyung-jin, Attorney at Law
Across six chapters and eighteen sections, the book examines digital agricultural infrastructure, remote sensing, crop diagnosis, yield forecasting, precision irrigation, genomics, molecular breeding, agricultural robotics, climate-smart agriculture, and global food security.
New English Edition
The Future of Forestry and Agroforestry
Kim Kyung-jin, Attorney at Law
Driven by Artificial Intelligence and Digital Innovation
Across five chapters and fifteen sections, the book follows satellites, drones, LiDAR, digital twins, forest-specific language models, wildfire and pest forecasting, forestry robotics, agroforestry, timber traceability, and forest carbon markets.
New English Edition
Smart Livestock Farming: AI Enters the Barn
Kim Kyung-jin, Attorney at Law
Sensors listen, cameras watch, and artificial intelligence helps farmers decide.
Across five chapters and fifteen sections, the book follows precision livestock farming from animal health and reproduction to robotic milking, virtual fencing, digital twins, methane reduction, welfare, and data ownership.
Table of Contents
Han Dong-hoon, Busan Buk-gu Gap: A Record of the 100 Days Before and After the Election (Mar. 26-Jul. 3, 2026)
Kim Kyung-jin
Table of Contents and 13 sections
From March 26 to July 3, 2026, this record follows the spring after expulsion, the Busan Buk-gu Gap by-election, victory as an independent, and the first bill submitted in the National Assembly.
Table of Contents
Artificial Intelligence and Medicine
Kim Kyung-jin, Attorney at Law
AI in clinical care, hospitals, education, and research
AI in medical imaging, risk prediction, treatment planning, hospital operations, education, and research, with patient safety, privacy, and accountability.
[AI Library] 11 The Atari Shock
Demis Hassabis, Father of Google's Artificial Intelligence
Part 4. DeepMind
11 The Atari Shock
Kim Kyung-ran, Kim Kyung-jin
AI that learns through games. Late one night in 2013, a strange mix of tension and fatigue hung in the air of DeepMind's small London office. Through the windows, the cold London streets were visible, but the eyes of Demis Hassabis and his colleagues were fixed solely on the flickering, rough pixels on their monitors. On the screen, an Atari 2600 console, a relic from the 1970s, was running.
There were no flashy graphics, no sweeping soundtrack. Just a crude 8-bit display. But what they were watching was not a simple retro game. It was a moment when the history of artificial intelligence, and indeed the very way humanity builds tools, was fundamentally changing. The AI agent they had built, a system that would later be called DQN (Deep Q-Network), was like a newborn baby.
The research team had not taught this AI a single rule of the game. They never told it which ships were enemies and which were allies, how to earn points, or even what it needed to do to clear a stage. The AI received only two things. One was pixel data from the screen, serving the same role as human eyes (visual data). The other was a signal indicating whether the game score (reward) was going up or down.
'Get a higher score.' That was the only instinct and command given to the AI. In the early stages, the AI was an absolute mess.
The paddle on the screen moved erratically, dodging incoming balls instead of hitting them, mashing buttons without purpose. It was worse than a three-year-old picking up a video game controller for the first time. But Hassabis and his team were not discouraged.
They were waiting for the moment when order would bloom from this random chaos. They watched with bated breath as the AI's brain, built on deep learning, combined with the carrot-and-stick mechanism of reinforcement learning. As time passed, something remarkable happened.
After thousands, tens of thousands of trials and errors, the AI began recognizing patterns in the pixels on screen by itself. 'When this white dot (the ball) comes
down, if I place my white bar (the paddle) underneath it, the number called the score goes up.' Without anyone teaching it, the AI inferred cause and effect on its own. This was an entirely different dimension from conventional computer programs that simply execute inputted commands.
It was 'learning.' The machine was reproducing the way humans learn about the world: the process of figuring things out through experience. This was the moment that proved the possibility of the 'general-purpose learning algorithm' Hassabis had long dreamed of.
An AI with hard-coded rules for a specific game, whether chess or Go, can do nothing beyond that game. Deep Blue, which defeated a chess champion, cannot even understand the rules of tic-tac-toe. But DQN was different. The exact same code that played 'Space Invaders' played 'Pong' and 'Breakout' (brick-breaking) without a single modification.
All they did was show it the pixels on the screen, and the AI figured out the physics and rules of each world on its own. This was the first step in AI's evolution from a specialist in closed worlds to a learner in open worlds. The tunneling strategy the AI discovered on its own in 'Breakout.' The highlight of this project came during the 'Breakout' game test.
In 2013, just before Google acquired DeepMind, Hassabis needed to prove DeepMind's value to Larry Page and Google's executives through this demo. And the moment of proof caught even AI researchers off guard. About ten minutes into training, DQN began catching the ball with decent skill. In human terms, it was at the level of a fairly competent amateur.
It returned every ball without missing, breaking bricks one by one. The researchers nodded. 'It's working well. It's reached human level.'
They thought this was a sufficient achievement. But the AI did not stop. Two hours into training, the AI's movements on screen shifted in a subtle way.
It was no longer satisfied with simply returning the ball. It began using a technique where it hit the ball precisely off the edge of the paddle to angle its trajectory. Four hours in, a scene unfolded that no one had taught it and no one had predicted.
The AI relentlessly targeted one end of the brick wall. It kept sending the ball to the same spot and finally punched through a vertical gap at the edge of the wall.
Once the hole opened, the AI sent the ball through that narrow gap. The ball entered the space behind the wall and ricocheted furiously between the wall and the ceiling. With a rapid-fire clatter, bricks collapsed in droves, and the score skyrocketed.
It was the 'tunneling' strategy. The secret technique known only to expert players, and the AI had found it on its own. The researchers watching froze in silence, then erupted in cheers.
This was not just about scoring points. It was evidence that the AI had engaged in 'strategic thinking.' 'Returning the ball right in front of me is the safe move, but if I take the risk and punch through one side of the wall, I can reap enormous rewards later.'
The AI grasped this complex chain of cause and effect and this long-term plan solely from the movement of pixels and changes on the scoreboard. Reflecting on this moment, Hassabis said, 'It was as if we were watching the AI think.' It was not code entered by a programmer. It was emergent intelligence, born from data and experience.
The 'tunneling' incident gave the DeepMind team conviction. Conviction that they were on the right path, and faith that someday this intelligence could punch tunnels through the massive walls of unsolved scientific problems, far beyond games. The 2013 Nature paper and the shock to the AI community. DeepMind's achievement was first unveiled at the 2013 NIPS (now NeurIPS) Deep Learning Workshop, then expanded and published as the cover story of the prestigious scientific journal Nature in 2015.
The paper was titled 'Human-level control through deep reinforcement learning.' The shock to the scientific world, and the AI community in particular, can only be described as exactly that: shock. The mainstream of AI academia at the time was still accustomed to the approach of having humans meticulously design rules by hand.
Others held the dominant view that deep learning excelled at static tasks like image recognition but could not work in dynamic environments requiring continuous decision-making, like games.
Reinforcement learning was theoretically elegant, but the consensus was that it was too unstable to solve complex real-world problems. Then a small startup in London shattered every assumption. The reason this paper was so shocking was its 'generality.'
Their single algorithm network (DQN) outperformed human experts in 29 out of 49 Atari 2600 games. Boxing, Video Pinball, Space Invaders: games with completely different rules and objectives, all learned by the same brain. This was a powerful signal that AI could evolve from a special-purpose tool into a general-purpose one.
The fact that a computer science paper, and one about a game-playing AI at that, made the cover of Nature was itself extraordinary. It symbolized that AI research had moved beyond mere engineering and entered the domain of natural science, exploring the very nature of intelligence. Researchers around the world analyzed DQN's source code with excitement, and Google bet a fortune to secure this potential early.
The Atari shock ended the winter of AI research and became the detonator that ignited the golden age of deep learning. And at its center was Hassabis's relentless quest to 'solve intelligence.' Why games? Calculating difficulty and symbolic value. Games were not mere entertainment.
They were the most sophisticated 'laboratory' humanity had ever invented for measuring and training intelligence. Hassabis often said, 'Games are a microcosm of the real world.' The real world is far too complex, too noisy, and results take far too long to materialize. Teaching an AI to trade stocks or making a robot walk in the real world involves enormous cost and risk. But games are different. They are safe, infinitely repeatable, and have clear objectives: victory or a score.
Hassabis mathematically calculated game difficulty and designed the stages of AI development accordingly. Atari games were the stage for processing pixel information on a two-dimensional plane. This was about verifying early visual cortex-level information processing.
Go was the pinnacle of 'Perfect Information Games.' On the Go board, there is no hidden information. Both players see every move, competing through pure calculation, intuition, and pattern recognition. Because the number of possible positions in Go exceeds the number of atoms in the universe, it was
a testing ground that asked whether AI could conquer the domain of 'intuition,' something beyond raw computational power. Hassabis was already looking beyond Go. The real world is not like a Go board where all information is visible.
You must make decisions without knowing what the other side is thinking, what tomorrow's weather will be, or what lies behind the drawn curtain. This is called an 'Imperfect Information Game.' Poker and StarCraft fall into this category. After proving 'perception' with Atari and 'intuition' with AlphaGo, Hassabis planned to push AI into the world most resembling real-life uncertainty: StarCraft II.
This was not about building an AI that was good at games. It was an essential training process for creating an AI capable of forming scientific hypotheses and designing experiments in an uncertain real world: a 'scientist AI.' The conquest of StarCraft II (AlphaStar): real-time strategy in an imperfect information environment. In January 2019, DeepMind stunned the world again. This time the stage was not a Go board but StarCraft II, a real-time strategy simulation game (RTS).
If AlphaGo was the ruler of static, turn-based games, 'AlphaStar' had to be a battlefield commander making hundreds of decisions in split seconds. StarCraft II was a nightmarish challenge for AI. First, there is the 'Fog of War.'
Players can only see areas where their own units are present. They cannot see what the opponent is building or where they are moving their forces. The AI must constantly gather information through scouting, predicting and inferring the opponent's unseen actions. This is a highly intelligent act: forming hypotheses and testing them under incomplete information.
Second, there is the pressure of 'real-time.' In Go, you are given time to think about your next move, but StarCraft never pauses. Third, there is the problem of a long time horizon. A decision to produce one extra worker early in the game can determine the outcome of a massive battle twenty minutes later.
The AI had to look thousands, tens of thousands of frames into the future and calculate the butterfly effects of its current actions.
AlphaStar scored overwhelming victories against professional gamers 'TLO' (Dario Wunsch) and 'MaNa' (Grzegorz Komincz). Many assumed AlphaStar won through superhuman clicking speed (APM), but DeepMind had limited the AI's click rate to human levels. What AlphaStar demonstrated was not raw reaction speed.
It was chilling 'judgment.' AlphaStar used scouting to identify the opponent's build and fluidly adjusted its army composition in response. When at a disadvantage, it retreated decisively; when it spotted a gap in the opponent's defense, it struck without hesitation.
The 'Blink Stalker' micro demonstrated against MaNa was a pinnacle of precision beyond human imitation, but what was even more striking was its strategic vision spanning the entire battlefield. There were limitations, too. When MaNa constantly harassed AlphaStar from outside its field of vision, AlphaStar was visibly flustered.
But DeepMind quickly corrected this, and AlphaStar ultimately reached Grandmaster level. For Hassabis, AlphaStar's victory was powerful evidence that AI could formulate complex strategies and achieve objectives in an uncertain, complex, and real-time-changing environment, in other words, the 'real world.' Atari Breakout game screen.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.