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] 25 What Is AGI?
Demis Hassabis, Father of Google's Artificial Intelligence
Part 9. The Final Gate Toward AGI
25 What Is AGI?
Kim Kyung-ran, Kim Kyung-jin
In January 2026, two people sat side by side on a stage at the Davos World Economic Forum. Demis Hassabis of Google DeepMind and Dario Amodei of Anthropic. The moderator asked: "What do you think AGI is? And when will it arrive?"
The two gave answers that pointed in the same direction yet diverged in subtle ways. Inside that gap lies the most fundamental debate in the AI industry. Hassabis answered like this.
"AGI is a system that implements all the cognitive capabilities humans can exhibit." The key words in this definition are 'all' and 'cognitive.' It encompasses reasoning, creativity, planning, problem-solving, and understanding of the physical world.
Hassabis elaborated on this definition in a 2025 conversation with Lex Fridman. AGI is not a system that excels in any single domain. It must consistently demonstrate human-level or above performance across every cognitive domain. He said it should be at a level where "a team of experts could test it for a couple of months and still struggle to find weaknesses."
What Hassabis emphasizes is 'consistency.' Current AI systems can solve math olympiad gold-medal-level problems while failing at everyday judgments that a five-year-old handles easily. He called this "jagged intelligence."
For AGI, this jaggedness must disappear. It should play chess, write poetry, formulate new physics hypotheses, and cook in an unfamiliar kitchen. Hassabis went further, proposing a distinctive thought experiment as a litmus test for AGI.
"If you dropped that system into the year 1900 and gave it only the physics knowledge of the time, could it discover the theory of relativity on its own?" Or: "Could it invent a game as deep and elegant as Go by itself?" That is Hassabis's standard for true general intelligence.
Dario Amodei's approach is different. He doesn't like the term AGI itself. In a 15,000-word essay published in October 2024, titled "Machines of Loving Grace,"
he deliberately used the phrase "powerful AI" instead. Amodei's definition goes like this: "AI that is smarter than a Nobel Prize winner in virtually every relevant field."
A system that surpasses the highest-level experts regardless of domain, whether biology, programming, math, engineering, or writing. He compared it to "a nation of geniuses living inside a data center." Where does the difference between the two lie? Hassabis takes as his reference point the architecture of the human brain, meaning every function that the brain can perform as a general-purpose learning machine.
His is a theoretical and structural definition. Amodei's standard is based on outcomes and results. The benchmark is practical performance that surpasses the best human experts in specific fields.
If Hassabis's definition answers "what should it be able to do," Amodei's answers "how well should it do it." This difference affects their timelines too. Amodei believes AI meeting his criteria could appear in 2026 or 2027. Hassabis is more cautious.
His estimate is a 50 percent probability by 2030. This gap stems from the difference in definitions. Because Amodei's standard focuses on superhuman performance in specific tasks, it seems reachable on the current scaling trajectory alone.
Hassabis's standard demands consistent generality across all cognitive domains, which requires new breakthroughs beyond the fundamental limits of current technology. What's interesting is that neither man completely rejects the other's perspective. On the Davos stage, Hassabis said, "Our views aren't that far apart."
Amodei likewise acknowledged in his 2026 essay "The Adolescence of Technology" that "powerful AI could come within one to two years, but it could also be considerably further away." There is agreement on the ultimate destination. They differ only in how they interpret the path and the speed of getting there.
This debate matters for a concrete reason. How you define AGI changes the priorities of safety research, shifts the timing of regulation, and influences investment decisions worth tens of trillions of dollars. Setting a high bar like Hassabis does sends the message that "there is still a long way to go."
Setting a performance-based standard like Amodei does becomes a warning that "it's coming soon." Whichever side is right, both definitions share one premise: what we are building now is something without precedent in
human history, and precise language is needed to understand its implications. Coding, Math, Natural Science, Physical AI (Robotics). In May 2025, at a Google I/O fireside chat, Hassabis sat next to Google co-founder Sergey Brin. When the moderator asked about the limits of current AI systems, Hassabis drew a clear line.
"If AI can do something that 90 percent of people cannot, that is an economically important milestone, and it matters from a product standpoint too. But let's not call that AGI. That should be called 'typical human intelligence' level." This statement shows how broad AGI's scope really is. The AGI Hassabis envisions cuts across multiple cognitive domains.
Coding is the domain advancing fastest among them. As of 2025, AI already performs the majority of programming work that builds Anthropic's own products, and within Google DeepMind, AI is used extensively for code generation and debugging. Hassabis acknowledges the possibility of a "self-improvement loop" in AI's coding and math capabilities.
That loop has already begun: AI writes better code, and that code builds better AI. But he insists this alone cannot be called AGI. Math is the second front.
DeepMind's AlphaProof and AlphaGeometry have solved problems at the International Mathematical Olympiad level. AlphaEvolve went a step further, opening an era where AI designs algorithms themselves. Hassabis takes pride in these achievements while clearly recognizing their limits.
Current systems excel at proving existing conjectures but have not acquired the ability to propose new conjectures on their own. Playing Go at a world champion level and inventing the game of Go are entirely different orders of intelligence. Hassabis sees the latter ability, the capacity for creative invention, as a core benchmark for AGI.
Natural science is the third domain, and the one closest to Hassabis's heart. AlphaFold's solution to the protein folding problem, unsolved for 50 years, proved that AI could serve as a tool for scientific discovery. AlphaGenome has begun mapping the relationship between genetic code and function, and GNoME has discovered the crystal structures of more than 2.2 million new materials.
Yet despite all these achievements, current AI is a tool that 'helps' scientists, not a 'scientist' that formulates and tests hypotheses on its own. This is why one of the two preconditions for AGI that Hassabis outlined on the Google DeepMind podcast in December 2025 is "automated experimentation." AI must close the complete loop of scientific research: posing questions, designing experiments, interpreting results, and revising hypotheses anew. DeepMind's announcement that it would open its first fully automated research laboratory in the UK in 2026 is the realization of this vision.
The fourth domain is the most challenging: physical AI, meaning robotics. On the Davos stage, Hassabis said, "I include physical AI, robotics, all of that in my definition of AGI." Current large language models operate within the digital world.
They read text, analyze images, and generate code. But picking up a cup in the real world, cooking in a kitchen you've never visited, or walking across uneven terrain demands an entirely different kind of intelligence. It requires real-time understanding of and response to physical phenomena like gravity, friction, inertia, and the deformation of flexible objects. This is why Hassabis names the "World Model" as another precondition for AGI. DeepMind's Genie generates interactive 3D environments from text prompts, and SIMA performs tasks within those environments, moving objects, avoiding obstacles, and finding targets.
Veo, the video model, reproduces fluid motion, light reflection, and material texture with remarkable accuracy, suggesting that AI can develop an "intuitive physics" understanding through passive observation alone, such as watching YouTube videos. This finding challenges a long-dominant assumption in AI research: the belief that understanding the physical world necessarily requires an embodied agent, like a robot with a body. In Hassabis's words, "The precondition for AGI is not becoming smarter but becoming capable of more actions."
No matter how fluently a chatbot converses, if it cannot observe, understand, and actually change the world, it is not general intelligence. Consistent capability that runs through every domain, from coding to math to natural science to the physical world. That is the full picture of AGI as Hassabis envisions it. Blank spaces remain to be filled before this picture is complete.
Hassabis: AGI Could Arrive "Within 5 to 10 Years." A prediction of 50 percent probability of reaching AGI by 2030. True to his background as a researcher with a doctorate in neuroscience, Hassabis prefers the language of probability. "Fifty percent." In July 2025, appearing on the Lex Fridman podcast, Hassabis put the probability of AGI emerging within the next five years, around 2030, at 50 percent.
A coin flip. And in December 2025, at the Axios AI+ Summit, he reaffirmed this prediction. "AGI, perhaps the most transformative moment in human history, is on the horizon."
Understanding the weight of this prediction requires context. In the AI industry, timeline predictions are a kind of strategic statement. Say it's too close and you attract investment but lose credibility if you fail to deliver.
Say it's too far and you stay safe but the market loses interest. Hassabis's "50 percent, 5 to 10 years" sits precisely in the middle of this spectrum. And that is intentional.
Consider the comparisons. Anthropic's Dario Amodei points to 2026 or 2027. Tesla's Elon Musk predicted that AI surpassing individual humans would appear in 2025, and that AI exceeding all humans combined would arrive around 2030. OpenAI's Sam Altman uses the vague phrase "a reasonably near future." Meta's Mark Zuckerberg says AI will write most code within 12 to 18 months.
On the other end, voices like Baidu's Robin Li predict "more than 10 years," and Meta's Yann LeCun argues "at least 10 years, probably much longer," insisting that new scientific discoveries are needed. Hassabis falls somewhere in the middle. What sets him apart from other CEOs is that he attaches conditions to his predictions. The number 50 percent rests on the premise "if the current trajectory holds." At the same time, it leaves room for uncertainty: "there may be obstacles we don't yet know about." This is the language of a scientist.
A habit of speaking in probabilities rather than certainties. What evidence led Hassabis to this prediction? He cites several specific advances. The rapid progress of the Gemini series is one. In his July 2025 conversation with Lex Fridman, Hassabis said, "Gemini 3 is an absolutely stunning model, and it exactly matches what I expected and the trajectory we've been on over the past few years."
Improvements in reasoning capability, expansion of multimodal understanding, and advances in agentic systems were the specific areas of progress he identified.
DeepMind's DeepThink runs multiple reasoning processes in parallel and then cross-validates them against one another. Hassabis described it as "reasoning on steroids." Test-time compute, the approach of investing additional computation at the point when AI generates an answer, is a methodology DeepMind has pioneered since the AlphaGo era.
The same principle that simulated thousands of scenarios before making a move in Go is now being applied to language models. There is an important caveat to Hassabis's 50 percent prediction. The other 50 percent, the equal probability that AGI will not arrive by 2030.
The reasons are clear. Current systems lack fundamental capabilities: reasoning, memory, understanding of the physical world, and creative invention. In a January 2025 interview with Alex Kantrowitz, Hassabis declared that "no lab will reach AGI this year" and criticized those who overhype the technology. Coming from a Nobel laureate and frontline researcher, this statement is a rare voice of restraint in the overheated AGI debate.
In December 2025, Hassabis's prediction was updated in one respect. When Amodei offered "2026 to 2027" on the Davos stage, Hassabis responded, "Our disagreement isn't that large." He added, "My timeline is just a bit longer."
This subtle convergence is an attempt to balance the pressure of competition against scientific caution. The market wants fast timelines, investors want certainty, but the reality of research can betray predictions. In the end, what Hassabis's 50 percent prediction says is this.
AGI will come. The question is not "if" but "when." That "when" could be five years from now, ten years from now, or even longer.
"One or two more breakthroughs on the level of the Transformer or AlphaGo are needed." At Google I/O in May 2025, Hassabis summarized the distance to AGI in a single sentence: "To get all the way to something like AGI may require one or two more new breakthroughs." He repeated the thought on the Lex Fridman podcast in July.
"To get the kind of consistency across all the areas you'd expect from general intelligence, we probably need one or two more things in reasoning, memory, and perhaps ideas like world models." This statement looks modest, but its implications are enormous. Hassabis references two historical precedents.
One is the Transformer architecture published by the Google Brain team in 2017. A single paper titled "Attention Is All You Need" completely transformed natural language processing and became the foundation for modern AI models, including the GPT series and Gemini. The other is AlphaGo in 2016.
The combination of deep neural networks, reinforcement learning, and Monte Carlo tree search, a combination that had not existed before, conquered the supposedly impregnable game of Go. In both cases, the advance was not incremental improvement but a qualitative leap beyond the existing framework. When Hassabis says "one or two more are needed," where might those breakthroughs come from? He repeatedly mentions three areas.
The first is reasoning. Current large language models generate responses based on pattern recognition. They appear to solve math problems and perform logical reasoning, but whether the process works the same way as human logical thought remains unclear.
DeepMind's DeepThink is one approach to this problem. Exploring multiple reasoning paths in parallel and cross-verifying them is an extension of the method AlphaGo used to search for moves in Go. Hassabis called this "part of the breakthrough."
Part, not the whole. The next stage of reasoning, the ability to independently generate and test new hypotheses, is a domain not yet reached. The second is memory. Current AI models forget the contents of a conversation once it ends.
Every interaction starts from a blank slate. Human intelligence is built on continual learning, accumulating experience and applying past lessons to new situations. Bridging this gap may require a new architecture that integrates episodic memory, semantic memory, and procedural memory.
This is precisely where Hassabis's neuroscience background shines. The hippocampus and episodic memory, the subject of his doctoral thesis, represent a core brain function that AI has not yet replicated. The third is the world model. On the Google DeepMind podcast in December 2025, Hassabis specified two preconditions for AGI. The first was the world model: the ability for AI to genuinely understand physics and space. If a language model can
tell stories, a world model can construct environments. But the decisive step that lets AI participate in reality is experimentation. Genie generates scenarios in real time based on context, SIMA performs tasks within them, and whether the outcome is success or failure, the results become learning material for the AI itself. This is the cognitive closed loop that Hassabis envisions.
Skeptical views of Hassabis's diagnosis exist. A December 2025 analysis on the Effective Altruism forum posed a sharp question: "The unresolved research problems Hassabis lists, hierarchical reinforcement learning, world models, continual learning, creative idea generation, are in fact decades-old problems. Search an academic database and you can find papers on hierarchical reinforcement learning from the early 1990s.
What grounds are there for confidence that these old problems will really be solved within three to five years?" This criticism touches an important point. Most investment in the current AI industry is concentrated on scaling large language models, and relatively few resources are being directed toward discovering the new science Hassabis says is needed.
Hassabis resolves this tension with the principle of "doing both at the same time." Maximizing the technology you have now while simultaneously preparing for the next innovation coming in six months to a year. This is the mindset of a chess player: making the best possible move right in front of you while already mapping out the board ten moves ahead.
The same strategy worked in AlphaGo. Pushing known moves to their limits while boldly throwing in moves nobody had considered. Sergey Brin offered an interesting historical analogy here.
"If you look at the history of N-body problem simulations, algorithms getting smarter probably made a bigger difference than computers getting faster. And right now, both are improving at the same time." The breakthrough Hassabis says is essential lies precisely on the algorithm side. Not building bigger data centers or stacking more GPUs, but a fundamentally new understanding of how intelligence works.
When and where these "one or two breakthroughs" will appear, and what form they will take, nobody knows. Nobody in 2015 knew that the Transformer paper would come in 2017. Nobody foresaw AlphaGo's move 37. Hassabis himself acknowledges this unpredictability.
Yet he says his team has "several promising ideas in preparation, and we want to merge them into the Gemini mainline." Throughout the history of science, breakthroughs have always come to those who were prepared. Hassabis has been preparing for 40 years.
The question he carried in front of the chessboard, "Can't this intelligence be put to better use?" has now reached its final turning point. A roadmap toward AGI.
Kim Kyung-jin
Attorney · Former Member of the National Assembly · AI Policy Researcher
© 2026 Kim Kyung-jin. All rights reserved.