A book that interprets MIT Professor Ray Kurzweil's famous book, 'The Singularity is Near'!
"The Singularity is Near" is a must-read for Ivy League universities and White House officials in the eastern United States!

This book is a more easily explained and supplemented version of the book published in June 2024 by MIT Professor Ray Kurzweil, a world-renowned futurist and renowned AI researcher.
For the general reader, there are many scientific and computer terms, so I tried to explain them in Korean as easily as possible.
This is the point where artificial intelligence (AI) resembles the human brain and shows the possibility of thinking similarly to humans.
To properly utilize general artificial intelligence (AGI), we need to properly understand the human brain.
Professor Ray Kurzweil defines the point at which the human brain and AI synergize to create a new human culture as 'singularity.'
The original book presented how singularity actually occurs in real life.

AI will change the world as it becomes more advanced and develops exponentially.
High-performance computing will become cheaper, human biology will be better understood, and nanoscale engineering will become possible.
In particular, as AI capabilities advance, information accessibility becomes smoother, and human intelligence becomes more closely integrated with AI, a new world will open up.
This event is the singularity that Ray Kurzweil talks about.
In this section, the fears that are being raised about the AI ​​era come to the forefront.
In other words, there is a fear that AI could overwhelm human capabilities and threaten human jobs and even survival itself.
However, Kurzweil would like to point out that in his view, this problem is purely human.

The term singularity is borrowed from mathematics and physics.
In mathematics, a black hole is a point at which a function becomes undefined when divided by zero, and in physics, a black hole is a point at the center of a black hole where the general laws of physics break down to an infinitely dense point.
However, it is important to remember that this term, as used in AI, is used metaphorically.
The reason we use the singularity metaphorically is because the progress of AI cannot keep up with the current level of human intelligence.
However, once the transition occurs, human technology will improve its cognitive abilities quickly enough to coexist with AI.

The debate between Ray Kurzweil and Marvin Minsky about the future of AI is famous.
Two fundamentally different perspectives confront each other regarding how machines can achieve human intelligence.
The core of the debate is 'computational power' and 'the right algorithm'.
Ray Kurzweil, a proponent of the singularity, argues that to reach human-level complexity, AI would need massive computational power—more than 10¹⁴ calculations per second—to approach the number of neuron connections and their interactions in the brain.
The argument is that by increasing processing power, machines will be able to implement the complex operations of human cognition.
He also emphasized the exponential growth of hardware performance based on Moore's Law, and predicted that if computers could process enough calculations per second, they would be able to perform the same intellectual tasks as humans.
Furthermore, it was thought that human intelligence would arise from "brute force"—that is, if enough data was processed and enough connections were modeled, human-level intelligence would be achieved.
His vision aligns with the current trajectory of machine learning models like GPT and PaLM.
In contrast, Marvin Minsky, who pioneered AI deep learning, argued the exact opposite.
He said that intelligence cannot be achieved through calculations alone, and instead, the most important thing is the quality of the algorithm.
For Minsky, what matters is not how much processing power a machine has, but how well it can mimic the abstract thinking and problem-solving skills of humans—in other words, whether it executes the right algorithms.
Minsky believed that AI could achieve human-level intelligence, even with a processor the size of a Pentium chip (a desktop processor released in 1993), if only the right algorithms were designed.
The idea is that it comes from symbolic AI, the ability to manipulate symbols, infer concepts, and derive logic.
At that time, neither side could prevail.
At the time, neither Kurzweil nor Minsky had the tools they needed to prove their claims.
Kurzweil lacked the computational power (computing source) to prove that human-like intelligence could be achieved, and Minsky lacked the algorithm to prove that the same results could be achieved using symbolic reasoning alone.

Interestingly, the enormous computational power deployed in deep learning (GPT, PaLM, PaLM-E) validates Kurzweil's argument that scale and data are crucial.
The performance of deep learning-based AI, which relies on massive data sets and billions or trillions of parameters, depends on computational power.
Nevertheless, there is still much room for algorithmic efficiency and innovation.
Astronomical computational power and sophisticated algorithms are both essential.
Neither one alone is sufficient to achieve full human-level intelligence.
Combining the computational power and sophisticated inference architectures Kurzweil emphasizes could allow AI to reach human-level reasoning capabilities.

~ 'AI model with sentence creativity'

The human brain evolved to help us survive in complex environments where pattern recognition, decision-making, and creativity are more important than number crunching.
The reason is as follows:
First, the priority of evolution.
Humans had to identify predators, find food, navigate the terrain, and communicate with others.
These tasks required pattern recognition (e.g., spotting a tiger in the bushes or finding edible plants) rather than solving equations.
Fast, flexible thinking was more useful for survival than slow, detailed calculations.
Second, the structure of the brain.
The brain is built for adaptability.
The brain operates as a highly interconnected network of neurons that excels at fuzzy logic (understanding incomplete or uncertain information) rather than precise calculations.
For example, the brain can infer groups of objects without having to memorize all the items.
That is, we can approximate a 'close enough' solution.
Third, compare neurons to transistors.
Neurons (brain cells) are excellent at recognizing complex patterns and associations, but they operate much more slowly than the transistors in a computer.
Transistors are specifically designed to process precise numbers, while neurons are built for flexibility and learning through experience.
Fourth, there are limitations to working memory.
The human brain has a limited working memory, capable of storing only a small amount of information at a time (usually 5 to 9 items).
Therefore, it is difficult to perform complex calculations.
Fifth, development of alternative tools.
Instead of developing advanced computational abilities, humans invented tools such as abacuses, calculators, and computers to handle complex calculations for them.
Thanks to this, the brain can focus on creativity, reasoning, and innovation.
In short, the human brain is designed primarily for adapting to uncertain and dynamic environments, recognizing patterns, and solving practical problems, not for computation.
For accurate calculations, we rely on AI or computers specifically designed for this purpose.

In particular, the statement that “human intelligence is a bundle of various cognitive abilities” has many implications.
It emphasizes that human intelligence is a collection of abilities that vary from person to person.
This provides important implications for the development of artificial intelligence that mimics the human brain.
These unique aspects of human intelligence inspire the development of future specialized AI.
One person may be a mathematical genius but lack social skills, or another may have a great photographic memory but lack creativity.
Likewise, current AI systems are designed to excel in specific areas, rather than being “all-purpose.”
For example, DeepMind's AlphaZero is excellent at chess and Go, but cannot engage in conversation.
ChatGPT-4.o is good at language processing and text generation, but it can't drive a car or analyze visual data.
Therefore, specialized AI development can advance in various directions. It's not about turning AI into an all-powerful "super brain."
It is developed as a collection of specialized systems that each demonstrate outstanding capabilities in their respective fields.
With this in mind, let's speculate on the future direction of AI development.

First, it is the construction of a specialized model.
Just as humans have unique strengths, AI should also focus on specific tasks where it excels.
These include medical diagnostic AI that interprets X-rays or identifies diseases, and creative AI that composes music or designs works of art.
Specialized AI will complement each other and work together to solve complex problems, just like human experts.
Second, there's capability integration. This type integrates specialized AI systems into a cohesive framework.
For example, this is how AI combines language processing (e.g., ChatGPT4.o), vision (e.g., computer vision AI), and decision-making (e.g., AlphaZero) to focus on real-world problems such as autonomous medical surgery support.
Third, overcome weaknesses with AI.
Even humans who excel in one area have limitations in others. AI can complement human weaknesses by complementing areas where individuals struggle.
People with poor memory can use AI personal assistants to compensate. Rather than replacing human intelligence, AI can be developed to bridge gaps in cognitive abilities in specific areas and even augment them.
Fourth, maximizing synergy.
For example, imagine a medical AI system.
One part of the system is trained to analyze medical images (radiology).
Another part is understanding patient records and communicating with doctors and patients.
Then, a predictive model is used to recommend a treatment.
Each part individually excels in a specific area.
Combining these systems would be similar to how human doctors integrate their cognitive abilities to diagnose and treat patients.
As an example of applying the concept of 'bundling' human intelligence to AI, it can be said that imitating the diversity and expertise of human talent is essential for AI development.
~Advances in Big Data and Artificial Intelligence~

Since around 2000, computers have gotten faster at a fairly regular rate.
Every 1.4 years, you can buy a computer that is about twice as fast as your previous one for the same cost.
It's similar to how smartphones and laptops improve in performance every two years.
In particular, the speed of AI learning becomes much faster.
Since 2010, AI model learning speed has doubled every 5.7 months.
In other words, AI processing power doubled in less than half a year.
In terms of development speed, AI is much faster than computers.
From 1952 to 2010, computer capabilities increased approximately 75-fold.
However, from 2010 to 2021, AI learning speeds didn't just increase by 75 times, but by roughly 10 billion times. Why is AI growing so rapidly? It's not because we've invented better computers.
First, parallel computing.
A method has been developed to solve problems faster by using many computer chips together simultaneously.
Connectionism is that.
It's like 100 people working together to put together a big puzzle, not just one person.
Second, as deep learning becomes more useful due to big data, a lot of capital is being invested.
Thanks to the internet, smartphones, and social media, the world is flooded with massive amounts of data.
Data is the training tool that makes AI smarter.
Investors and companies are investing heavily in AI research and development, which has the potential to transform industries.

Traditionally, doctors analyze images to detect signs of disease.
However, it is very difficult to identify early-stage cancer in photographs because the differences between healthy and cancerous tissue are very subtle.
However, AI trained on large data sets learns patterns that are too subtle for humans to detect. By learning thousands of X-rays or MRI images showing both healthy and diseased tissue, AI systems can spot early signs of cancer better than human doctors.
This is because it analyzes massive amounts of data to identify patterns and insights that humans miss.

~Three things AI lacks~

Google is attempting to leap forward in the artificial intelligence (AI) race once again by unveiling 'Gemma 3' on March 12 (local time).
Since the rise of China's DeepSec, economic feasibility and computational efficiency have emerged as important issues in the AI ​​industry, in addition to performance.
In response, Google introduced Gemma 3, an open-source large language model (LLM) that delivers high performance even in a single graphics processing unit (GPU) environment.
AI models can generally perform more complex calculations as the number of parameters increases, but this also slows down processing and increases hardware requirements.
However, Gemma 3 succeeded in reducing the computational burden by maintaining performance by compressing the information learned from the large model and transferring it to a smaller model.
As a result, while DeepSeek R1 uses 671 billion parameters and can optionally activate 37 billion parameters when needed, Gemma3 performs well with only 1 to 27 billion parameters.
Multimodal AI is still in its infancy, but it is one of the technologies expected to fundamentally change the way AI and humans interact.
By mimicking the way humans process senses, it can process information in a more intuitive way, which can be applied in various fields such as medical diagnosis, autonomous driving, and emotional analysis.
In the medical field, a more precise diagnosis is provided by analyzing a patient's symptom records and medical images together.
When applied to autonomous vehicles, it can process road signs, pedestrians, vehicle movements, and even the driver's voice commands in real time.
Furthermore, when applied to industrial robots, they can evolve beyond following work instructions to learning and adapting to their environment on their own.

--- From the text