The sequel to the 100,000 bestseller in the AI ​​field!
ChatGPT and LLM, AI corporate wars and semiconductors,
Knowledge and insights about technology in one book!

“How on earth does ChatGPT work?”
“How will the AI ​​and semiconductor industries fare in the future?”
“How will my work and life change?”
The book that gives the most accurate and refreshing answer to this question right now!

The authors of "AI Knowledge Even Non-Majors Can Understand," which received the love of 100,000 readers, have returned with a follow-up book covering ChatGPT.
This book explains the principles of ChatGPT, a leading generative AI, in an easy and interesting way with illustrations.
Rather than simply introducing the technology, it provides specific information on how the GPT model, the core structure of ChatGPT, learns information, generates sentences, and understands meaning.
Complex concepts such as large language models (LLMs), transformers, and GPU-accelerated environments are explained in a way that anyone can understand.

In particular, the author, as an AI engineer and LLM expert in the field, tells vivid stories from the field based on his practical experience.
Beyond theory, it covers the actual technologies that make up ChatGPT, the strategies of global companies, and the ethical and social issues facing AI.
This book covers all the technologies and trends related to ChatGPT, along with the people and vision behind them. This is an introductory and essential guide to ChatGPT for everyone, from those new to AI, to practitioners seeking to leverage generative AI, to developers seeking a deeper understanding of the technology.

Many people expect ChatGPT to significantly change the landscape of global businesses and usher in a new era.
In his book The Structure of Scientific Revolutions, philosopher of science Thomas Kuhn proposed the concept of a paradigm that transcends existing frameworks, and many believe that ChatGPT will be the beginning of a new paradigm.
Many experts predict that generative AI technologies will transform our daily lives and society in ways we never imagined.
--- From "Chapter 1: GPT-4 Surpasses Humans"

GPT, created by OpenAI, was a typical language model.
It trained on about 7,000 books and had about 117 million parameters, similar to Burt.
Initially, its use was ambiguous.
It was difficult to find a suitable use for a model that simply generated sentences.
Less than a year later, OpenAI announces GPT-2.
What's most surprising is that the number of parameters has increased to 1.5 billion, which is 10 times larger than the previous version.
The training data also grew tenfold.
The following year, GPT-3 was even more impressive.
The number of parameters has increased to 175 billion, which is a whopping 100 times larger than the previous version.
From this point on, competition among companies over model size begins in earnest.
As the model grew larger, papers began to appear one after another showing that it performed well in various tasks, and even papers showing that it had the ability to emerge out of nowhere.
Things that didn't happen when the model was small suddenly became smarter once it got past a certain size.
You will begin to show an unknown ability to easily understand knowledge that was not taught to you.
At the same time, expectations for GPT have also increased.
And finally, ChatGPT, which we all know so well, will be unveiled to the world in November 2022.
ChatGPT is causing a huge stir in the world.
In March 2023, the world will truly enter the era of artificial intelligence with the arrival of GPT-4, which boasts even more impressive performance.
--- From "Chapter 3: The Secret Recipe for Completing Chat GPT"

The biggest innovation of the LLM was that it required learning a huge amount of text without any human intervention.
The point was that a pre-learning model could be completed by learning the countless pieces of information accumulated by humanity so far without any separate processing.
If we had to assign correct answers to each piece of data or categorize it separately, today's LLM would not have been possible.
…The simplest and most effective method among them is DPO, proposed by a research team at Stanford University. DPO is an optimization method that directly reflects human preferences.
Translated into Korean, it roughly translates to "direct preference optimization." DPO contributes significantly to simplifying complex processes and increasing efficiency.
--- From "Chapter 4: Optimization of Ultra-Large Models"

o1 is a model that behaves as if it had done a good job of prompt engineering.
o1 introduces a thinking process that gradually refines the prompt.
In the first response, you supplement the prompt, ask the question again, and then use that response to further supplement the prompt.
In this way, we continue to refine the prompt through several stages, ultimately producing the best answer.
So, it's similar to automating all the prompt engineering that was originally done by people.
…we have a thought process to come up with a better answer, and in this process, we explore various possibilities.
We carefully consider opposing views while also raising our own counter-arguments.
Best of all, this entire process is no longer done by humans.
The model automatically proceeds during the inference process.
--- From "Chapter 5: The Magic of Prompt Engineering"

Let's take a quick look at the corporate valuations of several AI startups.
OpenAI is the undisputed number one.
Although it is only a startup that was founded about 9 years ago, its corporate value is over 430 trillion won in our currency.
This is the extent of the problem for a startup that is still running at a loss in the tens of billions of won and has not even gone public yet.
Second place is Antropic.
This is a company created by people from OpenAI.
It is evaluated as a company with competitiveness comparable to OpenAI.
Third place goes to Elon Musk's xAI.
Although there is a gap with OpenAI, it has become a unicorn with a corporate value of over 70 trillion won in our money.
Elon Musk was the one who first founded OpenAI, so the influence that one person has had on the AI ​​industry is truly enormous.
And all the companies I've mentioned so far are American companies.
The top three companies are all located in Silicon Valley.
The United States is the world's unparalleled economic power, and the gap with Europe is growing day by day.

--- From "Chapter 6: Global Companies' Competition for a 1,000 Trillion Won Market"

NVIDIA has long wanted GPUs to be used in a variety of fields beyond gaming.
And we start researching new programming models that could make GPUs more versatile.
In 2006, a platform called CUDA was finally born.
NVIDIA opened the way to GPU coding using the familiar C++ language with CUDA.
It is no longer a separate unit for each purpose, but rather a unified architecture that allows you to freely code using the language you normally use for research.
This has made GPUs more accessible to scientists and researchers.
… In 2009, Stanford University finally conducted the first experiment to introduce GPUs to artificial neural networks.
At this time, the paper revealed that learning could be done up to 70 times faster, surprising numerous researchers around the world.
From this point on, a new history of NVIDIA begins.
After that, everyone, regardless of gender, began to adopt GPUs in AI research. With GPUs, processing speeds were dozens of times faster than before, so there was no reason not to use them.
As excellent research results continue to emerge in the field of artificial intelligence, GPUs are also establishing themselves as representative equipment for artificial intelligence.
All of this was possible thanks to the power of the CUDA platform, which enabled researchers to easily write GPU code.

--- From "Chapter 7: Nvidia and the Semiconductor War"