Today, I will provide knowledge services on two topics.
First, let’s talk about AI. Looking back at December, it seems like the entire AI field suddenly launched a year-end sprint, with many groundbreaking products being released all at once. Normally, each of these products could dominate the trending topics for days, but with everything being launched in December, it’s hard to focus on the key points.
So today, let’s briefly review the major events in the AI field this December.
The starting point of this AI sprint was on December 3rd, when AI scientist Fei-Fei Li’s company, World Labs, released its first AI system. They defined this AI as a Large World Model (LWM). Fei-Fei Li announced her startup in April, made the company public in September, and launched the product in early December—an astonishing speed.
Then, just one day later, on December 4th, Google DeepMind released Genie2, which they defined as a large foundational world model. Yes, just one word different from Fei-Fei Li’s Large World Model.
Almost simultaneously, OpenAI’s Sam Altman, in an interview, announced that from December 5th onwards, they would release new products and demos for 12 consecutive days. Yes, 12 days in a row. And on the third day, they released the upgraded version of Sora—Sora Turbo.
On December 11th, OpenAI officially announced a comprehensive integration with Apple. Less than a day later, Google released its next-generation large model, Gemini 2.0, which they defined as being created specifically for AI agents.
There were many other actions from other companies, but as we mentioned earlier, the top AI companies around the world launched a series of groundbreaking products in December, just like a year-end performance sprint.
The key point we want to highlight isn’t the products themselves, but the clear trend reflected in this wave of AI releases. That trend is: World Models. Fei-Fei Li’s World Labs is positioned around spatial intelligence, which is a form of world model. Similarly, Google’s new product, Genie2, also claims to be a world model.
These two AIs have very similar functionalities: both can take an image as input and generate a complete 3D environment. For example, if you show them a picture of a concert, they can generate a 3D model of the concert, and you can explore this model like playing a 3D game.
In addition to Google and Fei-Fei Li’s company, OpenAI has also made a significant push into the field of humanoid robots this year, investing in related companies. The key underlying technology for humanoid robots is the world model.
In the coming year, you might see more media reports about world models. So today, let’s take a closer look at the world model to lay the groundwork for understanding this trend that could shape the future.
Many people first heard of the world model through Yang Likun. When ChatGPT first became popular and all the media were promoting it, the French AI scientist Yang Likun stated that OpenAI’s large language models had limited capabilities and that to achieve AGI (Artificial General Intelligence), we needed to follow the world model approach.
It’s important to emphasize Yang Likun’s role. He is one of the most senior, influential, and contributing scientists in the AI field. As early as 1988, he worked at the famous Bell Labs, where he was a peer to Shannon, the father of information theory, and Shockley, the inventor of the transistor.
According to Fei-Fei Li’s memoirs, Yang Likun provided her with important insights. During his time at Bell Labs, Yang Likun started training neural networks to recognize signatures. He obtained over 7,200 handwritten scanned copies in various fonts from the U.S. Postal Service and used them to train neural networks. Later, this research was adopted by banks for use in ATMs to read numbers on checks. This was in 1993, 31 years ago, and Yang Likun’s invention was one of the earliest real-world applications of neural networks.
You don’t need to remember the details, but it’s important to know that Yang Likun is one of the pioneers of the world model and a key figure in this field.
So, what exactly is a world model? Simply put, a world model can directly observe reality, abstract common sense from it, and then use that common sense for prediction and action.
Note that this approach is different from current mainstream large language models.
Most of the AI applications you see today, like GPT, Gemini, and Baidu’s Wenxin Yiyan, operate on data. You need large amounts of data to train them. Although many large models now support multimodal input (audio, images, videos), the essence of the process is still converting this information into data and having AI understand it.
A world model, on the other hand, directly observes the world and abstracts common sense. This process is more in line with how humans learn.
For example, Yang Likun gave an example in his memoir “The Road to Science.” Suppose you show an AI a picture of a little girl with a cake in front of her and several lit candles on the cake. You then ask the AI, “What will the girl do next?”
This is a difficult question for AI. It needs to understand concepts like birthdays, the custom of eating cake on birthdays, the need to light candles, make a wish, and blow them out.
According to Yang Likun’s thoughts in his book, a typical large model would only perform pixel-level computations, trying to predict how the pixels would change in the next moment. It might generate an image with several overlapping shadows of the little girl, moving in all directions. But a world model would understand the common sense of the situation and make the correct prediction.
Of course, Yang Likun’s “The Road to Science” was published in 2021, and by the following year, GPT exploded in popularity. Many of the capabilities of large language models now rival those of the world model Yang Likun envisioned.
However, the problem is that training large language models requires vast amounts of text data, and information on the internet is limited, making this data increasingly expensive. Also, with the increase in deepfake events (the use of AI to create fake information), this may lead to stricter regulation of the large-scale use of user data for training. In contrast, world models don’t rely solely on internet data, and the amount of training data required is much smaller. From this perspective, the development of world models may catch up to that of large language models.
So, how is the development of world models going now? Let’s take a look at the results released by Fei-Fei Li and Google.
Firstly, both of these models achieved a critical breakthrough in content consistency. For example, in the 3D environment generated by AI, if you walk from south to north and then back again, the images on either side of your path remain unchanged. The tree you saw the first time is exactly the same as the second time. This consistency is a big breakthrough.
From a functional perspective, according to Princeton AI Innovation Center founder Wang Mengdi, Google’s Genie2 and Fei-Fei Li’s AI appear similar, but in essence, they are different. Genie2 is based on pixel prediction, and it may not fully understand the physical laws of the real world. In contrast, Fei-Fei Li’s model uses information from the images to infer the relative relationships of objects and then generates a 3D model. As a result, Wang Mengdi believes that Fei-Fei Li’s model is more aligned with the true concept of a world model.
Alright, that’s all for the world model discussion. To summarize, we covered three main points: First, the essence of a world model is to abstract common sense from reality and use that common sense for prediction and action. Second, the most important scientist in this field is Yang Likun, and one of the key practitioners is Fei-Fei Li’s World Labs. Third, as the difficulty of obtaining internet data increases, the development of world models might catch up to large language models.
Now, let’s move on to the second topic for today—a lighter one. Recently, I came across a statistic showing that the owners of Korean fried chicken restaurants are getting younger. In 2020, 32% of fried chicken restaurant owners in Korea were between 20 and 30 years old. But by 2023, that figure had risen to 56%.
Why? Because the unemployment rate among young people in Korea has risen. According to The Korea Herald, in the first half of this year, there were 4.058 million university graduates in Korea who were part of the “economically inactive population,” the highest number since 1999. This group is defined as people aged 15 and above who are neither employed nor actively seeking work.
Interestingly, if you look at history, you’ll find that whenever there’s an employment crisis, the number of fried chicken restaurants in Korea surges.
In 2020, the Korea Land Institute released a “Land Issue Report,” which included data on fried chicken restaurants in Korea over the past 20 years. It found that the surge in fried chicken restaurants was closely related to financial crises and employment crises. The first surge happened during the Asian Financial Crisis in the late 1990s, when many companies closed down, and unemployed office workers turned to selling fried chicken. The second surge occurred after the 2008 global financial crisis, when a large number of workers entered the fried chicken industry.
This raises an interesting question: why fried chicken restaurants? Korea’s fried chicken market is highly saturated. There are more fried chicken restaurants in Korea than McDonald’s outlets worldwide. It’s said that in Seoul, there’s a fried chicken restaurant every 10 meters. So why do so many people choose to open fried chicken restaurants instead of other businesses?
The key lies in the “low barriers to entry.” Fried chicken restaurants require a relatively small investment and can generate revenue quickly. They also offer flexibility: you can hire people to work for you, and if the restaurant doesn’t succeed, the loss is manageable.
So, is this trend likely to continue in the future? Perhaps yes. With the increasing number of young people seeking jobs in Korea and the prevalence of financially struggling families, the demand for low-barrier businesses like fried chicken will likely remain high.
Here’s the English translation:
Secondly, the vitality of Korean fried chicken restaurants largely comes from Korea’s sports and entertainment industries.
For example, the first large-scale growth of fried chicken restaurants in Korea occurred during the 2002 World Cup. At that time, Korea and Japan were jointly hosting the World Cup. Many fried chicken restaurants took this opportunity to install large-screen TVs in their shops, attracting customers to watch the games while eating fried chicken and drinking beer. According to data released by the Korea Fried Chicken Franchise Industry Association, during Korea’s matches, the daily sales of fried chicken could reach 1.875 million pieces, more than 1.5 times the usual amount.
At the same time, fried chicken is also a very important element in Korean dramas. For example, in 2013, the popular drama My Love from the Star featured the famous line by the female lead, Cheon Song-yi: “It’s snowing, how can there be no fried chicken and beer?” This made the combination of Korean fried chicken and beer quickly popular across Asia. Another example is the 2019 Korean movie Extreme Job, which tells the story of an undercover drug squad that, in order to complete a mission, takes over a fried chicken restaurant. The film was later remade domestically, with the fried chicken restaurant changed to a crayfish restaurant.
Lastly, Korean society also provides a lot of support for fried chicken restaurants. For example, there are many courses and books in Korea specifically designed to teach people how to open a fried chicken restaurant. One of the most popular books is titled From Tomorrow, You’ll Be the Owner of a Chicken Restaurant. Additionally, some Korean fried chicken chain brands have opened training centers.
For example, the training center of the Korean chain restaurant group GENESIS is called “Fried Chicken University.” The courses offered include cooking fried chicken, marketing and promotion, delivery services, store management skills, and more. In July of this year, the Genesis BBQ Group launched a support program called the “Youth Smile Program,” setting aside 20 billion Korean Won to help young people open businesses, especially fried chicken restaurants.
In other words, many young Koreans view opening a fried chicken restaurant as a fallback option not just because they love fried chicken, but because it is a relatively certain and reliable business choice.
To summarize, today we covered two topics:
First, the new developments in the AI field. In December, major AI companies launched new products, with world models being a key trend. As the cost of training large language models increases, world models may receive more attention.
Second, why have fried chicken restaurants become a fallback option for unemployed youth in Korea? Due to factors such as low cost, high sales potential, and strong social support, fried chicken restaurants have become one of the most suitable entrepreneurial projects for young Koreans.
That’s all for today.
