GPT Astra Develops Game in 13 Minutes

As of the current recording, OpenAI's new AI model, Astra, has not been released to general paid users. Users of the basic paid service (costing $22 per month) cannot use Astra; access is only granted to users of the pro version, which costs $100 or more per month.

The game creation process by Astra, demonstrated by Director Kim Deok-jin, began with a single natural language command specifying requirements such as farm life, mole interruptions, farm tool upgrades, and the goal of a large farm. Astra performed the task for 13 minutes and 22 seconds, completing a game that reflected the requirements. Kim Deok-jin stated, "I wrote a single line asking to make it accessible to everyone. Then I ran it. The AI worked for 13 minutes and 22 seconds."

His perspective is that while AI provides basic completeness, delicate improvements and final touches require expert intervention.

One of the biggest problems with existing AI models is their 'black box' nature, where their operation is opaque. Not knowing how AI reaches a conclusion has increased user anxiety and distrust. Recently developed AIs are striving to solve this problem by attempting to visually disclose their inference processes to users.

Director Kim Deok-jin explained, "If you leverage this, you can conversely study that field very well." However, criticism has also been raised that the inference tasks provided by AI may not always perfectly match the actual process, leading to ongoing debate about the incompleteness of AI data.

In the past, separate video generation AI was needed, but now AI can create videos through coding and even compose music on its own. Director Kim Deok-jin revealed that AI independently judges and handles tasks not explicitly commanded, such as inserting the Understanding logo or finding old photos on YouTube. Such AI utilizes image data, automatically applies motion graphic effects, and even processes and inserts copyrighted music. He said, "In the past, there was a separate AI for making videos, but now it can code and create videos or even music on its own."

According to this clause, OpenAI is now able to utilize other cloud infrastructures such as AWS, not just Microsoft. Director Kim Deok-jin mentioned, "There was a contract that if we reached AGI, the exclusivity would be lifted, and we could work with others."

The increase in computing resources enables continuous AI development across pre-training, reinforcement learning, safety enhancement, and inference services.

AI's ability to grasp context has been shown to critically influence its performance. When tested without any contextual information, AI's ARC (Abstraction and Reasoning Corpus) score was only 62.7%. However, when contextual information about previous work was provided, the score significantly increased to 99.9%. This demonstrates that memory and context processing capabilities are key drivers of GPU and memory demand.

A rebuttal to the existing memory/SSD limitation theory is being raised, with predictions that as AI models' data learning increases, demand for storage devices and GPUs will grow even further. Director Kim Deok-jin stated that despite feeding AI models significantly more data than before, and not yet even half of the total, performance has greatly improved. This suggests that AI will become even smarter in the future, leading to a surge in demand for hardware like storage devices.

In South Korea, a project is underway to compensate for the incompleteness of the Hangul standard specification. Domestic developers are using AI to develop code that can read and edit HWP files. This project is evaluated as a real-world example demonstrating AI's ability to understand and process complex document formats.

Director Kim Deok-jin revealed that he personally built a reservation system for his own schedule management. He uses a tool created with Claude AI, linking it to his calendar and personal rules. This system works by having the AI coordinate and manage reservations based on information input through KakaoTalk chat captures. This is a practical example of AI effectively replacing traditional secretarial tasks, demonstrating the possibility of personalized task automation. Director Kim stated, "I created 'scheduling an appointment with Kim Deok-jin' myself. I use a tool I personally made with Claude, linking it to my calendar and my rules."

Unexpected results can occur when users give AI the instruction 'simply'. AI may interpret 'simply' as 'less effort and lower quality output', producing results different from the user's intention. This shows a gap between the simplicity intended by the user and the way AI interprets it.

In the AI era, standardized outputs can be easily generated by anyone. Consequently, the importance of individuals with professional insight and a deep understanding of the 'context of the business' is growing. The ability to understand how AI operates, its processes, and to adjust them to achieve optimal results has emerged as an essential core competency. Director Kim Deok-jin said, "Ultimately, people who are good at the context of the business we always talk about, who can adapt and use AI according to its level of intelligence, will become important."

When using high-performance AI models like Astra, cost is an important consideration. AI models incur costs based on token usage, but there is currently a problem where AI does not transparently inform users of specific token consumption. Director Kim Deok-jin pointed out that even though users have a choice between a 'very smart brain' and a 'slightly less smart one,' token usage is not clearly disclosed, forcing users to rely on their intuition for cost efficiency when selecting models.

Current AI technology has reached a stage where AI itself begins to develop some models. A cyclical structure is forming where AI receives feedback and data from people and then creates new models. Thanks to this self-development capability, the major update cycle for AI models has been shortened to every three months. Director Kim Deok-jin explained, "Now, it has started to create AI models itself. Some of them. AI creates AI." This shows a repetitive pattern of companies striving to optimize AI models through distillation methods, and AI's attempts to learn independently to overcome this.