The terrifying cost of AI development: A distilled model costing 1/2,000,000 of the estimated $100 million development cost of Chat-4.
The Rapid Advancement of Generative AI in Software Development
In recent years, the use of generative AI in software development has been advancing rapidly. Generative AI, which understands and processes human language, also demonstrates exceptional capabilities in creating computer programs. Engineers can leverage AI appropriately to enhance productivity.
On December 20, 2024, U.S.-based OpenAI announced its latest AI model, "o3" (pronounced "O-three"). Compared to the previous "GPT-4o" (Four-O), this model significantly improves logical reasoning capabilities. In particular, it achieved top-class performance in automated program generation tests. According to OpenAI's report, "o3" ranks among the top 0.05% of competitive programmers worldwide, possessing skills equivalent to 175th place globally.
Furthermore, the open-source AI technology developed by DeepSeek, a Chinese AI company known for its low-cost development, is causing a global sensation. By utilizing a technique called "simplified knowledge distillation," DeepSeek has developed a model at one-tenth the cost of traditional AI models. Despite using lower-end semiconductors, some reports suggest that it outperforms the AI models of major U.S. companies like OpenAI. However, given that it is developed in China, its security and reliability remain uncertain (potential risks exist).
Note:
"Knowledge distillation" is a technique that transfers and compresses knowledge from a large model into a smaller one, improving inference speed, reducing memory usage, and lowering costs while maintaining performance.
DeepSeek’s open-source AI technology is particularly noteworthy. A research team from Stanford University and the University of Washington developed an AI model called "S1" using a distilled model for less than $50 (approx. 7,560 JPY). Reports indicate that in mathematics and coding tests, S1 performed similarly to ChatGPT’s "O1" and DeepSeek’s "R1". Shockingly, its development cost was 1/2,000,000 of ChatGPT-4’s estimated $100 million development budget. This marks the arrival of an extraordinary era.
For AI developers, this means that anyone can now start AI development from this foundation. Eventually, AI itself will become a fundamental infrastructure of society.
The Growing Use of Generative AI in Japan
According to the December 2023 "DX Promotion Survey" by Mitsubishi Research Institute, 31.8% of Japanese companies are already using generative AI for programming. Alongside external information gathering, document summarization & translation, and internal information search, programming is among the key applications.
U.S. research firm Gartner predicts that the percentage of software engineers using AI code assistants will rise from less than 10% in 2023 to 75% by 2028.
The degree of AI adoption varies by company, but leading Japanese IT firms such as CyberAgent, ZOZO, and Money Forward are actively utilizing AI. For example, Findy, a Tokyo-based IT engineer matching company, has been using GitHub Copilot since 2023. Around 90% of its 50-person engineering team uses it daily, allowing AI to replace approximately 20 business days’ worth of work per month. This has resulted in a productivity increase equivalent to having top engineers earning 12–15 million JPY annually assisting for half a day each.
The Future of Software Engineering in the AI Era
With AI’s advancement, the role of engineers is shifting from simple coding to architectural development, which involves designing entire systems. During the transition to AGI (Artificial General Intelligence), the division of labor between humans and AI remains crucial. However, in the coming ASI (Artificial Super Intelligence) era, only top-tier engineers may be needed, and up to 80% of coding engineers may become obsolete.
A 2023 survey found that 92% of U.S. software developers are already using AI tools, and by 2028, 75% of engineers worldwide are expected to use AI coding assistants. The same trend is accelerating in Japan.
Currently, AI is most commonly used for code autocompletion—when an engineer types part of the code, AI accurately completes the rest. Additionally, AI is being applied to software requirements definition, prototype (mockup & prototype) creation, and development idea brainstorming.
In 2018, Japan’s Ministry of Economy, Trade, and Industry (METI) warned of the "2025 Cliff," where legacy systems could hinder DX (Digital Transformation), potentially causing an annual economic loss of up to 12 trillion JPY. While demand for engineers remains high during the AI transition period due to system updates, cloud adaptation, and AI adoption, this is expected to change in the future.
Going forward, engineers must shift their focus from routine programming to architectural design. The DevOps methodology, which integrates development and operations, is gaining attention, emphasizing robust system design from the early stages of development.
However, Japan faces another challenge: engineer salaries remain significantly lower than in the U.S. and South Korea. Since software development can be done remotely, the outflow of Japan’s top talent to overseas markets seems inevitable.
Embracing the AI Era
The evolution of generative AI is set to drastically transform software development. Engineers must leverage AI as a tool and evolve toward roles that create new value.
The AI era is undoubtedly coming within the next few years. I strongly encourage you to embrace AI as a weapon and pursue software development, intelligence development, and skill enhancement that add value to the future.