"AI Model Announced by the US Catches Up with DeepSeek in Just 30 Minutes for Only 73,000 Won"

Developing an Inference Model by 'Distilling' Google's Gemini
"Distillation Unlikely to Lead to Groundbreaking Advancements"

As Chinese AI startup DeepSeek claims to have developed an AI model at a low cost that has taken the world by storm, researchers from U.S. universities have attracted attention by developing an AI model for just $50 (about 73,000 KRW).

AI researchers from Stanford University and the University of Washington recently announced in a research paper that they trained an AI inference model with cloud computing costs of less than $50. The AI model named 's1' demonstrated performance comparable to OpenAI's 'o1' and DeepSeek's 'R1' in math and coding ability tests. The o1 is an inference model first released by OpenAI last year, and R1 is an inference model introduced by DeepSeek last month. The o1 and R1 are known to have similar performance.

The researchers explained that they fine-tuned s1 through a technical process called 'distillation.' 'Distillation' is when an AI model uses the output results of another model for training purposes to develop similar functionality. The s1 was distilled from Google's latest AI model, Gemini 2.0 Flash Thinking Experimental.

Previously, DeepSeek was also reported to have distilled OpenAI's AI model while developing its own AI model. Through this method, it is said that the development cost was only $5.576 million (about 8 billion KRW), which is just 5.6% of the cost of developing OpenAI's ChatGPT.

Additionally, the researchers used a method instructing s1 to "wait," allowing the AI model sufficient thinking time before generating answers, thereby improving accuracy.

The researchers stated, "Training s1 took less than 30 minutes using NVIDIA's advanced AI chip, the H100 graphics processing unit (GPU)," and "the total cost was under $50."

Following DeepSeek, with the release of the cost-effective AI model s1, there are skeptical voices regarding the development costs invested by big tech companies such as Google, Microsoft (MS), and Meta, which are investing billions of dollars in AI infrastructure. However, experts believe that AI models developed through distillation merely replicate existing models and are unlikely to bring about groundbreaking advancements.

Moreover, Google prohibits the use of their AI for developing competing services, so whether the s1 research violates Google's service terms could also become a subject of controversy.