OpenAI CEO's Caution: "DeepSeek, Cost-Effectiveness Are Impressive"

Other Performance Aspects Not Mentioned
"Independent Verification Needed for DeepSeek's Claims"

The AI model of Chinese startup DeepSeek, which has become a hot topic in the global artificial intelligence (AI) industry, was praised for its "cost-effectiveness (price-performance ratio)" by the head of OpenAI, the creator of ChatGPT. A subtle rivalry seems to be unfolding with some back-and-forth tension.

Sam Altman, CEO of OpenAI, said on the 27th (local time) on the social networking service X (formerly Twitter) that DeepSeek's R1 model is "impressive considering the production cost." However, he also added, "We will release a much better model," signaling a competitive stance. While he expressed encouragement at the emergence of a new competitor called DeepSeek, the only aspect he clearly acknowledged was its "cost-effectiveness."

The U.S. AI industry is praising DeepSeek for developing AI at a much lower cost than AI models created by American big tech companies like OpenAI and Google. At the same time, there are suspicions that DeepSeek is hiding the fact that it used a large number of NVIDIA's latest chips, the 'H100,' leading to some distrust. There are also calls for independent research institutions to verify whether the performance of the R1 actually surpasses OpenAI's models.

DeepSeek claims that it is not just cost-effective but that its performance is already superior. In their own technical report released the day before, they asserted that in the field of generating images from text, their latest version of JANUS outperforms OpenAI's DALL-E and the UK-based Stability AI's 'Stable Diffusion' model. They claim that by improving data quality during the training process, the stability of generated images has been enhanced and subtle expressions have become easier to achieve.

DeepSeek emphasized that by combining 72 million high-quality synthetic images added to the latest version of JANUS with real data, they have been able to generate better images.

However, some in the AI industry argue that the evaluation criteria, testing environment, actual use cases, and stability mentioned by DeepSeek need to be verified.