The Chinese AI startup DeepSeek, based in Hangzhou, is causing a stir in the AI industry. In collaboration with the renowned Tsinghua University in Beijing, the company has developed an innovative method for optimizing language models. This aims to tailor the models' responses more precisely to human needs while generating them faster.
The method developed by DeepSeek combines two previously separate approaches: Generative Reward Modelling (GRM) and Self-Principled Critique Tuning. GRM allows the model to generate its own reward signals instead of relying on external evaluation data. The complementary Self-Principled Critique Tuning allows the model to evaluate its own responses based on self-learned principles. This combination acts like an internal quality control, improving the accuracy and relevance of the responses.
The stated goal of this combined method is to develop language models that respond faster and more accurately to open-ended questions. DeepSeek is not only concerned with exceeding technological benchmarks but, above all, with meeting the actual needs of users. The developers emphasize that the models should deliver not only logically correct but also socially appropriate responses.
Initial test results, published in a paper on the Arxiv platform, indicate the potential of the new method. The DeepSeek models are said to have outperformed existing methods and can compete with established reward models. Although these results are promising, it remains to be seen how the method will perform in practice.
The publication of the new optimization method coincides with the upcoming release of DeepSeek R2, the successor to the already successful language model R1. Rumors about the imminent release of R2 have been circulating for some time, but official confirmation from DeepSeek is still pending. In contrast to US AI startups like OpenAI or Anthropic, which communicate their developments offensively, DeepSeek pursues a strategy of understatement and focuses on research and open source.
DeepSeek has already published code repositories in the past and announced that it will develop with "full transparency" in the future. The newly introduced GRM models are also to be made open source. However, a concrete date for the release has not yet been given. It remains to be seen to what extent DeepSeek will implement its transparency promise in practice.
DeepSeek is celebrated in China as a beacon of hope in the field of Artificial Intelligence. Founder Liang Wenfeng attended a meeting with tech entrepreneurs at the end of February, to which President Xi Jinping personally invited him. This underlines the importance that China attaches to technological independence in the AI sector. DeepSeek exemplifies China's ambition to play a leading role in the global AI competition.