Deepseek improves artificial intelligence model with Tsinghua University

China “starting” from deep artificial intelligence I have presented a new way to improve the ability of large linguistic models (LLM) to provide better and faster results In front of her competitors.

Deepseek caused madness in January when he appeared on the scene with R1, which is an artificial intelligence model (AI) with Chatbot, according to the company, was cheaper and worked as well as work SU, main competitor, Chatgpt de Openai.

In cooperation with researchers from the Chinese University of Tsinghwa, Dibsic says in The last article was published last Friday That has developed a technique to improve the same artificial models. Essential technology is artificial intelligence to develop your own rules to judge the content and then use them to improve your answers.

Usually, to improve artificial intelligence, it is necessary to increase the volume of models during training, which requires a lot of human effort and arithmetic strength. instead of, Desseek created a system with a built -in “judge” Who holds artificial intelligence responses in real time. When a question is asked, the judge compares the expected answer of artificial intelligence with the basic rules of Amnesty International and with what should be a good answer. If the answer is similar, then artificial intelligence receives a positive response This helps you improve.

Deepseek calls this automatic improvement system Deepseek-Grm. The researchers assert that this will help the models to work better than competitors such as Google Gemini, Llama de Meta and GPT-4O of Openai. Desseek plans to present these advanced artificial intelligence models As an open source programBut he did not give the deadlines.

Documentation occurs when it is rumored that desseek is It is about to reveal the latest chatbot R2. The company has not indicated any general comments on this topic.