Forget the chat? The lamp of China is working on smart, self-reform AI model. Mint

After shaking Silicon Valley with the AI ​​model earlier this year, the Chinese Startup is working on another innovation to help reduce the lampsac operating costs. The company, led by Liang Wenfeng, is working with researchers at Singihua University, developing a new approach called Generative Reward Modeling (GRM), which rewards the AI ​​model to follow human preferences.

The new approach, previously manifested in a pre-print paper (via Bloomberg), discusses the use of a technique called self-primary critical tuning (SPCT) to make the AI ​​model smart and is more efficient than a self-refusal manner.

Chinese Startup is calling these new models Deepsek-GRM and, like its previous model, is planning to release them based on an open source. Deepsek says its new AI model defeated Google’s Gemini 1.5 Pro, Meta’s Lama 3.1 and GPT -4o of Openai in benchmark score.

What is Deepsek? What was it on Western AI Chatbots?

Deepsek is a Chinese AI startup that operates a chatbot with the same name with two underlying models, Deepsek V3 (Frontier Model like GPT -4 and Mithun 2.0) and Dipsek R1 (Reasoning Model Openi O1).

Deepsek’s AI models went public in January when Chatbot beat Chatgpt as the most popular app on Apple and Google’s app store. As the demand for Deepsek’s chatbott increased, the considering that Western companies were ahead of China in the AI ​​race, quickly collapsed, wiping the value of about 1 trillion dollars of technical shares such as NVidia and Microsoft.

The AI ​​model of Dipsek, built on a shosting budget, also rejected the notion that investment would require billions of dollars in investment to manufacture the AI ​​model.

Deepsek used a machine learning technique called Mixer of Mixter of Experts (MOE) to make its model more efficient. The same technique was used by Mata when she released her Lama 4 Maverick and Lama 4 Scout models on Saturday.

While Deepsek has not said that when it releases its new AI model, some reports suggest that the Dipsek R2 model can be released in early May, as AI is launched by Western companies.