The Chinese artificial intelligence model Depseek-R1 learns more and better when it receives “rewards” to solve problems, but those stimuli require human intervention, so that approach can be expensive and also limit its growth capacity.
It was verified by a team of researchers and technologists, among which are responsible for the Chinese company that launched this open artificial intelligence model (AI), which analyzed their potentialities and limitations; Today they publish the results of their work in Nature magazine.
Teaching AI models to reason in the same way that humans is a challenge, and researchers corroborated that large -scale language models (LLM) are already demonstrating certain reasoning capabilities, although that training requires important computational resources.
You are interested: Deepseek spears new improved version of its AI model, compatible with Chinese chips
AI models begin to reason
The Deepseek-R1 model includes an additional training stage under human supervision to improve the reasoning process, and uses a “reinforcement” learning system instead of human examples to develop the reasoning steps, which according to the researchers and responsible for the company reduces the costs and complexity of the training.
Notwithstanding, however in the article that some of the limitations of the current version of that model of AI, including that combines two languages, Chinese and English, or that is only optimized for those languages.
They also cite, as a limitation, that there are some tasks in which its model showed no important improvements, such as software engineering, and stressed that future research must focus on improving these ‘reward’ processes to guarantee the reliability of reasoning and tasks performed by this AI.
The researchers showed that the model obtains good results in mathematical, biology, physical or chemical tests, in programming competitions, and concluded that training AI to reason with less human intervention is possible, which opens the door to get models capable of growing, more powerful and cheaper, although there are still many challenges to be resolved.
With EFE information
Subscribe to our YouTube channel and do not miss our content
