We talk about how the Chinese company DeepSeek has trained its AI model R1, which is good at thinking tasks like solving problems and writing code.
R1 is available for anyone to download and has been grabbed over 10.9 million times. The training of R1 cost around $300,000, which is cheaper than what big companies usually spend on similar AIs.
DeepSeek used a method called pure reinforcement learning to train R1, and it has been formally reviewed and approved by other experts as safe and helpful. R1 learned from existing online data, which might have some AI-generated text.
The tests show that R1 performs well compared to other AIs, especially in terms of cost-effectiveness. Other scientists are now using DeepSeek's ideas to create better AIs for various applications.
