A team of researchers at the University of California, Berkeley, successfully re -created core technology behind Deepsek AI for a surprisingly low cost of $ 30. Ph.D. Candidates Ziai PAN, team managed to replicate the abilities of learning the reinforcement of Deepsek R1-Zero using a small language model with only 3 billion parameters.
Regardless of its modest shape, AI demonstrated self-support and search capabilities, allowing it to recur its reactions. To test the skills of solving their problem, researchers used the countdown game, the players using a mathematical puzzle required to reach the target number using arithmetic operations. Initially, AI made random estimates, but through learning reinforcement, it improved its ability to modify and customize its answers.
The PAN team experimented with various model sizes, given that the 500 million-parameter model struggled to refine the reactions, while the 1.5 billion-parameter model began to include modification techniques. Once the AI displayed a significant improvement in skilling problems efficiently, after scaling for 3 to 7 billion parameters.
The power of this entertainment raises questions about the cost of AI development. Currently, Openai charges a fee of $ 15 per million per million through its API, while the Deepsek offers a lower rate of $ 0.55 per million. However, AI researcher Nathan Lambert has expressed doubts about Deepsek’s ability, arguing that its operating costs could exceed $ 500 million annually $ 1 billion.
Additionally, Deepsek has been banned in some parts of America due to concerns over data privacy and national security. While the questions remain, Berkeley’s findings show that the high-demonstration AI models may be more accessible in the near future-which disrupt the dominance of AI giants such as Openi, Google and Microsoft.
Thanks for reading..
