US Scientists Replicate deepseek AI Model for $30

Researchers at the University of California, Berkeley, have successfully replicated the core technology behind DeepSeek’s R1 model for just $30.

Led by PhD student Jiayi Pan, the team developed “TinyZero,” a smaller-scale version of DeepSeek’s AI, utilizing reinforcement learning techniques.

This approach allows the AI to improve its performance by learning from its mistakes over time.

TinyZero was trained using the numbers game from the television show “Countdown,” where players combine random numbers to reach a target value. According to Independent UK.

Initially producing random answers, the AI gradually learned to understand the rules and correct its mistakes, demonstrating the potential of reinforcement learning in developing efficient AI models.

This achievement challenges the prevailing notion that developing advanced AI models requires substantial financial investments and extensive computational resources.

DeepSeek previously claimed that training its main model cost less than $6 million, significantly lower than the budgets of many Western AI companies.

The Berkeley team’s success suggests that effective AI development can be achieved with minimal costs, potentially democratizing access to AI technology.

While TinyZero (GitHub link) is not as complex or scalable as larger models like DeepSeek’s R1, it serves as a proof of concept, indicating that advanced AI capabilities can be developed with limited resources.

This development could have significant implications for the AI industry, encouraging more cost-effective and accessible approaches to AI research and application.

For a more in-depth understanding, you can watch the following video: