A team of researchers at the University of California, Berkeley, has recreated the core technology behind DeepSeek AI at a surprisingly low cost of just $30. Led by Ph.D. candidate Jiayi Pan, the team was able to replicate the reinforcement learning capabilities of DeepSeek R1-Zero using a small language model with just 3 billion parameters.
Despite its modest size, the AI demonstrated self-verification and search capabilities, allowing it to improve its answers after repetition. To test its problem-solving skills, the researchers used the game Countdown, a mathematical puzzle that requires players to reach a target number using arithmetic operations. Initially, the AI made random guesses, but through reinforcement learning, it improved its ability to revise and optimize its answers.
Pan's team experimented with different model sizes, noting that a 500 million-parameter model struggled to improve its answers, while a 1,5 billion-parameter model began to incorporate revision techniques. When 3 to 7 billion parameters were added, the AI showed significant improvement in solving problems efficiently.
The affordability of this build raises questions about the cost of developing AI. Currently, OpenAI charges $15 per million tokens through its API, while DeepSeek is asking $0,55 per million tokens. However, AI researcher Nathan Lambert has expressed skepticism about DeepSeek’s affordability, arguing that its operating costs could run as high as $500 million to over $1 billion per year.
Additionally, concerns about data privacy and national security led to DeepSeek being banned in parts of the US.
Some reports suggest that DeepSeek may have been trained using OpenAI's ChatGPT.
While questions remain, the Berkeley findings show us that high-performance AI models could become much more accessible in the near future, disrupting the dominance of large AI companies like OpenAI, Google, and Microsoft.
https://github.com/Jiayi-Pan/TinyZero
George is still wondering what he is doing here….
