Berkeley Researchers Replicate DeepSeek’s AI with $30

Researchers from the University of California, Berkeley managed to replicate the core functions of Chinese artificial intelligence (AI) enigma DeepSeek for only a measly USD30.00 price tag.

This comes as the rise of Chinese AI startup DeepSeek continues to capture global attention, especially after its flagship product, R1-Zero, surpassed ChatGPT on the App Store.

This achievement cemented the company’s position as a major player in AI, and has been with much controversy.

DeepSeek’s mysterious rise to the AI playing field left it open to facing challenges ranging from cyberattacks on its website to skepticism over its claimed reliance on Nvidia’s export-restricted H800 chips.

This also led the researchers at the University of California, Berkeley, to seemingly throw another wrench into DeepSeek’s narrative by replicating its AI’s core functionality for a fraction of the cost.

DeepSeek, an open source project on the surface, has drawn both admiration and doubt for its ability to develop cutting-edge AI technology at a claimed $5 million training cost for its 671-billion-parameter model.

However, the company’s AI prowess remains evident, with reports claiming its model operates at $0.55 per million input tokens—significantly cheaper than OpenAI’s $15 per million tokens for its APIs.

Berkeley researchers, however, have cast doubt on the necessity of such high costs. Led by Ph.D. candidate Jiayi Pan, the team successfully replicated DeepSeek R1-Zero’s capabilities for a mere $30.

They achieved this by using a smaller 3-billion-parameter model and applying reinforcement learning to develop self-verification and advanced search abilities.

Their experiment focused on solving arithmetic challenges, showcasing that sophisticated AI reasoning can be achieved at minimal expense.

Jiayi Pan emphasized the shift to cut costs on AI development and to further democratize its access, stating that their work demonstrates the potential for affordable reinforcement learning scaling research. (GFB)