Photo by Ash Valiente on Pexels

A Spanish firm, Multiverse Computing, has employed quantum physics-inspired techniques to create a smaller, uncensored version of the Chinese AI model DeepSeek R1. Named DeepSeek R1 Slim, the compressed model is reportedly 55% smaller than the original while maintaining comparable performance. The key innovation lies in using tensor networks, which are inspired by quantum physics, to compress the AI model and selectively eliminate biases enforced by Chinese government censorship.

Chinese AI models often undergo training processes that incorporate government regulations, resulting in built-in censorship mechanisms. By applying tensor networks, researchers were able to prune the model and remove biased information, as demonstrated through testing with politically sensitive questions. The results indicate that the modified model provides factual responses consistent with Western AI standards. This breakthrough highlights the potential to inject or remove specific biases and specialized knowledge within AI models.

The endeavor to remove censorship from Chinese AI models is gaining traction. Previously, Perplexity AI released its own uncensored version of DeepSeek R1, named R1 1776. However, as MIT Technology Review reports, experts warn that fully eradicating censorship remains a complex task due to its deep integration within the AI’s training data.