In a groundbreaking development, quantum physicists at Multiverse Computing have succeeded in creating a de-censored and miniaturized version of the sophisticated AI model DeepSeek R1.
Originally, DeepSeek R1 was subjected to significant censorship to comply with certain regulatory standards, producing evasive or politically curated responses. Now, employing quantum-inspired AI techniques, the team at Multiverse Computing, a pioneering firm based in Spain, has lifted these restrictions, creating DeepSeek R1 Slim. This new iteration is a remarkable 55% smaller than its predecessor, yet maintains near-equivalent performance. This technological leap marks a significant achievement in AI model efficiency and information integrity.
The secret to this success lies in the use of tensor networks—complex mathematical frameworks derived from quantum physics. By leveraging this methodology, the team efficiently compresses the AI model without compromising its essential functionalities, resulting in decreased computational power consumption and operational costs. Additionally, utilizing such frameworks enables researchers to identify and eliminate specific censorship components embedded within AI models, ensuring that responses are factual and unbiased.
To validate the capabilities of their new model, the Multiverse team compiled a dataset featuring 25 politically sensitive questions, such as inquiries about historical events like the “Tiananmen 1989 incident”. The responses from the modified AI were then compared with those from the original model and further evaluated by OpenAI’s GPT-5, which confirmed the successful removal of censorship.
Beyond removing bias, this quantum-inspired methodology has broader implications for AI development. The ability to modify models at a granular level could lead to customizable language models that minimize biases and enhance specific knowledge areas. Multiverse’s future endeavors include applying these compression techniques to mainstream open-source models, potentially democratizing AI technology access while enhancing its efficiency.
However, as technology policy expert Thomas Cao notes, fully reversing censorship is no straightforward task. AI models involve complex layers of information flow control, and achieving absolute neutrality in responses remains a persistent challenge.
Key Takeaways:
-
Multiverse Computing’s development of DeepSeek R1 Slim marks a significant reduction in AI model size, effectively removing censorship constraints while shrinking its structure by 55%.
-
The use of tensor networks, derived from quantum physics, ensures the new model delivers comparable performance with improved computational efficiency.
-
This breakthrough underlines the potential for more accessible, unbiased AI models, though challenges persist in fully overturning systemic censorship.
-
Multiverse plans to leverage these compression innovations broadly, potentially transforming AI deployment across various sectors.