Academics bypass DeepSeek to discuss Tiananmen Square openly

Earlier this year, a Chinese AI chatbot named DeepSeek caused a stir in Silicon Valley with the launch of a new AI model capable of rivaling OpenAIs ChatGPT while requiring far less computing power. The model, called DeepSeek R1, quickly gained attention for its efficiency, triggering a market selloff that erased $1 trillion from the AI-driven tech boom in January.

However, R1 had a major limitation: it strictly adhered to Chinas censorship rules. The AI refused to respond to sensitive topics, including the 1989 Tiananmen Square events or comparisons between President Xi Jinping and Winnie-the-Pooh.

Now, researchers at the Spanish quantum computing company Multiverse claim to have bypassed these restrictions, according to MIT Technology Review. In addition to removing the models censorship constraints, the team reports they have reduced its size by 55%, making it even more efficient without sacrificing performance.

Although DeepSeek had released smaller versions of R1, these compressed models may offer greater compute efficiency but none of them fully stack up to R1, according to a Multiverse blog post. Using a proprietary method called CompatifAI, the team says they successfully eliminated the models limitations while keeping its capabilities intact. CompatifAI works by removing parameters that minimally impact overall performance, including behaviors related to censorship, using quantum-inspired tensor networks to manipulate large data grids.

Tests showed that the compressed model experienced only minor losses in accuracy. Unlike the original R1, the modified AI now freely responds to questions that were previously blocked, such as analyzing the implications of Xi Jinpings removal of presidential term limits or describing what happened in Tiananmen Square in 1989. It even handles whimsical comparisons, like whether someone resembles Winnie-the-Pooh.

Experts note that China is producing some of the most influential open-source AI models, shaping the global information landscape through built-in censorship. Yet, much of the training data may already be affected by these restrictions, posing ongoing challenges for AI development worldwide.

Author: Sophia Brooks