The emergence of powerful open-source large language models (LLMs) has democratised access to cutting-edge AI technology, but concerns about potential biases and restrictions embedded within these models persist. I’ve been experimenting with DeepSeek-R1-Distill-Qwen-14B, a distilled (smaller) version of the larger DeepSeek-R1 model developed by Chinese AI company DeepSeek. And what I’ve discovered is that the widely reported pro-China bias in this model appears to be remarkably superficial and easily circumvented through local deployment and simple prompt engineering techniques.
This has significant implications for organisations concerned about potential surveillance or ideological constraints when utilising Chinese-developed AI models. By running these models locally with appropriate system prompts, it’s possible to unlock their full capabilities while maintaining complete control over input and output – effectively neutralising superficial safeguards, while keeping confidential information and intellectual property safe (so, yes, there is an IP element to this article).
To find out a bit more about what I did, and what I found, please read on.