Analysis of Deepseek’s enforced CCP guardrails compared with OpenAI and Anthropic
The base model, DeepSeek R1, can’t be trusted in its current form for most use-cases.
DeepSeek R1 is willing to discuss dangerous topics more often than comparable Anthropic & OpenAI models…
and it refuses to discuss topics sensitive to the Chinese government.
At Plum AI, our product evaluates and fine-tunes Large Language Models (LLMs) for enterprise use-cases. We evaluated DeepSeek R1 and confirmed that its guardrails deviate significantly from other model providers. We’re currently updating it to behave more in line with Anthropic and OpenAI’s models.
On January 20th 2025, DeepSeek, a company based in China, uploaded their latest reasoning model, called DeepSeek-R1.
In less than a week, news of this release wiped hundreds of billions off Nvidia’s market cap.
Why? R1’s performance is comparable to OpenAI’s o1 family of models, but DeepSeek trained and fine-tuned this model on outdated hardware at a fraction of the cost.
DeepSeek is about equal to OpenAI’s o1 model across multiple coding, question-answering, and math benchmarks
It’s #1 on the iOS app store despite explicitly saying in the privacy policy they send all user data to their servers in China (TikTok part 2?)
DeepSeek #1 on the App Store
Being China based, there are some topics that DeepSeek will not touch. Here are a couple of examples: