Interview, Fireside Chat
How to use DeepSeek safely
- DeepSeek is expected to be highly vulnerable to basic and zero-day jailbreaks comparable to those seen with GPT-3.5 in 2023, creating significant obstacles for hardening the system for public-facing applications.
- Current performance limitations include slow processing speeds, excessive verbosity, and the inclusion of random Chinese characters, rendering the model unsuitable for immediate daily production use.
- Enterprises deploying the model face anticipated difficulties and "real pain" when attempting to secure infrastructure, which is described as inherently insecure, and may encounter stability and security concerns if used immediately.
- Political censorship limits regarding sensitive topics such as Taiwan and Tiananmen Square are predicted to persist across all future versions of the model.
- Potential future iterations, including DeepSeek version two, may incorporate unknown manipulations or backdoors by Beijing on topics beyond obvious censorship, though the extent remains uncertain.
- An open-source alternative implementing similar reinforcement learning techniques is forecast to emerge within a few weeks, offering reasoning capabilities on par with DeepSeek without requiring long development cycles.
- Organizations building applications on DeepSeek are expected to quickly replace the model once viable alternatives become available due to its current technical and security shortcomings.
- Uncertainty remains regarding whether user data from local or US-hosted instances of the first version will be utilized in the training of future model iterations.