The article says that DeepSeek was easier to unalign to obey the users instruction. It has less refusals and they make that sound like a bad thing.
If anything, it’s a glowing positive praise for the model. Looks like Western media is starting a campaign to gaslight people into thinking that users being able to tune the model to work the way they want is somehow a negative.
They created “ccp-narrative-bench” to measure political bias and it is exactly what you expect it to be. It is described on page 53 of this document: https://www.nist.gov/system/files/documents/2025/09/30/CAISI_Evaluation_of_DeepSeek_AI_Models.pdf
Also aren’t these the guys who put a backdoor in a cryptographic algorithm?
that’s absolutely fucking hilarious