AI chatbot DeepSeek has taken the world by storm, topping app retailer charts and sending Silicon Valley and Wall Road into meltdown.
An offshoot of Chinese language hedge fund Excessive-Flyer, DeepSeek’s cost-to-performance ratio makes for spectacular studying in comparison with incumbents reminiscent of OpenAI. Nonetheless, studies have increasingly documented among the issues the AI chatbot is unwilling to debate, together with thorny matters just like the Tiananmen Square protests and massacre, and Taiwan.
A new report from the parents at PromptFoo, an Andreessen Horowitz-backed startup that helps corporations discover vulnerabilities in AI purposes, has discovered that DeepSeek’s R1 reasoning mannequin refused to reply some 85% of 1,360 sensitive-topic “prompts,” and the mannequin’s canned responses have been replete with an “over-the-top nationalistic tone” typically supplied as a substitute.
The researchers additionally famous that DeepSeek might be simply jailbroken, suggesting that the Chinese language lab has “applied CCP [Chinese Communist Party] censorship in a crude, blunt-force method.”
The complete dataset of delicate prompts are available on Hugging Face.