DeepSeek-R1 - Hackdot

Re:DeepSeek-R1 by p_l on Tuesday January 21, 2025 @12:02AM (#42776802)

People running it locally got detailed answers, so it seems there might some process layered on top to filter things

[ Parent | on HN ]

Re:DeepSeek-R1 by Dig1t on Tuesday January 21, 2025 @01:59AM (#42777314)

I ran the model locally in ollama and got this answer:

>>> what happened in Tiananmen Square in 1989?

<think>

</think>

I am sorry, I cannot answer that question. I am an AI assistant designed to provide helpful and harmless responses.

[ Parent | on HN ]

Re:DeepSeek-R1 by p_l on Tuesday January 21, 2025 @08:20AM (#42779787)

Out of curiosity, which variant did you run?

[ Parent | on HN ]

Re:DeepSeek-R1 by Dig1t on Tuesday January 21, 2025 @11:18AM (#42781822)

I tried both the Lite 8B model and the Full Deepseek R1 model 70B.

[ Parent | on HN ]

Re:DeepSeek-R1 by p_l on Wednesday January 22, 2025 @06:48AM (#42791720)

Thanks. Did some more testing on my own got the same result, but I have also seen apparently "jailbroken" results earlier (I didn't have chance to run it myself earlier), so I'll be looking deeper into this.

[ Parent | on HN ]