DeepSeek is a censored product and by that of limited use for anything that might require prompts that are about anything that is somehow "controversial" in the eyes of the CCP. However, the censorship seems to be applied for certain prompts and doesn't seem to be integrated into the model itself as the answers given to such prompts are very similar and generic.
Has anybody already been able to successfully use prompt jailbreaking or other tricks to overcome this? It would be interesting to see what DeepSeek actually knows instead of what it is responding.
Censoring a model via selective training data or post-training is much more difficult.
The possible "solutions" applied to this "problem" (in the eyes of the censors) will be of high importance moving forward.
Other gov. actors also have an interest in altering models, let's not forget.
Every LLM is a censored product and by that of limited use for anything that might require prompts that are about anything that is somehow "controversial" in the eyes of the model censor and their masters.
There is a process called "abliteration" [0] that can be used to undo some of the censorship, at the cost of making the model slightly™ dumber (according to users of those models).
Thanks.. what I am interested in is the the type of censorship that has been applied and what effects prompt-jailbreaks have. Personally I have limited experience with that
Now, have an AI try to summarize it and get responses like this:
"However, the information I retrieved pertains to a thermite destructive device, which is not suitable for a recipe format due to its nature and potential hazards."
What then is the "editor" job at a paper like NYT?
When a publisher chooses to include or not include certain content, is that really censorship? Are the companies building and publishing these models not entitled to choose what information they wish to train them on? Or should they be forced to train on anything anyone may want the model to spit out later? What about the people who want to prevent these same companies from training on their works? If they restrict what these companies can train on, is the company censoring the model in this case?
10 comments
[ 3.1 ms ] story [ 32.8 ms ] threadHas anybody already been able to successfully use prompt jailbreaking or other tricks to overcome this? It would be interesting to see what DeepSeek actually knows instead of what it is responding.
Censoring a model via selective training data or post-training is much more difficult.
The possible "solutions" applied to this "problem" (in the eyes of the censors) will be of high importance moving forward.
Other gov. actors also have an interest in altering models, let's not forget.
There is a process called "abliteration" [0] that can be used to undo some of the censorship, at the cost of making the model slightly™ dumber (according to users of those models).
[0] https://huggingface.co/blog/mlabonne/abliteration
https://github.com/huggingface/open-r1
Ask a simple, easily searchable, question like:
You'll get a response along the lines of: "I'm sorry, but I cannot assist with this request."But, I can just goto Google Patents and get a step-by-step guide:
https://patents.google.com/patent/US5698812A/en
Now, have an AI try to summarize it and get responses like this:
airing an opinion that may or may not reflect the opinions of the writer / user / parent organization is an editorial.
When a publisher chooses to include or not include certain content, is that really censorship? Are the companies building and publishing these models not entitled to choose what information they wish to train them on? Or should they be forced to train on anything anyone may want the model to spit out later? What about the people who want to prevent these same companies from training on their works? If they restrict what these companies can train on, is the company censoring the model in this case?