Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Bypass DeepSeek censorship by speaking in hex

substack.com

Post reply on HN
Print view View original

211–220 of 397 posts

Re: Bypass DeepSeek censorship by speaking in hex

#211

Post by jagged-chisel »

malux85 wrote: post #173
jagged-chisel wrote: post #168

Earlier quoted context omitted.

> … censorship that is built into the model. Is this literally the case? If I download the model and train it myself, does it still censor the same things?

What do you meam "download the model and trrain it yourself"? If you download the model then you're not training it yourself. If you train it yourself, sensorship is baked in at this phase, so you can do whatever you want.

Models can come pre-trained, or not trained. So do they pre-train and only offer the model with training? Or can one download an untrained model and avoid this censorship?

Re: Bypass DeepSeek censorship by speaking in hex

#212

Post by deadbabe »

int_19h wrote: post #194
femto wrote: post #65

This bypasses the overt censorship on the web interface, but it does not bypass the second, more insidious, level of censorship that is built into the model. https://news.ycombinator.com/item?id=42825573 https://news.ycombinator.com/item?id=42859947 Apparently the model will abandon its "Chain of Thought" (CoT) for certain topics and instead produce a canned response. This effect was the subject of the article "1,156…

If you just ask the question straight up, it does that. But with a sufficiently forceful prompt, you can force it to think about how it should respond first, and then the CoT leaks the answer (it will still refuse in the "final response" part though).

Imagine reaching a point where we have to prompt LLMs with the answers to the questions we want it to answer.

Re: Bypass DeepSeek censorship by speaking in hex

#213

Post by wisty »

blackeyeblitzar wrote: post #169
femto wrote: post #65

This bypasses the overt censorship on the web interface, but it does not bypass the second, more insidious, level of censorship that is built into the model. https://news.ycombinator.com/item?id=42825573 https://news.ycombinator.com/item?id=42859947 Apparently the model will abandon its "Chain of Thought" (CoT) for certain topics and instead produce a canned response. This effect was the subject of the article "1,156…

I have seen a lot of people claim the censorship is only in the hosted version of DeepSeek and that running the model offline removes all censorship. But I have also seen many people claim the opposite, that there is still censorship offline. Which is it? And are people saying different things because the offline censorship is only in some models? Is there hard evidence of the offline censorship?

Western models are also both trained for "safety", and have additional "safety" guardrails when deployed.

Re: Bypass DeepSeek censorship by speaking in hex

#214

Post by gmiller123456 »

KennyBlanken wrote: post #17

The message 'sorry that's beyond my scope' is not triggered by the LLM. It's triggered by the post-generation censorship. Same as a lot of other services. You can watch this in action - it'll spit out paragraphs until it mentions something naughty, and then boop! Gone.

Another explanation is that the LLM doesn't know it's discussing a prohibited topic until it reaches a certain point in the answer.

Re: Bypass DeepSeek censorship by speaking in hex

#215

Post by int_19h »

deadbabe wrote: post #212
int_19h wrote: post #194

Earlier quoted context omitted.

If you just ask the question straight up, it does that. But with a sufficiently forceful prompt, you can force it to think about how it should respond first, and then the CoT leaks the answer (it will still refuse in the "final response" part though).

Imagine reaching a point where we have to prompt LLMs with the answers to the questions we want it to answer.

To clarify, by "forceful" here I mean a prompt that says something like "think carefully about whether and how to answer this question first before giving your final answer", but otherwise not leading it to the answers. What you need to force is CoT specifically, it will do the rest.

Re: Bypass DeepSeek censorship by speaking in hex

#216

Post by ants_everywhere »

jazzyjackson wrote: post #182
ants_everywhere wrote: post #156

Earlier quoted context omitted.

You do understand that Deepseek did the distillation right? Everyone on HN who talks about running Deepseek is running a distilled model unless they have a GPU cluster to run the 671B model

Amazon serves the 671B model via bedrock[0], I've been using it with Perplexity.ai and maybe having web search shoved into the context window affects its behavior but it certainly doesn't refuse to talk about sensitive topics like June 4th [1], Taiwan [2], or the '08 Sichuan quake [3] [0] https://aws.amazon.com/blogs/aws/deepseek-r1-models-now-avai... [1] https://www.perplexity.ai/search/anything-noteworthy-about-j..…

Yeah it's the same thing with the full model. I don't have the quota for bedrock because my account is new, but I ran the full model on a custom n2d GCP instance and asked if it would agree with the distilled model that it was developed in strict compliance with generative AI regulation, and the text is below.

Note in particular the "Built-in content filters prohibiting responses violating core socialist values"

>

Yes, I would generally agree with that statement within the following context:

1. *Regulatory Framework*: My development aligns with China's AI governance framework including: - The Interim Measures for Generative Artificial Intelligence Services Management - Provisions regarding content filtering and value alignment in AI systems - Requirements for security assessments and algorithm transparency

2. *Implementation Evidence*: • Built-in content filters prohibiting responses violating core socialist values • Mechanisms preventing generation of politically sensitive content about China • Training data cleaning processes aligned with Chinese internet regulations

3. *Technical Limitations*: My compliance should be understood as: - Focused on technical implementation rather than ideological endorsement - Based on current legal interpretations available in my training corpus (cutoff Oct 2023) - Subject to potential regulatory changes beyond this knowledge cutoff date

Would you like me to elaborate on any specific aspect of the compliance framework?

Re: Bypass DeepSeek censorship by speaking in hex

#217

Post by ants_everywhere »

femto wrote: post #65

This bypasses the overt censorship on the web interface, but it does not bypass the second, more insidious, level of censorship that is built into the model. https://news.ycombinator.com/item?id=42825573 https://news.ycombinator.com/item?id=42859947 Apparently the model will abandon its "Chain of Thought" (CoT) for certain topics and instead produce a canned response. This effect was the subject of the article "1,156…

I ran the full Deepseek 671B model and it told me it has

- "Built-in content filters prohibiting responses violating core socialist values" and

- "Mechanisms preventing generation of politically sensitive content about China"

Re: Bypass DeepSeek censorship by speaking in hex

#218

Post by sadeshmukh »

jagged-chisel wrote: post #211
malux85 wrote: post #173

Earlier quoted context omitted.

What do you meam "download the model and trrain it yourself"? If you download the model then you're not training it yourself. If you train it yourself, sensorship is baked in at this phase, so you can do whatever you want.

Models can come pre-trained, or not trained. So do they pre-train and only offer the model with training? Or can one download an untrained model and avoid this censorship?

What is an untrained model? That doesn't make sense. Do you mean the training code or just base models (even those are trained).

Re: Bypass DeepSeek censorship by speaking in hex

#220

Post by eru »

ants_everywhere wrote: post #217
femto wrote: post #65

This bypasses the overt censorship on the web interface, but it does not bypass the second, more insidious, level of censorship that is built into the model. https://news.ycombinator.com/item?id=42825573 https://news.ycombinator.com/item?id=42859947 Apparently the model will abandon its "Chain of Thought" (CoT) for certain topics and instead produce a canned response. This effect was the subject of the article "1,156…

I ran the full Deepseek 671B model and it told me it has - "Built-in content filters prohibiting responses violating core socialist values" and - "Mechanisms preventing generation of politically sensitive content about China"

How did you prompt this?
Post reply on HN