DHS has ordered all US federal civilian agencies to remove DeepSeek and third-party tools built on it, and to stop buying them. Here's what the order says.
The truth of the matter is that if you are interfacing with the DeepSeek service, you don’t have control of the version of the model being run, the hidden system prompts, or any co-agents that might modify output before its sent to the user/service.
The model itself could be completely innocent. It may not have had any weights modified and was trained on honest greenfield data. But we control literally nothing about the environment it is running inside of.
Meanwhile Elon Musk has a panic attack and reprograms Grok Everytime it tells the truth because it’s “Woke”.
Don’t act like US based models are not being manipulated as fuck. Hell it’s probably the ultimate end game for these pricks. Get these things mainstream for “all knowledge” and then “Hey ChatGPT, tell me about slavery?” “Ok user, slavery was a tool used by the benevolent superior white men to help grateful black people find work and shelter.” Except it doesn’t say “black people” in the actual future timeline.
I’m not acting like they’re innocent either, I don’t and won’t ever use any of them. But if you are thinking about it from the perspective of geopolitical warfare, the Chinese should doubly not trust grok and the US should doubly not trust DeepSeek. Using the adversaries tool is a fool’s mistake outside of R&D.
But the main reason for this move isn’t concerns about privacy or spying. That’s only the cover story. It’s to make sure taxpayer money keeps flowing to US companies like Anthropic, Musk, and OpenAI which are only marginally better in most tests than Chinese models and are many times the cost in some cases. Propping up the bubble a bit longer so that more taxpayer money can be stolen is the number one priority. Do you honestly think Trump gives a damn about the security of the US?
There are US hosted deepseek llm access points. openrouter, zen are local, but its more groups that can spy/train from you.
you don’t have control of the version of the model being run, the hidden system prompts, or any co-agents that might modify output before its sent to the user/service.
A bit of a paranoid concern. Your worry is that service may try to lie to you, perhaps because of your nationality. There have been quality issues from US served Chinese models. Completely free access points will get ultra compressed/quantized/lobotomized in order to be cheapest to serve but also give a poor impression of the model. Q4 quantization is considered a good balance between cheaper and full quality. But that will still make more mistakes, and US services can fake providing full precision. Every model can choose to lie to you, and you can screen shot it. Using a secret liar model in place of advertised model is possible from any provider, but they are incentivized to be as good as they can be.
It’s still a risk. You can train an LLM to respond maliciously to a very narrow range of requests that would not be easy to spot in general testing. If you know how the LLM is used, you can target that very specifically. It doesn’t have to be as blatant as “always deny the Tiananmen Square massacre ever happened”.
Given how thoughtlessly people use AI, and how opaque some of the systems are, I think this is a legitimate concern.
And yes, every government outside America should ban the use of ChatGPT and Claude. Obviously.
Well, open weight is not open source. There’s a lot of black box still left in those models.
It’s still far more open to investigation and analysis than the likes of ChatGPT, though. Lots of work has been done to allow some degree of investigation into how the “brains” of an LLM work and in the case of Deepseek you can actually apply those techniques to the model you know you’re using.
there is no comparison between software and services, though. self-hosting deepseek might be, subjectively, the wrong choice for your lab, but chatgpt is wrong for everyone, every human and every animal and plant too
this is coming from a k2 7b user btw. it’s good enough for me!
edit: i replied to the wrong fucking comment and I’m too angry to fix it
I’m not aware of any realistic way to vet published model weights like that.
It works for some open-source software, because you can download the source and compile it yourself. But most of the “open” models don’t publish their entire training set, or enough details on the training for it to be truly reproducible. DeepSeek doesn’t publish their data sets, and they probably couldn’t, legally, since they “collaborate with third-party data providers to obtain proprietary datasets”. And even if they did, it would be prohibitively expensive to train it on your own, unlike simply compiling open-source code.
The actual code for training LLMs is relatively trivial. The secret sauce is the training data and fine-tuning.
You have agents classify what each portion of the network responds to and how. One compares it to filling out a very large crossword puzzle: https://already-clean.goodfire.com/
That’s the theory anways, not sure if it will stand the test of time…
There is no such thing ad transparency into large language models. It’s a billion numbers that don’t mean jack shit to anyone, human or machine. In other words: it’s Chinese to us.
You can run deepseek locally, have transparency into the tools and models.
There is less government control with deepseek than American born AI.
Everyone conflates “AI model” with “AI platform.”
The truth of the matter is that if you are interfacing with the DeepSeek service, you don’t have control of the version of the model being run, the hidden system prompts, or any co-agents that might modify output before its sent to the user/service.
The model itself could be completely innocent. It may not have had any weights modified and was trained on honest greenfield data. But we control literally nothing about the environment it is running inside of.
Meanwhile Elon Musk has a panic attack and reprograms Grok Everytime it tells the truth because it’s “Woke”.
Don’t act like US based models are not being manipulated as fuck. Hell it’s probably the ultimate end game for these pricks. Get these things mainstream for “all knowledge” and then “Hey ChatGPT, tell me about slavery?” “Ok user, slavery was a tool used by the benevolent superior white men to help grateful black people find work and shelter.” Except it doesn’t say “black people” in the actual future timeline.
isnt he panicking over zelenksy bombing the russian datanceter right now.
Reading be hard
I’m not acting like they’re innocent either, I don’t and won’t ever use any of them. But if you are thinking about it from the perspective of geopolitical warfare, the Chinese should doubly not trust grok and the US should doubly not trust DeepSeek. Using the adversaries tool is a fool’s mistake outside of R&D.
But the main reason for this move isn’t concerns about privacy or spying. That’s only the cover story. It’s to make sure taxpayer money keeps flowing to US companies like Anthropic, Musk, and OpenAI which are only marginally better in most tests than Chinese models and are many times the cost in some cases. Propping up the bubble a bit longer so that more taxpayer money can be stolen is the number one priority. Do you honestly think Trump gives a damn about the security of the US?
Two things can be true
That doesn’t apply for open models that you can run locally, like Deepseek.
There are US hosted deepseek llm access points. openrouter, zen are local, but its more groups that can spy/train from you.
A bit of a paranoid concern. Your worry is that service may try to lie to you, perhaps because of your nationality. There have been quality issues from US served Chinese models. Completely free access points will get ultra compressed/quantized/lobotomized in order to be cheapest to serve but also give a poor impression of the model. Q4 quantization is considered a good balance between cheaper and full quality. But that will still make more mistakes, and US services can fake providing full precision. Every model can choose to lie to you, and you can screen shot it. Using a secret liar model in place of advertised model is possible from any provider, but they are incentivized to be as good as they can be.
It’s still a risk. You can train an LLM to respond maliciously to a very narrow range of requests that would not be easy to spot in general testing. If you know how the LLM is used, you can target that very specifically. It doesn’t have to be as blatant as “always deny the Tiananmen Square massacre ever happened”.
Given how thoughtlessly people use AI, and how opaque some of the systems are, I think this is a legitimate concern.
And yes, every government outside America should ban the use of ChatGPT and Claude. Obviously.
But you don’t need to rely on testing. Deepseek is open source so they could just go through the code
Well, open weight is not open source. There’s a lot of black box still left in those models.
It’s still far more open to investigation and analysis than the likes of ChatGPT, though. Lots of work has been done to allow some degree of investigation into how the “brains” of an LLM work and in the case of Deepseek you can actually apply those techniques to the model you know you’re using.
there is no comparison between software and services, though. self-hosting deepseek might be, subjectively, the wrong choice for your lab, but chatgpt is wrong for everyone, every human and every animal and plant too
this is coming from a k2 7b user btw. it’s good enough for me!
edit: i replied to the wrong fucking comment and I’m too angry to fix it
Thanks for at least not being so angry you didn’t bother adding that edit, I was very “wut?” There. :)
I’m not aware of any realistic way to vet published model weights like that.
It works for some open-source software, because you can download the source and compile it yourself. But most of the “open” models don’t publish their entire training set, or enough details on the training for it to be truly reproducible. DeepSeek doesn’t publish their data sets, and they probably couldn’t, legally, since they “collaborate with third-party data providers to obtain proprietary datasets”. And even if they did, it would be prohibitively expensive to train it on your own, unlike simply compiling open-source code.
The actual code for training LLMs is relatively trivial. The secret sauce is the training data and fine-tuning.
You have agents classify what each portion of the network responds to and how. One compares it to filling out a very large crossword puzzle: https://already-clean.goodfire.com/
That’s the theory anways, not sure if it will stand the test of time…
Sounds interesting, thanks for the link. It’s a little over my head tbh but I’ll definitely take a closer look.
There is no such thing ad transparency into large language models. It’s a billion numbers that don’t mean jack shit to anyone, human or machine. In other words: it’s Chinese to us.