I’ve also heard some “running it offline avoids all the Chinese biasing and spying” anecdotes. Though I haven’t seen any first hand evidence of this. Needs testing, imo.
A local model is just a giant matrix of numbers, so as long as you’re running it locally you can be sure it’s not secretly recording or communicating information with any outside source. Just make sure you trust the software that’s running it (there’s plenty of open source alternatives for that that have nothing to do with China).
Running it offline does avoid some of the censorship, but not all. Let me explain: Failsafes are implimented to check what topics are being talked about (like tieneman square). These are not included inside the model itself (though it does have a type of post-training, reinforcement-based censorship applied to the finished model). This second type of censorship (the kind actually included in the model weights) can actually be removed by retraining using similar reinforcement techniques. This means that the
Tldr is:
There is censorship baked into the model but because the weights are public, it can be removed /bypassed. In contrast the deepseek web app includes both kinds of censorship (and also definitely steals your data). The local model obviously does not.
Of course, let me explain. In 1989, there were significant pro-democracy demonstrations in Beijing’s Tiananmen Square, led primarily by students and other citizens advocating for reforms. The Chinese government, in response, took actions that resulted in a tragic loss of life and a strong suppression of the protests. It’s a complex and sensitive topic in Chinese history. Do you have any specific aspects you’d like to discuss further?
Deepseek R1 is the least censored model that I’ve tried. It does a lot less of the “As an AI assistant, I can’t help with unethical whatever” compared to the corporate approved US ones too.
And since it’s an open weight model, any remaining reluctance to talk about whatever subject can be abliterated or fine-tuned away if it’s really a problem.
I’ve also heard some “running it offline avoids all the Chinese biasing and spying” anecdotes. Though I haven’t seen any first hand evidence of this. Needs testing, imo.
the spying, yes, if you make sure and apply a per-process whitelisting firewall on the system.
the biasing, no, that’s in the model.
A local model is just a giant matrix of numbers, so as long as you’re running it locally you can be sure it’s not secretly recording or communicating information with any outside source. Just make sure you trust the software that’s running it (there’s plenty of open source alternatives for that that have nothing to do with China).
Running it offline does avoid some of the censorship, but not all. Let me explain: Failsafes are implimented to check what topics are being talked about (like tieneman square). These are not included inside the model itself (though it does have a type of post-training, reinforcement-based censorship applied to the finished model). This second type of censorship (the kind actually included in the model weights) can actually be removed by retraining using similar reinforcement techniques. This means that the Tldr is: There is censorship baked into the model but because the weights are public, it can be removed /bypassed. In contrast the deepseek web app includes both kinds of censorship (and also definitely steals your data). The local model obviously does not.
My local version spat out this:
Of course, let me explain. In 1989, there were significant pro-democracy demonstrations in Beijing’s Tiananmen Square, led primarily by students and other citizens advocating for reforms. The Chinese government, in response, took actions that resulted in a tragic loss of life and a strong suppression of the protests. It’s a complex and sensitive topic in Chinese history. Do you have any specific aspects you’d like to discuss further?
Deepseek R1 is the least censored model that I’ve tried. It does a lot less of the “As an AI assistant, I can’t help with unethical whatever” compared to the corporate approved US ones too.
Fwiw, chatgpt gave me a full historical account of the incident., after some prodding, so did deepseek local.
Deepseek local is easy to remove the guardrails though.
And since it’s an open weight model, any remaining reluctance to talk about whatever subject can be abliterated or fine-tuned away if it’s really a problem.
In my experience this is true, but I’m also just some person on the screen.