Like with all the other Chinese hysteria, I think the American versions present a far more pressing national security issue. People complaining that China will use our data against us while American companies already are but they don't wanna do a damn thing about that (other than try to force them to push their horrible racism/sexism/etc).
I saw some article claiming Deepseek was sending unencrypted data to ByteDance servers, but near impossible to know if it was made up shit, was true but doesn't mean anything let alone it presenting the doom of all mankind as TikTok was claimed to be.
Which, people realize Turmp didn't give a shit about TikTok as far as being a security threat, right? Like he doesn't even understand what that means in that regard (people know this is a guy that had taken thousands of classified documents and was showing them to random people including foreign politicians, to brag and shit, right?).
The last one though is interesting. Seems he might be willing to throw the TechBros under the bus if he can get the tech cheaper. China should really be going hard at promising to undercut all the American tech companies (Turmp doesn't actually like them, he only likes Musk at the moment, but doubt that lasts especially if Turmp starts sabotaging Musk's endeavors like electric cars), it likely wouldn't be any worse for Americans in the long-run, especially if it lowers hostility between the US and China. But knowing him, he'd just work a deal for the US to get the Uyghurs and force them to build Turmp branded tech for cheap while he sends his sons and Nazi period trackers to do what the Chinese are supposedly doing.
There are plenty of companies that are hosting DeepSeek. Microsoft is one of them.
You say that like Microsoft doesn't host malicous code or AIs can be used for such purposes.
Service used undocumented APIs and other tricks to bypass safety guardrails.
arstechnica.com
What code?
Edit: there is no code when you download an LLM from like the huggingface repo. There are only weights and biases. Its kinda like an extremely large excel spreadsheet which you download to load using your favourite spreadsheet program.
But there is no code. I run Chinese made LLM's here at home. There is nothing in the files that I download that can call home.
When people are talking about banning Deepseek from devices, they are talking about the Deepseek App. Which just presents a chatting interface to the user infront of Chinese servers hosting a copy of the DeepSeek LLM. But anyone can host a DeepSeek LLM.
If you control the server, you can run any LLM on it without any security issue stemming from the LLM download you made.
Its not just the code that's the issue. Those biases and weights inherently have their own issues. I also think that's a bit of simplification as the LLM inherently will dictate whatever means you use to interact with it and the weights an biases manipulate the output. Plus, AI models are used to write code, it absolutely contains code because it will have been trained on that (even if not specifically for that reason, but incidentally as part of it being trained on language models that will contain discussion of code). As mentioned in that article I linked, that certainly can include the ability for the AI to write malicious code. Heck, Microsoft is suing for breach of terms of service since the AI fundamentally was capable of being used in that manner and it was just their terms of service that actually was violated (since, after all Microsoft was hosting the model that was capable of generating that code).