81–90 of 146 posts
Plot twist: all these people sharing on twitter yet another creative way of mentioning Xi and Tiananmen in a conversation without triggering the protection (count to 11 in roman numbers, leetspeak etc) were in fact collecting the training data for the nextgen LLM-based protection. Well played!
Reminder on various DeepSeek problems: 1. DeepSeek is full of propaganda/censorship ( https://arstechnica.com/ai/2025/01/the-questions-the-chinese... ) 2. They already had a serious security and privacy issue when they left their database wide open and leaked everyone’s chat history ( https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepse... ) 3. Multiple teams of security researchers found code that links Deep…
>1. DeepSeek is full of propaganda/censorship ( https://arstechnica.com/ai/2025/01/the-questions-the-chinese ...) Your source doesn't even mention "propaganda". Moreover, while censorship is concerning is concerning, I don't see how that practically affects users. If I want to know how to center a div, who cares if it's cagey about what happened in 1989? "Western" AI is also arguably full of "propaganda/censorship".…
Clearly a false equivalence. You think government propaganda compelled by a dictatorship with access to a military and nukes is the same thing?
Here is my story. I needed to buy central console for my car (purchased it a while ago in used cars lot). Went to Amazon and made my selection. Next thing is I see is the warning: this particular console will not fit you car which is MAKE: XXXY, MODEL: YYYY, YEAR: ZZZZ. How's that for data sharing.
At some point you have entered your car data while searching for another car part on Amazon. Amazon caches this information.
Shouldn't this be the other way around? TikTok has the most user data for any LLM to train with. I bet they will make a killing with it, unless of course the CCP decrees that they share it for free.
What useful textual user data do you see coming from TikTok? All the text seems very low quality, to the point where I naively assume that including it in training data would decrease performance.
Some missing context is that the data is shared via the DeepSeek app's use of ByteDance analytics/configuration frameworks. So not a backroom deal where DeekSeek handed over the chat history for its user base, but rather ongoing analytics data being sent from the DeepSeek mobile app. Here's the SecurityScoreCard article that brought attention to this: https://securityscorecard.com/blog/a-deep-peek-at-deepseek/#... Be…
We analyzed the iOS app[1] and observed similar traffic as well as a number of basic security issues (hardcoded encryption keys, use of 3DES and some traffic over HTTP). [1] https://www.nowsecure.com/blog/2025/02/06/nowsecure-uncovers...
Shouldn't this be the other way around? TikTok has the most user data for any LLM to train with. I bet they will make a killing with it, unless of course the CCP decrees that they share it for free.
Firstly, Bytedance is far more than just Tiktok. Secondly, most data in China is shared among most companies anyway, because, firstly, the government (not necessarily CCP) orders most companies to share data with "technological leaders" and "strategically important" companies, and secondly because computer security is mostly an alien concept to Chinese. Copyright (broadly speaking, most restrictions on unrestricted d…
that's the main reason
i don't know how the situation is elsewhere, but in China, 2/3 of startups expose their databases on public network with a password 'abc123'
Earlier quoted context omitted.
Mark Zuckerberg famously had a term for people like you who trusted him.
In case this is not clear enough: you'd better come up with some real arguments with concrete evidence, or move on. Nobody has time for meaningless speculation.
When you have a history of doing greasy shit, you don't get the benefit of the doubt.
Earlier quoted context omitted.
You could literally talk about beyblade toys on a whatsapp video call and you'll be getting Amazon ads for dem blades the next day.
WhatsApp is secured by Signal encryption