3 ms·

I had an idea a few years ago around anonymity, but now it seems even more possible with things like billion parameter LLMs. A service where you enter text that

I had an idea a few years ago around anonymity, but now it seems even more possible with things like billion parameter LLMs. A service where you enter text that you want to post on Reddit/Twitter/etc. but you instead give it to your locally-running model, which is trained to output the text with the following transformations: 1. Writing style change 2. Degree of extra/reduced fluff 3. Degree of typos 4. Typing style change (e.g. using semicolons a lot, misusing commas consistently, conventions like S.O.S vs. SOS) All in an effort to further anonymize your text. The idea is that people have different, consistent grammatical habits and mistakes, writing, concision, etc. You create a "profile" for each identity you want, usually one per account. Another interesting concept to explore is platform style. Have you ever noticed that everyone on Reddit sort of sounds the same? People on HN sort of sound the same. I think there's an emergent platform-based profile that top comments hover around, because adhering to this style is more likely to get upvotes. The hive mind is real!

warner25 4y ago
I just want to say that this is a really interesting comment and set of ideas. I'm a PhD student in the area of privacy and anonymity, so I think about this stuff a lot. To your point about the emergent style of Reddit and HN comments, developed through the timely feedback of upvotes, I've often wondered how tools like suggested or predictive text are changing writing styles; like when I'm tapping out a quick email or text message on my Android phone and the keyboard suggests the next word that I probably wouldn't have come up with on my own, but it sounds good enough or better so I accept it. I thought the analysis that people have done on the writing style of Satoshi Nakamoto was the most persuasive evidence of his real identity, and I think you're correct that someone could now mask or morph their writing style in a very convincing way with these tools.
- throwaway4837 4y ago
  Incidentally, I first learned about Stylometry when reading about people using it to decipher the identity of Satoshi Nakamoto, so I'm happy you brought it up. I had my suspicions that bot farms for Reddit (even before 2019) were doing something similar to this to generate massive amounts of fake account activity to make subreddits appear popular. Now it seems like a piece of cake to do this with things like GPT-3.
O__________O 4y ago
My understanding is that a lot of the APT units use methods like these to make attribution harder and/or make it appear another entity is responsible. I have done a startup related to anonymity and identity protection, and at least at the time, it was amazingly hard sell, rapidly killed it off. Generally speaking, people simply don’t care about anonymity, privacy, etc. As for same style topic for cultural fit, yes, thought it would be funny to communicate with YC using same style as PG; one of the partners now even oddly talks very similar to him, but didn’t in the past.

warner25 4y ago

I just want to say that this is a really interesting comment and set of ideas. I'm a PhD student in the area of privacy and anonymity, so I think about this stuff a lot. To your point about the emergent style of Reddit and HN comments, developed through the timely feedback of upvotes, I've often wondered how tools like suggested or predictive text are changing writing styles; like when I'm tapping out a quick email or text message on my Android phone and the keyboard suggests the next word that I probably wouldn't have come up with on my own, but it sounds good enough or better so I accept it. I thought the analysis that people have done on the writing style of Satoshi Nakamoto was the most persuasive evidence of his real identity, and I think you're correct that someone could now mask or morph their writing style in a very convincing way with these tools.

throwaway4837 4y ago
Incidentally, I first learned about Stylometry when reading about people using it to decipher the identity of Satoshi Nakamoto, so I'm happy you brought it up. I had my suspicions that bot farms for Reddit (even before 2019) were doing something similar to this to generate massive amounts of fake account activity to make subreddits appear popular. Now it seems like a piece of cake to do this with things like GPT-3.

O__________O 4y ago

My understanding is that a lot of the APT units use methods like these to make attribution harder and/or make it appear another entity is responsible. I have done a startup related to anonymity and identity protection, and at least at the time, it was amazingly hard sell, rapidly killed it off. Generally speaking, people simply don’t care about anonymity, privacy, etc. As for same style topic for cultural fit, yes, thought it would be funny to communicate with YC using same style as PG; one of the partners now even oddly talks very similar to him, but didn’t in the past.