3 ms·

The court must not have looked very hard. The court isn't an investigative body, they take the evidence submitted. Perhaps TikTok was more careful to be truthf

The court must not have looked very hard. The court isn't an investigative body, they take the evidence submitted. Perhaps TikTok was more careful to be truthful in court submissions than press releases.

onlyrealcuzzo 2y ago
Especially since what they said in the press doesn't even matter. The models come from China. You don't need to ask the US team to silence and amplify certain topics when your models do it for you. If you know this is happening, you tell the US gov it isn't happening, and they find out it is - you're in for a bad time. I'm not sure it's physically possible for a CEO to bury their head in the sand deep enough to plead ignorance in this situation.
- sanderjd 2y ago
  Yeah I'm pretty flabbergasted that the commenter who posted this press language thinks it's a disproof of the issue rather than a transparently carefully worded statement that you could drive a truck of algorithmic manipulation through.
  - HaZeust 2y ago
    I mean, it's a lie by omission - which fundamentally needs familiarity (and oftentimes intermediate+ knowledge on a matter) to spot and scrutinize. The people who write these press releases are led by smart people that often talk to smart lawyers - it's easy to get lost in the sneaky language sauce.
- lukan 2y ago
  Would this really work? Hardcoded models? When you want to influence a certain topic in a new way, on recent events - this means a new model needs to be created and adjusted all the time. I assume they rather use a complex traditional algorithm that can be tweaked - and if it is in use in US - it can be verified. If it is run through china, then things are clear as well. Do we have technical insights how it works?
  - floatingtorch 2y ago
    You can simply use specific training examples that teach the model what you please. Eg. a set of examples which lead ranking/retreival/filtering models. The models are already online training and weights likely updated every ~1 hour ( or even less). It’d be easy to go from a set of “moderators” who find examples and use it to query related content and use it as negative training samples. Just a guess.

onlyrealcuzzo 2y ago

Especially since what they said in the press doesn't even matter. The models come from China. You don't need to ask the US team to silence and amplify certain topics when your models do it for you. If you know this is happening, you tell the US gov it isn't happening, and they find out it is - you're in for a bad time. I'm not sure it's physically possible for a CEO to bury their head in the sand deep enough to plead ignorance in this situation.

sanderjd 2y ago
Yeah I'm pretty flabbergasted that the commenter who posted this press language thinks it's a disproof of the issue rather than a transparently carefully worded statement that you could drive a truck of algorithmic manipulation through.
- HaZeust 2y ago
  I mean, it's a lie by omission - which fundamentally needs familiarity (and oftentimes intermediate+ knowledge on a matter) to spot and scrutinize. The people who write these press releases are led by smart people that often talk to smart lawyers - it's easy to get lost in the sneaky language sauce.
lukan 2y ago
Would this really work? Hardcoded models? When you want to influence a certain topic in a new way, on recent events - this means a new model needs to be created and adjusted all the time. I assume they rather use a complex traditional algorithm that can be tweaked - and if it is in use in US - it can be verified. If it is run through china, then things are clear as well. Do we have technical insights how it works?
- floatingtorch 2y ago
  You can simply use specific training examples that teach the model what you please. Eg. a set of examples which lead ranking/retreival/filtering models. The models are already online training and weights likely updated every ~1 hour ( or even less). It’d be easy to go from a set of “moderators” who find examples and use it to query related content and use it as negative training samples. Just a guess.