News for hackers. Stuff that matters.
Re:Purple Llama by Meta AI by dwighttk on Thursday December 07, 2023 @11:57AM (#38558744)
>can only assume someone read it, DIDNT EVEN WATCH THE VIDEO,

You are picturing Facebook employing enough people that they can investigate each flag personally for 15 minutes before making a decision?

Nearly every person you know would have to work for Facebook.

Re:Purple Llama by Meta AI by ForkMeOnTinder on Thursday December 07, 2023 @12:39PM (#38559305)
It wouldn't take 15 minutes to investigate. That's just how long the auto_deny_appeal task took to work its way through some overloaded job queue.
Re:Purple Llama by Meta AI by Ambroos on Thursday December 07, 2023 @12:43PM (#38559359)
I worked on Facebook copyright claims etc for two years, which uses the same systems as the reports and support cases at FB.

I can't say it's the case for OPs case specifically, but I absolutely saw code that automatically closed tickets in a specific queue after a random(15-75) minutes to avoid being consistent with the close time so it wouldn't look too suspicious or automated to users.

Re:Purple Llama by Meta AI by sroussey on Thursday December 07, 2023 @01:17PM (#38559771)
This “random” timing is even required when shutting down child porn for similar reasons. The Microsoft SDK for their mandated by congress service explicitly says so.
Re:Purple Llama by Meta AI by black_puppydog on Thursday December 07, 2023 @01:18PM (#38559782)
100% unsurprising, and yet 100% scandalous.
Re:Purple Llama by Meta AI by coldtea on Thursday December 07, 2023 @01:12PM (#38559706)
>It wouldn't take 15 minutes to investigate.

If they actually took the effort to investigate as needed? It would take them even more.

Expecting them to actually sit and watch the video and understand meme/joke talk (or take you at face value when you say it's fine)? That's, like, crazy talk.

Whatever size the team is, they have millions of flagged messages to go through every day, and hundreds of thousands of appeals. If most of that wasn't automated or done as quickly and summarily as possible, they'd never do it.

Re:Purple Llama by Meta AI by themdonuts on Thursday December 07, 2023 @01:32PM (#38559939)
Could very well be! But also let's not forget this type of task is outsourced to external companies with employees spread around the world. To understand OP's comment was a joke would require some sort of internet culture which we just can't be sure every employee on these companies has.
Re:Purple Llama by Meta AI by orly01 on Thursday December 07, 2023 @12:48PM (#38559414)
I agree with you, no way a human reviewed it.

But this implies that people at facebook believe so much in their AI that there is no way at all to appeal what it does to a human eventually. Not even for doing learning reinforcement they have human people to review eventually some post that a person keep saying the AI is flagging incorrectly.

Either they trust too much in the AI or they are incompetent.

Re:Purple Llama by Meta AI by dragonwriter on Thursday December 07, 2023 @12:53PM (#38559465)
> But this implies that people at facebook believe so much in their AI that there is no way at all to appeal what it does to a human eventually

No, it means that management has decided that the cost of assuring human review isn't worth the benefit. That doesn't mean they trust the AI particularly, it could just mean that they don’t see avoid false positives on detecting unwanted content as worth much cost to avoid.

Re:Purple Llama by Meta AI by orly01 on Thursday December 07, 2023 @01:01PM (#38559556)
Yep, that's why I said either that, or they are incompetent.

Not caring at all about false positives, which by the way are very common, enters the category of incompetence for me.

Re:Purple Llama by Meta AI by dragonwriter on Thursday December 07, 2023 @01:10PM (#38559669)
Someone having different goals than you would like then to have is a very different thing than incompetence.
Re:Purple Llama by Meta AI by AnthonyMouse on Thursday December 07, 2023 @01:39PM (#38560029)
If you employ someone to do a job and your goal is to have them do the job effectively and their goal is to get paid without doing the work, arguing about whether this is incompetence or something else is irrelevant and they need to be fired regardless.
Re:Purple Llama by Meta AI by dragonwriter on Thursday December 07, 2023 @01:51PM (#38560180)
Yes, but your complaint is that the job people at facebook are paid to do isn't the one you want them to be paid to do, not that they aren't doing what they are actually paid to do effectively.

Misalignment of Meta's interests with yours, not incompetence.

Re:Purple Llama by Meta AI by AnthonyMouse on Thursday December 07, 2023 @04:37PM (#38562143)
It's not Facebook's employees who need to be fired, it's Facebook.
Re:Purple Llama by Meta AI by r3trohack3r on Thursday December 07, 2023 @01:22PM (#38559829)
Facebook has decided to act as the proxy and archivist for a large portion of the world's social communication. As part of that work, they have personally taken on the responsibility of moderating all social communication going through their platform.

As you point out, making decisions about what people should and should not be allowed to say at the scale Facebook is attempting would require an impractical workforce.

There is absolutely no way Facebook's approach to communication is scalable. It's not financially viable. It's not ethically viable. It's not morally viable. It's not legally viable.

It's not just a Facebook problem. Many platforms for social communication aren't really viable at the scale they're trying to operate.

I'm skeptical that a global-scale AI working in the shadows is going to be a viable solution here. Each user, and each community's, definition of "desired moderation" is different.

As open-source AI improves, my hope is we start seeing LLMs capable of being trained against your personal moderation actions on an ongoing basis. Your LLM decides what content you want to see, and what content you don't. And, instead of it just "disappearing" when your LLM assistant moderates it, the content is hidden but still available for you to review and correct its moderation decisions.

Re:Purple Llama by Meta AI by Joeri on Thursday December 07, 2023 @03:12PM (#38561211)
For the reality of just how difficult moderation is and how little time moderators have to make a call, why not enjoy a game of moderator mayhem? https://moderatormayhem.engine.is/
Re:Purple Llama by Meta AI by mcfedr on Thursday December 07, 2023 @03:35PM (#38561477)
Fun game! Wouldn't want the job!