6 ms·

Am I the only one wondering how any half capable AI classifier would be unable to identify a literal beheading video or man having s*x with a fish? And especial

Am I the only one wondering how any half capable AI classifier would be unable to identify a literal beheading video or man having s*x with a fish? And especially at a cost that would justify using manual labellers? I get that these human agents are most likely providing labelled training data as opposed to actual filtering but still, surely on more edge-case footage where the model is not confident: not things that could very easily and confidently be classified as inappropriate.

Culonavirus 3y ago
How do you know that already isn't happening? When we're talking about FB, X, Youtube, we're talking about (tens of?) billions of multimedia posts a year. "Edge-case" will still be a massive amount. A quick Google tells me about 1% of the population are psychopaths. There are 80 million people on the planet that have no empathy or conscience. And that's just one angle we could look at this from. The point is, there's a truckload of fucked up shit on the net and AI isn't that good yet.
- justinl33 3y ago
  > The point is, there's a truckload of fucked up shit on the net and AI isn't that good yet. I would beg to differ. AI is definitely good enough to identify fucked up things. Feed GPT-4 even a subtly fcked up question, or fcked up image and it'll be onto you immediately When I said edge cases, I wasn't arguing that there wouldn't be any edge cases - I'm saying the human agents were being shown videos that were not edge cases - i.e sexual abuse, gore - if it's bad enough to traumatize a human being, it's also high signal enough for a model to pick it up. Yes, human agents should be used for labelling ACTUAL edge cases where a model might be confused - i.e is this a nerf gun fight or a real gun fight, is this person nude or is it just their shirt... not a plain beheading video.
schroeding 3y ago
I would very much hope that the available training data for "man fucking fish" is pretty sparse, maybe too sparse to build a training set. (I'm not as optimistic about the sparseness of beheading videos, though)
ogurechny 3y ago
M*n stick their p*p*s into aquatic me*t?!
dudeinjapan 3y ago
First they came for the fish f*ckers, and I did not speak out— Because I did not f*ck fish.

Culonavirus 3y ago

How do you know that already isn't happening? When we're talking about FB, X, Youtube, we're talking about (tens of?) billions of multimedia posts a year. "Edge-case" will still be a massive amount. A quick Google tells me about 1% of the population are psychopaths. There are 80 million people on the planet that have no empathy or conscience. And that's just one angle we could look at this from. The point is, there's a truckload of fucked up shit on the net and AI isn't that good yet.

justinl33 3y ago
> The point is, there's a truckload of fucked up shit on the net and AI isn't that good yet. I would beg to differ. AI is definitely good enough to identify fucked up things. Feed GPT-4 even a subtly fcked up question, or fcked up image and it'll be onto you immediately When I said edge cases, I wasn't arguing that there wouldn't be any edge cases - I'm saying the human agents were being shown videos that were not edge cases - i.e sexual abuse, gore - if it's bad enough to traumatize a human being, it's also high signal enough for a model to pick it up. Yes, human agents should be used for labelling ACTUAL edge cases where a model might be confused - i.e is this a nerf gun fight or a real gun fight, is this person nude or is it just their shirt... not a plain beheading video.

schroeding 3y ago

I would very much hope that the available training data for "man fucking fish" is pretty sparse, maybe too sparse to build a training set. (I'm not as optimistic about the sparseness of beheading videos, though)

ogurechny 3y ago

M*n stick their p*p*s into aquatic me*t?!

dudeinjapan 3y ago

First they came for the fish f*ckers, and I did not speak out— Because I did not f*ck fish.