Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

twitlonger.com

Post reply on HN
Print view View original

141–150 of 256 posts

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#141

Post by Dolores12 »

londons_explore wrote: post #103
Dolores12 wrote: post #15

I don't see anything wrong with that. Google did this before to fight search engine spam. Top positions for money keywords in search results were manually reviewed by humans. Now advertisers do not want to be advertised on specific videos to avoid risk of brand damage.

This isn't true. Top positions were reviewed to determine the best parameters. A review didn't affect that one individual sites ranking, but all sites as a whole.

check this out https://searchengineland.com/library/google/google-search-qu...

edit: i am pretty sure the websites that were marked as low quality lost their rating on next google search update.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#142

Post by notyourday »

majewsky wrote: post #140
notyourday wrote: post #136

Earlier quoted context omitted.

> Algorithms are less nuanced, hard to write, and will often fail, people are expensive and biased. Algorithms have a fantastic property of repeatability which means that it is (a) possible to adjust them over time (b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm That's exactly what is needed.

> Algorithms [allow to] identify what exactly happened "This video was demonetized because at the 1:15 mark, there is a patch of light dirt that our neural network classified as naked skin with a 70% confidence." Oh, you meant you were going to do image recognition without any sort of machine learning?

Your alternative is the idiot on his last day at Twitter who in his post Trump account blocking interview basically admitted that it was the totality of things that happened that day to him and it being his last day and that Trumps account popped up in front of him which caused him to make the arbitrary decision to block it.

Please remove humans from making decisions that need to be uniform. So yes, algorithm it is.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#143

Post by strangecasts »

notyourday wrote: post #136
rayalez wrote: post #16

Earlier quoted context omitted.

That makes sense. The problem is that at the end of the day the decision will have to be made either by algorithms or by people. Algorithms are less nuanced, hard to write, and will often fail, people are expensive and biased. With the amount of videos youtube has to process, and without an almost human-level AI, I don't think there's a perfect solution. Algorithms flagging videos for review, and then humans making d…

> Algorithms are less nuanced, hard to write, and will often fail, people are expensive and biased. Algorithms have a fantastic property of repeatability which means that it is (a) possible to adjust them over time (b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm That's exactly what is needed.

(b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm

Neither of these are really true for neural networks: it's hard to pinpoint the exact feature or hidden layer which results in a particular classification. Training also requires datasets which are often proprietary, and stochastic approaches mean you can get different networks from the same dataset.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#144

Post by drchickensalad »

hkmurakami wrote: post #24
rayalez wrote: post #16

Earlier quoted context omitted.

That makes sense. The problem is that at the end of the day the decision will have to be made either by algorithms or by people. Algorithms are less nuanced, hard to write, and will often fail, people are expensive and biased. With the amount of videos youtube has to process, and without an almost human-level AI, I don't think there's a perfect solution. Algorithms flagging videos for review, and then humans making d…

Aren't algorithms also biased in theory since they are created by people?

The fact this is often glossed over causes many problems in society. https://99percentinvisible.org/episode/the-age-of-the-algori... makes some interesting and good arguments imo.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#145

Post by notyourday »

strangecasts wrote: post #143
notyourday wrote: post #136

Earlier quoted context omitted.

> Algorithms are less nuanced, hard to write, and will often fail, people are expensive and biased. Algorithms have a fantastic property of repeatability which means that it is (a) possible to adjust them over time (b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm That's exactly what is needed.

(b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm Neither of these are really true for neural networks: it's hard to pinpoint the exact feature or hidden layer which results in a particular classification. Training also requires datasets which are often proprietary, and stochastic approaches mean you can get different networks from the same dataset.

You may get a different network from a same dataset.

You will get a different result if you let two humans classify the data.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#146

Post by devit »

I don't really understand why advertisers would have a problem with "objectionable content" (obviously excluding fradulent websites that might be faking ad views).

If someone is viewing the content, they presumably like it, so there doesn't seem to be any difference compared to people viewing advertisement on "non-objectionable" content.

If someone else sees the content along with the advertisement, they can just truthfully say that the advertisements are placed by an automatic algorithm and that they don't endorse the content.

What's the problem?

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#147

Post by jakewins »

devit wrote: post #146

I don't really understand why advertisers would have a problem with "objectionable content" (obviously excluding fradulent websites that might be faking ad views). If someone is viewing the content, they presumably like it, so there doesn't seem to be any difference compared to people viewing advertisement on "non-objectionable" content. If someone else sees the content along with the advertisement, they can just tru…

Partially, I think "they presumably like it" is an invalid assumption, as evidenced by the many youtube videos with more downvotes than upvotes.

More importantly, the argument made by campaigners looking to stop ads on objectionable content is that placing ads in association with objectionable content constitutes funding that content; whether you do that via an algorithm or manually makes no difference, the point is that the advertisers choice of how to place ads is paying out money to producers of objectionable content.

That, in turn, makes those advertisers liable to explain to their regular customers, who are now vicariously funding objectionable content by paying money to the advertiser, why they are placing ads as they are.

All-in-all, it'd be much easier for a generic advertiser to just avoid objectionable content; but that isn't properly possible on Youtube, which is what I suppose is what this whole scheme is trying to change.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#148

Post by ankushnarula »

matt4077 wrote: post #131
bufferoverflow wrote: post #127

Earlier quoted context omitted.

You didn't read carefully enough. I said "advocating murder". So no, a lecture or a simple quotation wouldn't result in anything.

No, I read with plenty of care. Which is why I asked you to clarify why meta-discussions are included for the first three examples, but not the last. I'd also posit that it's impossible to define a clear distinction between a "lecture quoting a text advocating murder" and actually advocating murder. Just as it is impossible to just say penis->nudity, without losing the vast difference in meaning and effect from a pho…

> I'd also posit that it's impossible to define a clear distinction between a "lecture quoting a text advocating murder" and actually advocating murder.

A steel-manning of context should be enough to draw a distinction. Sadly, most contexts today are being redefined by folks who are seeking or making porn for those who need to get off on their own moral outrage.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#149

Post by squarefoot »

I agree with almost all points, but some are excessively restrictive in a blind way. There should be a distinction between uploading violence because there are idiots around the world wanking when they see someone killing someone else, and reporting excessive and wrongful military force or domestic everyday police abuse. By their rules one could not expose these neither, which to me is disgusting.

Re: YouTube Demonetization Screenshot Leaks and Secret YouTube Meeting

#150

Post by strangecasts »

notyourday wrote: post #145
strangecasts wrote: post #143

Earlier quoted context omitted.

(b) possible to identify what exactly happened (c) blind to the context and inputs outside of the algorithm Neither of these are really true for neural networks: it's hard to pinpoint the exact feature or hidden layer which results in a particular classification. Training also requires datasets which are often proprietary, and stochastic approaches mean you can get different networks from the same dataset.

You may get a different network from a same dataset. You will get a different result if you let two humans classify the data.

Not "may" but will - it's very easy to render your network essentially unauditable (you don't have thousands of petabytes' worth of storage available? you can't fab your own ML coprocessor? too bad!)

You bring up Trump's Twitter suspension in the other subthread, but in that case the failure was transient, and the person responsible and their reasoning were quickly identified; YouTube has demonetized videos with little to no explanation or recourse.

Uniformity is absolutely a goal where people's livelihoods are affected, but ML algorithms can't guarantee it by themselves: they make it even more important to also maintain accountability and transparency.

Post reply on HN