9 ms·

> When you are as big as Google, very improbable failures happen all the time, pointing them out doesn't tell the story of how well they're doing unless your ex

> When you are as big as Google, very improbable failures happen all the time, pointing them out doesn't tell the story of how well they're doing unless your expectation is constant perfection Just a thought experiment, does Youtube need to be as big as it is? Can’t Youtube say “given the current accuracy/precision of ML based solutions and how much our manual operations can scale, we are going to rate limit the content that circulates to match our ability of giving the best judgment on what videos to take down and what videos to keep”. Physical stores have maximum occupancy limits, and with coronavirus some even shown proactive rate limiting to say “hey, we can’t ensure transmission control above a certain density”. Those are corporations too, those also have fiduciary duty. I think “growth at all costs” is a major driver of these types of failures, especially when costs are externalized and distributed to the userbase; one annoying notification, one abuse feature change, one harmful policy decision at a time. > Go find rates, not anecdotes, then you might have something Youtube has near perfect knowledge of the warm data of their operations, and use it to optimize user frustration rather than minimize it, while we only get to know others are being abused when news make it to Hacker News. The inability to find the rates is inherent to the problem. Short of sneaking into Youtube data centers, there is only so much data you can find without Youtube’s say so. Which brings the question, for such a massive policy layer operation, why is this data not made transparent?

colechristensen 6y ago
>Just a thought experiment, does Youtube need to be as big as it is? A few scattered moderation complaints compared to wholesale rate limiting of incoming content... I can only imagine the madness of content producers if YouTube tried that. >Youtube has near perfect knowledge of the warm data of their operations, and use it to optimize user frustration rather than minimize it YouTube content moderation isn't human spaceflight. Having a goal of never failing is, well, a very inexperienced way of thinking about failure. On the order of half a million hours of new video content a day goes on YouTube, or something like 6 million uploads. Let's just pretend that the moderation decision is only made once on upload, you make 6 million decisions a day, how many ridiculous decisions are you going to make every day? Do you think that number is ever going to be zero? The reality is that there is an exponential relationship. Let's just say you can double your cost to cut failures in half. If it's not that relationship it'll be something like it. You can just keep doing it. Doubling your cost over and over, and you're still going to fail all the time. When a "one in a million" chance means something happens several times a day, well, you have to have a different attitude about failure. And when you're just about the only game in town, whatever you do to mitigate failure is going to attract people trying to screw with you to upset you and your users, to take advantage. Not only are you trying not to fail every day, you're trying to avoid people exploiting those attempts. It's just reality that you have a popular service, there is going to be some nonzero number of very upset users with just cause to be upset. That's just reality.
- acituan 6y ago
  > A few scattered moderation complaints compared to wholesale rate limiting of incoming content I'm sure there is already some per-user upload rate limits in place to prevent gross abuse. It is a matter of playing with these parameters. Rate limits don't have to be distributed mindlessly equally globally. Just like in Hacker News or in Twitter, certain proxy variables like the account age and karma scores are used to assess abuse potential and certain capabilities of the platform is unlocked to the user. Most probably something like this also exists. Again, it is a matter of playing with these parameters. Additionally, "few scattered complaints" suffer from selection bias in telling what is actually going on. Not all policy (mis)applications transform into a complaint. If I had to guess, the ratio could be at least an order of magnitude but obviously we have no data on this. > Do you think that number is ever going to be zero? Zero complaints is a straw-man end-state. Of course there is going to be diminishing returns. You can aim to minimize user complaints at 80% to save from diminishing returns, or you can aim to maximize platform usage and ad revenue and take user frustration into account only to the extent it affects platform usage maximization. You can operate on either goals while paying attention to diminishing returns. Very simplistically, there is a relationship such that "for every x dollars we don't spend on reducing abuse, we lose y dollars in user frustration". Who decides the satisficing ratio between x and y? This optimization is done asymmetrically by these types of services and in general there is no way for us to say "wait a second, turns out you were abusing your users/letting them get abused more than you should while making a smidge more profit out of it" with such information asymmetry.
  - colechristensen 6y ago
    >Zero complaints is a straw-man end-state. The comment of yours I was responding to >use it to optimize user frustration rather than minimize it It is not at all a straw man. You are criticizing nonzero optimization and encouraging minimization. To minimize is to have a goal of zero. What other definitions do minimize and optimize have in the dichotomy you are trying to make here?
    - acituan 6y ago

colechristensen 6y ago

>Just a thought experiment, does Youtube need to be as big as it is? A few scattered moderation complaints compared to wholesale rate limiting of incoming content... I can only imagine the madness of content producers if YouTube tried that. >Youtube has near perfect knowledge of the warm data of their operations, and use it to optimize user frustration rather than minimize it YouTube content moderation isn't human spaceflight. Having a goal of never failing is, well, a very inexperienced way of thinking about failure. On the order of half a million hours of new video content a day goes on YouTube, or something like 6 million uploads. Let's just pretend that the moderation decision is only made once on upload, you make 6 million decisions a day, how many ridiculous decisions are you going to make every day? Do you think that number is ever going to be zero? The reality is that there is an exponential relationship. Let's just say you can double your cost to cut failures in half. If it's not that relationship it'll be something like it. You can just keep doing it. Doubling your cost over and over, and you're still going to fail all the time. When a "one in a million" chance means something happens several times a day, well, you have to have a different attitude about failure. And when you're just about the only game in town, whatever you do to mitigate failure is going to attract people trying to screw with you to upset you and your users, to take advantage. Not only are you trying not to fail every day, you're trying to avoid people exploiting those attempts. It's just reality that you have a popular service, there is going to be some nonzero number of very upset users with just cause to be upset. That's just reality.

acituan 6y ago
> A few scattered moderation complaints compared to wholesale rate limiting of incoming content I'm sure there is already some per-user upload rate limits in place to prevent gross abuse. It is a matter of playing with these parameters. Rate limits don't have to be distributed mindlessly equally globally. Just like in Hacker News or in Twitter, certain proxy variables like the account age and karma scores are used to assess abuse potential and certain capabilities of the platform is unlocked to the user. Most probably something like this also exists. Again, it is a matter of playing with these parameters. Additionally, "few scattered complaints" suffer from selection bias in telling what is actually going on. Not all policy (mis)applications transform into a complaint. If I had to guess, the ratio could be at least an order of magnitude but obviously we have no data on this. > Do you think that number is ever going to be zero? Zero complaints is a straw-man end-state. Of course there is going to be diminishing returns. You can aim to minimize user complaints at 80% to save from diminishing returns, or you can aim to maximize platform usage and ad revenue and take user frustration into account only to the extent it affects platform usage maximization. You can operate on either goals while paying attention to diminishing returns. Very simplistically, there is a relationship such that "for every x dollars we don't spend on reducing abuse, we lose y dollars in user frustration". Who decides the satisficing ratio between x and y? This optimization is done asymmetrically by these types of services and in general there is no way for us to say "wait a second, turns out you were abusing your users/letting them get abused more than you should while making a smidge more profit out of it" with such information asymmetry.
- colechristensen 6y ago
  >Zero complaints is a straw-man end-state. The comment of yours I was responding to >use it to optimize user frustration rather than minimize it It is not at all a straw man. You are criticizing nonzero optimization and encouraging minimization. To minimize is to have a goal of zero. What other definitions do minimize and optimize have in the dichotomy you are trying to make here?
  - acituan 6y ago