3 ms·

What fraction of videos are actually problematic? What if it's one in a thousand, and they're reviewing 10x over that? I mean, they also have $120B in the bank.

What fraction of videos are actually problematic? What if it's one in a thousand, and they're reviewing 10x over that? I mean, they also have $120B in the bank. If they thought more reviewers would fix things, I'm pretty sure they'd get more reviewers.

Godel_unicode 7y ago
Those numbers only matter if you're very good at finding the problematic videos via some automatic means. You are vastly underestimating the size of the problem. See also: https://en.m.wikipedia.org/wiki/Base_rate_fallacy https://en.m.wikipedia.org/wiki/Base_rate_fallacy
- killjoywashere 7y ago
  How do you know I'm vastly underestimating the size of the problem? Do you have a source? Random sampling from the ingest stream sounds like a good start for your base rate. Once you have some metadata about bad actors, you can also start hunting their videos and compare your random sample collection stream to the ever-improving bad-actor collection stream.
  - Godel_unicode 7y ago
    To be clear, "size" here refers not to the volume of undesirable content, but rather to the total volume of videos submitted to YouTube. In order to use automatic classification, one needs to examine every new video submitted. YouTube gets (last I heard) double-digit hours of video submitted per second. You have to scan an appreciable portion of the video to determine if it's objectionable. Edit: the common suggestion that you can just ban the few bad actors effectively ignores decades of research into spam detection. Accounts are too easy to create, you have to go after content.
    - killjoywashere 7y ago
      Let's say I have a river. I can only examine 1 drop under a microscope at a time. How do I find the cholera in the river? I take drops and look at them under the microscope. Over and over. I may not sample the whole river, but I can definitely build a representative sample. I can buy more microscopes and hire a bunch of microscopists and increase the sampling rate. Eventually, I'll figure out the amount of cholera per cubic liter, or accept that it's below the limit of detection. In this case, you're arguing that the bad thing is above the limit of detection, so I should be able to find some using random sampling. The metadata I refer to could be all sorts of things. Fingerprints of key frames in the videos, patterns of IP addresses, cities where those streams are coming from, browser fingerprints, etc. If the videos are drops of water, then you can imagine they start as rain. Every drop falls from the sky, somewhere in the river's basin. You can't possibly examine every drop before it gets to the ground. But you can sample them in the river. In this case, I bet Google does, in fact, monitor the browser fingerprints of uploaders. Monitors all sorts of data around the upload event. Every single time. And does use all sorts of fancy algorithms to aid in assessment. And, and, they sample the river, with actual human microscopists/moderators, because they surely assume new patterns of plague are continuously evolving.
      - Godel_unicode 7y ago
        First, the goal here isn't to find the cholera. We know there's cholera in this river, the goal is to remove it before it gets to the village and drunk by the town's people. Same method though, you can only remove one dropper of water per person sampling. Now imagine you have not a river, but the bering strait. As Napoleon said, quantity has a quality all it's own. Please, please read about the base rate fallacy.
        killjoywashere 7y ago
        You can't remove all the bad videos any more than you can remove all the cholera from the river. Do you know how we protect populations from cholera? * isolate the source and provide supportive care (infected people). * make sure the source is downstream of the population (livestock). * On inland waterways where multiple cities depend on the same river, you filter the water until your sampling demonstrates it is below the limit of infectivity. It took a while to figure out the limit of infectivity for cholera. You still do the work. It's going to take a while to figure out the limit of infectivity for videos, github, etc. Also, you really have to let go of this base rate fallacy bit. The methods I describe are entirely general. Also, since you keep talking about the base rate fallacy, please understand I live in the realm of diagnostics. I understand the problem. So, since you brought it up: The difference between the Bering strait and the Mississippi river (not far from where I grew up) is trivial on the scale of cholera. Flow rate of Bering strait: ~ 2.1e4 m3/s. Flow rate of Mississippi: ~ 1.7e4 m3/s. Volume of a V. cholera: 1.17e-19 m3. In either case, the difference is 21 orders of magnitude. If the Bering strait was the lifestream of all 8 billion humans alive today, finding a vibrio in the flow would be equivalent to finding a single person generating half a picosecond of violent video. The point is not finding all of them. The point is sampling well enough that you can calculate the limit of infectivity. Once you can do that, you can start engineering solutions at social scale, e.g. move the village, move the cattle, install filtration. Maybe even fund development of new filtration systems.

Godel_unicode 7y ago

Those numbers only matter if you're very good at finding the problematic videos via some automatic means. You are vastly underestimating the size of the problem. See also: https://en.m.wikipedia.org/wiki/Base_rate_fallacy https://en.m.wikipedia.org/wiki/Base_rate_fallacy

killjoywashere 7y ago
How do you know I'm vastly underestimating the size of the problem? Do you have a source? Random sampling from the ingest stream sounds like a good start for your base rate. Once you have some metadata about bad actors, you can also start hunting their videos and compare your random sample collection stream to the ever-improving bad-actor collection stream.
- Godel_unicode 7y ago
  To be clear, "size" here refers not to the volume of undesirable content, but rather to the total volume of videos submitted to YouTube. In order to use automatic classification, one needs to examine every new video submitted. YouTube gets (last I heard) double-digit hours of video submitted per second. You have to scan an appreciable portion of the video to determine if it's objectionable. Edit: the common suggestion that you can just ban the few bad actors effectively ignores decades of research into spam detection. Accounts are too easy to create, you have to go after content.
  - killjoywashere 7y ago
    Let's say I have a river. I can only examine 1 drop under a microscope at a time. How do I find the cholera in the river? I take drops and look at them under the microscope. Over and over. I may not sample the whole river, but I can definitely build a representative sample. I can buy more microscopes and hire a bunch of microscopists and increase the sampling rate. Eventually, I'll figure out the amount of cholera per cubic liter, or accept that it's below the limit of detection. In this case, you're arguing that the bad thing is above the limit of detection, so I should be able to find some using random sampling. The metadata I refer to could be all sorts of things. Fingerprints of key frames in the videos, patterns of IP addresses, cities where those streams are coming from, browser fingerprints, etc. If the videos are drops of water, then you can imagine they start as rain. Every drop falls from the sky, somewhere in the river's basin. You can't possibly examine every drop before it gets to the ground. But you can sample them in the river. In this case, I bet Google does, in fact, monitor the browser fingerprints of uploaders. Monitors all sorts of data around the upload event. Every single time. And does use all sorts of fancy algorithms to aid in assessment. And, and, they sample the river, with actual human microscopists/moderators, because they surely assume new patterns of plague are continuously evolving.
    - Godel_unicode 7y ago
      First, the goal here isn't to find the cholera. We know there's cholera in this river, the goal is to remove it before it gets to the village and drunk by the town's people. Same method though, you can only remove one dropper of water per person sampling. Now imagine you have not a river, but the bering strait. As Napoleon said, quantity has a quality all it's own. Please, please read about the base rate fallacy.
      - killjoywashere 7y ago
        You can't remove all the bad videos any more than you can remove all the cholera from the river. Do you know how we protect populations from cholera? * isolate the source and provide supportive care (infected people). * make sure the source is downstream of the population (livestock). * On inland waterways where multiple cities depend on the same river, you filter the water until your sampling demonstrates it is below the limit of infectivity. It took a while to figure out the limit of infectivity for cholera. You still do the work. It's going to take a while to figure out the limit of infectivity for videos, github, etc. Also, you really have to let go of this base rate fallacy bit. The methods I describe are entirely general. Also, since you keep talking about the base rate fallacy, please understand I live in the realm of diagnostics. I understand the problem. So, since you brought it up: The difference between the Bering strait and the Mississippi river (not far from where I grew up) is trivial on the scale of cholera. Flow rate of Bering strait: ~ 2.1e4 m3/s. Flow rate of Mississippi: ~ 1.7e4 m3/s. Volume of a V. cholera: 1.17e-19 m3. In either case, the difference is 21 orders of magnitude. If the Bering strait was the lifestream of all 8 billion humans alive today, finding a vibrio in the flow would be equivalent to finding a single person generating half a picosecond of violent video. The point is not finding all of them. The point is sampling well enough that you can calculate the limit of infectivity. Once you can do that, you can start engineering solutions at social scale, e.g. move the village, move the cattle, install filtration. Maybe even fund development of new filtration systems.