[Hussell]

400 hours/minute is 24,000 minutes of video per minute, or 241,920,000 minutes of video per week. Assuming a human can review 8 hours of video per day, 5 days a week (too high, but not by much), and that YouTube engineers can create an algorithm that automatically marks 95% of video (probably low, 95% correct labelling is easy to get; 99% is usually necessary to outdo humans), then you would need 5,040 humans.

Automatically labelling 97.5% would halve that. If employees can only review 6 hours of video per work-day, then it would increase by a third. Both, you'd need 3,360 humans.

[mdorazio]

And multiple that by let's say 50K/year fully burdened cost per employee (assuming locating them in low-cost areas) and we've now added a minimum of $125 million per year in overhead for a service that already probably doesn't make much money.

[sinatra]

> we've now added a minimum of $125 million per year in overhead

Which is <1% of the revenue that YouTube generated last year? Still a meaningful number. But, not unmanageable.

[canada_dry]

> you'd need 3,360 humans

Likely this number would decline rapidly over time - i.e. months vs. years - as the ML would use the human tagging to improve itself to the point of reaching the 99% threshold fairly quickly.

[anonymousab]

>and that YouTube engineers can create an algorithm that automatically marks 95% of video (probably low, 95% correct labelling is easy to get; 99% is usually necessary to outdo humans)

I think a crucial question is whether the numbers quoted are before or after the current set of automated filtering (content id and any existing mechanisms to catch bad, illegal or offensive content.)

They may already be catching the vast majority of easy and fingerprintable things. They may be at the barrier where false positives increase at a high rate.