> 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.
> Rate limits don't have to be distributed mindlessly equally globally.
Yet that's what the grandparent comment is suggesting:
> 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”.
YT already does this abuse-prevention, they just lean towards letting more content on rather than less. The grandparent suggests that they start limiting uploads in a way that more regular uploaders (eg. a few videos a day) would feel, which would warrant some explanation or statement from the company.
> they just lean towards letting more content on rather than less
Come on. It is more than a lean. Growth is an explicit, fractal goal at every level.
> The grandparent suggests that they start limiting uploads in a way that more regular uploaders (eg. a few videos a day) would feel
I didn't give implementation details on the rate limit, because as I preambled, this was a thought experiment, not a design document. They whole point is to underline the fact that, there is a cost of making the review process better (whether automatic or manual) and to what extent corners should be cut versus this cost should be paid seems to be informed by the objective to grow at any cost.
>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?
That's not what minimize means at all, hence my straw man argument. Minimize means reducing the amount as much as possible while still satisfying other parameters of your equation. Zero makes the whole operation nonviable as we agree due to diminishing returns, so it doesn't satisfy that definition.
The dichotomy of optimizing user frustration and minimizing user frustration is basically a question of what parameter gets what valence in the whole objective function. Maximizing profit at the expense of user frustration is optimizing user frustration, meaning having the perfect amount of user frustration in which their leaving the platform and the cost of reducing the frustration is on balance. Minimizing user frustration means doing this the other way around and having a product that, as Google claims, serves "the user first". If you understood that to be zero, I hope it is clear now. At any rate my point doesn't hinge on what the minimal point is. It is clear that any goal other than optimizing user frustration will be an improvement to current situation as far as what users get out of the transaction, and it is unlikely to be adopted.
> Minimize means reducing the amount as much as possible while still satisfying other parameters of your equation.
> basically a question of what parameter gets what valence in the whole objective function
You're splitting extremely fine hairs there, your difference between minimize and optimize seems to be not much more than how much you like a term "satisfying other parameters of your equation".
>Minimize means reducing the amount as much as possible while still satisfying other parameters of your equation.
You have to stop after "as much as possible". Adding a "while still" makes it optimization, your value judgment doesn't get to determine what does and does not fit inside the "while still" of minimize.
Once somebody calls me out for a logical fallacy based on their own difficult to understand definitions of seemingly common words... if that can't be resolved there really isn't a point in continuing, it's not like we could really communicate anything much less make arguments if we can't agree what minimize means.
> You have to stop after "as much as possible". Adding a "while still" makes it optimization, your value judgment doesn't get to determine what does and does not fit inside the "while still" of minimize.
I'll try to simplify it. Imagine a function with dependent and independent variables. Choosing an independent variable so that the dependent variable is the minimum is minimizing. Imagine the plotting of the equation, we are still trying to be on the curve. That's what "as much as possible" means. You are saying minimizing has to be when that dependent variable is zero. That is only possible if the equation crosses the x axis. For our case, it doesn't because diminishing returns make that impossible. Hence my calling strawman.
> You're splitting extremely fine hairs there
The thing I've been trying to separate is not optimization and minimization because minimization is a particular type of optimization. I'm trying to distinguish what particular variable is given valence while solving the whole system of equations. Because in a case where two of those variables are correlated (user frustration & revenue), picking one over the other chooses different optima for both (optimizing for user frustration vs optimizing for revenue yields different optimum values for their correlated counter-parts; for revenue and user frustration respectively). I want to emphasize, these variables are not independent with respect to each other. So there is no simple "revenue is max and frustration is min" solution for which we could say "youtube should just optimize the whole thing". There is picking to be made and max revenue is picked over min frustration, always.
Picking up on me using words "optimize" vs "minimize" for two different variables appears to me as further strawmanning. I don't ascribe malintent, but what we have been discussing has been irrelevant to the conclusion of the original post from the beginning.
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.