3 ms·

> My presumption is that by far the single biggest signal on whether to display a post on the FB feed to more people is how many people like and share. That's a

> My presumption is that by far the single biggest signal on whether to display a post on the FB feed to more people is how many people like and share. That's a people problem at the core, not an algorithm problem. It's still an algorithm problem at its core, because the algorithm is what determines what the "single biggest signal" it will use for amplification is. There's no reason why maximum shareability should be the main factor, it could be something else like some proxy for erudition, for instance.

intended 8y ago
There’s no proxy for content quality. Simply: it’s all content and it’s impact depends on the people who take the information in and then apply their beliefs and understanding to it. Some content is simple - recipes for example. Others are impossible to parse: “they want to reduce hate speech” translates as “they want to remove conservative/right wing speech” (this actually happened in a conversation yesterday. Suggesting that the term hate speech is being transformed/co opted) Engagement in contrast, is simple, reliable and measurable.
- CharlesColeman 8y ago
  > There’s no proxy for content quality. I don't think that's true, it's just that the proxies are imperfect. But the specific alternative metric wasn't that important to my point. > Engagement in contrast, is simple, reliable and measurable. That's the root of the problem: the algorithms have been designed to use "simple, reliable, and [easily] measurable" things as their whole input, and then narrowly A/B tested to maximize some other easily measurable output. Looked at holistically, it's too simple to do an acceptable job. If you took a similar attitude, and applied it to create an algorithm for another discipline, like medicine for instance, it'd be pretty clear it'd be a disaster.
jtbayly 8y ago
Even if I accept your claim, the parent was suggesting getting rid of the algorithm and just letting it be human driven. Human driven means most shared is what gets most seen. How is that different than the current algorithm? It doesn’t change anything. That’s my point. However, this shows that the root problem really is human. You can tell because you are suggesting designing an algorithm to fix a pre-existing problem. Without algorithms of any sort, this problem is there. I don’t think it can be solved by algorithm.

intended 8y ago

There’s no proxy for content quality. Simply: it’s all content and it’s impact depends on the people who take the information in and then apply their beliefs and understanding to it. Some content is simple - recipes for example. Others are impossible to parse: “they want to reduce hate speech” translates as “they want to remove conservative/right wing speech” (this actually happened in a conversation yesterday. Suggesting that the term hate speech is being transformed/co opted) Engagement in contrast, is simple, reliable and measurable.

CharlesColeman 8y ago
> There’s no proxy for content quality. I don't think that's true, it's just that the proxies are imperfect. But the specific alternative metric wasn't that important to my point. > Engagement in contrast, is simple, reliable and measurable. That's the root of the problem: the algorithms have been designed to use "simple, reliable, and [easily] measurable" things as their whole input, and then narrowly A/B tested to maximize some other easily measurable output. Looked at holistically, it's too simple to do an acceptable job. If you took a similar attitude, and applied it to create an algorithm for another discipline, like medicine for instance, it'd be pretty clear it'd be a disaster.

jtbayly 8y ago

Even if I accept your claim, the parent was suggesting getting rid of the algorithm and just letting it be human driven. Human driven means most shared is what gets most seen. How is that different than the current algorithm? It doesn’t change anything. That’s my point. However, this shows that the root problem really is human. You can tell because you are suggesting designing an algorithm to fix a pre-existing problem. Without algorithms of any sort, this problem is there. I don’t think it can be solved by algorithm.