What are we trying to build?

[celtschk]

@YLearn

@celtschk

Taking the Mac vs. Windows issue as an example, let’s say you are generally a Mac user and therefore the AI optimizes your view for that. But then, you have to use Windows at your job, and you happen to have a question on it. So you post on your catch-all site, which has learned that you value the answers from Mac supporters. Now you will prominently see the answers saying “don’t use Windows, Mac can do that much better, here’s how.” Which may match your sentiment as Mac user, but certainly doesn’t help you with your problem as you simply don’t have the choice to use Mac.

Saying the site should post your questions to a particular group just because that is the one you are most active in is just plain silly. Let’s extend your example and say you are now remodeling your bathroom and post a question with regards to “how to replace a faucet.” Would you also expect that to get posted to Mac users? No, because that makes no sense.

I believe you overestimate what machine learning can do. Your faucet example is off-limits because there are unlikely to be any questions that mention both a faucet and a Mac computer, but there should be plenty of questions that mention both Windows and Mac (“I liked this on Mac, can I get this on Windows?”, ”Does an equivalent of this Windows software exist on a Mac?”).

Questions should be evaluated by the site for the content of the post when it is categorized. So your primarily Windows question will get “Windows” based tags and be seen by users who want have been determined to want to see Windows content. You aren’t likely to get “don’t user Windows, Mac can do that much better, here’s how” from those users.

You are aware that current machine learning is not strong AI? Heck, there are some humans I wouldn’t trust to get this right!

@celtschk

In short, relying on machine learning takes away autonomy from you, as you no longer decide yourself which content you see, the machine does for you. Sure, it does by predicting you, but when it predicts wrong, you’ll have a hard time fighting those decisions (if you even notice that your bad experience is because of the machine learning, and not because the community on that site generally is anti-Windows).

Hmm. I get it, you believe machine learning is bad. Machine learning is all around you already in some form or another. Most of it is built on older frameworks, but any site or service that provides your recommendations or suggestions is already doing it.

No, I don’t think machine learning is bad. Machine learning is a great way to automate complex tasks. But there are certain tasks I don’t want to be automated, and for those I also don’t want machine learning.

Now a good use for machine learning on codidact could be a tag suggestion mechanism. That is, the algorithm looks at what humans do with the tags, and then when someone writes a new question, it figures out what tags most likely apply, and the human asking the question then has the option to accept those tags, or to change them before posting. New users would likely take the tags as is. If the tags are wrong, experienced human users will notice it and retag, which gives the algorithm more data to learn from. Experienced users will also review the tags, and change them right on the spot if they don’t fit.

The point is, unlike in your scenario, the user is in control. And the machine simplifies it, but does not bypass the user.

The key to doing it well is to allow the users to tune their own “views” of what they want presented. Pick any of the major streaming music services and you will see this in action. Is it always perfect? Certainly not and you just need to give feedback when it isn’t. And when it predicts right (if you even notice that your good experience is because of machine learning), you get the benefit of finding other topics of interest to you, meeting more people that are interested in the same things as you, etc.

Streaming services form your music taste just as much as they react on it, if not more. It is a known psychological effect that the more you listen to a certain type of music, the more you like it (the same is already true for music radio stations). You may consider that a good or a bad thing, but it certainly means that music streaming services are a particularly bad example. Also, music streaming services replace radio stations where you have even less influence on the music you hear.

Also, music streaming services almost certainly tag the music style by hand. The machine algorithm doesn’t figure it out, it uses that information to

Anyway, I don’t usually use music streaming services, and while I do listen to music on YouTube, I rarely use their auto-generated playlists.

Then again, even when I am a member of community X (let’s say Mac from your example), when I visit the “Mac Users” specific site there is still plenty of “noise” that has no interest to me. The filtering there isn’t perfect either, and often less so as the filtering is purely done by human knowledge and thought processes, which are far from perfect. And there is nothing (besides maybe the “Hot Topics” list) that brings me anything outside of the Mac “filter bubble” I have placed myself into.

This is different because I know I’m on a Mac site, and therefore I know that everything I see will have a Mac bias. And I know that I can easily see a different view by going to the Windows site.

With your machine algorithm, how am I going to see the Windows side if I want to?

@celtschk

You surely have heard the term “filter bubble” before. What you are proposing is basically the building of such filter bubbles. Filter bubbles are harmful, exactly because you don’t notice them. It’s not your decision to see or not see something, it’s the machine’s decision.

You can make this argument about any sort of filtering you choose, even if it is filtering you choose for yourself. And clearly the wrong filters can be dangerous, but filtering is a daily part of life for the human condition. Humans take in far too much information for our brains to process every detail all the time

Filtering you chose is different in that you can always choose to change your filter. A machine learned filter is fixed; you have no choice. Well, you can try to actively work against that filter, but that is hard work, much harder than just switching to a different site, or adding/removing a tag from your preferred tags.

Filtering happens and most of it isn’t your own conscious decision. Your brain does it. The media sources you pay attention can do it. Your cultural group can do it. Where you live can do it. Your circle of friends, your workplace, your [insert whatever you want here] can do it as well.

That’s not an argument for making things worse. That’s an argument for making things better. That’s not an argument for letting the machine chose the community for you. It’s an argument for being member of several communities, so you get several different viewpoints.

The smaller you make your potential source(s) of information, the more likely you are to be putting yourself into a filter bubble. Did you not notice that topic specific communities also form their own filter bubble as well? If you didn’t notice, maybe you should be concerned.

For the topic specific site, I see the topic right in the title. I’m constantly reminded that I’m at a place that is biased in a certain way. When a machine learning algorithm determines what I see, it is intransparent what selection I get to see.

Also,the machine learning algorithm cannot adapt to the fact that at different timesI am interested in different topics.

Earlier today I browsed math.SE, but not worldbuilding.SE. On other times I browse worldbuilding.SE but not math.SE. How is the algorithm supposed to know that now I’m interested in seeing math topics, two hours later I’ll be interested in worlbuilding topics, and tomorrow I’ll be interested in programming topics?

Well, I guess I could get different accounts, and use them specifically for different topics, but then I effectively get back to topic sites, except with significantly higher effort. And without the option to effortlessly start browsing a new topic, as I’ll first have to train the machine learning algorithm on a new account that I’m now interested in that other topic.