* Let me adjust the relative weight that the current video gets vs my past views (i.e. do I want to see related videos to this one, or ones that YouTube generally thinks I'll click on)
* Let me adjust the weighting of thumbs up vs. watch time.
* Let me configure the homepage
* Let me make the entire recommender panel be just subscriptions or items from a particular list
* Let me replace the recommender entirely with something of my own devising, called back through a webhook
These things would make YouTube much more useful for me. I'm not going to YouTube just to kill time, and I pay them $13/mo. They're not getting any ad revenue from me, so why am I stuck with their recommender that only cares what it thinks will cause me to spend the most time on YouTube. I am not interested in spending the most time on YouTube. I'm interested in getting the most out of it in the minimum time.
YouTube, Spotify, Snapchat, Google all have horrid recommendation systems. I mistakenly clicked a link to a YouTube clip of a TV show once and closed it immediately before even watching 2 seconds of it, but now my home page and the "watch next" bar are filled with clips from that show. I'm convinced that the dislike button and "I'm not interested in these" option in the hamburger menu are actually defunct because as far as I can tell they have absolutely no effect.
Spotify is just as bad. I once made the mistake of liking an instrumental track on one of the radio stations I listen to. Now a good 25-30% of the songs that come up are just the instrumental versions, not the songs themselves. No matter how many of them I dislike they keep appearing. The same goes for remixes.
There's no notion of "nuance" in these systems and it just makes me think that they were only ever evaluated internally. The models might perform well when they're tested on all 1 billion YouTube users, but on an individual level they're really awful.
As an aside, I've never given them any of my data so this is purely anecdotal, but I frequently hear (in both positive and negative contexts) about how great Facebook's recommendation systems are on both Facebook and Instagram. I think that's what investor confidence in them as an ad company has stayed so high even through their controversies. I honestly don't think I've ever heard anyone praise Google's recommendation systems, by contrast.
> There's no notion of "nuance" in these systems and it just makes me think that they were only ever evaluated internally. The models might perform well when they're tested on all 1 billion YouTube users, but on an individual level they're really awful.
Well, that is all that really matters to them in the end, isn't it? A more accurate recommendation system might be better for many individuals but not drive engagement, whereas what they are going after may be lousy for the individual but drive more engagement overall. It's the classic playing to the lowest common denominator. Getting enough people to keep viewing ads is all that matters.
There has been an explosion of Fanta flavors and canned pasta sauce because they realized that optimizing for the median consumer is a lot worse than optimizing for a small handfull of cluster centroid users.
Recommender systems don't optimize for the median consumer though. They optimize for clusters, much as pasta sauces now do. That's why my recommendations probably include a lot more super smash Bros content than yours do.
At a certain level of volume, more transparency should be LEGALLY required for multiple companies (auto personalization / recommendation engines / etc.). At least the ability to opt-out. Just hard to define. Could you legally define “bubbling” or super duper smart recommendation algorithms?
Big tech knows this is the direction IMO.
It’s my guess why they randomly started pressing so hard for screen time usage reminders.
They’re selling addiction to kids (and adults).
YouTube, FB, Instagram, Twitter Moments all have liability when it starts to click with everyone.
It’s a serious issue that is totally unregulated and not understood by most.
I built Pony [0], an email platform that delivers once a day, to see if it is possible for a platform that doesn't manipulate user psychology to succeed.
In the UI I eschew every traditional user manipulation technique I could think of: there are no notifications, there are no unread message counts. There even isn’t read/unread message state. It's launching now, so finally I get to find out: can a platform that doesn't vie for attention, that lets users create their own unguided, unprompted experience succeed?
> - Weekly/Fortnightly/monthly delivery. I want a way to say to an old friend "hey, lets keep up with each other every fortnight.
This is probably the most requested feature right now. I want this too but decided I had to keep it as simple as possible for the launch. But this will definitely be fleshed out as there becomes room to evolve the concept. Turning those delivery time radio buttons into check boxes will be exciting. I, for one, would like a delivery mode that syncs up with sunrise/midday/sunset at your geolocation. This will be fun to explore.
> - Offline mode. I want to be able to take 5 letters to a park near my house, turn off wifi, and just write with a clear head.
This is already in the works. I want this too. I even have a ServiceWorker installed for another reason (https://stackoverflow.com/a/55245528/741970) but will leverage it for proper offline functionality as well.
Thank you. That's a great way to put it. I also liked when someone recently described it as an "anti-technology". Why is it often easy to describe something in terms of what it's not?
Because people think in terms of responding to stimulus. It is a lot easier to see pain points and to say, "I don't like that" because the problem is concrete. To describe something in terms of what you want, you've got to do the work of imagining something new -- and thats hard.
I'm not seeing evidence of that from his profile. I see evidence that he's finding threads that are tangentially related to a problem and posting a solution which he built.
As someone who has been looking for something very close to this and has thought about building it, I'm grateful that he's posting it here.
I want a university or public library to manage the largest online database of videos (which is defining our culture to a great extent), not a big company like Google. A big company can provide the hardware, though.
To play the devil's advocate, isn't technological "addiction" the inevitable future, so why even fight it? The more engaging technology becomes, the more people will naturally(and perhaps rightfully) choose it over reality. Isn't trying to curb the addiction merely an attempt to postpone the future?
A distinction without a difference. Some experiences promote better outcomes than others. As a society we promote those that are demonstrably better vs those that are not. Why should this new terrain - digital - be exempt?
> They're not getting any ad revenue from me, so why am I stuck with their recommender that only cares what it thinks will cause me to spend the most time on YouTube.
I can think of two reasons.
1. If YT recommends good content to people, and make them watch YT more, they are more likely to keep paying the monthly subscription fee. So I'm not sure that the recommender is completely useless for their subscription business. I do agree that for subscribers, the weight of engagement metrics should probably be lowered in favor of user satisfaction or some other metrics.
2. YouTube is an extremely complex system developed over many years, and it's very expensive and error-prone to try and change anything. For many years, it was monetized through ads; so it was designed around that goal. The monthly subscription is relatively new, and it's still probably a tiny fraction of the ad revenues, so many features of YouTube are just left untouched even when they don't make much sense for monthly subscribers.
IMO it's got more to do with the fact that providing a power user feature such as "run a regex search on the closed captions" would implicitly define an API which needs to be maintained. It could complicate the backend forever or block certain optimizations even if the feature is used by a tiny percentage of users who'd be outraged when it's removed.
Any UX designer just had a heart attack reading that. Yes having a bazillion knobs and buttons is cool for hackernews folks, but as a user, it's just confusing. There's a reason iOS is so popular, and Pixel devices are creating an iOS like simple UI.
I do agree that 1-2 or the stuff mentioned above would be nice, but most of it comes off as a programmer trying to tweak the hell out of a system, ignoring that 99.99% of users actually won't benefit from ANY of that.
Being able to hide complexity can easily solve this problem. A user interface can come in multiple flavors. The default interface is typically simple and suitable for 95% of people. The other 5% need things like stats for nerds, advanced options, etc. Making advanced options limited defeats the purpose of advanced options. If you have too many casual users clicking through to advanced options, it means you did a bad job separating concerns in the first place.
But the question is, how much engineering hour and effort do you want to put into a feature that will benefit a very tiny percent of your users, vs putting that time into trying to make the experience that 99% of users have better?
What is even more confusing is having UI elements that don't work as you expect (not interested, dislike button etc).
The search on youtube is also completely broken, I have not monetised any of my videos and the penalties for that makes it so that if I search for the exact title + my channel name the video will show up on something like page 5.
Also, complex != confusing. You can have very simple UIs that are confusing as hell if you're not used to them or you need a specific issue. I usually have trouble helping my mother on her iphone because I'm not used to it and it has absolutely no guides or help functions if you can't find what you need. Usually it's a "magical gesture" that is documented on some obscure web page.
Considering that pretty much every huge service goes into the direction of less control over time (Amazon, Netflix are other examples), I assume that for other people this has to somehow be great and actually increase their engagement. For me (and I assume many others on HN) it simply makes the experience worse and recommendations less relevant.
It is for me and, (again, my assumption) probably for most of HN. When I don't like stuff I'm just annoyed by it taking up space and if I somehow end up watching it, it's only for the few seconds it takes me to close it.
Moreover, Youtube became increasingly resistant to user configuration. And the discussion doesn't touch on user configuration. I'd say the reason is that there's been uniform switch across multiple platform to auto-recommender model. Why? Part may be bureaucracies assuming they need control, part is a "the average user is a total moron who can't configure and will just accept defaults anyway" consensus (not entirely false but not entirely true), part of it is the rise of machine learning makes having supposedly strong recommendation engine be equated with having value.
And the situation with a completely unconfigurable recommendation is that they can't show people what they want. They can only show people "what people can be conditioned to like, what grabs some average more." Which is going to be polarizing content.
There is no additional cost to Google for playing additional videos. They have interconnects and they aren't full, so there is no reason not to add more content.
It's only the last mile that has bandwidth constraints really.
I think most of the suggestion aren't realistic. You can project the same desire for customizability to any software product. For Hacker News:
- Let me configure the order of posts based on voting patterns (hot, newest, controversial)
- Let me adjust the weighted score of posts based on the age of the account.
- Let me configure the home page.
> Let me replace the recommender entirely with something of my own devising, called back through a webhook.
Is there any popular consumer product that allows custom recommendation based on a webhook? Opening up a core part of your product to an untrusted and probably unreliable third-party is not a good idea.
> why am I stuck with their recommender that only cares what it thinks will cause me to spend the most time on YouTube.
This is addressed by the article:
> The first is this notion that it’s somehow in our interests for the recommendations to shift people in this direction because it boosts watch time or what have you. I can say categorically that’s not the way that our recommendation systems are designed.
I wouldn’t mind having these customization methods you suggest organized under different “smart playlists” with different recommendation methods and levels. Some are just chronological subscriptions, others try to dig up new and interesting/different content.
Part of me feels like these tools with "recommendation algorithms" who's parameters and behavior are opaque and not adjustable by the user are meant to put the people that wrote them in control of the user's mind.
What they are "meant" to do is increase engagement metrics. If the recommender is smart enough, the content diverse enough, and the userbase big enough, the recommender will learn how to exert psychological influence over the user as a means to increase engagement metrics.
None of this is exactly deliberate, it's just that the recommender will figure out that if you can get someone angry about something, or entranced by some conspiracy, they're going to be very engaged.
The problem is, going to a social media site with a specific intention, it's an attentional battle. They are going to show you the most entrancing thing they can come up with above the fold. For example, on youtube.com, you can choose to some degree what sections are shown, but the unconfigurable recommender is always at the top and cannot be moved. "Saved for later" is always a couple of screenfuls down no matter what you do.
This is hostile technology. YouTube is not made to be a tool for you to use. It's made so that you're a tool for it to use. It pokes and prods at you with its recommendations to serve its own interests. It does not care at all about yours as long as it doesn't piss you off so much that you stop going there.
Many times a recommendation system is opaque and not adjustable because the people who made it don't know how it make it transparent and adjustable. A good example of such a system was given in Andrew Ng's old machine learning MOOC. I'll describe it below.
Suppose we want to do a movie recommendation system, and we have a bunch of data consisting of anonymous user identifiers and for each user a list of movies they have seen and their rating of that movie.
Imagine that we had a list of movie attributes that we thought might be relevant to whether or not someone would like the movie. This might be things like running time, if it is funny, if it has car chases, if it has romance, if there are horses in it, if there is profanity, and so on. Let's say we've got 50 of these attributes.
Now suppose we had for each movie a vector of length 50 whose components were how well the movie fit each of those attributes, on a -1 to 1 scale.
If we had that movie attribute data, then what we could try is modeling the users as if each user had a 50 component vector telling how important each attribute is for that user, and for each user we could go over the list of movie ratings and try to find a set of weights for the components of that user's vector such that the dot product of their vector with a given movie's attribute vector correlates with that user's rating.
That works pretty well, but there are two practical problems. First, we need to come up with that attribute list for the movies. Second, once we've got our attributes someone has to go through each movie and figure out the vector.
So forget about that approach for a moment. Suppose instead we came up with a list of attributes, somehow, but instead of figuring them out for the movies and then inferring the user weights, suppose we told the users the attributes, and asked them how important each was, and assumed that the users are actually right about what is important to them?
Then maybe we could take all the movies, and try to assign attribute weights to them that lead to consistent predictions of user scores!
It turns out that works, but we still have the problem of guessing what attributes matter, and depend on the users actually somewhat knowing what makes them like a movie.
So...if we knew the movie attribute weights we could infer the user preference weights, and if we knew the user preference weights we could infer the movie attribute weights.
The brilliant solution is to get rid of the attribute labels. All we decide on is the number of attributes! So we might decide that there are going to be 50 attributes, and we can assign the movies random weights for all those attributes. We can assign the users random preferences for the attributes.
Then you can go through an iterative processes where you tune the movie attribute weights to better predict user scores, and you tune the user preferences to better match the movies. This ends up converging to a set of movie attribute weights and user preference weights that do a good job of reproducing the user's scores for the movies, and makes good recommendations...
...and you have no idea when it is done what the attributes mean!
All I really want is for Youtube to stop recommending me videos that I've already watched, or at least show the watched indicator on them consistently. It has always annoyed me to no end how Youtube seems to "forget" about videos that I've watched after a while, even though they're all still right there in my history.
* Let me turn it off entirely.
* Let me adjust the relative weight that the current video gets vs my past views (i.e. do I want to see related videos to this one, or ones that YouTube generally thinks I'll click on)
* Let me adjust the weighting of thumbs up vs. watch time.
* Let me configure the homepage
* Let me make the entire recommender panel be just subscriptions or items from a particular list
* Let me replace the recommender entirely with something of my own devising, called back through a webhook
These things would make YouTube much more useful for me. I'm not going to YouTube just to kill time, and I pay them $13/mo. They're not getting any ad revenue from me, so why am I stuck with their recommender that only cares what it thinks will cause me to spend the most time on YouTube. I am not interested in spending the most time on YouTube. I'm interested in getting the most out of it in the minimum time.