I agree. I think all data can tell us is what people _are_ watching, not what people would get the most enjoyment out of watching.

It's like if a chef based a restaurant menu on the average diet of the average person. You may find neat patterns and be able to hyperstimulate some drive in people, but that's a very different thing than carefully designing a new or engaging thing people didn't know they wanted.

It's sort of a guaranteed averaging feedback loop.

> I think all data can tell us is what people _are_ watching, not what people would get the most enjoyment out of watching.

Just to be clear/nitpicky, you're talking about an algorithm that starts with the catalog Netflix could afford to license at some moment in time, and then maximizes the viewer's enjoyment for that limited set of content.

That's probably a vastly different level of enjoyment than starting with, say, the catalog of available content listed in IMDB and maximizing for the same thing.

Surely Netflix starts with the larger catalogue of all possible content when deciding what content to licence and also what content to make.

Sure. But they are still at best delivering "the most enjoyable experience where the enjoyment is bounded by the content that Netflix can afford to license."

A user can hit those constraints when they search for something that exists but is not available on Netflix.

That point becomes noteworthy when that same content (or portions of it) are available elsewhere on the net. In my experience this is typically through:

a) downloading a torrent of the content that is missing from Netflix

b) watching all or some portion of that content on Youtube

> You may find neat patterns and be able to hyperstimulate some drive in people

This is pretty much what fast-food chains do methinks? And I think that's an apt comparison to what people are lamenting with Netflix above.

I have a strong suspicion some of the newest crop of Netflix/Hulu/Amazon shows are already doing that to some degree.

It only tells what people are watching from an extremely limited data set. I would have much different viewing patterns if Netflix streaming was anywhere near what their DVD selection was. You could watch all manner of obscure films that aren't worth buying and streaming services aren't going to carry.

I'm glad I was able to watch much of my movie bucket list before Netflix decided to let the DVD inventory decline.

There's definitely a DVD rental market for "every dvd that is or ever was in print."

Yup. The endgame for this type of content if there is one is probably dynamically generated shows that adapt story lines and characters to need your mood and desires, even the unconscious ones.

This is a very insightful comment.

Just simply occasionally injecting something randomly chosen from the catalogue would probably do wonders. I do find that these systems as they currently exist tend to sate your appetite for a particular thing until you are fatigued by it and no longer wish to engage. Maybe it's a sign of how rudimentary their recommendation systems really are, as you say. My use of Netflix and Youtube have both dropped off precipitously because of this over-feeding of my appetite for certains types of content.

However, I've also gone back to Netflix DVD. They have such a massive catalogue of movies and it seems easier to just skim through genres and find weird gems.

You can inject your own novelty. Youtube does the same thing and after a while I just clear out its recommendations and start over.

The non-customized (viewed in incognito) main page of Youtube that is the recommendation for the average viewer is even worse, nobody should ever touch that horror show.

Ublock origin with all filters enabled is the only thing that makes Youtube usable in my opinion.

Youtube is a horrific service. The recommendations are terrible, the UX sucks, too many ads, they dont know how to use the sreen real-estate.

I freaking hate youtube.

WRT to real-estate, I see that the video I am wanting is taking up ~20% of the screen, then there is a list of completely unrelated content, and I have to click to see the date of the submitted video.

The UX is different on every single device. Phone, machine, ipad, browsers differ... its just a mess.

Yeah, the front page of Youtube is an abomination. Reddit isn't much better.

I think it would rather help if there was two buttons; "watched" and "hide". Both remove the item permanently from the visible catalog for you, "watched" adds it to a profile-private list of things you have watched.

This would help me to quickly sort out if I would consider watching something and I could reduce my watching list to the things I'm actually currently watching instead of abusing it for "might watch at some point".

If I have watched something I don't want it to be recommended to be and if I don't like something I don't want to see it again.

> Just simply occasionally injecting something randomly chosen from the catalogue would probably do wonders

This is a standard ML practice, otherwise ad networks would be stuck showing you the same ads (even though it feels that way sometimes). I'm willing to wager large sums of money that Netflix has tested EVERY idea mentioned anywhere in this thread and many others you couldn't guess.

You're probably right, but then the recommendations seems to just boil down to what other people are watching, what Netflix has recently produced and things in the same genre as what I'm watching a lot of at the moment. Seems like a pretty rudimentary output for such a sophisticated system.

I think there would probably be value in curated "playlists" done by actual people.

Steam does this. I honestly didn't get a ton of use out of it, but I only really followed Jim Sterling, Total Biscuit and TB's 30 FPS Locked curation list. The latter was the only one I found particularly useful because it turns out I have pretty different tastes from both of them and just enjoyed their critique.

It seems to be evolving the same way that Hollywood and the AAA game industry did. Which optimizing for safety as budgets increase. So it's mostly just the same stuff over and over - things that have been previously proven to work. As a result you get movies/games that you don't hate, but also don't love.

Surely the point of a recommendation system is that it is just that, a recommendation?

I find a lot of my joy in watching netflix is randomly browsing through the catalogue and selecting a random film that I would never watch if I had to either schedule watching on TV or pay to watch at the movies.

The fact that Netflix brings together a wide range of different movies and television programs in one place where for no cost I can watch something you may not like has opened a world of media I would never have found before.

I found the opposite, where I'm willing to try things I initially didn't think I would like, especially foreign language series.

Absolutely, watching a stand up comic show from Poland in Polish via English subtitles is a hoot just in itself. Could never do that prior to the Netflix streaming era.

Also highly recommend the Norwegian “Occupied” series.

Okkupert is fantastic. Also recommend

Le Chalet Suburra Dark The Break Marcella The Fall Happy Valley

This is a good observation and I share the sentiment.

AI/ML is great at optimizing what you're currently doing. While it can also find correlations and connections that you may miss without it, it's unlikely that it will make any radical recommendations.

IMO this is ok because AI/ML can evolve into that capacity - you have to take baby steps after all.

But in the interim, it's definitely something to watch out for - you don't want to take away the element or principal of randomness out of life - it's at the core of how we operate, innovate, and (almost quite literally) how we evolve.

It would be interesting to see ML recommendations of the form "There is a 75% chance you'll LOVE this movie, 25% chance you'll hate it" rather than only "98% chance you'll like this movie". Right now those 75% movies are actively hidden, as 75% is considered a bad prediction.

In other words, rather than optimizing for zero risk of disliking something, allow the user to optimize for the possibility of truly loving something, or being surprised. Especially on Netflix, where the risk is you stop part way through and switch to something else!

I'd like to see Netflix invest more in making a great browsing discovery experience. I feel like their UI is really optimized to the watch something quick metric now. It seems like search is pretty well implemented these days and both YouTube and Netflix recommendations have come a long ways but it still seems hard to replicate that browsing experience that I get at the local bookstore with some rudimentary sections and staff picks.

I do agree with what you say. On the flip side of that, I also think Netflix has been greatly beneficial in offering the opportunity to both produce and distribute movies that would have had a hard time getting to market otherwise with the current studio oligarchy. Going away from the recommendation topic, this enhances the types of movies available, which is direly needed in the age of superhero dominance.

We have lost a lot in no longer looking to our friends, or trusted people for content we might like.

I still get loads of recommendations from friends - often for shows on Netflix but friends are still a pretty big source of inspiration for what to watch.

Right! The web has really started to flatten everything...some of it seems good, but the things we've lost really suck.

Agreed - I do miss the days of Netflix's star ratings. I would be able to browse content I wouldn't necessarily be drawn to, but would consume anyway because, 'oh wow - tons of people gave it a great rating.'

I get why they've moved to 'this content matches [x%] to your tastes', but I'm watching similar stuff over and over again without any insight into whether other content is also good, unless I want to spend hours pouring through IMBD and Rotten Tomatoes.

There is the filter bubble paradox - the stuff that we need to learn is what we have never encountered before, but the ML filters can only learn from our past opinions.

If the problem is the recommendation system, you can still get recommends the old-fashioned way, like by r/NetflixBestOf. A glaring deficiency for me in Netflix's dashboard are the presence of various consumer opinions and reviews.

If I want decent content, I have to wait until it rolls around on the MGM or TNT cable channel, none of which streams.