As always, Matt Levine has a readable, insightful take that covers many different ways of seeing the world from the same set of facts and several inconvenient distinctions that various biased parties would prefer to be left out of the retelling: https://www.bloomberg.com/opinion/articles/2024-07-26/andrew...
Not really, reliably detecting a user's preferred languages has been a persistent Hard Problem in tech since the start of the internet. Every proposed alternative solution ends up having vastly more false positives due to browsers/people incorrectly setting of default preferences so companies begrudgingly default to geographic heuristics knowing it is a terrible experience for an outlier group of people.
Why wouldn’t you just ask them, and particularly for media like this that has a native language, default to that? I don’t want software to think for me about what I want to see, if I want something different I’ll change it.
It's probably Microservices(TM)... The black box responsible for rendering the HTML is some other black box to the UI or whatever. The UI offers you locale options, but the renderer that fetches video titles hasn't been configured to respect this, probably uses locale from geolocation and some overpaid genius said "always use machine translation if locale doesn't match video title language"...
The homepage of google.com is also localized. I remember noticing that even when requesting and getting the English locale, the tooltip for the doodle was still in my region's language... wahey!
It's very easy to construct strawman examples where the worst ways to travel are compared to some abstract "best" way of learning as if that proves anything.
Reading up on something is a great way to discover the "high order bits" but it's very hard, apart from being in person, to ever pick up the "low order bits". I recently had a friend visit Australia and notice that attitudes towards mild speeding were very different from the US, not something you ever could have found from hours of trawling on the internet. One of the hundreds of different observations he made on the trip.
Every travel opportunity for me has used these low order bits to propel huge amounts of reading to fill in the missing high order bits that mesh with it. On a recent trip to South Korea, I became obessessed with the South Korean presentation (or rather, the lacuna) of the country's history 1955-1987. I went to countless history museums around Seoul just so I could see what they wanted attendees to know about Korea between the day-by-day recap of the Korean War and the miracle of K-Pop and industrialization. It was interesting the degree of frankness each museum had but all of them made me delve much more into the scholarly writing to see what was pointedly omitted.
The distinction is, if there is room provided for one, is it expected to be provided by the landlord or the tenant, not whether a typical apartment has one as an amenity.
I've said for decades that Google is terrible at search in every area except Google Search. Youtube search? Terrible! Chrome history search? Abysmal! Gmail search? Atrocious! Google Maps Search? At some point, standing in a middle of a mall searching for "coffee" returned only 3 SERPs despite me standing in front of a coffeeshop that I could not get to show up.
>I've said for decades that Google is terrible at search in every area except Google Search.
From my point of view Google Search is terrible, too. Is hard to find relevant results, you mostly get results optimized to make money, or junk. You have to explore tens or hundreds of results to find the needle in the haystack.
I find YouTube search to be serviceable. At least it has decent filtering and sorting options. Gmail search is just OK, but I haven't found anything much better. Chrome history search, though, is completely worthless. Especially since it got merged into that myactivity thing that is utter garbage, completely non-functional for any purpose. There's so much potential in searching a complete history of everything you've ever personally seen online, and it would make Chrome more sticky. Incredible fumble by Google here.
Youtube search does a baffling thing where it shows you 5 SERPs, then a bunch of unrelated things it thinks you like, then another 5 SERPs. It used to only show you the top 5 SERPs before switching to "suggested videos" for the rest of the scroll. Truly a terrible product when that was the design.
Youtube is not in the business of giving you accurate search results or information. It's now in the business of getting you to watch any video, related or not to your query, in order to serve you ads.
> It's now in the business of getting you to watch any video, related or not to your query, in order to serve you ads.
Youtube was in this business from day 1. Even before Google. Youtube was never going to be anything other than an ad-platform with videos to lure in the products.
Vid.me tried to be a video platform with videos to lure in users, but it went bankrupt, because nobody wanted to pay and nobody wanted to watch ads.
It is a very crude method for injecting diversity into search results (and the browsing experience). It can't be turned off and still shows up even if very specific search terms are used.
Hard to believe it is the best possible video search implementation for their ad serving goals.
They fear tiktok is outcompeting them with even more aggressive attention hijacking, I guess, so they can't resist showing up something "This wasn't what you were looking at but can I get you to click it?"
the unrelated videos it shows me are so far from anything I'm interested in that I can only conclude it's showing both of us the same stuff, just lowest common denominator popularity.
>videos I'm also interested in, sometimes more than what I'm searching for
therefore, based on my argument, you must have horrible taste
I never understood why the "collaborative filtering" approach never took off with most review options. Google Maps shows you what the average person thinks is a good restaurant, meaning the rich get richer faster and tiny statistical noise converts to durable competitive advantage.
Instead, I'd love for Google to understand me well enough to show me which restaurants I would disproportionately love compared to other people based on its understanding of my taste profiles. That way, the love can be shared amongst a much wider base of restaurants and each distinctive restaurant could find its 10,000 true fans.
On top of that, it actually gives me an incentive to rate things. Right now, you only rate from some vague sense of public service instead of "this can actively improve your experience with our product".
It's not just Google Maps, Netflix used to operate on the model of deep personalization that they've slowly de-emphasized over the years. I'm still waiting for Letterboxd to introduce a feature to give me personalized film recs based on the over 1000 ratings I've given it over the years as a paying customer but they seem in no hurry to do so. Amazon used to take your purchase history into account when ordering search results but I think that's also been significantly de-emphasized.
About the only arena this is widespread is streaming music services like Spotify.
I have a theory: They realized the right approach is to focus purely on the yes/no of what you choose to consume, rather than trying to optimize the consumption experience itself.
Remember how YouTube and Netflix used to let you rate things on 1-5 stars? That disappeared in favor of a simple up/down vote.
Most services are driven by two metrics: consumption time and paid subscriptions. How much you enjoy consuming something does not directly impact those metrics. The providers realized the real goal is to find the minimum possibly thing you will consume and then serve you everything above that line.
Trying to find the closest match possible was actually the wrong goal, it pushed you to rank things and set standards for yourself. The best thing for them was for you to focus on simple binary decisions rather than curating the best experience.
They are better off having you begrudgingly consume 3 things rather than excited consuming 2.
The algorithmic suggestion model is to find the cutoff line of what you're willing to consume and then surface everything above that line ranked on how likely you are to actually push the consume button, rather than on how much you'll enjoy it. The majority of which (due to the nature of a bell curve) is barely above that line.
I think Netflix realized that reducing ratings to a simple thumbs up/down was a bad idea after all. A while back they introduced the ability to give double thumbs up which, if you can treat non-rating as a kind of rating, means they're using a four point scale: thumbs down, no rating, thumbs up, double thumbs up.
Netflix are right that 5-stars is too many, it translates to a 6 point scale when you include non-rating, and I don't think there is a consistent view on what "3 stars" means, and how it's different to either 4 stars or 2 stars ( depending on the person ).
For some people 3 stars is an acceptable rating, closer to 4 stars than 2 stars. For others, 3 stars is a bad rating, closer to 2 stars than 5 stars. And for others still, it doesn't give signal beyond what a non-rating would be, it's "I don't have a strong opinion about this".
Effectively chopping out the 3-star rating, leaves it with a better a scale of:
- Excellent, I want to put effort into seeking out similar content
- Fine, I'd be happy to watch more like it
- Bad, I didn't enjoy this
- Terrible, I want to put effort into avoiding this
With the implicit:
- I have no opinion on this
But since it's not a survey, it doesn't need to be explicit, that's coded into not rating it instead.
These are comparable to a 5 point Likert scale:
"I enjoy this content"
- Strongly agree
- Agree
- Neither Agree nor Disagree
- Disagree
- Strongly Disagree
The current Netflix scale effectively merges Disagree and Strongly Disagree, and for matters of taste that may well be fine.
It would be interesting to conduct social science with a similar scale with merged Disagree and Strongly disagree to see if that gave it any better consistency.
When given a 5-star choice “very bad/bad/ok-ish/good/very good”, I rarely pick one of the extremes.
I suspect there are others who rarely click “bad” or “good”.
Because of that, I think you first need to train a model on scaling each user’s judgments to a common unit. That likely won’t work well for users that you have little data on.
So, it’s quite possible that a ML model trained on a 3-way choice “very bad or bad/OK-ish/good or very good” won’t do much worse than on given the full 5-way choice.
I think it also is likely that users will be less likely to click on a question the more choices you give them (that certainly is the case if the number of choices gets very high as in having to separately rate a movie’s acting, scenery, plot, etc)
Combined, that may mean given users less choice leads to better recommendations.
I’m sure Netflix has looked at their data well and knows more about that, though.
I apply my own meaning to the 5-star rating, and find it to work really well:
1 = The movie was so bad I didn't/couldn't finish watching it.
2 = I watched it all, but didn't enjoy it and wouldn't recommend it to anyone.
3 = The movie was worth watching once, but I have no interest in watching it again.
4 = I enjoyed it, and would enjoy watching it again if it came up. I'd recommend it.
5 = a great movie -- I could enjoy watching it many times, and highly recommend it.
> The current Netflix scale effectively merges Disagree and Strongly Disagree, and for matters of taste that may well be fine.
I'm a bit skeptical about this.
To me there's a big difference between "This didn't spark joy" and "I actively hated this": I might dislike a poorly-made sequel of a movie I previously enjoyed, but I never ever want to see baby seals getting clubbed to death again.
Every series has that one bad episode you have to struggle through during a full rewatch. Very few series have an episode bad enough that it'll make you quit watching the series entirely, and ruin any chance at a future rewatch.
They're only useless in that they aren't displayed for your peers, but that was always the least-useful function.
Being able to see a counter that reads as "Twenty-three thousand other people also didn't like this video!" doesn't serve me in any meaningful way; I don't go to Youtube to seek validation of my opinion, so that counter has no value to me. (For the same reason, the thumbs-up counter also has no value to me.)
But my ratings remain useful in that the algorithm still uses the individualized ratings I provide to help present stuff that I might actually want to watch.
As we all know, investors and advertisers love growth; Youtube thrives and grows and gathers/burns money fastest when more people use it more. The algorithm is designed to encourage viewership. Viewership makes number go up in the ways that the money-people care about.
Presenting stuff to me that I don't want to watch makes the number go up -- at best -- slower. The algorithm seeks to avoid that situation (remember, number must only go up).
Personally rating videos helps the machine make number go up in ways that benefit me directly.
---
Try to think of it less like a rating of a product on Amazon or of an eBay seller; try not to think of it as an avenue for publicly-displayed praise or admonishment. It's not that. (Maybe it once was -- I seem to recall thumbs-up and thumbs-down counts being shown under each thumbnail on the main feed a million years ago. But it is not that, and it has not been for quite a long time.)
Instead, think of it as one way in which to steer and direct your personalized recommendation algorithm to give you more of the content you enjoy seeing, and less of what you're not as fond of.
Use it as a solely self-serving function in which you push the buttons to receive more of the candy you like, and less of of the candy that you don't like.
I have literally not rated anything at all, ever since YouTube removed dislikes, and my recommendations are working fine. Ratings indicate(d) if a given video was likely to be a waste of my time or not, and in an age of AI slop, this feature is more desirable than ever.
Someone should make a SponsorBlock/Dearrow-type addon to flag AI slop.
You only assume recommendations are based on ratings, but you don't know. And I have seen your metaphorical green grass, because actual ratings were a thing up until about 4 years ago, remember?
>I don't find any of that on my end.
Good for you. The true crime genre has been hit hard by AI slop.
> And I have seen your metaphorical green grass, because actual ratings were a thing up until about 4 years ago, remember?
I remember this conjecture of yours (that ratings unilaterally ceased to matter as soon as they stopped being displayed to users) very well.
And unlike you, I can see over to the other side of the fence -- in the present day -- at a whim: All I have to do is fire up YouTube in a private session on a disused device. It's fucking awful over there; it's complete bedlam.
Same point as always: That it definitely doesn't have to be that way at all.
(I can't make you take the blinders off and use that utterly useless, vestigial Thumbs Down button, though. You're free to live your life with as blindly and with much suffering as you wish, no matter what anyone else thinks.)
I mean, if you read about how current industry-standard recommendation systems work, this is pretty bang on, I think? (I am not a data scientist/ML person, as a disclaimer.)
If e.g. retention correlates to watch time (or some other metric like "diversity of content enageged with"), then you will optimize for the short list of metrics that show high correlation. The incentive to have a top-tier experience that gets the customer what they want and then back off the platform is not aligned with the goal of maintaining subscription revenue.
You want them to watch the next thing, not the best thing.
I think Spotify and other streaming services have a problem very similar to the restaurants. Take an artist with a 40 year career and a dozen acclaimed albums and bags of songs almost everyone loves, and when that artist comes up it is always the same one or two songs. The most played songs, causing feedback and making the problem worse. In my mind, one of the core reasons for asking for recommendations is to discover something different, which means ignoring or maybe even penalizing popularity, because you are likely already familiar with the popular by definition.
I found Spotify surprisingly good at recommending new music. Not amazing, but considering how low the bar is thanks to other services like Netflix I'veveen pleasantly surprised.
For example it recommended a band with just a hundred monthly plays which I loved. Almost all bands it recommends has less than 10k monthly plays, so not huge "safe bets", and most are quite decent.
Netflix's DVD recommendations worked this way. It identified cohorts with similar categorical preferences and recommended content other people in the group enjoyed.
If the service actually shows you things you want to see, then you're less likely to click on ads (or "sponsored results") which you also don't want to see.
Perhaps more importantly, if such organic growth is possible, it lowers the incentive for businesses to buy ads.
>Instead, I'd love for Google to understand me well enough to show me which restaurants I would disproportionately love compared to other people based on its understanding of my taste profiles.
I don't want for Google to collect data on me, build a profile and "understand" me. I want Google just to return relevant search results.
I was part of the team that built exactly this. It launched in 2010. Some Googlers of that era are probably still annoyed at all the internal advertising we did to get people to seed the data. This is one of the launch announcements:
https://maps.googleblog.com/2010/11/discover-yours-local-rec...
> Google Maps shows you what the average person thinks is a good restaurant
I'm fairly sure this isn't true. At least, I still get (notably better) results searching while signed in. Couldn't tell you what the mechanism for that is these days, though. But at least back in 2010, the personalization layer was wired into ranking. You can see in the screenshots how we surfaced justifications for the rankings as well.
Pretty much immediately after launch, Google+ took over the company, the entire social network we had was made obsolete because it didn't require Real Names(tm), and a number of people who objected (including me) took down all our pseudonymous reviews. Most of the team got split off into various other projects, many in support of Google+. As best as I can tell the product was almost immediately put into maintenance mode, or at least headcount for it plummeted like 90%. Half of my local team ended up founding Niantic, later much better known for making Pokemon Go.
As for why collaborative filtering didn't take off, I can offer a few reasons. One is that honestly, the vast majority of people don't rate enough things to be able to get a lot of signal out of it. Internally we had great coverage in SF, London, New York, Tokyo, and Zurich since Geo had teams in all those places and we pushed hard to get people to rate everything, but it dropped off in a hurry elsewhere. The data eventually fills up, but it takes a while. I'm told we had 3x the volume of new reviews that Yelp had at the time, but Yelp mostly only covered the US, while Google Maps was worldwide, so density was quite low for a long time. It was probably 5-10 years before I started hearing business owners consistently talk about their Google reviews before their Yelp reviews.
Another thing is that people are really bad at using the whole rating scale. On a 1-5 scale, you'll probably find that 80% of the reviews are either 1 or 5 stars. Even more so in a real life situation where you meet the humans involved. While you can math your away around that a bit, at that point you're not getting a ton more signal than just thumbs up/down (anecdotally I've heard that's why Netflix moved away from 5 stars). And then at that point, you might be getting better signal from "were you motivated enough to rate this at all?", which is why there's the emphasis on review counts. Many people just won't review things badly unless things have gone terribly wrong. I sat in on a few UX interviews, and it was really enlightening to hear users talk about their motivations for rating things, many of which were way different than mine.
BTW I'm familiar with linkrot, but I just discovered link poisoning.
I was reading the blog post on my Android phone and saw the Maps links to Firefly and Home Restaurant. So I tapped the Home Restaurant link and it took me to the Google Maps app in my normal home position with my home in the center. I thought for a moment that maybe it confused Home restaurant with my home.
So I tapped the Back button and nothing happened. Tapped it several more times with no luck. Finally I used the ||| button and swiped Maps up to kill it.
Then I tried the Firefly link, with the same results.
On the web, both links work fine, but someone forgot to test that these old links still work on Android.
Turns out that Home Restaurant is closed, but Firefly is alive and well. Their menu looks tasty, and the FAQ is something to behold:
> About the only arena this is widespread is streaming music services like Spotify
And even they can't get it right, and will give me promoted content before they give me anything related to my tastes. Pandora is the only recommendation engine that actually gives me what I would consider to be valid results. Shame they refuse to improve their audio quality, or I'd jump ship to Pandora. Until then, I'll keep using their free tier to curate Spotify playlists.
related to your letterboxd suggestion, https://couchmoney.tv is quite good! it uses trakt instead of letterboxd but it's given me quite a few good suggestions. their FAQ describes a similar approach to what you've been talking about, it tries to find movies and tv you like disproportionately like.
> Instead, I'd love for Google to understand me well enough to show me which restaurants I would disproportionately love compared to other people based on its understanding of my taste profiles. That way, the love can be shared amongst a much wider base of restaurants and each distinctive restaurant could find its 10,000 true fans.
This kind of ties into "but your computer is broadcasting a cookie and you're being tracked" paranoia though.
People have been convinced by uninformed twaddle that somehow folk are looking through their screen at them to see what they're doing and that this is bad, but it also means you get fed an awful lot of adverts that really don't fit your demographic.
I don't mind if advertisers or supermarkets are profiling me based on things I like. You want to show me things I like? Good. The flip side is I'd prefer you not to show me things I don't like.
Youtube seems to be hilariously bad at this latter part, and all I get are adverts for a bank I'm already with and have been for 30 years, adverts for online gambling sites which I will never be interested in, adverts for Google's AI slop which I will never be interested, and adverts for online grammar-checking services that don't work in the UK because they convert everything into some weird North American creole dialect, which - again - I will never use.
Yes, take a look at my restaurant-using profile. Recommend stuff I like.
> Instead, I'd love for Google to understand me well enough to show me which restaurants I would disproportionately love compared to other people based on its understanding of my taste profiles.
I mean... this sounds like the perfect use case for a third party app like "My taste restaurant finder"? There are undoubtedly apps out there like this.
I don't think Google Maps (a general purpose maps app) should try to be everything for everyone. It's good enough for what it is.
The reason is money. Google (in spite of what they would have you believe) does not show you what is "good" for you, it shows you what it gets paid to show you (paid in various, sometimes very complicated ways).
I am sad that Google services are so popular, because it makes the world a little bit worse for everyone. This includes not only Google Maps, but also Gmail (did you know that Google is quite active at censoring your E-mail and you will never see certain E-mails?).
I would really like to see more competition, ideally without the ever-present enshittification (I'm pretty sure Apple Maps will go down the drain, too, because KPIs and money).
Because if you can't sell launches on the open market, your own launches become exponentially more expensive. Cost sharing allows for the economies of scale that let governments piggyback.
reply