My biggest problem with YT recommendations is that they dramatically overweight recent videos when generating the recommendations. I mean, I can spend a week doing nothing but watching (well, listening) to videos of classical music performances, and my recommendations will be full of classical music, exactly as you'd expect. Then I watch one random one-off thing, like "15 funny pitbull fails" and suddenly ALL of my recommendations are animal videos. Then I have to go manually edit my feed history and remove "15 funny pitbull fails" to get my recommendations back to normal.
All the AI stuff is really still incredibly dumb. Same for Amazon: Buy a mixer and suddenly you will be haunted everywhere by ads offering more mixers.
All the AI stuff is really still incredibly dumb. Same for Amazon: Buy a mixer and suddenly you will be haunted everywhere by ads offering more mixers
Do we know it's actually dumb?
Like, given that you just bought a mixer, perhaps in an absolute sense, you're unlikely to buy a second mixer. But relative to everything else that you might buy, perhaps you are very likely to buy a second mixer!
There are huge financial incentives to get these ads right. I'm more inclined to guess that Amazon is doing something right than that they're leaving money on the table.
Just spitballing here, but if you just bought a mixer, your odds of buying another mixer is lower than that of buying a cook book or some other related stuff -- which Amazon could totally figure out mining the other stuff people bought at the same time or shortly after a mixer.
It's still not necessarily the better idea for them.
Some percentage of people who just bought a mixer will get a product they aren't satisfied with, for whatever reason. That population has a much higher change of buying another mixer than other people. And mixers are probably far more expensive products.
I have no idea if, in an absolute sense, there's more chance you'll buy a cookbook rather than another mixer, but it wouldn't surprise me either way - on the one hand, obviously most people don't need two mixers. On the other hand, plenty of people who buy a mixer will never buy a cookbook in their life, but literally 100% of people who bought a mixer have proven that they were at some recent point in the market for a mixer. I'm not sure which effect wins out here.
There are huge financial incentives to get these ads right.
Not really. All you have to do is convince companies that machine targeted ads are worth it. Apparently that is orders of magnitudes easier than actually recommending anything relevant.
Which isn't hard to imagine either, there aren't much competition after all. So the primary purpose of targeted ads is that it functions as an imaginary hammer you can use to squash new competition with.
Alternativly there are many matching and complementing products for a mixer (or any other products for that matter)
Just bought a mixer, how about: A new set of mixing bowls, alternative whisks and hooks, perhaps a set of measuring cups, a kitchen scale?
Just bought a new smartphone, how about: an external battery, screen protector (that fits the specific model), a fancy charging dock?
It really doesn't seems like it would take much intelligence to create these sets of complementing products and then use them for further ad targeting in the future, you could even use AI and ML to create the sets in the first place so you can fill out your buzzword qouta.
You think that's dumb, but I bet the marketing data behind that says otherwise.
You think "I already have a mixer, idiot!" but in reality the chance that you are interested in buying a mixer just went from 1% to 2%. (Because you want another one, because you want to gift one, because you returned yours, etc.)
The last time I saw an Amazon employee weigh in, he said the ads are in fact not effective but the ad team isn't measuring it well. In particular, he claimed that he showed them that after adjusting for returns, the order-again rate for a particular expensive item was effectively zero, they asked him to repeat his study for another product, and he wandered off and found work to do for his own team.
My skepticism level remains high. This is billions of dollars, with easy access to tons of data.
I'm not sure if put more stock in the government successfully hiding a fake moon landing, or analysts in one of history's most successful companies not running the numbers.
I think the problem here is that "recommendations that maximize expected marketing revenue" and "recommendations that maximize user experience" are not the same. Most of us would prefer Google to present us the latter, but they have considerable financial incentives to present the former instead.
The mixer probability maybe went from 1% to 2% but there must be 100s of other products they could show that have much higher probability. Maybe a mixing bowl or a cookbook?
> All the AI stuff is really still incredibly dumb.
and scarily brilliant at the same time. I was looking up Newtek some time ago of the Amiga Video Toaster fame. Googled Wil Wheaton promo video (Will worked for Newtek :o), Penn Jillette (also promoted it) and finally Kiki Stockhammer (spokesperson). Here is where things get interesting. Google spewed 1997 "How To Have Cybersex On The Internet" https://www.youtube.com/watch?v=PBDKEn-TeQg as the result for Kiki haha, why would it do that? Turns out this "instructional video" was made using Video Toaster and features one of the iconic special effect transitions
so what does it have to do with Kiki Stockhammer? This transition is Kiki doing poses on the green screen. Google search somehow connected this information together and was able to recognize Kiki Stockhammers silhouette on random youtube meme video :o !
This isnt the first time I got image search results that could only happen if Google was analyzing actual image content. Microchip part searches often return pictures of random boards featuring that particular IC despite no bom list in the html, Google is OCRing images and recognizes part numbers. Same with faces, cars etc.
Read the Youtube recommendation paper (https://ai.google/research/pubs/pub45530) and it will become very clear why it recommends a ton of high-engagement clickbaity content based off of a minimal set of recent watches.
> More recently, people told us they were getting too many similar recommendations, like seeing endless cookie videos after watching just one recipe for snickerdoodles.
Just checked my feed of recommended videos and it is still full of Tetris and excavator videos that I was into lately. It would be awesome to be surprised with some extraordinary lectures or documentaries similar to stuff I watched 5 years ago, or so.
I have the opposite problem. I'm only interested in new videos for the most part, and my recommendations are filled with videos from months ago. It would be much better if they would give us a flag or some date range options. Obviously this increases UI complexity, but maybe it doesn't need to be that prominent?
Things like this were a constant source of frustration at google. The philosophy there was to streamline the experience for showing ads to the masses rather than provide any features for power users. Feature requests like this were brushed off by labeling the requester "not the target user".
I guess we all know this by now, but google doesn't care about your being able to use its products in a reasonable way. In some sense, this makes sense since I bet the overlap of power users and people with ad blockers is pretty high.
> In some sense, this makes sense since I bet the overlap of power users and people with ad blockers is pretty high.
Maybe you should speak for yourself? I do not run ad blockers (much less anything targeting Google ads!) and the experience I get as a power user on Google services is no better than anyone else's - although, to be fair, it's not noticeably worse, either. (And mind you, the average ad supplier on the Internet is a hell of a lot shadier than Google, Facebook or any of the big "household name" tech companies!) I'm sure that plenty of other people here can confirm my POV.
Google's trouble engaging with power users as of late is entirely self-inflicted. There's quite simply zero doubt about that.
It always annoys me that giving users options they can control is viewed as complex but search engines that return whatever they deem as important aren't viewed as complex.
I think when the GP says new, he meant newest entry in their history. His problem is that watching one unusual video immediately and dramatically biases the recommendation engine at the expense of videos that match his long term viewing habits.
What you’re after already exists, the “recently uploaded” section on the home page.
It's a combination of that and focusing on recommending videos in depth on whatever topic it thinks you're most interested in. I don't want to watch fifty videos about swords. Three is plenty for now. I want to watch a bunch of different videos and shallowly explore a lot of topics, and the YT recommendation system is extraordinarily incompetent at facilitating this exploration mode.
It's terrible. I have to completely log out when I watch videos which put me to sleep a couple times /week because I don't want my suggestions filled with dry history lectures.
> I have to completely log out when I watch videos which put me to sleep a couple times /week because I don't want my suggestions filled with dry history lectures.
Why not just pause your watch history (or, in case you forget, remove the items after the fact)?
One of the first things mentioned on that Github page is the following warning:
WARNING: PUTTING NEWPIPE OR ANY FORK OF IT INTO GOOGLE PLAYSTORE VIOLATES THEIR TERMS OF CONDITIONS.
Newpipe is available on F-Droid but that version is sometimes a bit old. The project maintainers suggest the following order of preference for getting access to the latest version:
In order to get this new version, you can:
- Build a debug APK yourself. This is the fastest way to get new features on your device, but is much more complicated, so we recommend using one of the other methods.
- Download the APK from releases and install it.
- Update via F-droid. This is the slowest method of getting updates, as F-Droid must recognize changes, build the APK itself, sign it, then push the update to users.
That's used to be the case for me, but since like a month I see more older content(also YT has removed date of when the video was published from right hand recommendations).
If I want something new, escape my bubble, I use other platforms like bitchute or peertube. Sometimes it's pretty scary but it's getting better.
I do this. it has pluses and minuses. the plus is I'm not tracked. the minus is YouTube can't recommend. Sure of course if it's a cat video then I dont want YouTube recommendations but I watch almost all videos in incognito mode to avoid tracking. I'm not sure which ones I want YouTube to know and which I don't.
I tried FB video recently and was amazed by its predictions. It was pulling up compelling video from stuff I liked a decade ago that I was more than willing to view, it was kind of uncanny.
