Hacker News Reader

Comment by jypepin | Hacker News Reader

Normally when I see this kind of announcement, I think "It will take long before they learn and nail it like the main competition" - in this case Spotify. Spotify seems to be doing a great job curating new music or building playlist specifically for me according to what I listen to, and I can only assume it's 100% automated obviously.

Google is so strong with ML today, and not stopping, I don't know what will allow smaller competitors to stay differentiated. I just had a call with a G recruiter, and pretty much 100% of the roles they have in all their European offices are ML-related roles, on most products. And they hire without ML experience, which I think is a sign of how much they are investing in ML.

This is no news, but having this call and seeing how little non-ML they had was really eye opening on how much this is G's focus now, and it can only be exponential I assume.

View on HN

What you’re saying seems like it must be true... Google must be so strong at ML. How could it be otherwise?

And yet, YouTube recommendations are terrible lowest common denominator nonsense. It’s just “more of the same” for some surface level definition of “same”. There doesn’t seem to be any playfulness to the learning algorithm at all. No Bayesian experiments. There’s no flex.

Amazon too... so much data, so many smart people.

Why can’t these companies lead me on a garden path of delights? Why can’t they take my hand and walk with me through the full content space? Why do I continually feel that there are vast delightful segments of the catalog I will never see unless I query for them?

Not just YouTube, Google Play Music's recommendations are truly terrible. I'm regularly being recommended whatever is popular, rather than anything I'll actually listen to. The only time I've ever witnessed Google Play get a recommendation right is when a band I've already added to my library releases a new album.

It irks me that after being a subscriber for many years, Google Play still doesn't have anything like Spotify's Discover Weekly. The Google Play Music recommendations are just so god-damn awful. I don't know if it's actually the case, but they come across as ads, not recommendations i.e. Some record label paid lots of money to be "recommended" frequently.

You need to use the thumbs up to train it, Also if you play the artist's radio it does play similar artists and i have found some new artists that way. But there is no structure to it, the home page only suggests things that are already in my library, mostly i doubt anybody cares about play music at google.

I thumbs down bad recommendations every time I try them. I must admit I don't thumbs up music frequently, but absolutely everything I like I add to my library. I've been doing this since Google Play Music was first launched.

One glaring issue is that you can only easily thumb-down individual songs. There used to be a "Not Interested" button hidden the 3-dot drop down menu, I used that all the time but it never seemed to work - probably why they removed it.

im going to have to disagree here. ive trained my google play by hitting the like button on songs i like. My music taste is wide ranging but often quite selective within a genre and its really blown me away with how good its been with providing me with subtley new music i.e just similar enough to interest me but different enough that i dont get bored. This also feeds into my weekly 'new music that i might like' menu. I really dont get suggested anything thats in the charts anymore much to my pleasure!

So I guess this actually trumps my previous point, which was music is easier than video. But if you say now google is garbage at recommending music too... then I guess they are not as great as I thought on ML.

Can I ask what music?

It's been pretty good for me with new releases. I mean, it's not perfect but I listen to mostly indie rock and it's turned me on to a lot of stuff I enjoy, and it never suggests top 40. I also never use the thumbs up/down.

Perhaps one of the problems is that the music I listen isn't well defined in terms of genre - songs on the same album can vary significantly. However, I guess mostly post hardcore and progressive rock - there are bands/artists in these genres I like, but others I have absolutely no interest in. However, I also have lots of "rock", punk rock and some pop rock stuff in my library. I mostly get a lot of metal and hardcore recommendations which I'm simply not interested in. It's also not unusual for me to see rap and R&B recommendations, which I'm unlikely to be interested in (although some songs I like may be defined as such).

people with non-mainstream tastes are not statistically significant and in the age of ML optimization we can expect recommendations for them becoming even worse

One of the many problems with machine learning, and using it for recommendations, is that it completely ignores how humans think and behave. Just because I watch a few videos of a certain type today, and actively seek them out, doesn't mean that I want to see similar videos again next week. Maybe I was just in the mood for that type of content today, tomorrow may be something completely different.

That being said, YouTubes recommendation algorithm is mostly garbage. I don't believe it has any insight into the actual content or quality of the videos it recommends.

Right!

YouTube simply knows what you'll find tempting to watch by being just a bit more radical than you currently are, in a spiral.

https://www.nytimes.com/2018/03/10/opinion/sunday/youtube-po...

Which is, indeed, garbage...

We see you recently bought a dishwasher. Here are some ads for more dishwashers.

Anecdotal, but I recently wanted to make bread. I watched 4-5 videos to get the gist of it, made bread, all is good.

Except it's not. About 70-80% of my recommandations on YouTube are about bread now.

Again, anecdotal, but I can't get rid of Bill Buhr recommendations. I've tried numerous times to help YouTube by telling the algorithm that: "No, I don't want to watch Bill Buhr".

But no, he's immune.

That's because there is so much variation in what you watch on youtube - from listening to music, to movie trailers, to sports highlights, to howtos about cooking, etc.

The music we listen to is way easier to analyse, even if we have some variation in what we like. Sometimes im in the mood for classical, sometimes its rap. Easy to differentiate and offer me playlists for each genre.

> And yet, YouTube recommendations are terrible lowest common denominator nonsense.

You might be assuming that YouTube recommendations are optimized for your enjoyment. In fact, they might be optimized to maximize advertising revenue.

It is known (or has been frequently alleged) that the advertising revenue tier that videos are eligible for influences their position in rankings, for example. Some videos earn more from ads than others, and YouTube ranks the high-earning videos above low-ranking videos purely for that reason. It's very possible that the "lower common denominator nonsense" videos are the popular ones, with broadest appeal, that earn the most advertising revenue for Google.

Last.fm (which tracks played music) could already recommend me some good music ten years ago... It could even tell why it recommended something ("because you like X and Y").

They even had a fricken desktop app for Linux to play their personalized radio! Today, well the basic tracking funtionality still works...

Weirdly it seems to be because of scale - YouTube is just so big with such a high volume of videos, user data, and new videos, they can't yet run the ML algorithms they'd really want to over the whole corpus.

https://www.quora.com/How-does-YouTubes-recommendation-algor...

So meanwhile we get emotionally-stunted video recommendations from a savant toddler. It's really harmful.

That just sounds like a poor choice of a first step; why not use a random sample, then apply their sophisticated algorithm on that?

>And yet, YouTube recommendations are terrible lowest common denominator nonsense. It’s just “more of the same” for some surface level definition of “same”. There doesn’t seem to be any playfulness to the learning algorithm at all. No Bayesian experiments. There’s no flex.

Because the goal is not to suggest the videos that you're most likely to like. The goal is to maximize revenue. That pirated rip of a How It's Made episode will make them next to nothing, but that 10 minute partnered video "DO GIRLS POOP?!?!??!" will make them how ever many thousands of dollars the creator payed to promote it + ad sharing.

You are making a good point - Youtube recommendations ARE garbage. But I wonder how harder recommending videos compared to music. My guess is that doing data analysis on music in simpler because you only have to analyse the music data itself (frequency, bpm, etc) + related artists, etc.

I'm sure doing movies recommendation is easier than youtube videos. For example I watched "this is america" recently. I liked the video, but mostly because of the video, not the music. How does youtube understands that I want more of this type of imagery and not sound? Very hard.

That might also explain why we services like Spotify doing a good job at discovery, but I can't think of a service doing the same for videos.

But I agree with you, lots of example of google not doing a good job with ML. At the same time, it seems like they are doing a good job with ML on other (maybe simpler) things, such as gmail/inbox, search, etc.

> Google is so strong with ML today, and not stopping, I don't know what will allow smaller competitors to stay differentiated.

If they use the same algorithm YouTube uses to pick videos for me, Spotify doesn't have to worry at all. It just gives me the same things over and over and I never discovered anything worthy from their video suggestions. I have to actively search to find anything new.

> Google is so strong with ML today, and not stopping, I don't know what will allow smaller competitors to stay differentiated. I just had a call with a G recruiter, and pretty much 100% of the roles they have in all their European offices are ML-related roles, on most products. And they hire without ML experience, which I think is a sign of how much they are investing in ML.

Weird. I'm having the opposite experience, where I'm not having luck finding an ML team, even with research experience. What teams/products are hiring for ML-related roles without experience?

email me and I'll send you the email of the recruiter I told to.

They bought Songza a while ago. The playlist for certain times of day, events etc. all are available on GPM on my phone, but not on AndroidTV. It is ridiculous.

Why are Google Apps always so unfinished and not thought through?

So what are they actually doing with all this ML, or is it just an effort to keep people with those skills from working for competition?

Reply on news.ycombinator.com