00:00like storytime this is great cool so
00:04yeah before we talk shop just out of
00:08curiosity how many of you work more or
00:12less in product just to get a sense okay
00:15sort of majority of the room awesome and
00:17how many people either use discover
00:19weekly or pretty much aware of what it
00:22is and what it what try see amazing okay
00:24so I think what I want to do tonight
00:28normally you know it's a new product I
00:31have been going out and talking about it
00:33and normally when I do that I talk about
00:35Oh aren't we clever here's this thing we
00:38made here's a bit about how the
00:40algorithm works and look lots of big
00:42numbers you know the end but I thought
00:45since this was a room of fellow product
00:47people what I'd rather do with the
00:49slides tonight is tell you a little bit
00:53about sort of the context around the
00:55product and the end the landscape tell
00:57you a bit about my background and where
00:59I was coming from when I joined Spotify
01:00and then really follow through how
01:04internally it came together and some of
01:06the steps along the way towards making
01:07it and then ideally we want to leave
01:10like at least 20 minutes at the end for
01:11questions so if you do one though gory
01:13details of the algorithm I can I can
01:15spill my guts then but yeah that's what
01:18I wanted to share they cool so let's
01:21begin hope everyone is had as much wine
01:23as me so yeah before we start I think
01:28it's worth stepping back for a second
01:29asking you know why does Spotify care
01:32about music discovery why do we invest
01:34in it why do a lot of our competitors do
01:37the same why do entire startups spin up
01:40around the concept of doing a better job
01:42of music discovery and there's a lot of
01:44reasons for that but for me
01:46fundamentally there's two big ones
01:48the first is technical technological and
01:52the second is is emotional so the first
01:54reason I think everyone in this room
01:56will know all about it is really over
01:58the last decade we've of course seen
02:00this huge shift away from an ownership
02:02model either you walk into a real store
02:04or you enter a virtual store and you buy
02:06some some music or some media and bring
02:10it home with you announce yours - the
02:11access model where you basically have
02:14you know in the case of music you
02:16basically have the entire history of
02:17recorded music at your fingertips and in
02:21this scenario when you've got the
02:24infinite jukebox you can't really just
02:26slap a search box on it and say alright
02:27job done have fun good luck you could
02:31but it misses a huge opportunity which
02:33is to make sense of all this abundance
02:35I think the promise of really smart
02:37music discovery technology is that in in
02:41a more comprehensive way than ever
02:43before we can always connect the right
02:45people to the right music or if you flip
02:47it around from the artists perspective
02:49you should always be able to find the
02:50right ears wherever they are in the
02:52world for whatever it is you're doing
02:53whether that's 20 people or 20 million
02:55people and so really when we think of
02:58the fact that we are currently
02:59rebuilding the music industry and the
03:01music economy and how how music travels
03:03throughout the world you know we feel
03:06that this is not only a cool opportunity
03:07to do make connections that haven't been
03:09able to be made before but also a
03:12responsibility when it comes to the
03:14people actually making music so it's the
03:17first thing sorry the second thing is is
03:19a feeling which I've illustrated in an
03:23old-timey way here this is an old
03:25feeling it's as old as music itself and
03:26I think anyone who cares even a little
03:29bit about music knows that the thrill of
03:31that moment when you first get turned on
03:34to a new song or a new band
03:36and the reason that that feeling is
03:41important is essentially in this age of
03:43access over ownership just pure catalog
03:46you know a new release some back catalog
03:48that's commoditized so you can basically
03:50if you just want to hear the new justin
03:51bieber album start to finish you can go
03:53to a lot of places for that but what
03:55isn't a commodity is your ability to
03:58create a compelling user experience and
04:01to retain you know the loyalty of your
04:03users by making them feel like this on
04:06the regular and so that's a really
04:08powerful feeling and I think the better
04:09we can get at you know at at being you
04:14know at creating these kind of scenarios
04:16the better chance we stand to to attract
04:19and retain users so it's really
04:20important for that reason to so I've
04:23been I've been scratching away at music
04:25discovery and music tech for over a
04:29mostly over in London actually where I
04:31lived until the beginning of last year I
04:33started back in 2005 at last.fm along
04:37with Johan is here somewhere he joins
04:41we were both early engineers there and
04:44last.fm was a pioneer in using a
04:47semi-automated collection of what people
04:50were listening to to do collaborative
04:52filtering and recommendations on on
04:54music listening data to create online
04:56radio stations and to connect you to
04:58other like-minded music fans and so I
05:02spent six years there and actually
05:04helped carve out the product layer there
05:05I was a lead engineer and we actually
05:07just had the founders and the engineers
05:09sort of typical ten years ago startup
05:11story so I eventually moved over into
05:13product and was head of web product
05:15there from there I took a bit of time
05:19off but then ended up spinning up a
05:20London office for a company called the
05:21echo nest which was if you imagine the
05:23sort of science he bits of last.fm they
05:25were like that writ large to MIT Media
05:29Lab guys started it and they had an
05:31incredibly rich API and said the music
05:33they crawled the web they did a bunch of
05:35other things um they're now actually
05:36part of Spotify as it turns out so it
05:38comes full circle but I joined them in
05:412011 doing kind of R&D special projects
05:43for them briefly experimented with music
05:47and booze discovery this one kind of got
05:50away from me we can talk about a Q&A if
05:53you're curious but then I would have
05:56projects for the echo nest I actually
05:58ran my own startup for for two and a
06:00half years and it was called this is my
06:01jam and it was basically I wanted to
06:05play with the opposite product impulses
06:07from everything I'd done till then it
06:08was always real-time big data lots of
06:11listening crunch the numbers and I
06:12really wanted experiment with you know
06:15slow instead of fast with hand-picked
06:18songs so this is my jam was a music
06:20social network where you got a profile
06:22and the only thing you could do with
06:23that was put one song on it at a time
06:25and it was meant to be your favorite
06:26right now and then you follow people get
06:28a playlist of their favourite so that
06:29changes slowly you update it only when
06:31you find something really great so that
06:33was sort of me playing with with
06:34constraints and seeing how far I could
06:36push the kind of slow hand-picked human
06:40music Tech but fast forward to the end
06:44of or the yeah the end of 2014 the
06:46beginning of last year and the phone
06:49rang and it was my friends that Spotify
06:52and they said hey we've got the
06:54super-talented machine learning and
06:56discovery team at in a New York office
07:00and they need a senior product person
07:04how would you like to do music discovery
07:05at Spotify scale and basically I was on
07:08the plane a month later and moved to New
07:11York and the first team I started
07:15working with there they like I said
07:18they're a bunch of the machine learning
07:19engineers and so they maintained a bunch
07:22of really interesting models of listen
07:23taste and and all the songs on Spotify
07:25and their main window to the user was
07:29what we call the discover screen and so
07:32this is it's still like this today you
07:34go into the Browse section of the app
07:36and then you tap on discover and then
07:38you basically get a grid it's a little
07:39bit like Netflix where it says hey
07:40because you're into this kind of thing
07:42and it's organized around strips of
07:44albums and it goes all the way down
07:45here's some top racks here's so many
07:46releases here's similar to artists and
07:48things like that and when I came across
07:51this I quickly recognized two things one
07:54that there was some really solid
07:56engineering behind this - it was
08:00dramatically under serving the Spotify
08:02user base it just wasn't really getting
08:04any love not enough people were using it
08:07and so there were a couple theories
08:09about why that might be the first theory
08:11says Spotify is growing we're becoming
08:14more and more of a mainstream product
08:16and turns out you know music discovery
08:19it's kind of a niche music nerd thing
08:21not that many people actually want to go
08:23crate digging and discover music you
08:25know there's a lot of other ways to do
08:26that and and this is just for power
08:28users so that was one line of reasoning
08:29the second line of reasoning which I
08:31believe did more was that something
08:34about the packaging or the delivery of
08:36this was wrong and I had a hunch that
08:38even if you care a little bit about
08:41music discovery you don't wake up on a
08:42Sunday and go like time to discover now
08:45and then where shall discover click you
08:48know and then spend time going into each
08:50back up so I felt like actually it was
08:53more just about the delivery mechanism
08:54and the amount of you know the you
08:56really had to lean in to use this and
08:58get value out of it even though the
09:00quality was good so I quickly canvassed
09:03around and I kind of said to the teams
09:05like what else we got going on like
09:07because everyone has hack projects
09:08inside you know fancy zone things so
09:10there was one other thing that caught my
09:13eye which was but when I joined they had
09:17do you remember 2014 I think was the
09:19first big year for Spotify year in music
09:21which is their kind of year-end review
09:23of what you've been listening to and
09:25they made a big website for each users
09:26and that's actually driven by our
09:28marketing team and they're they're
09:29awesome the fact that they could build a
09:31product like that is amazing and they
09:33had collaborated with a few people from
09:36this team on idea they're like we want
09:38to give people a gift at the end of this
09:40process and so this team said hey well
09:43maybe we could create them like a
09:45playlist of music they should check out
09:46in 2015 and so they created this thing
09:49at the end of the web experience called
09:52play it forward and it would generate
09:53this playlist and it wasn't perfect it
09:57had a lot of rough edges but as soon as
10:00I saw this two things struck me right
10:02the first was you know if there was any
10:05feature on Spotify that has mass appeal
10:07that almost everyone on Spotify knows
10:09how to use notice how to engage with
10:11knows how to manage and you know spin
10:14off of it's the playlist and I thought
10:15this is kind of our native this is the
10:17native unit of Spotify and in fact since
10:20what if I started users have curated
10:22more than two billion playlists so we
10:24actually have two billion of these on
10:26the service I think that number is even
10:27bigger now and then the second thing was
10:31I talked about these machine learning
10:34models the one that had sort of become
10:36the main model that we were using to
10:37power the discover screen and other
10:38stuff the irony was is it it was
10:41actually programmed by playlist saves
10:44not by by listening and collaborative
10:47filtering so essentially we'd created a
10:49model by looking at what songs people
10:51are adding to playlists and learning
10:53about the relationships between music by
10:55what real people on Spotify were
10:56curating next to one another and this
10:59was actually the first time anyone had
11:01ever thought this whole model is based
11:04what if we tried to make a playlist
11:05rather than using it to recommend
11:07artists artist or album to album that
11:10so immediately I was like okay there
11:12might be something here
11:13and so alongside the other work we're
11:17I spun up just a really small group it
11:19was essentially one engineer lead
11:20engineer who was passionate about this
11:22playlist' idea it was me and it was a
11:25bit of time from Paul you who's a really
11:27that's about if I and the three of us
11:29started quickly just jamming on this
11:31idea but we gave ourselves one
11:33constraint we said whatever we make it
11:35can't have any custom UI or extra bits
11:38let's see how far we can get with this
11:40idea just as a vanilla Spotify playlist
11:43so that was kind of the constraint we
11:45imposed and it was a bit practical at
11:48the time we didn't have a lot of
11:49front-end client developers on the group
11:51that's changed now but you know that was
11:53motivated by real constraints too so we
11:55gave ourselves that constraint and right
11:58away really in the first two three weeks
12:00a bunch of significant iterations and
12:03decisions happened so oh I I collect
12:06humans versus robots
12:08images because I often get asked to be
12:11on these panels that are like curators
12:13versus algorithms and it's it is not an
12:17accurate description of how anyone does
12:20and I hate this false dichotomy so
12:23anyway this is my current favorite I
12:25think it's hilarious so so we started we
12:30started iterating and the first thing we
12:31started looking at was how long should
12:33this playlist be right we're gonna fill
12:35it with music we think a user might like
12:37and I think the first version had like
12:3950 tracks or something and what we
12:40discovered is is staring at a playlist
12:42of 50 songs that are by design
12:44unfamiliar to you is kind of
12:46intimidating and so we kept shortening
12:49and shortening it and you know often the
12:52engineer was like oh but I love
12:53discovery I want as much as possible and
12:55we kept going let's make it shorter and
12:57finally we we landed on this two hours
13:00idea and 30 songs and sort of said no
13:02longer than that and it was interesting
13:05because somewhere around the 35 song
13:07mark all of a sudden it looked more like
13:08a playlist a friend might send you or a
13:10person made and it was just this I can't
13:13even you know the shift was really
13:14subtle but we were like okay this is a
13:15good life this feels like a Hugh
13:17you know less intimidating approachable
13:19length you could imagine listening to
13:20two hours of music it's not it's not
13:22crazy and then we thought well how do we
13:25update it maybe when you listen to it it
13:27changes maybe depending on how often you
13:30listen to it we update it maybe you
13:31click a thing and then we were like no
13:33we said vanilla playlist we can't add
13:36any special features and also we
13:38realized we had the opportunity to maybe
13:40create a bit of a ritual around it we
13:42thought if it was based on how often you
13:43consumed it it would be different for
13:45everyone it might be confusing and feel
13:47so we said two hours of music how often
13:50do you really want to hear two hours of
13:51new music that takes energy right
13:53listening the stuff that you aren't
13:54familiar with it actually does you know
13:56take a bit out of you
13:57so we thought weekly and then we're
14:00searching for a name and Paul the
14:01designer was like well you know you
14:05might discover weekly with this plan
14:07anyway so we hit on the names that was
14:09good so within a couple weeks we had
14:11like the size and we had this idea of
14:13every Monday we would we would refresh
14:15it for you so that started to feel
14:19pretty good then we had another
14:21challenge which was we decided this
14:23thing's called discover weekly so says
14:24discover weekly by Spotify and then
14:26playlists have a short description it's
14:29like two lines of text where you can be
14:31like this is a blah blah blah but most
14:33people don't read it on mobile it's
14:34hidden behind the swipe and we realized
14:37we needed to communicate that this
14:38wasn't just another playlist from
14:39Spotify like we have acoustic
14:41coffeehouse we have good morning we have
14:43get pumped we have wrapped caviar we
14:46have all these playlists people might
14:48just think it was another one of those
14:49so we needed to convey that it was
14:51personalized and it was just for you
14:53and we were kind of running out of real
14:54estate within our constraint to do that
14:56so we went for the oldest trick in the
14:58book which is why don't we put your face
15:00on it so we we have a many of our users
15:04are connected to Facebook and we
15:06realized we could get their profile
15:07photo and then at the time our design
15:10team brand team was working on this new
15:11kind of duo toned image style which we
15:14actually rolled out last year it's kind
15:15of our new our current aesthetic and we
15:17realized it was actually trivial in code
15:19to apply this color palette to your
15:22profile photo and so basically we end up
15:25creating you as a cover star where
15:27possible and then if we you are
15:28unconnected to Facebook you get Buzz
15:30Aldrin walking on them
15:31which felt appropriately discovery and
15:34we actually tested this just to be
15:35rigorous about it and turns out you are
15:3717% more likely to engage with the
15:39playlist if it has your smiling face on
15:42it it wasn't the perfect approach we
15:44actually want to improve this some
15:46people did think oh my god Spotify
15:49picked me to be the face of the discover
15:54we also yeah we had various other
15:57complaints our legal team had to work
15:58pretty hard when this first came out
15:59there was a guy saying you have used my
16:02artwork without permission for discover
16:03weekly and of course his Facebook
16:05profile photo was his artwork etc so
16:08it's not perfect and and we're still
16:11evolving our approach here but that was
16:13that was that felt like a breakthrough
16:14now we're like okay this is a different
16:16kind of player we've never done this
16:17before this is cool and then of course
16:19there was content right obviously this
16:22whole thing hinges aren't on our ability
16:23to make good recommendations and and
16:26fairly sum them to be truly great so
16:28this went through the most revisions and
16:30iterations we have a term in our product
16:34area at Spotify it's in charge of
16:35discovery and recommendations and it's
16:37we try to minimize the number of WTFs
16:40which is really just when we say WTF in
16:42this context we mean those tracks that
16:45stand out like what is this doing here
16:46you know this this is like we were going
16:48great and then this you know and humans
16:50are really good at spotting those right
16:52and so we went through a bunch of rounds
16:54setting like popularity thresholds a lot
16:57of you know a lot of iterations tricks
16:59trying to smooth over the WTF there's
17:01still many of these lurking and now that
17:04worth scale were of course able to run
17:05tons of experiments around solving them
17:07currently my main bugbear is Swedes tend
17:12to get a lot of Finnish and Danish rap
17:14and I don't know that much about inter
17:17scandinavian politics but if that's not
17:19cool according to Sweden
17:21whereas for Danes getting Swedish rap
17:23might be ok it's very anyway ongoing
17:26anyway so we did we did a lot of that
17:28and the important thing I think from a
17:31product starter note is during this
17:32whole it was really just a month but it
17:34was like almost every day we'd at least
17:35check in with each other we're pretty
17:38much doing this just within this small
17:39group and then eventually within the
17:41slightly larger team of
17:421214 people that that we were working in
17:45and this was really important because it
17:47gave us we weren't hiding it but it just
17:50gave us a safe environment to experiment
17:53in and we knew that we wouldn't really
17:55show it outside our group until we
17:57personally were satisfied that we'd met
17:59some kind of baseline quality level that
18:03that we you know we weren't sure in the
18:05early days of this that this was really
18:06a thing like maybe it was interesting
18:08but maybe it wasn't a product and so we
18:10wanted to make sure we could iterate to
18:11the point where we knew it was safe to
18:13share and in this case cuz of all the
18:15things I mentioned that came together
18:17pretty quickly and after about a month
18:19so I joined in January and by February I
18:23made mid to late Feb we actually had
18:27something that we're ready to share more
18:29widely so I said to my colleagues I'm
18:32new at Spotify but you know what's the
18:34next step how do you how the heck do you
18:36say I think we have a new product here
18:37and they're like oh just do an employee
18:39test just email the whole company and
18:41tell them you now have this new thing
18:43and then ask for feedback so I emailed
18:45the whole company in my my fifth week in
18:48the job I just was like hey guess what
18:52you know for the employee test discover
18:54weekly should magically appear we just
18:55added it without asking to the list of
18:58playlists of every employee and then
19:01have a link to get feedback and you know
19:03blah blah blah and to my pleasant
19:05surprise everyone freaked out everyone
19:08was like this is so good these are some
19:11next-level wrecks my favorite quote it's
19:13as if my secret music twin put it
19:15together everything in it is good so on
19:18one hand we were like great this is
19:20maybe a real thing on the other hand we
19:22are keenly aware that Spotify employees
19:24are not the general population and do
19:27not represent the average Spotify user
19:30we are for obvious reasons music nerds
19:33we over-index heavily on caring about
19:36this stuff so I was kind of worried that
19:38we might be falling into the trap of we
19:40made a thing for us and you know maybe
19:42maybe people don't care about discovery
19:44so rather quickly we decided all right
19:47the best way to test that out would be
19:49to quietly do the same trick on real
19:56a few weeks after the employee test we
19:58spun up a 1% test so about at that time
20:01about 700,000 people got this added to
20:05their playlist and I was pretty excited
20:07to once again be at a company where a 1%
20:10test involves almost a million people
20:11that was well so we put that out and the
20:16you know I helped a bit but really
20:17credit to the team here that the team I
20:20was working with it was really rigorous
20:21of other metrics from day one even with
20:23the employee test we actually had some
20:24dashboards up there you know almost
20:261,500 Spotify employees at that time and
20:29that's enough to actually start seeing
20:30some patterns in data so we have a
20:32framework we call reach retention and
20:34depth reach is okay if everyone who
20:36could have used it how many people
20:37actually streamed using that feature
20:40retention is you know week of a week or
20:43or by whatever interval you decide how
20:45many people actually returned to it and
20:46depth is how long are they spending how
20:48long are these sessions that you're
20:49actually you know it's great if people
20:51try it but if they play two songs and
20:52then they're out it's probably not that
20:54worthwhile so this was how we evaluate
20:56it but I was keen like I said it's easy
20:59to laugh at now but at the time we
21:01weren't entirely sure what we'd made or
21:03how to package it or how to call it I
21:05was thinking about how do we all on
21:06board people all this stuff and we
21:07really just wanted more candid feedback
21:09from these users that we were running
21:10this test on so we did something really
21:12guerrilla we had this description to
21:14play with in the in the playlists and in
21:17that we said hey you know discover
21:19weekly is a new experiment where we give
21:22you two hours of music to check out
21:24based on what you've been listening to
21:25tell us what you think and we had a link
21:28to a Google Form that was just your
21:30standard in a free Google Forms thing we
21:32just threw together as a team and it
21:34just had two questions on it rate the
21:37music in here out of five five meaning
21:38wow I found my new favorite song zero
21:40meaning it sucks and freeform text input
21:44any other thoughts typed them in here
21:45and it was crazy we got over 1500
21:49responses in our first two weeks we we
21:53were very proud of our Yelp rating of
21:55about four point three so you know I'd
21:57probably eat at this restaurant and for
22:02me the coolest thing was people really
22:03opened up about like hey this is how
22:05this makes me feel hey this is you know
22:08what I think this is
22:09and actually as we got closer to the
22:11full rollout I worked with our consumer
22:15marketing team product marketing team
22:17and we actually ended up basically using
22:20language that came from our own users
22:22telling us so a lot of the feedback was
22:24like it's like a friend made me a
22:26mixtape so we ended up using that in the
22:28messaging so actually a lot of the
22:29language in the press release and when
22:30we went to market with this thing came
22:32from what people told us they thought
22:34the feature was during this test so that
22:36was that felt really cool and again that
22:38was just fairly serendipitous we just
22:39wanted a bit more color to the you know
22:42cold hard numbers we were looking at but
22:45the numbers were good the numbers were
22:47crazy good in fact and normally it's
22:50fought a fire anywhere really as a
22:52product person when you're but to rule
22:54something out you think I'll write what
22:56are all the touch points what how do we
23:00on board existing users how do we on
23:01board new users maybe we should have
23:04explicit feedback like you should be
23:05able to say thumbs up thumbs down
23:06perhaps you know we knew there were a
23:09lot of things we wanted to build
23:10and we'd really done this test by
23:11cluding it we just without asking you
23:13know kind of impolitely
23:15we'd added it to people's playlists and
23:17ID until now I'd been thinking oh we
23:20would never ship it like that like that
23:21you can't just you know sneak it in
23:23there and hope they notice it and call
23:24that like a good a good product launch
23:27but then the response was so
23:28overwhelming and we were sitting on this
23:30thing and going like people would like
23:32this that that is exactly what we did we
23:36just thought you know what we can
23:38activate people later why don't we see
23:40what the whole world makes of this and
23:42just let word-of-mouth do its thing and
23:45so we launched it in June it was again
23:48just my fifth month there and we just
23:52added it to 75 million people's lists of
23:54playlists and went out the world said
23:57hey this is this new thing we're trying
23:59and brings discovery to you 2 hours of
24:01music once a week check it out we picked
24:04the weekly cadence you know just based
24:06on intuition and when we had to take a
24:08day of the week we thought well Mondays
24:10suck like why don't we why don't we put
24:12it there it seems like a logical start
24:14and yeah so one of the fun things that
24:17the team likes to do whenever we need
24:18you know a pick-me-up is you can search
24:20on Twitter for discover weekly and
24:22and oh also we yeah the response to get
24:27back to the only nerds care about
24:29discovery point we were able to disprove
24:32this with discover weekly I mean its
24:34first ten weeks it hit a billion tracks
24:36I can share a new number actually we
24:38ended 2015 with 2.5 billion tracks
24:42stream through discovery weekly so
24:45that's a big number and big numbers are
24:46cool we like those but for me the fact
24:48that by design all of those streams are
24:50our discoveries and that we're
24:52potentially helping artists find whole
24:54new audiences is really exciting but
24:58more than just the numbers again this
24:59user response has been crazy this person
25:02got really excited started crying
25:03because tomorrow's Monday and they got a
25:06new discovery weekly that's pretty cool
25:08it's so sad how exciting it is okay this
25:13so Florida manatees are no longer near
25:15extinction and there's a new discover
25:17weekly that's that's it's pretty awesome
25:21something that we think is is this one's
25:25funny because we don't actually ping you
25:27when it comes out you kind of have to
25:29like I said we shipped it without you
25:31know rolling tell most fo we'll get
25:32there but this is you know friends you
25:34know friends tell friends to check their
25:36discover weekly that's cool and yeah
25:39users are forming whole new habits I
25:41mean Goffman is now baby bathing every
25:45every Monday morning so that's that's
25:47pretty pretty heartening and now that
25:51it's out there you know now the real fun
25:54begins we are now that we're at scale
25:56we're running a bunch of different
25:57experiments around content how we choose
26:00it we've been partnering with our audio
26:02team to figure out how much does the
26:03sequence matter should it you know move
26:06coherently through genres and and and
26:09use some of what we know about the audio
26:10to shape the playlist tons of
26:13experiments there we're also looking at
26:15you know they've been very common
26:16request you know when a track shouldn't
26:18be in there when it is bad can I get rid
26:20of it doing things around letting you
26:24explore deep dives into different genres
26:26you like things like that we're thinking
26:28about but also just for context the
26:31product group that I look after it
26:33Spotify is now growing it's it's you
26:35forty or fifty people even and actually
26:38the discovere weekly team is still three
26:40three people four people and it's cuz it
26:43really was about you know taking these
26:45building blocks years of music science
26:46all that curation almost a decade of
26:50curation from Spotify users you know
26:51every song is in there because someone
26:53put it on a playlist the existing
26:56playlist infrastructure that we're
26:57reusing and so a lot of the struggles
26:59have discovered weekly we're just on
27:00scaling it it turns out when you have a
27:02system to store all the playlist on
27:05Spotify and every Sunday night you try
27:07to update seventy-five million of them
27:08in an eight hour time window interesting
27:11things happen so a lot of what we're
27:12working on now is taking the learnings
27:14from this and creating infrastructures
27:15so that other teams can create similar
27:17lists so that all of Spotify can benefit
27:20from some of the personalization
27:21learnings and really create those
27:23feedback loops right now we filter out
27:26your listening from discover weekly when
27:28we're building a picture of your taste
27:30because you know we don't want to cause
27:31presentation bias and have the snaky did
27:34but what we're gonna experiment with
27:36soon is when you save stuff and when you
27:38skip stuff we can feed that back into
27:40the model and hopefully make the recs
27:41even even sweeter so that's all I got
27:43for slides so any any questions
