Video Recommendations in Industry
In this episode, Kyle Polich sits down with Cory Zechmann, a content curator working in streaming television with 16 years of experience running the music blog "A Silence No Good." They explore the intersection of human curation and machine learning in content discovery, discussing the concept of "algatorial" curation—where algorithms and editorial expertise work together. Key topics include the cold start problem, why every metric is just a "proxy metric" for what users actually want, the challenge of filter bubbles, and the importance of balancing familiarity with discovery. Cory shares insights on why TikTok's algorithm works so well (clean data and massive interaction volume), the crucial role of homepage curation, and how human curators help by contextualizing content, cleaning data, and identifying positive feedback loops that algorithms might miss. The conversation covers practical challenges like measuring "surprise and delight," the content deluge created by democratized creation tools, and why trust in tech companies is essential for better personalization. Cory emphasizes that discovery is "a good type of friction" and explains how the CODE framework (Capture, Organize, Distill, Express, plus Analysis) guides professional curation work. Looking to the future, they discuss the need for systems thinking that creates narrative connections between content, the potential for conversational AI to help users articulate preferences, and why diverse perspectives beyond engineering are crucial for building effective discovery systems. Resources mentioned include the newsletter "Top Information Retrieval Papers of the Week" and Notebook LM for synthesizing research. Connect with Cory on Instagram and TikTok @silencenogood or at [silencenogood.net](http://silencenogood.net).
Guest
Cory Zechmann: a content curator working in streaming television with 16 years of experience running the music blog "A Silence No Good."
Transcript
Kyle Polich : Welcome to Data Skeptic a podcast exploring the methods use cases and consequences of recommender systems
Kyle Polich : Welcome to another installment of Data Skeptic Recomer Systems If there's a bias this podcast has it's towards the algorithmic and machine learning side of things Explain to me what it is you think a human can do that a machine can't Whether you're right or wrong that's an interesting question
Kyle Polich : Yet somehow especially when it comes to recommendations and this might be the final frontier even human curation surely plays a strong role in recommender systems now until forever In this episode I sit down with Corey Zeckman a content curator who's worked at a lot of streaming companies on problems like this as well as his own music blog Silence No Good He's at a unique position in industry seeing both the curation side and the heavy data-driven world of recommendation systems
Kyle Polich : We get into topics like what algatorial is and the role that human in the loop component still plays in most modern systems
Corey Zeman : My name is Corey Zeman I currently work at Sling TV and I have run a music blog for about 16 years in which I curate music
Kyle Polich : We'll start there Tell me a little bit more about the blog
Corey Zeman : So it started off in 2009 I had been pirating a lot of music on uh different torrent sites and felt bad about it And I was a college student and wanted to give back in some way And blogs were a thing at the time So I decided to start a music blog to promote mostly underground hip hop I'm in the Bay Area right now And I moved out to the Bay Area because I learned about underground
Corey Zeman : Hip hop and really wanted to help give those artists a voice And the music blog has evolved a lot It's not necessarily just hip hop anymore more electronic music house music funk music soul music But yeah it's given me the opportunity to go to a lot of conferences I go to South by Southwest every year I cover a bunch of music festivals and really understand culture not just from a computer but really engrossing myself in podcasting And it really pushed me into
Corey Zeman : To audio and podcasting and now into film and television where I'm curating professionally I think of a silence no good as my unprofessional curation because it's really not about the data It's not about the analytics It's just about the feeling you know just the story behind it Whereas my work in film and television and podcasting has been very data focused because it's more of this kind of pulled back view of curation and looking at other curators and trying to map out the
Corey Zeman : Industry more so than just looking and trying to discover the next best music artist
Kyle Polich : Where can listeners find the blog
Corey Zeman : Silencenogood.net or .com I got the .net first and then got the .com as soon as that was given up Otherwise TikTok Instagram are probably the places that I am most frequently using right now opposed to the blog The blog is less so more just across the industry And for me as well I like to write more so on TikTok and Instagram
Kyle Polich : Can you describe the role of a curator like you show up for work what do you spend your time on
Corey Zeman : So there's something called building a second brain It's actually for content creators They had this process called code It's captured organized distill and express Curation is a form of creation and it is the same process that I take in And I think this is also to be fair what machines in a very simplistic way are doing as well is that they're retrieving all of the data set that they can use for whatever playlist or whatever
Corey Zeman : Type of recommendation they want they're organizing it so filtering out certain types of things maybe business rules that you know you don't want again paid programming to be the easiest one to say and then ranking them so ranking them on how relevant or how likely that person will like that for a curator there's no personalization for the most part there are some tooling that you can have for personalization more so at companies than you know something on
Corey Zeman : Spotify quite yet but being able to organize and then distill all of it and ideally contextualize it right like to be able to have a title and a description and artwork and all of these different things because you can drop the best movie onto someone But if they don't have the context of like are there any famous directors or artists or actors you know what is the plot and all of these different things
Corey Zeman : Artwork is the best way to showcase that because especially on a visual or a streaming platform people don't really want to read too much but they do want to see you know great artwork and you know being able to AB test this
Corey Zeman : I remember Mr Beast talks about this a lot is that he on YouTube and YouTube now has this functionality and being able for the creators to AB test their artwork being able to AB test all these different things And that is basically what I'm doing but more from not from a creator standpoint but from the company's standpoint and being able to contextualize and get people to understand why people would want to consume this And then the last part that code really doesn't go into but is clearly important part
Corey Zeman : I analysis and we kind of have this loop of you know creating a playlist and putting it out to users but then seeing like how well did it resonate Because on some platforms true crime does better than fictional crime And and on other platforms it's the opposite So being able to kind of gauge what do our users want And how can I make the experience better by these daily or weekly cadences It's small incremental improvements in this case but being able to just make a better experience day to day
Kyle Polich : So when you work on something like the blog you're doing a curation but it's a very like personal one tastemaker kind of collection you know you were going and exploring artists and handpicking or whatever your process is it doesn't feel very algorithmic How do you merge these two worlds or do you compare and contrast what we can do with machine learning versus what you do in the blog work
Corey Zeman : Yeah with the blog that's kind of the problem is that I don't have a lot of technology to use in order to make the process not even necessarily more efficient but just more robust and a better experience So I track maybe 50 or 100 playlists a week or a month depending on their cadence And I go through them every week or every month depending on their cadence But I also
Corey Zeman : Go through like Discover Weekly which is OK And we can talk about that because I've learned some new things about Discover Weekly that I thought was interesting But other algorithmic playlists and the curation and the algorithmic playlists it's kind of a crapshoot A lot of them are even personalized to me obviously the more algorithmic ones And it feels like it's pulling from pretty basic metrics like just the vibe and the feel and
Corey Zeman : I get this a lot even in my job you don't necessarily want something that's similar in music or in film or in television You want something maybe completely different but have these distinct threads through them that it's hard for an algorithm to pick up Whereas humans they can do that but it's hard for them to personalize that So it's this combination of the two that I've I talked a lot about at this Spotify I coined it algatorio or just the mix
Corey Zeman : Of the two is really the sweet spot because there's a lot of things that I can't do that machines can do and there's a lot of things that machines can't yet do that I can do So
Kyle Polich : if I bring I guess my most arrogant machine learning self to the table I would say OK fine maybe you have some hidden insight Make me a feature and I'll put your feature in my model and that's all there is to it Is that kind of deflating the effort or how do you feel that describes the situation
Corey Zeman : I think eventually it comes to that but it is us working together and deciding how much this metric should be weighted versus this metric A lot of the times that I see is like you put a lot of these things into a model and they don't necessarily know oh this is what should be focused on here versus this what should be focused on here A lot of the times we look at a lot of positive feedback loops at my job and just it getting recommended this popularity bias or just something getting
Corey Zeman : Recommended way too much for good reasons in certain ways because you see it in the metrics that people are engaging in it But you can't really gauge their happiness yet You can't gauge do they really enjoy it versus are they just kind of passively scrolling by and they're like OK this is good enough I just see a lot of content whether it's in film and television on social media that it just tends to be kind of the same thing And it doesn't break you out of that filter bubble And there are ways
Corey Zeman : Ways in which it can do that and still bring you back to things you love and not just be about random discovery but discovery that brings you back to things that you love Recently I was listening to a podcast that talked about surprise and delight And it's definitely an industry term that I've heard working at Apple a while ago But it really does give that basically with filter bubbles and the cold start problem It's just this mix between
Corey Zeman : Making sure that people find things that are relevant to them but also pushing them outside of those boundaries And it's this push and pull that surprise and delight or filter bubbles or familiarity versus discovery or I think the industry term for recommendation systems is explore and exploit these types of things this balance that it varies so much from person to person even time to time for that person that it's hard just to
Corey Zeman : Put it into an algorithm or a bunch of algorithms a system in which it can really understand that point of view at the time for that person There's just so many variables that it can't even look into because of privacy you know there's a lot of things that we don't want these algorithms to know but we kind of do It's just we want to be able to trust them and trusting these algorithms a lot of tech companies have made it so that you really don't trust these algorithms because they're being used for different purposes than just recommending content
Kyle Polich : Well I really like this idea of surprise and delight as a metric Sounds great in a boardroom right we're going to maximize delight but now you hand that to the engineers and they say what does it mean mathematically that could be overthinking it maybe it's we just need to take the commander's intent to heart and that's what it is or maybe it is something more formal you can measure how scientific is surprise and delight
Corey Zeman : I think the surprise is the hard part is that you can delight people with things that they've already known but aren't necessarily thinking about It's that surprise part It's hard to mathematically gauge because you can't look at whatever emotions someone's feeling And it would be very easy to gauge if you had all the metrics on people if you could just really gauge how happy they were
Corey Zeman : Their cortisol levels their their heart rate that kind of thing If you could do that I think that's when you could really measure the surprise and delight But right now it is more of these implicit and explicit data that again is just very proxy data It's very far away from what people actually want But there are these guessing games of OK well
Corey Zeman : They liked this They watched this at least 90% So therefore it's a good approximation that they at least liked it to a certain extent Getting that to that surprise and delight even more so is a lot harder because even in explicit data like a like at least now Netflix has the double like or the love or whatever it's hard to understand how much they love it just because having too much explicit
Corey Zeman : Data it's a lot of friction and people will just stop using it That's a problem I see a lot with explicit data from the companies I've worked at is that any type of friction even if it's the smallest amount of friction people will just drop off and not utilize the feature What I see with explicit data being useful is negative feedback I do this a lot with TikTok and a lot with Instagram
Corey Zeman : There's already so much friction when I'm starting to not like something that I don't really mind giving feedback that this sucks versus when I'm enjoying something I don't want to have to like it I don't want to have to even save it I just want all that done for me And that is where it should be really showing that I like something
Kyle Polich : Well can you say a few things about the role of curation in the modern era and maybe a little bit about how AI is evolving that or not evolving it
Corey Zeman : I think more from a professional lens curation and machine learning have kind of been at odds with each other at least at the corporations I've worked on in the past And at the start I worked at TuneIn Radio and we were completely separate orgs and we did not talk really at all together But more and more so uh I've come to realize and I think tech companies as well is that the the melding of the two is really the most important
Corey Zeman : Part of how can I help these machine learning algorithms or these systems have a better understanding of what people want you know are there certain channels that may not be that helpful to users Like if it's paid programming should we recommended that to users Most likely not Most people don't want to necessarily listen to paid programming and those types of things It's like helping with whether it's filter bubbles or the Coldar problem or
Corey Zeman : These positive feedback loops making sure that we're looking at the data and contextualizing That's what really what my team does best is contextualizing The data isn't just numbers There's a lot of story behind it And it takes people who are constantly keeping up with trends in the industry and just what people want That's where we really come in and try to give that context to both data analytics to the engineering teams and products
Corey Zeman : Because they're working on just making a great product And whereas we are looking outside of that product and seeing what is the industry doing How are people really discovering new music now or new content now versus how they were previously It's really that more holistic viewpoint that we're trying to bring in Again there's a lot to learn just in within a company even outside of that company and really bringing that outside perspective in is really what we're trying to do
Kyle Polich : Well from my perspective I think if we rewind it to let's say the 90s there's been an encroachment of what engineering solutions are going to do So for example we don't need a person to clap out the beats per minute of a song we can use signal processing techniques if we care about beats per minute
Kyle Polich : I suspect nowadays we could reliably classify the genre even kind of specifically like not just hip hop but that probably divides into 20 subgenres I'm not aware of I think probably an algorithm could do some of that Where do humans maximize their contribution
Corey Zeman : Actually what you kind of brought up let me get to your question after this but how much machines can help humans There is so much drudge work I talk about and that of ordering like you know we'll have a romantic comedies collection And I don't want to have to order that for everyone Obviously that's almost impossible But what machines can do for me is organizing that and sorting and this is where where gatorial comes in is like I'll give a list of 100 different romantic comedies and the algorithms will
Corey Zeman : Personalized to each individual user And so we clean up the data is really what we're doing best at is that there's a lot of romantic comedies that are maybe not applicable to everybody but are still important that we include But then there are some that are just really outside of probably what most people want So we kind of filter that and make sure that the data set is clean enough at least for that perspective There might be some romantic comedy playlists that we want that are a lot more exhaustive and have hundreds or thousands of movies
Corey Zeman : Whereas some different types of playlists that are a little bit more fine tuned that are romantic horror comedies or whatever else it may be And the metadata is usually helpful but not always accurate And at least with recommendation systems now there are a lot of issues with it recommending something based off of bad metadata And that's where we can kind of come in and help clean that up It's just again giving that context to clearly just fixing things if the metadata is off
Kyle Polich : Will you use the word algatorial that I'm not sure every listener's gonna know although hearing it kind of self-defines in a nice way So you have to give us the dictionary definition but could you talk a little bit about how that plays out in your day job
Corey Zeman : I think a good way to put it with film and television but it's this mix of algorithms and editorial again having to deal with so much content exhaustion like I want some things automated I don't want to have to deal with it I don't want to have any brainpower that has anything to do with that versus augmenting some things and that I want to be a part of that kind of that human in the loop where we're playing back and forth with this machine learning
Corey Zeman : Algorithms and saying like OK I agree with this And this is where large language models come in And we're only starting to really get that into our systems where we can start talking to these systems and saying I don't know if I agree with this being there And I don't know if we should have this there I want to be able to talk to these algorithms so that I can either automate certain processes or augment them and really be that human in the loop looking at the data and saying clicks went up
Corey Zeman : Watch hours really went up high but it's mostly people just consuming the content that they've consumed before So that looks good in the metrics but we want to get them pushing into different things So I think this is where this algatorial term comes from is display between us And in its most simplest form it is having a playlist you know whether it's 50 or 100 movies or 1000 or 5000 songs and
Corey Zeman : The algorithm sorting that for a different type of person versus another one I listened to a podcast a long time ago with Gustav Saterstrom which is the soon to be co-CEO of Spotify and he talked about how there's songs to sing in the car you know we all like certain songs to sing in the car and I may have a completely different you know 5 1015 songs that I want to sing versus the 5 1015 songs that you want to sing So
Corey Zeman : Their editors will throw in maybe 1000 different songs that they've all deemed the best songs to sing in the car whether research it on blogs or they've looked at the data in some other way And then my playlist will have completely different 15 to 20 songs than your playlist but all of them are definitely songs to sing in the car that maybe a machine learning algorithm can't necessarily parse out
Kyle Polich : Well one topic that seems to come up in most of my interviews is the cold start problem Everyone doing any sort of recommender's got to deal with that For the most part machine learning people just throw up their hands like I don't know that's it is what it is It seems like maybe that specifically is an area where curation could raise the bar Do you see it the same way
Corey Zeman : Yeah yeah I think that's probably one of the best places just because if we don't have data on users it's very hard to predict what they want to consume next And I think if the Cold Star problem is is a continuous problem I've heard of it somewhat called also a warm up problem where you have a little bit of data on people and yet you still can't make those predictions that are novel yet relevant to people
Corey Zeman : And that's really where we can come in and bring in this cultural relevance A lot of what my job also is is is cultural relevance is looking at you know HBO looking at all of these different platforms and seeing like what is popping on Apple right now what is popping on all of these different platforms because our data might not show that yet but really outsource data like IMDb and all these other places does show at least a
Corey Zeman : A Little bit of an uptick of potential there So let's try it out and let's test it on our own platform to see if there is good content You know we talk about itemized cold start versus user cold start and they're kind of the same thing on different sides of the coin or however you would say it but
Corey Zeman : Yeah being able to figure out what people want a lot of it does come back to this popularity bias I don't think of popularity bias is necessarily a bad thing except for when it becomes this positive feedback loop that becomes a little bit too much But it is a good indicator of what people might like because people want to connect over this content And if you know there's these kind of water cooler moments where it's like I want to know the latest thing on Netflix so that I can talk about it with my friend or you know the latest true crime
Corey Zeman : On Hulu so that I have something to talk about at work Uh I think that's where we can kind of come in with a lot of outsource data that our recommendation systems don't quite have And even recommendation systems like I use Gemini and and and chat GBT to help me kind of pull in things But there's a lot of issues there just as far as copyright with them And whereas I don't have that issue of being able to look at IMDb data chat GPTs hit or miss can't always look at IMDb data
Kyle Polich : DeleteM makes it easy quick and safe to remove your personal data online at a time when surveillance and data breaches are common enough to make everybody vulnerable
Kyle Polich : Your data is a commodity Data brokers compile your name contact info Social Security number home address even information about your family members and they sell it online to anyone willing to pay This puts your data at risk for identity theft phishing attempts and harassment But there's a solution Delete Me removes your personal information from hundreds of data broker websites helping you protect yourself from all those threats
Kyle Polich : As someone with an active online presence privacy is really important to me I started using DeleteM because I wanted to take control of who had access to my information
Kyle Polich : The detailed reports they send show exactly what data they found and removed and it's honestly eye-opening to see just how much of it was out there It gives me real peace of mind knowing my digital footprint is being actively reduced
Kyle Polich : Take control of your data and keep your private life private by signing up for Delete Me Now at a special discount for our listeners Get 20% off of your Delete Me plan when you go to joindeleteme.com/data and use the promo code data at checkout The only way you get that 20% off is to go to joindeleteme.com/data and enter code data at checkout That's joindeleteme.com/data codedata
Kyle Polich : There's a maturity cycle I've seen a lot of organizations go through V1 let's hope it stands up and works and kind of does the job and you know you make as fast of iterations as you can improvements along the way At some point it sort of levels off Maybe there'll be some really you know outstanding new feature you deliver and there's a little hockey stick moment or a stair step
Kyle Polich : But for the most part those things kind of asymptote in a way and you're fighting for small margins on a lot of data And if you're gonna AB test as you were describing at least in my experience a lot of times you're measuring a signal that's hard to measure or there can be questions about what are the right metrics are we looking at everything Could you talk about your approach
Corey Zeman : Yeah I always talk about how every metric is a proxy metric in my viewpoint it might not be the best way of saying it but these are just trying to guess and understand what people are consuming and what people want to consume and
Corey Zeman : So Roblox has a a TED Talks podcast and they had an episode on Discovery And I think they said it very eloquently in that there are short term metrics like our people engaging in the world these kind of short term metrics of engagement on on a film and television platform it might be again clicks and watch hours and things down the funnel or time to video and things like that But these are all very short term metrics And that's good
Corey Zeman : For one way of analyzing it is on a daily or weekly cadence It's good to look at that and see how things if there are any outliers or things that are looking a little bit strange but really you want to go one step further It's retention and long term retention and seeing our people sticking on the platform even if we're you know serving up a bunch of ads versus not serving up a bunch of ads do ads really hurt the platform as much as maybe some
Corey Zeman : Someone would assume but and maybe in all actuality it doesn't hurt the platform that much So it's also that long term retention And then on Roblox they had this third layer that I agree with more So and this is more of the discovery is like on Roblox how they talked about is are people discovering new worlds or are they just staying in the same world that they've always done Because if you're just staying in the same world at least for certain platforms you're not really in touch with that platform You're just in touch with that community But if that platform is
Corey Zeman : Showcasing you different things It's breaking you out of this filter bubble That's really the third and final layer that really gets people to enjoy the product I'm a little biased towards discovery but I think it's a very important part that people want to push out of their filter bubbles not all the time maybe not in with news and certain things but if you give them the right path and there's you know an infinite amount of paths if you give them that right path they will push out of those familiarity things you know because it is comfortable
Corey Zeman : to just I'm watching Frasier right now And I never thought I would say that It seemed like such a crappy show even back then but it is such a nostalgic nice view of things And that even that I discovered through Amazon but it is this familiarity but also getting pushed out to the discovery and even rediscovery is a type of discovery that I really try to push on because a lot of people forget about a lot of the songs that they've known for tens if not 2030 years
Corey Zeman : That they want to come back to but you know you can't remember all the songs and even all the artists that you've listened to So finding that nostalgia is a clear like mix of surprise and delight because it's you know like you're surprised by a song that you haven't listened to in 30 years but also delighted because it's something familiar
Kyle Polich : Could we talk about a like first start homepage experience in um areas you're familiar with What's like the problem that you have to solve and how do you approach it
Corey Zeman : Most of my work is on the homepage And that's really the key to it is again without friction getting people to start discovering without having to scroll to 15 different subpages and going to search or whatever else It's like you're dropped into it We want to help you find new content right away And that's been most of my professional work It really needs to be on the landing page There are other subpages like you know if you have a holiday
Corey Zeman : A collection of movies and TV shows You can definitely showcase that and it won't necessarily be on the homepage But really the home pages are bread and butter because it's trying to get people where they are and people don't want to browse They do want to discover but they don't want to have to go through tons of different content just to be able to find what they want I want to talk a little bit later about the future of where where these systems are going and where these where discovery is going but
Corey Zeman : I think I heard it from I definitely heard it from an exec at Spotify and I believe at Netflix talk about how the perfect user experience would be one click of a button you know like you just click that button it's 100% maybe not 100% but 99% of what you want And I do agree to that to some degree I think a lot of users just want the least amount of friction and just want to go right into it
Corey Zeman : But I think there is that other side maybe you know a a decent percentage maybe not the majority that do want to discover The problem with a lot of these apps is that it's just tile after tile after tile It's just kind of the same old thing and having it more of an immersive experience maybe not a metaverse
Corey Zeman : Quite yet but more of like an engaging experience where you have a Harry Potter collection And you know it's it's a bit more immersive and that you see you know the the characters and you can kind of interact possibly with them is more of where we're we're trying to push into because everything is pretty standardized and having these tiles and it's you got to push out of that and figure out how to make things a little bit more entertaining especially on an entertainment platform
Kyle Polich : Well I've seen this show up in a lot of contexts where that hero image or whatever 6 things are above the fold that's 80% of your clicks for obvious reasons That also means there's a lot of responsibility in populating those spots and uh in some organizations a lot of hesitation a fear of change Do you have any secrets to getting anything shipped
Corey Zeman : I think it's seeing other people's point of view because there are a lot of different organizations within an organization that have different viewpoints and as valid a viewpoints as any of my viewpoints and it's seeing their viewpoint and weighing against yours It's not like my viewpoint is 100% right and their viewpoint is 0% right It's either my viewpoint is 40% right
Corey Zeman : And theirs is 60% right and I was wrong and I should definitely move in their direction or it's you know some other play on that you know it's and it's seeing that point of view that if you don't see it in that way a lot of the times you won't get buy-in from other parts of the organization So you can't get these features or these initiatives pushed just because understandably like we all want to be represented in this you know and
Corey Zeman : There are different representations There's a business representation There's a user representation There's creator representations All of them are valid and they should be mixed to some degree But in certain viewpoints obviously for the most part the user is the winner But yeah there does need to be this balance between all of our different types of users whether it is marketing whether it's creators whether
Corey Zeman : Its users and balancing that is really where you get initiatives pushed out and makes for a better user experience Well
Kyle Polich : the barrier to entry to be a creator has plummeted I don't want to say it's gone to zero quite yet but you know I I have nieces and nephews that at 8 years old started YouTube channels and are uploading stuff Pretty much anybody can do this We have a deluge of content
Kyle Polich : How does that impact discovery
Corey Zeman : It's probably the best thing for discovery but it also makes my job a lot more difficult I talk about this a lot because the most important part of my job and what I need is quantity Quantity is the first and foremost thing then it's diversity
Corey Zeman : Then it's quality like everything but I need quantity first because quantity begets diversity quantity begets then quality at you know more and more so And if if you don't have quantity amount of data then you can't really represent what people want in a more holistic way The problem also with quantity though is it makes my job and just really people's job a lot more difficult and that
Corey Zeman : It's kind of this in shittification There's especially with AI content I've gone through Sora's app and uh and and and Meta's AI app And I'm first of all it really is just shitty content But more and more so it's not that I I I do think there is a place for generative content I found actually a very good amount that are very good quality But that's what generative content
Corey Zeman : It will do is it will make a lot more shitty content and as well some really good content that is a lot more personalized to people and just a lot more representative of people who don't necessarily have the skills to create these big blockbuster movies or even songs that would take even 1015 years ago
Corey Zeman : Hundreds of thousands of dollars to create It's it's a beauty that people with just an idea can do this but it does come at a cost and that there's just a flood of of shit to be honest with you And that's really where my job and people who work in discovery it's becoming more and more of an issue So therefore there needs to be more and more people who are looking into the data who are looking into
Corey Zeman : What are we even doing with Discovery Where do we want to have guardrails and where do we want to open it up you know and for the most part you know I'm not in moderation I'm not trying to remove anything off of the platform I'm more of the amplification and I want to see like OK this should be amplified because it it really is relevant to a certain community and so we'll target that cohort
Corey Zeman : So that they know that this new TV show is out Whereas if we didn't have like cohorts and we didn't have this personalization they might not know about it because it's all this popular data and popular as I said before is super relevant to a lot of people but not to a small amount of people And you want to serve both the masses and the individual
Kyle Polich : What are some of the limits you see in discovery today What would you like to have be it you know more compute more GPUs better metadata what are the limiting factors you'd like to see improved
Corey Zeman : Content quality or like metadata quality and quality of metrics and understanding because a lot of the stuff I'm just triaging if this data looks correct because if we don't even know like OK if it's a unique user versus a user that can show a lot of different results and I think a lot of that is just understanding of good dashboards and things like that the technical way of saying it but I think what discovery needs is more systems thinking approach and that
Corey Zeman : A lot of what TikTok and Instagram does is that it it shows you you know a minute like oh here's something on Fortnite and here's something on this and this and it's like it's not connective It's kind of just making people scatterbrained Whereas a lot of these systems I mean this is what AI does probably better than humans is connecting all of these things together so that you can have
Corey Zeman : A narrative and what you're watching or you're reading like I'm learning a lot about recommendation systems And that's how I found out about your podcast or I think that was it was actually through just my understanding of data but being able to understand where my gaps are and what I'm learning or even not even necessarily learning but you know Lord of the Rings or fantasy like where
Corey Zeman : the cool new worlds in fantasy Like I want to be able to have that kind of connective tissue that I don't see right now in a lot of recommendation systems It's very just short term gratification versus like what am I trying to understand Like even if it is just trying to disassociate and just relax like
Corey Zeman : I want to be able to disassociate with the content that I want best And I think having that historical data and being able to talk with generative AI to be like hey what should I be watching next that you know I've seen all of these things This goes back to just having as much data on people but also being able to trust that the system is not going to use it in a nefarious way is is really where the future lies But it's also a
Corey Zeman : A very tight rope because I don't know a one tech company that I would trust with all of my data because of just all of these ways in which they can use it that you know it's a slippery slope and being able to use it in ways that will help the business but aren't necessarily helping that user find something new or something novel
Kyle Polich : And when you think about the future of discovery are you expecting I guess more of the same the techniques and approaches known today or do we need a methodology revolution
Corey Zeman : I always go back to collaborative filtering and content-based filtering It does seem to be like how do we define the item and how does that connect to the user and how do we define everybody else And how do those people connect with each other And I guess it is this more holistic approach of not just looking at user to user or user to item but more holistically and I don't know necessarily
Corey Zeman : That what that looks like but something that isn't just this simple connections I see a lot of recommendations like you'll like this song based off of you know the song you've already liked And it's you know I want something that's completely different If I'm trying to discover And what I was talking about earlier about Discover Weekly I found out that Discover Weekly listening to a podcast it initially was for just completely new content
Corey Zeman : And then they saw that there was a bug that showed some content that people liked and that's what people really really engaged with And once they removed that feature you know all these metrics dropped and it's like well of course it's going to drop because discovery is very hard
Corey Zeman : I think of discovery is actually a type of friction but it's a good type of friction versus there's a lot of friction that's useless or not necessarily good And I don't mean good and bad and necessarily the common sense but it's good to have that type of friction of finding new things
Kyle Polich : Macquarie work and listeners follow you online
Corey Zeman : You can follow me on Instagram and TikTok on Silence NoGood or on LinkedIn uh at my name But there is also one thing I want to I do want to give a recommendation All right let's do it
Corey Zeman : I wanted to do something maybe in film music television but this is a little bit more applicable And there's a newsletter I consume weekly called Top Information retrieval Papers of the week Are you familiar with it I'm
Kyle Polich : not but it sounds
Corey Zeman : good It's really good It's what other than podcasts like yours being able to understand how recommendation systems work It basically gives you maybe 8 or
Corey Zeman : 9 different academic papers on machine learning on recommendation system sometimes out of that And then what I'll do is take it into Notebook LM have that synthesize it And basically what I do is create a podcast out of it And that's what Notebook LM is known for but also interact with it And if I have questions about what is the difference between collaborative filtering versus content-based filtering
Corey Zeman : You know these papers and obviously don't hallucinate as much And this is the beauty of Notebook LM is that it's only that subset of data that it'll look at it won't look at outsource data so that it won't hallucinate on what these questions are It might not know the question and it will say that but that is a really good resource for what I've learned with recommendation systems
Kyle Polich : Sounds good I'll put that in the show notes for listeners to check out as well Corey thank you so much for taking the time to come on and share your work
Corey Zeman : Thank you