The Vergecast - In search of the perfect movie recommendation

Episode Date: July 28, 2024

On this episode of The Vergecast, we look at why TV and movie recommendations are so complicated, and whether AI might be able to make them better. If Spotify can build infinite playlists of music you’ll like, and YouTube and TikTok always seem to have the perfect thing ready to go, why can’t Netflix or Hulu or Max seem to get it right? If you want to know more about everything we discuss in this episode, here are a few links to get you started: Movievanders Reelgood The internet is a constant recommendations machine — but it needs you to make it work Netflix’s Greg Peters on a new culture memo and where ads, AI, and games fit in From Scientific America: How Recommendation Algorithms Work—And Why They May Miss the Mark From Google: Multimodal prompting with a 44-minute movie Email us at [e-mail] or call us at 866-VERGE11, we love hearing from you. Learn more about your ad choices. Visit podcastchoices.com/adchoices

Support for the show comes from Retool. Too many companies run critical operations on duct taped spreadsheets, Slack workflows, and whatever else they could cobble together. Not because they want to, but because building internal tools means weeks of waiting on someone else's backlog. That's where Retool comes in. Build custom internal tools just by describing what you need. Prompts something like,

Build me a revenue dashboard on our Salesforce data. And Retool actually builds it on your company's data, in your cloud with enterprise security built in. Go to retool.com slash vergecast. We all need to retool how we build software. Welcome to the Vergecast, the flagship podcast of interest clusters. I'm your friend David Pierce, and this is the third and final episode in our series

all about AI in the real world. We talk a lot about AI right now, including obviously on this show all the time, but it somehow all feels really abstract and big and kind of jargony and technical most of the time. So we wanted to find some places where AI is actually showing up in our real lives. So far, we've talked about using AI as a tool to give everyone perfect memory, and we've talked with an artist who uses AI to make us think differently about privacy

and surveillance and life itself. Today, I want to talk about movies. And more specifically, why it's so hard to figure out what to watch. This is a problem incredibly familiar to any way. one who gets their entertainment through streaming, right? You sit down, you turn on the TV, and then you spend somewhere between two and two thousand minutes looking for something to watch. No matter where you look, it's just stuff all the way down.

So I opened the Netflix app, and the first thing it shows me is Mucho Mucho Amor, the legend of Walter Mercado, which I feel like, no thanks, probably not for me. No offense to Walter Mercado, I just don't know who that is. Then we get my continue watching list, some new stuff on Netflix, top ten shows, because I watched Girls 5Eva into that, binge-worthy TV shows. Today's top picks, for me, like eight rows down, it's giving me personalized recommendations. Then we get some romantic favorites, my list, quirky TV shows, goofy sitcoms, your next watch, critically acclaimed TV shows.

These all sound like the same thing. What are we doing here? I don't get it. Let's try something else. Okay, we go to Hulu, and Hulu starts by taking a long time to load. And then we get Dress My Tour. No chance. me. We got a bunch of live stuff. I don't really care. TV for me, which dressed my tour again.

A bunch of crime documentaries. That's my wife. House shows, sure. Then movies for me. Continue watching Emmy nominees because I watched how I'm at your mother. Top shows. Another recommendation for the same top show. Trending, newly added TV, full series comedies, newly added movies. These are like just things that Hulu has. None of this really feels like it's actually directed towards me at all. Okay, now we try Max. Max first shows me a CNN show and then Hard Knocks with the New York Giants that I will watch and we'll be very into. Recommended for me is up top.

Appreciate that. Continue watching. Today's picks, which is actually Game of Thrones and Justice League and Kite Man and a spy documentary. This feels like me. Then a bunch of top tens, live, true crime, designing home, just at it. Okay, this feels closer to being me, but it's still just kind of like. here's a bunch of stuff that we have on Max.

That's the vibe here. Like, do you want to watch this? Here's some things. Like I said, it's all a lot of stuff, but it doesn't feel like anything here is just telling me what to watch. And actually, this seems to me like the kind of thing AI ought to be able to do really well. It's low stakes in the sense that if the model doesn't recommend the objectively perfect thing or if it even makes up a movie because these systems tend to make things up a lot,

that's not a huge deal, right? I just want to be able to tell a chatbot what I want and get some recommendations. To be fair, there are actually lots of ways to do this already. You can do it with ChatGPT. What's a good movie that's kind of like Mission Impossible or The Born Identity, but isn't either one of those movies? You might like Skyfall. It's a James Bond filter.

I've seen all the Bond movies. In that case, you might enjoy the man from you in CLE. It's a stylish spy film with plenty of action and a good sense of humor, similar to Mission Impossible and the Born Identity. There are also a bunch of plug-ins in Chat G-PT, the custom GPTs, that are explicitly designed for this, though they're mostly just kind of better formatting on top of the same answers. All right, I'm in Movies GPT, which calls itself a go-to movie encyclopedia.

So I'll just do, I just watched Crazy Stupid Love and loved it. What's a similar movie I probably haven't seen? Says, if you enjoyed Crazy Stupid Love, you might like the five-year engagement. This romantic comedy drama directed by Nicholas Stoller stars Jason Seagel and Emily Blunt follows the ups and downs of a couple's prolonged engagement, blending humor with heartfelt moments. The film explores relationship dynamics in a way that's both touching and funny, similar to Crazy Stupid Love. It's a pretty good answer. I'll take that. You can also use other AI systems like Google's Gemini.

What are some classic 1980s action movies that I need to see? The 1980s was a golden age for action movies, churning out some of the most iconic characters and quotable lines in cinema history. Here are a few classics you can't miss. Die Hard, 1988. Bruce Willis redefined the action hero with this game change. Okay, yeah, we don't need to hear all that, but it recommends Die Hard. Raiders of the Lost Ark Terminator, Predator, Beverly Hills Cop

and then it says this is just a small sampling of the many great action movies from the 1980s. Still, not bad. I don't know that I learned any new movies from that, but it's a decent start. These models will happily give you movie recommendations and you can ask for them in pretty natural language, which is cool.

So you don't have to just search for a movie or like a genre. You can actually say, I want a romantic movie that's under two hours that is funny and not sad and you get... How about crazy stupid love? It's a romantic comedy that's funny

and heartwarming without being set. I just watched that. Great recommendation, but can you give me a different one? Sure. How about ten things I hate about you? It's a funny and charming romantic comedy and it's just under two hours long. It's a modern take on Shakespeare's The Taming of the Shrew

and has a lot of great moments. Not bad, right? It's not exactly groundbreaking new title ideas there, like crazy stupid love and 10 things I hate about you, you'd find in one Google. But they're good recommendations. And there's something important in that interaction, I think, that is actually a signal of why AI might be a big deal here. So for years, I've been asking people about recommendations.

And for years, I've gotten roughly the same answers. Recommendations are hard to do really well, really, really hard. And the most effective way we've found over time is to basically say, okay, you like this set of things, which we know because you've either rated them highly or watched them a bunch of times or done something else that makes us confident that you like them. Maybe you gave it a five-star review. There are some other people out there who also like that same set of things. Oh, and they also like this other thing that you haven't seen. So we bet that you will like that thing too. It's simple. It's how recommendations have worked forever, but it does kind of work. But AI has the potential to approach

the problem really differently. So there are two components that generally AI seems to be like a good tool to solve, right? That's Diana Puppescue, a developer and entrepreneur who recently started an AI movie recommendations tool called Movie Vanders. We are pretty fans of Harry Potter. Right? And we said, we were joking, saying, you know, the movie has to choose you. It's not only that you are choosing a movie, right?

You know, so we said, oh, that sounds like the, all the vendors from, you know, the wound is choosing you as a wizard. So we said, okay, magic plus that kind of concept. Movie vendors sounds really good. This is how the name show up. I love a Harry Potter name. But back to the two components. The first thing AI does really well is just ingest and organized data. So, so, so much data.

All the genres and actor names and everything that you'd normally associate with the movie, but also all kinds of other stuff. Every feedback that people is maybe giving in, I don't know, Reddit, Twitter, Facebook, whatever, like, you know, social media, we give more information than we think, right? Every click we are doing, it's information that goes somewhere and it's somehow interacting with other information and producing some kind of behavior that could easily be incorporated into an artificial intelligence to learn from. The second component is that AI can begin to structure that data for you and turn all those articles and reviews and tweets and Reddit posts and everything else into something kind of like a usable database of information. So if I'm saying, I want to see a movie that has a dog and it's not sad, please,

because I'm a very sensitive person. And all these things somehow that is interacting with the information that is already sort of flying in Internet. and how those generative AIs are working, we are not exactly sure. The reason Diana and others are excited about AI for making sense of all this information about movies and shows is that otherwise it doesn't really exist. And to explain what I mean, we have to talk a little bit about data and metadata. As I've talked to folks in this space over the years, I've learned that there are basically four useful kinds of information about a show or a movie. The first is the metadata, which is just information about the thing itself.

Its name, who's in it, who directed it, the poster. That's the sort of stuff you can find on a search with IMDB. The second thing is watch data. The single best signal someone can give you about how they felt about something is how they watched it. Did they watch it all the way through? Did they turn it off right away? Did they watch it every night, 12 nights in a row? Did they get halfway in and then bail?

That's all really, really useful information. But that information doesn't exist at any real scale, at least not for streaming services. And it definitely doesn't exist publicly. Netflix knows what you do on Netflix, Hulu on Hulu, Max on Max, Peacock on Peacock, but they don't know about all the others. And watch data, even those lists of most popular titles that you see from these services sometimes, is as a result a very closely guarded secret. It's crucial data and no one wants to share it. The third thing is just stuff to recommend. This is sort of an underrated one, honestly, and not really an AI story at all.

But the recommendations you get are only going to be as good as the stuff there is to recommend, right? Even if you know exactly what I'll like, it only matters if you have that thing to offer me. And most streaming services have a few thousand shows and a few thousand movies, and honestly, sometimes not even that. And in theory, even if someone did have the whole available library of TV and movies ever, it's not a very big library. IMDB's library includes 685,000 movies and 265,000 TV shows, which is a lot. It's more than you'll ever watch in your life, but it is nothing compared to the amount of content being uploaded to TikTok and YouTube and Instagram reels every single day. The fourth piece of information about movies and shows that matters, and maybe the hardest, and maybe the hardest, and maybe. the most interesting for AI is what you might call traits, deep information about the thing itself.

Not just who's in it or whether it's a comedy or a thriller, but one or two or three levels deeper than that. Is it a sad movie or a happy movie? Does it move fast or does it move slow? Does it take a while to get into or does it capture you right away? Is the soundtrack awesome or does the soundtrack kind of suck? Is it a good movie to watch while you're looking at your phone? Is it a good movie to fall asleep too on the couch? Is it a movie you should only see in theaters? This stuff really matters, and it's information about the title that is hard to know unless you really understand the thing itself. I think the easiest way to understand what that might look like in practice is through music, which has actually gotten a lot of this stuff really right. Back in 2021, I interviewed

Gustav Soderstrom, who is now the co-president of Spotify, about the company's recommendation system. He actually described the whole idea of it kind of perfectly. If you actually have non-real time, if you have time shifted on-demand content, then you have the benefit of actually being able to machine listen to the content, transcribe it and understand it. Then the problem is actually quite similar to text recommendation if you think about it. Then you have a few things. One big difference from music. In music, we had these four and a half billion playlists where people have sat and manually organized the world's 70 million tracks into like how they go together. There is nothing.

like that for podcasts and not even on the competitive platforms, right? So we kind of had to bootstrap it somehow. And so what we started with was to build a what is called a knowledge graph, quite similar to an artist graph. Like this artist has this discography of these albums with these songs and, you know, belong to these genres. So we started building a knowledge graph, a traditional knowledge graph quite early on before we had any listening data to bootstrap it. And then we started adding listening data to that. So we could see that people that listen to this also listen to that. So we started with actually a pretty simple system of a sort of hard-coded graph. And then after a while, now that we have lots of listening data, we started inferring with more

machine learning-based techniques. Both the quote-unquote traditional collaborative filtering, people who listen to this also listen to that, which is kind of what the playlist were about. People who playlists this also playlists out. But we also started adding much more advanced techniques like actually machine listening to the context and summarizing them. so that we can do things like podcast, topic search. And there we're actually just, we're not just looking at the headline. We're actually listening to the content or listening, we're reading the content, is transcribed, summarizing it.

And then we put it into one of these embedding spaces, as they're often called, to understand how it's similar to other things. That's the magic right there. That combination is where everyone seems to agree the best stuff happens. And when Gustav talks about machine listening, by the way, he's talking about technology that Spotify has that can actually break a song into its component parts automatically. So with that technology, Spotify's computers can recognize

that the song is 110 beats per minute. It has a driving guitar line, there's a violin, a female lead singer, angsty vibes, and lots of other things. Tons and tons of different categories and classifiers for what this song actually is. And that actual understanding of what something is at a content level is really important in getting recommendations right.

And that just hasn't existed in video before. It hasn't really even been possible. But then just recently, something changed. We got to take a break and then I'll tell you about it. We'll be right back. Support for this show comes from Shopify. Every thriving successful business has to start somewhere. A good place to start is a relatively simple question.

What if, given the right tools, I really put my all into this. One tool that can help grow your sprouting business to new heights is Shopify. Millions of businesses around the world rely on Shopify for e-commerce. They offer a host of helpful tools you can take advantage of, from payment processing to analytics to website design. Their design studio includes hundreds of templates to help you create the exact website you've been envisioning for your business. If you're wondering, what if I need help? Then no worries, because you're never left to fend for yourself. Shopify's award-winning customer support is available 24-7.

It's time to turn those what-ifs into a thriving business with Shopify today. Sign up for your $1 per month trial today at Shopify.com slash vergecast. Go to Shopify.com slash vergecast. That's Shopify.com slash vergecast. Welcome back. Let me tell you about an AI breakthrough. So in February of this year, Google launched its new AI model. It's called Gemini 1.5. Gemini 1.5 was all the things that new AI models usually are,

faster, smarter, does well on benchmarks, all that good stuff. But its big new feature was a new context window, which basically refers to the amount of information that the model can take in and consider at a time during a single query. So like one token is a few letters. Gemini 1.5 could do a million tokens. That's like a book. You could just feed a book all at once. Or as in one demo that Google showed, it's a movie. The company demoed Gemini 1.5 with a 45-minute Buster Keaton movie called Sherlock Jr. The movie came out to 696,417 tokens, if you're curious, and just listen to how the demo goes. In Google AI Studio, we uploaded the video and asked, Find the moment when a piece of paper is removed from the person's pocket and tell me some

key information on it with the time code. Then the whole thing processes for a minute, 57 seconds to be exact, and it answers. The model gave us this response, explaining that the piece of paper is a pawn ticket from Goldman and company pawnbrokers with the date and cost, and it gave us this time code 1201. When we pulled up that time code, we found it was correct. The model had found the exact moment the piece of paper is removed from the person's pocket and it extracted text accurately.

I remember the first time I saw this demo and being totally blown away by it. This is an AI model processing an entire movie for the first time and identifying one thing inside of it in under a minute. There was one other demo in there also. Next, we gave it this drawing of a scene we were thinking of and asked, What is the time code when this happens? The drawing is like a crude stick figure drawing of what I think is a water tower, big round thing on two legs, person below getting sprayed by water. It's either a UFO abducting someone or a water tower.

Anyway, here's how it goes. The model returned this time code, 1534. We pulled that up and found that it was the correct scene. Like all generative models, responses vary and won't always be perfect. But notice how we didn't have to explain what was happening in the drawing. Simple drawings like this are a good way to test if the model can find something based on just a few abstract details like it did here. This is obviously just a tech demo, and most things are not this easy and this perfect.

But the possibilities with that idea are huge. Suddenly, we're heading towards a world in which you can upload a movie to an AI model, and that model can start to understand things about it. Granted, identifying a water tower is easier than figuring out the mood and vibe and pace of a movie, but this is still a big leap down that path. It is, by the way, a leap filled with copyright violations and complicated legal issues, but just for the sake of our episode today, I'm going to leave those to the side for right. now. So much of what's coming in AI is going to be decided in court. But until that happens,

this is what we're going to get. And this is where we come to the biggest, maybe thornyest question in the whole AI recommendations universe. What actually makes a good recommendation? And actually, no, there's a question even bigger. Why do you like what you like? There's an example I always give for that one, which is the crown. Like the first couple seasons of that, like, it was a great, I don't know why, but I loved it. It was great. But then, like, you know, I noticed on, like, whatever streaming service it was on, like, it started to recommend, like, British shows to me, which like, or British royalty, which, like, I have no interest in British royalty shows. There was something intangible, I don't know what, about the first couple of students of the crown that caught me. That's where it's trickier. That's David Sanderson. He's the CEO of a company called Real Good. Real Good is one of those products that tells you where you can stream a given show or a movie. But it's also becoming a really important. data provider in the streaming industry. Like I mentioned earlier, there just is no good universal source of data about TV shows and movies, what they are, who's in them, who made them, and where you can find them.

I know that sounds nuts, and it is, but it's true. That data just doesn't exist. So Real Good has its own consumer product that you can use, but it's also the database behind a lot of the other streaming search providers out there. But think about that example, The Crown, or any other show or movie you like. do you like it because of the star? Maybe, and maybe that means you'll watch anything with that person in it. That is an easy recommendation problem.

But how do you account for someone liking Breaking Bad but not Better Call Saul? Or being super into selling sunset, but really only liking the first season. If you liked Game of Thrones, does that mean you'll like House of the Dragon or any other show that has dragons? Or was it the palace intrigue that you liked? Or just the naked people or one of the actors, but not any of the other actors? or did the thing with the Starbucks Cup make you like it, or something else entirely? Like they say, there is no accounting for taste.

But frankly, I hate that, and I want AI to fix it for me. So Real Good has been experimenting with AI a lot recently, mostly as a way to try and explain whether you might like something. It's less proactive recommendations and more a way of you saying, I'm interested in this. Do you think I'm going to like it? In their app, you go to a title and you tap on the icon, and it takes a couple of seconds based on what it knows about you to analyze whether you're likely to enjoy it.

All right. So I'm in the Real Good app, and one of the recommendations it has for me is Eric, which I've never heard of before. Vincent, a grief-stricken father whose son goes missing, find solace through his friendship with Eric, the monster that lives under Edgar's bed. Sure, that's something. It looks like Benedict Cumberbatch is in this. Anyway, I hit the button that says, should I watch this? And it says, it's analyzing it just for me. And it says, with its low IMDB audience score of 6.9 out of 10 and real good score of 79 out of 100,

Eric may not be the best fit for you. However, this shows crime and drama genres align with your interests and mystery and drama. This shows suspenseful storyline about a father's search for his missing son could keep you engaged, but the monster under the bed element might not be to your liking. And then if I hit a button that says suggest similar, it just shows me a bunch of other things that I may or may not like, including The Bear and Presumed Innocent and Sunny and House of the Dragon, all of which I very much do like.

So maybe I need to watch Eric. When you think about the four parts of recommendations, metadata, watch data, stuff to watch, and traits, Real Good is a really interesting one. It has lots of metadata. That's the database it's been building for years. It has very little watch data because Netflix and the others don't share that. Real Good and every other platform like it, Just Watch, and likewise, and all these other recommendation systems are desperate for you to tell them which shows you like

and which ones you've seen, so that they can try and back into some of that watch data, but they'll just never have it like Netflix has it. On the flip side, Real Good has way more stuff to watch and more things to offer you, because its library includes lots of streaming services. And when it comes to traits, actually understanding the content itself,

well, that's complicated. Here's Pablo Alessia, who runs engineering and data for Real Good. It's not the same thing as looking at a genre or a tag, right? It's a completely different problem. It's a way more holistic view of looking at, that, okay, what are the archetypes? For example, I've seen some analysis, like, that's, like, made by actual people, like,

analysis of movies and shows where they're, like, looking at the archetypes and, like, the psychology of them, and that's really maybe the reason why we'd like movies, we'd like the archetypes of the characters, I mean. I would say LMs are probably the best equipped at looking at problems of that because you would need to look at such a breadth of, like, knowledge at the same time to be able to, like, do that analysis that humans do,

but I would say that we're pretty far away from getting to that level of analysis. Even in the best, most powerful scenario, Publis says he's skeptical that we're going to be able to teach LLMs to completely, deeply understand movies and shows and also understand why people like them. I know just people that do. LLMs are the way humans think. I don't think that's the case, but let's say it is.

Let's assume that for a second. We're still not like LLMs don't perceive the world in any way, shape, or from close to the way humans do, because we're giving them data in such a different way. and also the data that we give them is so limited and lacks so much context. Even if you give them a thousand people, they're still individual, a thousand pieces of people.

Just imagine the, like, think of the way you watch movies and shows, right? It's such an emotional, entertaining, right? You're, like, going through feelings as you're looking. It also depends on who you're watching the movie with and all this criteria that the LLM is not, like, it's not that it needs to be smarter. It's just like lives in a box, right? Versus us who live in the actual world, right?

So there's that part of the problem as well. that makes a lot harder for her to get to that point. Paolo also said he wonders how deep it's actually useful to go down that rabbit hole. Like if a model knows that you like a movie and there happens to be a bird in the shot, four minutes and six seconds into the movie, is that why you like the movie? No, obviously it's not why you like the movie. That level of detail is basically totally irrelevant for our purposes here.

That is such a solution in search of a problem. but if level one of understanding is just title and genre and cast, there's maybe a level two that might be just as useful. There's no right answer here, right? This is like the one we use. But the one that we're looking at, and I think is the second level, is the mood of the people of content. So if this is like in what mood the person is in when they want to watch this piece of content,

that would be the second dimension and the one that we're most actively looking at right now because it's, I mean, somewhat solvable, right? Like, you can go down the rabbit hole of, like, the archetypes of each character, but are we in a place where we can solve that quickly? Not at all, versus moods, obviously. It's something much easier to actually tackle. So I would say that's the next dimension. So you go genres and taggers, which are the ones we all know, like drama,

serial killer, that sort of thing. So one genre, ones are tagging, and the next one would be the mood the person is in when they're watching that piece of content. Notice that's both about the mood of the content itself and your mood when you watch it. I didn't really thought about it until Pablo mentioned it, but so much of the equation here, and really the reason recommendations are so hard in general, is that whenever I sit down on the couch to watch something, my mood changes everything. Yes, I like spy movies, but I don't

always want a spy movie. And yes, I like historical documentaries and silly rom-coms, but God help you if you try to get me to watch a World War II documentary on like a Tuesday night when I just want to fall asleep watching a food show. You know what I mean? I feel like it's possible. that recommendations will never be perfect and that they'll never know the exact right thing every time because I don't want the same thing every time. I'm human and the computer isn't

and that might always be a problem. So where does all that leave me in my quest to just find something great to watch? I'll get into that right after the break. Support for the show comes from Upwork. The days of doing it all, all by yourself, are over. There's no romance and burning out while you're trying to scale.

Instead, you can check out Upwork. Upwork helps grow your business by giving you fast access to specialize talent across more than 125 categories so you can fill skill gaps, launch projects faster, and scale without committing

to full-time headcount. And finding the right talent is easy. You can browse profiles, review past work, and get help scoping the role so you can get started quickly. Seriously, you could connect with the right freelancer

in just a few hours, especially when you sign up with Business Plus. Their AI powered shortlisting pairs you with the top 1% of talent in under six hours. No endless searcher required. You can visit Upwork.com right now to post your job for free. That's Upwork.com to connect with top talent ready to help your business grow. That's UPWRK.com. Upwork.com. Okay, we're back.

Let's go back to MovieVanders, that AI. movie recommendation service we were talking about, just for a quick second. I asked Diana Poppiskew how people are using the service so far. And she said there are basically two things people are doing. One is maybe you have already a movie in your mind, something that, some movie that you already seen. This has happened to me very often. And I know a lot of people like to rewatch movies that they've seen maybe years ago. And they've enjoyed very much, but they don't remember the name. They do remember certain elements of the movie like, like you said, right? It happened in New York.

It had like the main character and stuff like that, right? So they put that and find me, find that movie to me. So they're looking for one specific movie, but they don't know what it is. Exactly. They don't remember exactly the name, et cetera. So that would be the find me option. And then there is a recommender is like, I don't know exactly what I'm looking for. There's no specific movie that I want to see.

But I would like to have certain elements in it. You can be as deep as you want, right? You can give the better you formulate your request, the more specific is going to be that request. The better is going to be the outcome. So you don't have to fear that if I'm giving too many details, I'm not going to find anything. Because eventually you're going to be more accurate.

If that exists, of course. I mean, you can search for some, like for instance myself. I love because I have a small kid and I like Cars movie. So I was thinking, I like cars stories, like pretty good. But I don't like the fact that the characters are cars. Is there any movie that has that kind of story, but without cars, like with people? And actually the outcome from movie vendors is pretty good. And I was like, oh, yeah, yeah, sure.

I want to see all those movies. They were explaining, like, what's the plot? And it was like, hmm, that's true. It is indeed. So that's a very deep type of questions. Like I want this story, but not with cars, with people. And I was surprised to see that level of understanding of the artificial intelligence to the point where, I don't know, it did some comparison between plots and said, oh, there are common elements between these two movies. That last thing is pretty impressive, actually, and very telling.

If you go to Google right now and you just say, what are some movies like cars but with people, you're either going to get nothing use. or like a link to the movie cars. It'll probably be the movie cars. But the idea of AI is that it can comb through all the reviews, all the synopsies, all the tweets, and say, oh, this movie is talked about similarly to the way that people talk about cars, but it doesn't seem to have any cars in it. Maybe that'll work. That's the kind of thing that no system other than AI can do really well and do at scale.

In the first case, Diana mentioned, AI models are already actually pretty great. Frankly, for the same reason, because they've ingested all these recaps, reviews, and tweets and more, they can do a good job of guessing what you're looking for just by synthesizing information from all over the web. So when you're like, what's the Jim Carrey movie with the boat and the TV show? It gets to the Truman Show really fast. That's not recommendations so much, but it is pretty handy. Personally, I've been using these AI tools for more than a year to try and find stuff to watch, and I found a few tricks that really work.

The first is the show me stuff like option. I'll add three movies I like and tell the model to find other ones like it. That works great, which actually makes sense. These models have ingested all those lists of Best Spy movies or whatever and can easily compare the summaries that you find on IMDB, so matching things to like things is actually pretty easy. But the problem is it'll usually spit back really obvious examples. If I like Back to the Future, everything everywhere all at once,

and so I married an axe murderer. What other movies do you think I might like? You might enjoy Bill and Ted's excellent adventure, Hot Tubb Time Machine, and Scott Pilgrim versus the world. So in that case, what I'll do is add but only recommend lesser-known titles, or I'll ask only for underrated ones. And that usually weeds out some of the most obvious answers.

If I like Oceans 11, the gentleman, and he, what are some underrated movies that you think I might like? You might enjoy layer cake, inside man, and the town. These films have strong heist... I've seen all those. Give me more underrated examples. Sure. How about a simple plan? The bank job and the way of the gun. These movies have gripping stories and clever twists that should appeal to you. Those are good, thank you.

The other thing I do now is try and ask just for the vibes that I'm looking for. Like, here's one of my go-to prompts these days. Can you recommend a good rom-com that's ideally under 90 minutes and doesn't require much brainpower? Absolutely. Here's a perfect rom-com for you. Win a date with Tad Hamilton, 2004. It clocks in at a breezy 84 minutes and is a light-hearted watch. It follows Rosalie, Kate Bosworth, a small-town video store clerk who enters a contest to win a date with the megastar Tad Hamilton, Josh D'Mell.

Is it any good? Win a date with Tad Hamilton is a classic rom-com that's more about charm than plot. It has predictable elements, but is still enjoyable for its lighthearted humor and performances by Kate Bosworth and Tofer Grace. It kind of works. I don't know if these are like the platonic ideal of perfect recommendations, and a lot of these have been terrible movies. But I've had good luck with this so far. And as I said earlier, I like this use of AI because it's so low stakes. If it recommends a movie I've seen, I just tell it. I've seen it and ask for another one, and actually adding that data to my search often gets better results. If it recommends a movie that doesn't exist, which, to be fair, has only ever

happened to me once or twice in the last year or so, that becomes obvious very quickly, and I just move on. I found a lot of good stuff to watch this way, and it's taken a lot of the aimless browsing out of my evenings, and I really appreciate it just for that alone. What it doesn't do well is match to the just watches and real goods and this, like, what service can I do it on? piece of things, so I have to get a recommendation and then go to another thing to find where I can actually watch it. But those are easy problems to solve. And things like movie vendors are actually starting to put those two pieces together. But after talking to lots of people about this, do you want to know my number one best recommendations recommendation? It's to do as much

of your watching as possible on as few services as possible. Because the thing I heard from every single person that I talked to is that the best predictor of what you're going to watch next is what you've watched before. If you want great recommendations, you need a watch history. This is why YouTube and TikTok have such better recommendations, because they know what you've watched. You watch a lot. You watch it very quickly. And these systems are able to just build this internal flywheel of what you do and don't like in a way that is much harder when you're sprinkling, viewing across 10 different streaming services. If you have the energy and diligence, which I frankly don't, maintaining your profile on something like Real Good or Just Watched or Letterbox

where you can keep track of everything you've watched everywhere is also an excellent path towards good recommendations. That stuff serves itself. The more data you put in, especially with these strong signals of high reviews and saying you loved something and all that stuff, that matters a lot. I think eventually we might get AI that understands us as people

and understands movies and shows in a deep and rich way and does, in fact, recommend the perfect thing or at least the perfect handful of things every time we turn on the TV. But I get the sense that that's still a ways off. The Gemini 1.5 demo is a demo. That's not how we're going to start querying movies anytime soon, I don't think. For now, if you want Netflix or Peacock or whoever to find you something great,

the best thing you can do on those services is watch stuff you like, all the way through over and over and over and over again. It sounds pretty low-tech, but I got to say, it's not the worst homework in the world. All right, that is it for the Vergecast today, and that is it for our AI miniseries. Thanks to everyone who is on the show, and thank you, as always, for listening. As always, if you have thoughts, questions, feelings, or movies that you want me to watch, because that is still the best recommendation system out there, you can always email us at Vergecast at theverge.com, or call the hotline, 866, Verge11.

We truly, truly love hearing from you. This show is produced by Andrew Marino, Liam James, and this episode was edited by Zander Adams. The Vergecast is a Verge production and part of the Vox Media Podcast Network. We'll be back on Tuesday and Friday this week with our regularly scheduled programming. I'm actually out for the next two weeks, but we have some really fun stuff teed up for you. And also there is just a lot of news going on. So keep it locked. We'll see you then.

Rock and roll.