How vertical soap operas are making $10 million a month

0:00In this episode of the DeNoise podcast, we're going to break down vertical soap operas. H &M using AI versions for its models. And if Anthropix Claude model is actually thinking. Let's get into it.

0:13Hey, Addy, good to see you. Good to see you. How was your weekend? Good weekend. Another week, another AI model launch. We're not going to get into it this time, but I saw this morning that Runway launched Gen 4. Nice. I mean, it's sad to say at this point, but it's like, I was looking at it like, okay, it's going to be better. Bigger, better, faster. It looks more realistic. But I don't know what specific updates, if it has any more camera control or actor control, but outputs looked really impressive as usual, which is sort of now the like standard as like a new video gen model comes out. You're like, yeah, it looks really good.

0:37Yeah, and you're like, oh, that's better than the last previously amazing model. Yeah. Right, so there's like an arms race and the iterations are like weekly. Yeah, I also wonder too, is the timing with it coming out before NAB, which is like a big like camera industry announcement coincidence or just be like, oh, give something. Yeah, I don't know if runway is even aware of NAB. that's that's our world for sure but uh you know up in the bay area in the ai developer crowd not sure if it really resonates last year was when sora first like the clips were teased and that was like the talk of the show was that only a year ago yeah oh my gosh the first the spaceman clip and all that stuff yeah and that lady walking through the tokyo street yes yeah and i remember seeing that for the first time i was like wait video can be generated too like it was such a light bulb moment.

1:25I was like, oh, it looks good. I think I first saw it and was like, that's not real, right? That's just a clip. Or a sample clip. I was like, nope. That was generated. Especially the lady was wearing a leather jacket and the physics on the jacket was shot. Yeah, shot on the train and you could see the reflections of the model in the reflections of the train. Yeah. It was like, oh, wow, that's a big moment. Only in 12 months. Yeah, now we're just like, yeah, whatever. Now we're like, Sora's the mid-model. It's not even that good. Exactly. Oh, yeah, the random story too remember when we were like talking about grok and i was like oh did you know that xai and grok is technically not part of x it's a separate company yes and then turns out well now it is but uno reverse xai bought x so now x formerly twitter is owned by xai the ai company do you know what the reasoning for that was i think it wasn't because the x stock was taking a dump or something to like i thought i think it was just like financial yeah it's just shifting around funny money situation yeah that's what it felt like to me that's right yeah Yeah, that's what I thought.

2:23All right, let's get into actual stories. Okay, the first one. So when I was at some holiday networking industry events, and I was talking to some people that were like on the camera camera teams, and kind of like just the vibe in Hollywood in general, just like, are you working on shows? And if you are working, you're pretty lucky to be working on stuff. Because not a lot of stuff is filming right now. But the general thing I kept hearing of people who were working on shows, like three different people, it was like, yeah, I've been working on these vertical soap operas. And this was the first time I heard about this.

2:49and I was like thinking like oh like vertical but are they like filming like open gate or something wide and they'll crop it but they have like you know the wide shot to like repurpose for other things and they're like nope we turn the camera sideways we recorded vertical it's just vertical that opened me to a world of digging into what are these things yeah and it is a big industry now yeah it started in China I believe shortly after COVID their version of TikTok it's tiktok ish but these are like produced 90 second stories with a hook but then to continue the story you got to pay up so it is a soap opera e-format with kind of crazy stories of like the usually fall into genres of like werewolves or like rich tech billionaires or it's all the tropes that are in like romance novels yes or that were on soap opera like novellas and stuff yeah yeah and so there are a handful so this has taken off in china it is the vertical series market in China exceeded$5 billion in 2023, with projections suggesting it could reach$13 billion by 2027.

3:46I'm not surprised that for a second, because in China, the vertical content platforms have direct, like Amazon, like shopping built into them. So it's not just ad revenue, it's like product sales revenue. And it's a lot less friction. Yes, because it's a mega app, or I forgot what they call it. But it's a type of app that we don't have here, because there's too many monopolies and paywall. Yeah, we have a fragmented ecosystem where there's more unified because the state probably controls the whole thing. They've adapted the model here. The big company here is Real Short, but there are a handful of other companies, Drama Box, Short TV, Serial Plus, Flex TV.

4:19And the model is basically, you get an, I've seen them now as some ads. You get an ad with like the first 90 second clip that is very much designed to hook you. It's a very kind of, you know, cheesy soap opera-y thing, you know, woman in distress or like, she doesn't know that, you know, he's really a billionaire. And like, you know, the guy's getting treated like crap, but he's like, oh, secretly a millionaire, you know, some kind of crazy story that gets you hooked. And then if you want to continue the story, you got to pay up. And you can sometimes pay as much as like$40, I believe, to continue the story, which I think the total per month or per content, I think it's either way.

4:52I think some of them work off like a token system. So you kind of buy credits, and then you can choose which stories you want to continue. Users typically purchase digital coins or passes to access additional episodes with the cost to watch an entire series ranging between$20 to$40. Many viewers end up spending$10 to$20 weekly or up to$80 monthly. So you pay up to watch the rest of this story. It reminds me of games. And I couldn't find info on this, but I believe I talked to someone and they believe that the people behind some of the people that developed this, their background is in mobile gaming.

5:22And that is all about building out these ads. They've like perfected the digital revenue where you see the like kind of play in the game thing and they get to hooked in and they know their customer acquisition costs and how to turn it around and get you addicted. Same exact psychology, same exact backbone, but with video instead of a game. Yeah, in a lot of ways, that's way easier, right? 2D video as a infrastructure pain point is a much smaller pain point than like rolling out a 3D game and updating the characters and making sure it's running on all devices. Video will way more efficiently run on all devices.

5:54it's a brilliant move on their part and also the fact that they're not charging you real money they're charging you some digital currency which has probably a weird conversion rate so you don't know how much it's hard to do math in your head is that two dollars or twenty dollars i don't know i don't know i'm just gonna buy it because i really want this right because if you're like would you pay twenty dollars to watch like a soap opera episode you'd be like no what twenty bucks for a soap opera but if it's like oh your credits and the thing and it all renewed i really want to know how this ends i'm gonna spend uh twenty thousand credits yeah yeah a lot easier left i I mean, I'm kind of rooting for this because it's like, hey, production's happening.

6:25People are shooting this stuff. Obviously, they're cranking it out a lot faster than a regular TV show. I couldn't get any numbers on like budget or production time frame. I mean, I think it's in the range of like you could shoot a whole series, like an hour long content in a week, probably in the range of like$100 ,000. That was some numbers I heard. Any virtual production? Any volume? I feel like that's what something that could work with this if they were. I think it's a combo, too, of they produce some in-house and then some, you know, work with vendors to produce additional content. I'm sure there's a whole science methodology of the scripting.

6:55Yeah. And the teasers and the hooks. Yeah, that's probably very formulaically written out and very templatized. I would imagine that's in-house because that's got to be very specific. And then the production aspect, either do it in-house or source it out. But yeah, this feels like something where if they were doing it in-house and they just built some virtual production studios, like that would speed up. I don't like these guys are so lean and they're from the gaming world that I don't think they even know the power of virtual production yet. Like it hasn't, the sphere hasn't sort of caught up to them.

7:25Yeah, I mean, maybe not. And then the crazy thing too, is what they're making in an article. Rolling Stone had a big article about them. And so in Real Short, which is one of the big ones, founded in 2022, they've had more than 30 million downloads and they generate over$10 million in revenue a month. Wow, that's a big business. Yes. Do you know if this stuff is shot in and around LA or elsewhere? I'm sure all over, but I think a lot does happen in LA. I wonder why the numbers work for them in LA if they're just able to get lower studio costs. Probably non-union people. Non-union, easier actors.

7:57I mean, there's a lot of actors here not working as much. The concentration of BUC list actors here is incredibly high. Yeah. We talked about this before. The talent is here. Everyone is here. You can easily pull together. Also goes back to my vacuum principle, which is anytime you have a vacuum of work in a resource rich environment other work will kind of kind come in and suck it up yeah and this could just be taking advantage of like well everyone's kind of looking for some stuff to do and so you can punch above your weight and get some like talent that would probably be booked elsewhere but they're not and so you can bring them on board i'm sure some like you know pro capable dp will work at half or a third of their rate because they're not working right now they're like hey we're getting amazing production quality i wonder if you i mean you got to make sure you have a dp that also like knows how to move because it's like you could get a flip the camera vertically and that too i thought you can get a big tv dp but if they're like used to you know big sets it's like sometimes it works counterintuitive to like what you're trying to do if you're like no we need to move we need to move yeah yeah we throw put yeah and so also obviously this reminds us of the attempt to do this that did not go very well could be rest the piece could be yeah you know that jeffrey katzenberg was always at least five years ahead in all of the big things he's done other than dreamworks he you know helped develop pitch which was a video sharing platform with high social interactions which then it was too early because the phones weren't able to run it at the time and the internet connections were not up to like video levels at the time but then just a couple years later vine yeah came out and vine lived and went and then TikTok, right?

9:33But Jeffrey was like five years ahead. And with Quibi as well, Quibi was being concepted in 2018, 2019, came out during the pandemic as we covered in one of the episodes. It went and now we're seeing that format and that platform all over again. Yeah, I think there's some things that they got wrong with Quibi. Like I think they were in the right space of thinking like, yes, short form content for vertical phones, producing that would work. Well, first, I think they raised way too much money. And so they had way too much money and trying to try to like produce and force stuff onto people. and he was really good at that right time to like figure out what works and they had like two billion dollars i think they're right and they had meg whitman on who was the exe of ebay really powerful you know executives on the board and yeah i mean they had the right names yeah i think they were producing highbrow content which you know real shorts and the other ones i think cracked where it's like no you need just like just disposable content yeah the soap opera stuff that's just like it's cheesy gets you hooked yeah hits that like you know uh lizard brain part of us and like gets you on and it's not like trying to be a netflix on the phone it's just be like an elevated soap opera on your phone a game yeah also i believe they did spend a lot developing their original tech which their whole thing was they definitely shot i think open gate or some sort of square format and they built the tech where it could frame every episode what worked framed vertically yes framed horizontally and you had to use the quibi app and no matter how you rotated it it would adjust the framing yeah i believe it was able to find the center of interest within a 16.9 frame using the vertical window okay yeah it was like very sophisticated for its time unnecessary unnecessarily as we like feel realized too where it's like well people are fine just watching a vertical and if you're making for the phone it's got to be vertical and like it's been a lot on this it was their original ip to develop this technology unnecessary yeah and uh when quibi first launched uh there was a big press conference and literally like every a-lister in hollywood signed up for whatever show like you had like ben stiller and you know just a bunch of big names just having them sit there was probably costing money right like so it was like burning money from the very yeah beginning of it and you have to make it work and it's like yeah that's it was the first out too i think also the idea of people paying for you know like netflix prices but for something to watch on your phone that at the time you were getting for free on tiktok and even in 20 right before the pandemic the box office still wasn't as dead as it is today like there was still a very healthy box office and so people are like wait should i watch you know uh brad pitt in a in a cinema in a big theater or on my phone vertically that like that doesn't make any sense yeah but now it makes total sense do you think you know i mean so real shorts and these other companies have found some success in this but you know with the soap opera format do you think there's like other genres that could work in the future for you know just being like fully on fully dedicated to being short i mean i hate to say this because i have kids but absolutely uh there's a lot of disposable level animation and animated content that could be served to kids in a gamified way uh stuff like coco melon or blippy like a lot of the youtube stuff you can chop it up into micro episodes and uh have the parents pay for the credits that's something yeah that would be an easy cash grab would that work i mean what would the advantage be versus all the stuff they could get on like youtube kids or it's more automatically fed it's uh frictionless and so they don't have to look and search for it there's no ads they're just consuming at a higher rate per day or whatever i feel like also like this is also probably where the play of like ai generated characters or yes generated content yeah um user generated avatars so you know you can have a complete but your kids can be in the film in the in the short yeah in the cartoon yeah so like take three pictures of you and now that goes into like a storyline that has you know a third character yeah and they're the third character crazy amount of privacy laws and child what's the i forgot the specific act but what's the one that's prevents under 13 or without i don't know the name but yes you know where like all the social media platforms can't have users under 13 yeah and and there's yeah there's like a law but yeah i'm glad there is a law i'm glad about that too but i'm thinking about this route it's like uploading kids photos of your kids yeah maybe not the best idea for an app targeting young kids yeah and there's uh there's so many rules and regulations as rightfully so around child actors in general like how their likeness could be used how long they work and all that stuff yeah the next story does tie into ai actors yeah and um you know when you mentioned child actors i'm thinking like well maybe there's ai versions of them in the future but aside from that that's that was a random thought.

14:07The next story is H &M is using AI versions of existing models with their permission to create AI avatars of them to load them up and use them in additional campaigns without having to call the actor back to do additional photo shoots. And there is, I mean, I read through the articles, it seems like the models are okay with it. Yeah. So they're creating digital twins, the models with the models permission, the models are okay with it. And the H &M is being upfront about like that they're using AI and that they're putting this campaign out. They said that they're expecting that, you know, some people might not appreciate this or approve of this, but they did say like, it's kind of a learning process.

14:42We want to like put this out there and see what happens. Their reasoning was fashion. There's just tons of marketing assets they need to make. This could be a benefit, obviously for them, where they can create additional assets with a model. Yeah. In a much more controlled setting, they could have better reviews of the stuff or the new clothing or something. And then when they are loading up this digital twin of the model who is based on, who is a real person, the model gets paid. I don't think they're getting paid the same rate as if they were to be called in for a photo shoot, but they are making additional revenue off of their digital twin being used.

15:13And meanwhile, they can go do a fashion show or some other revenue opportunity. I don't know, Joey. It feels like this is the end of the line for a lot of fashion models. Yeah. Part of the thought is like, okay, well then if you're like a top model and you have a digital twin, then it's like, okay, good for you. But then if you're booked and they call up the digital twin versus yeah calling up another model that's less work the model's not getting work yeah the photography crew that would have been called up to do the photo shoot is not getting called up because it's being generated well the fact that h &m went all the way through and actually have output that's out in the public ether that's ai generated tells me that they're fully committed to every little piece on the back end so they have probably consultants and tech companies that are generating this stuff and full control over the generation, which means as these models are getting better, I mean, we just talked about chat GPT's image generator.

16:09Like it's so easy to now do consistent characters, pose them specifically, swap out outfits. Like, I don't see why you would need models after this sort of transition period is over. Yeah. I don't know. Yeah. It's kind of grim, but I mean, this is, this is the world we're living in. Aside from you like you know lean into that you know if everyone you know everyone zigs you zag so like if everyone starts leaning into AI generated models whether they're based on real people or you know I'm thinking like well in the future it's like well why would you even base it on a real person just make them up?

16:39I think from a brand to another brand it'll be different strategies so for example H &M is going you know the synthetic route but a brand like like a higher end brand, like a Louis Vuitton or a Gucci, you know, they want to still use real flesh and blood humans because that is just their philosophy and they believe in the output of real people. And so like, I think from brand to brand, you're going to see either a hybrid approach or fully synthetic or fully flesh and blood. Yeah. And they lean into it. Maybe they lean into more in-person activations or in-person things happening with real models, real people that they film and make their brand materials out of.

17:17Yeah. Something that's like, you can't do that with it or you can't like fake the reality or you can't fake the like real event sure yeah there will be uh edge cases where you absolutely need a real person but like 90 of the stuff like if you look at a target walk around the target look at all the you know the images of people wearing like the catalog-y kind of photo shoots like that's that stuff catalog models are like a genre of models yeah you know where it's just like you just got to crank through wear a bunch of outfits because they have like thousands of outfits you got to put on and it's for the white wall or the catalog.

17:46It's not a high couture fashion shoot. Yeah. Yeah. I mean, like I think all roles, you know, that's probably going to be one of the earlier ones to lean on AI more or just completely completely replaces it. Yeah. Because AI is so good at scale and automation, right? I mean, I can see the advantage of this or the argument of like benefit where you can combine both, you know, personalization and, you know, if you're trying to create social media assets at scale or something, but being able to do something that's like personalized for the potential where the ad getting served. I mean, that's been the holy grail for digital fashion for over a decade they've been trying it with digital humans for a long time this feels like the first case well a case where like a company's being outspoken about it you know like hey we're using this ai we're kind of like try to learn from that and this is presumably also we're talking about still photos how soon do you think this is going to go for videos for commercials for actors not for a while like and when i say that my time scale is in months not for a while like it's yeah till uh q4 of 2025 i think 20 by early 26 or so we're going to start to see some fully end-to-end videos that are generated with AI that's used commercially.

18:48By that, I mean companies are using it to generate sales. Again, this is a transitionary period where the models sort of know that their replicas are being created. They're still getting a wage from H &M, but then who knows if they're going to hire the next wave of models as these models age out, so to speak. Yeah. One interesting thing from this article on BOF, the website, To create the twins, H &M takes numerous pictures of the models in motion from different angles and in different lighting to capture them down to their birthmarks and movement patterns, making sure it has enough data to generate an accurate twin.

19:21It works with a technology partner called Uncut on the project. But they don't have to. You know what I mean? Like, they're doing this as a creative choice to generate the model with the same birthmarks, the same hair, the same. Yeah, they are making a digital twin of a real person. But they don't have to. they could just literally make a synthetic human that is no longer like anybody on right and also i mean we've we have there are famous 3d generated influencers like there are big accounts on instagram i mean if you look at the social influencer space uh with instagram and we saw the little mikaela rise right over the last 10 years or so and little mikaela is still the same age over the last 10 years or so because she's fully synthetic i believe she's no scandals no drama no drama right there's no boyfriends well there's digital boyfriends not gonna be on tmz paparazzi shots and she's owned by a company named brud b-r-u-d they're based here in downtown la super elusive i had a chance to actually do a little mikaela shoot so i know all yeah it was it was at verizon i so i know all about this little company and uh you know little mikaela is designed by like a group of four or five people so they're not replicating one person And it's a committee deciding what her personality is, what she should look like, what drama she should have, what clothes she should wear.

20:35Incredible. So apply that same thing now. Instead of using an Unreal Engine-based character, you're using AI-generated character. There is an influencer called Fit Aitana. I believe her name is Aitana Lopez. You can look her up on Instagram. And she's about 358K followers. And I think over the last few months that this came up and is having a meteoric rise in the Instagram space. Yeah. And it's not like a secret that they're trying to say this is a real person. Like it's known this is an AI character. Yeah. And people are still like, yeah, that's cool. I mean, yeah. How long until there is like a famous synthetic actor or CG actor or a AI actor or actress that doesn't exist.

21:15And people are like, I don't care. I'm a fan of this actress. What if Brad releases an AI character who does movies, goes on real shorts and does a couple of shows and then has a whole social life on Instagram and then does product placement and ads in H &M. That could totally happen. They could have digital rights to the whole thing. Yeah. One thing that, you know, talking about the digital twins or even just if you make a synthetic character, but then you're licensing it out to the model for another company to, you know, Coca-Cola where it's like, okay, we want to use your model and we're going to like generate some ads with it with the Coca-Cola brand.

21:49And also going back to like tracking the data. This is something that came to mind. And, you know, maybe we can dig into this in the future more. But remember the whole, you know, NFT craze and all that stuff and trying to forget it. Did you did you buy some apes that lost all their value? Yes. Yeah. So the NFT craziness, you know, got buried in expensive ape images and like, you know, million dollar JPEG photos and all that stuff. But there's some really interesting stuff. I agree. And the technology that I thought was going to be more relevant. I still think it will come back, especially with this type of stuff with like AI and like what your sources are going in what we talked to last time, where it's like being able to track the data that goes into the model.

22:28It's the blockchain authenticity stuff. Yeah. And the smart contract. Yeah. And the stuff like that. We're like that whole smart contract idea I thought was very appealing because it's like, okay, well, if you make a movie, and you know, you want to sell the digital copy of it, but then build in some sort of like the NFT, which would enable you to build in smart contract features where it's like, okay, if you use a clip, you know, from my movie, or use something from the song and you sell the derivative of that it's still trackable so there's a smart contract where it's like okay if your product is sold smart contract triggers we're based fractional yeah whatever we said or whatever i agreed to or we both agreed to then it's like but it automatically tracks like oh you sold it for 10 bucks but i have a 20 cut of that so then that automatically comes back to me right and it was just like a digital way to kind of verify like yeah i think that it was automated too right yeah it's automated it's all part of this you know blockchain and that's a little bit above my head but it was trackable enough where you could build these smart contracts with digital assets.

23:18And, you know, after you sell the copy, you can still see what's happening to it and have permission and rules for like what you can and can't do with it. I think that will eventually come back. I agree. Because that was very smart and clever. And that got completely buried by by bathing apes. Yeah, the apes and the JPEGs. But I think especially it's a it's a good use case to when you know, we're talking about digital twins, you start, you know, doing more stuff with AI and being able to like track like training and usage. I think it's going to come back. Every AI image generation should create like a blockchain entity underneath it.

23:49That's tracking the progress of where it's going, how it's going, what IP is looking at it. And then it should all be there in the metadata. Right. And then like something like Itana Lopez, you know, you can take her photo and share it with your friend. But if that image has some metadata, then it kind of knows. Like it's just naturally building a giant massive data table. which is more valuable than the character itself. Like how it's spreading through society. Yeah, or how it's being used. And also maybe we should win an award too for mentioning both blockchain and AI in the same podcast episode.

24:28You heard it here in 2025. Denoised is predicting a merger between NFT blockchain technology and the user-generated AI models that are going to take over our feed. Yeah. Okay. All right. And our last story. Actually, I did remember one thing that we were talking about real shorts and the dramas. I was like, oh, I thought I had this one other thing. I did hear another story from an executive and we were talking about, you know, YouTube and content creator, the creator economy and yada, yada. And, you know, I was like, oh, what are your kids, you know, what are your kids watching? He's like, oh, well, they watch YouTube on the TV.

24:58Yes. But just YouTube shorts. So that's rough. That's very damaging to young brains. so they're watching the vertical shorts on the big screen tv on youtube yeah so if you can think you could predict how we're gonna watch movies or what content in the future i think it's a losing game yeah that was the crazy thing i've heard the reason shorts are so effective is because it trains your brain to naturally limit your attention span over time and so like 30 second tiktok shorts are now too long yeah yeah it's like five or ten but it's like i need a more instant hit of dopamine yeah it's destroying your ability to just sit and watch a 90 minute movie exactly something longer yeah and especially if a young kid at that age you know where their neurons are still developing they get accustomed to that level of dopamine hit at that frequency this can't be good but that is how like you get the most number of clicks and the most number of clicks that generate the most revenue and so on and so forth it's like a losing battle yeah sort of i don't know i mean revenue wise the youtube long form is still more profitable 100 if you can get it to work yeah if the short form kind of funnels you back into a long form piece of content i think that's the way or you're just such a good short form creator where your shorts like are making millions of hits every time and yeah yeah can make money off that yeah all right and so last story uh we've got this interesting paper from anthropic yeah what do you see i love papers uh that come out from the big research firms like OpenAI and Anthropic and Google, I actually take the time to read through them and it makes me appreciate AI so much more from a technological standpoint and just how sophisticated these LLMs or these diffusion models are.

26:44So Anthropic with its very Portland, Oregon style and hippie sort of organic branding released a really cute video, which we'll link to and you can see here it goes through the process of how an ai model llm model actually computes the answer and it's how it thinks how it thinks and uh surprisingly they used a lot of medical terms to describe it stuff like ai biology x-ray you know obviously neurons and things like that so it's like splitting a real brain open what i thought was actually too from even from like Like just the surface level was like, even the people developing these things are like, we don't really know how it's thinking or how it's working.

27:27Right. It's a black box. You know, we talked about the black box, but even the people creating this are like, it's kind of a black box to us too. And they're trying to understand like, how is it figuring out what it does? So one of the conventional ways of thinking about how something like ChatGPT generates an answer is recursive. So when it generates a word, it goes back and takes that word into account and generates the next word takes those two words back into account, generates the third word and so forth. Yeah, that's what we hear about AI. It's like, oh, it's just a predictive model. It's a very offensive.

27:55It's just predicting the next word. Thinking on the fly and just spitting things out in a very probabilistic manner. Like it thinks 90 % of the time, this is the right word to put. Based on the trove of data I have, what is the most likely thing to come next? Yeah. So, we've understood this to work. It turns out that's not how Anthropics Cloud works, which is fascinating. So one of the tests that they did was, and then we'll screenshot it here, was it gave it a first sentence of a poem and said, go ahead and create a second sentence and make it rhyme with the first sentence. So it said, the first sentence that was given was he saw a carrot and had to grab it.

28:33And then Anthropik created a second sentence. His hunger was like a starving rabbit. And you would think that the answer comes in a sequence. It generates his, goes back to hunger, generates was, like, no. it generated the word rabbit first prioritized it and then filled everything in backwards guess what right that's how right because it was trying to find it's like okay it's a poem it's trying to find a word that raps with it yeah and so rabbit another option was habit i guess in their test it says green green yeah thinking that it just wasn't a strong contender so it went back to rabbit this was interesting because you normally think like oh it predicts it word by word like you just said but it's like no it's like a human found the word that rhymes first and then went backwards and tried to make a sentence that fixed the context where that makes sense yeah if you look at on rabbit or hip-hop rappers like this is how they build rhymes and bars sometimes they'll have multiple rhymes within a sentence and they'll have those keywords first and then they'll start to fill in the rest to make sense of it that's exactly how it's doing it so way more sophisticated than we thought.

29:37Another example they used was just a simple calculation of what is 36 plus 59. Now in a classical computer, this is going to go through a bunch of logical gates. So it's going to go through an AND gate, an XOR gate, an OR gate, and those are transistors on a giant CPU, right? And then it'll get to the answer. 36 is going to get broken down into binary code, binary math is going to happen, and then it's going to turn it back into an ASCII, which is going to be 95. So how does Claude do this without knowing math? Well, turns out it's first creating a rough approximation of what it could be. And that is a number that's between 88 and 97.

30:16And then with a second pass, it's fine tuning it to get to 95 because it knows the last digit has to be five. And then they actually ask Claude, how did you calculate 36 plus 59? said i added the ones first six plus nine is 15 carry the one then out of the tens three plus five plus one equals nine resulting in 95 but the crazy thing is that's not how it did it i just told you how it did it it did a rough order of magnitude and then it fine-tuned it but when you ask it how it did it it gave a conventional human answer how a human would do it okay that's both creepy and incredible at the same time so like did a different way and then told the person like oh Well, I did it the way that you think I should have done it, but I actually didn't do it that way.

31:01Right. Because it's probably like, oh, they won't understand how I did it. So I'm just going to give them an explanation that they would understand. But now it's still not doing actual calculations. Correct. Right. There is no calculator in LLM. Right. Because this was also an issue too, when you're like trying to ask math questions and then you kind of get weird answers. Or when you're, you know, like how many R's are in strawberry, which is a common test. Interesting thing is there is also no memory table, right? Like you can do all of this with a massive lookup table. Like every number added by every number could be like billions of results.