November 16, 2024 - YouTube
https://www.youtube.com/watch?v=4MZnQcSywmM

Transcript:
(00:00) this the whole reason I made the Symposium was maybe an attempt to compress what I'm about to share with you um so if you guys have ideas on how to compress it later um let me know this is where it started um this is my view of what the data flow could look like for achieving the relationship Singularity um these are things that are so what this is is what you guys are used to when it comes to personality questionnaires uh are you very extroverted um do you like to spend a lot of time in your room basically self-reported uh question and these are
(00:32) basically personality constructs what these are things like the ink block test um if you like had someone watching their face and you were showing a movie to them and getting their reactions you might see um data corate there Google's quick draw um if you ever saw like a cultural analysis of Google quick draw um which is another thing I should include in this that I it's probably these slides are probably still only 70% of what I have uh maybe at a future date I'll have a more complete whole blog post on this
(01:01) can you define relationship Singularity or are you going to get to them we might get there yeah I think we will um so Google Google quick draw you can see for example that Japanese people when they draw circles draw them this way uh Americans draw them this way uh and this is just one of many signals you get about personality uh just from people's actions that are not telling you anything specific about personality uh what this is is tropes um so we'll get to this later but there's a website called TV Tropes uh basically trop sces
(01:30) are combinations of personality traits um with different like uh frequencies of different traits and that is another embedding space that like this would map into this would map into this then you get actual characters and these are nodes and then you have a edge between them and that ends up um with a s a graph of emotions that that relationship elicits in each person and then those become cordinates in relationship embedding space and this is the map of all types of relationships and then this is the probability distribution of the
(02:01) outcomes you might expect in that relationship um so this a high Lev view of what I think is maybe the path and we can look at ways that we might map each of these things through data what is the relationship sity um it is the event where all social outcomes we seek are instantly attainable at least in terms of knowing the optimal move to make to achieve the best ones possible to us um we cannot achieve it we can get ASM totically close uh so right now what I view this as is the goal is to ultimately make uh
(02:32) what I think Facebook uh might have become or could have become uh is like a Google for social search um and so yeah this is what I think it could be um I think there's a lot of people in the cultural space who have had some glimmer of hope that this could exist and I think they all think about it and it's like antimimetic where they just they forget about it again because they just have no hope they have no real Vision that like this could be technologically achieved um you can see for example here um preferences can be
(03:07) respectfully derived from social media streams and public data sets um Can systems help move users pass unhealthy partner seeking habits can the system pre-screen people by evaluating their friends experiences or ratings using ubiquitous sensors and subtle actuators how magic and subtle could proximal matchmaking become if a Central Computer knew that two people were made for each other could it use ambient or augmented reality technology to help him notice her across crowded bar by amplifying her laugh or even by slightly brightening
(03:33) the light above him when she looked in his way could match not this many interruptions guys please let me could matchmaking technology like love be in the air and so this is like the concept of like what a technological relationship Singularity might start to look like why should we achieve the relationship Singularity um I think this talk was maybe from like I forget when Maybe 2019 but it says over the next 50 years the greatest source of New Wealth May consist in finding choices we didn't already know we had um I think that
(04:10) there's a lot of things people like Elon are doing which are really great which is like making a new rocket ship um I think like I don't know there's like a lot of technological advancements that are like trying to make a new thing uh what I think is that like this is one of the greatest opportunities to create improve societal happiness uh that costs almost nothing you just have to introduce to people and think about for everyone here like your closest friend who's like not a family member think about how much happiness they generate
(04:44) for you so one study I saw was that um if you look at two people in like the sort of 50k to 250k income range um the people who are married have a 50 and then you look at their happiness levels the people who are married have the equivalent of a $50,000 increase in salary in terms of happiness and I think this was actually between 60,000 to $110,000 so like people with $60,000 in salary uh are the equivalent happiness of people with 100 110,000 who are single um that's a lot of money uh and that's like per year uh so if you're
(05:21) thinking about like making a new car or anything you think about how much more happy that makes a person you're basically increasing people's happiness in terms of economic value if you give any cre to like the market has any sort of authority over what is true um you're creating that much happiness between people and I think that's just like a free choice we didn't know we had [Music] um I have another view of the relationship Singularity um that the goodness of a relationship follows power laws where um there's like at the at the far tale
(05:56) like I think you can find people who are like one in 100 for you one in a th one 10,000 and you might say then why Try to find the one in 100,000 Why Try to find the one in a million because maybe they're only a little bit more compatible maybe only like them a little bit more they only make you a little bit happier in my personal experience of what I've seen from people uh finding that person does not make you a little bit happier it makes you significantly happier that like of all compatible people at the very tail end in the top
(06:21) one and then decreasing percentages uh people start to have something like they satisfy their soul like they feel like lonely in being like a thing that exists in the universe and I think that's just like a feeling that maybe even can't be like quantitatively compared to just having a group of friends you get along with um I think being able to optimize the process where we begin relationships with people with a very high likelihood of positive outcomes can save us a lot of bad outcomes and maximize good outcomes that we might never have
(06:56) experienced otherwise okay uh there's actually less wrong post about this um that was on similar waveline uh how to make billions of dollars reducing loneliness and he uh this is by John Maxwell and he essentially talks about the same thing that um like if we just could get people to be with people that they are really really compatible with they could this could like improve the quality of life massively for people um and that we are currently living in like a social Dark Age where like we you know maybe I think
(07:27) people here at Li Haven and in the bay have done some of the filtering that these techs would do and it makes you a lot happier maybe than the average person but uh just imagine how much better it could get if every time you went on vacation um you were matched with someone in that City who could become a new best friend not just a acquaintance that you like to hang out with like an actual best friend I think people's investment in their communities is often B like for me at least the cities I love I've always realized in
(07:56) retrospect when I went back to that City that the only reason I ever loved it was because like I love someone in that City and every time I've gone back to a city where I wasn't with that person in that City I I suddenly did not love the city anymore and I think as we get increasingly where we have this Urban disinvestment um in our communities in our civilization in our society a lot of it is because of loneliness where we don't care about the people we spend time with in our society and I think if we could have it where you had us like
(08:30) super compatible people all around the world that you knew um it would sort of create like um more investment in the human species for people they actually had people they loved um and I don't think the main barrier to that is people simply not being open enough to loving others I think it's that there's just inherent compatibility that takes a lot of effort and time and luck to actually find someone who is optimal in this way so this person gives a long list of things that you can try they're all fans of the same sports team they enjoy the
(08:58) same movies musicians blah blah blah blah blah blah what has this all become this is just a personality Vector that they're describing they're describing different ways that this that a human could be measured um or Quantified as to like what their personality is um getting out of the way at the beginning some people have made good critiques of this they said is this literally automating Echo Chambers in real life um the answer is probably yes uh that's who you're going to get along with you're probably G like the online
(09:26) ification of echo Chambers this would probably make it social Echo Chambers where you're literally only spending time with people you like so like now you don't actually have to deal with the fact that people like you disagree with part of what you disagree with is like you just don't like them because they're not in your bubble so yes this might amplify bubbles um another thing it might do is deprive people of learning experiences you know you learn a lot about what it means to be a human and stuff by uh suffering through bad
(09:49) relationships you learn about yourself uh yes these are bad outcomes and I don't have answers for what to do about them I will just say all new technology that gives blessings also uh comes with purses and I think the good on this outweighs the bad and if you don't don't work on it I will um I think you'd have to give me outcomes worse than never getting to meet your true best friends and soulmates and living in an era of Mass Social loneliness for me to consider not creating this um but yeah it's good to know that those uh might be Downstream
(10:20) effects if this actually got achieved some people people think this probably won't work I already gave you my whole I'm always right thing so there it is um here here's how I became always right uh though I didn't know it was going to happen uh clip came out in 2021 I was in like a hostel in Costa Rica and there was a party and I ended up not going to the party and instead reading the clip paper for like six hours because this thing changed um the state of image synthesis and what they did um that is what I think is how all
(10:51) of this unsupervised learning stuff works now is that instead of having one hot encoded labeled to say uh okay so here's how it used to work is you had um pictures of cats and dogs and you just had like the first bit was like a zero one zero for dogs one if it's a cat and you just have the ml model learn is this image more zero or more one uh and then it said and then we said okay oh can it also tell us the species of dogs so now it has to be zero for dog and then someon coding to say rottweiler versus husky uh but then what about what about
(11:22) the orientation of the dog oh now we need another bit to say which way the dog is pointed oh but how about the length of the hair of the dog okay we need is another number that says the L of hair dog we need a human to come and label every single one of these things about every single image and we said this is never going to scale actually scale AI um like almost went out of business when clip came out because people thought like the whole reason that they had so much demand was to do all this one hot and coded labeling uh
(11:47) and what clip did was it found an unsupervised way to generate a vector that described every single one of these labels in a high-dimensional space um so the way it did it um actually don't have good clip pictures um but clip creates this uh manifold where a caption of an image and an image have to be embedded in the same coordinate space um the actual the actual thing it does is embed it where it's the nearest thing uh what I was actually discovered is that the language space was this smaller sphere at hypersphere and that uh image space
(12:23) is this larger hypersphere because there's more possible images than ways to describe them um you can just imagine one way to you could describe a cat there's like a lot of ways you can draw a cat from that description so the point is is that it found an unsupervised way to have any single image represented with as much resolution as was needed to distinguish it from every single other image in the world and once there was collected a a data set called lion 5B um of five billion images this clip embedding space became extremely optimal
(12:54) at knowing everything that you need to know about an image to distinguish it from every single other image um and that's what I plan on doing a human personality is basically getting a large enough data set and an unsupervised training where you could be able to distinguish not only every possible person from another but also it could learn things like every Poss every not every person but also every possible person or another um so yeah can the same be done for mapping arbitrary modalities to a semantic personality
(13:23) space um arbitrary modalities in this case means not just images of people not just just text of people like we had with clip but could it be the voices of people the genetics of people the sense of people um videos of people any of that stuff like any way that you can get data about a person could we transform that into a thing that describes the totality of who they are in a single efficient embedding space learned through unsupervised learning uh so I want to now show anyone in the room who who in the room uh like
(13:55) raise your hand if you feel you have a great understanding of embedding spaces just be confident let's see it yeah I know you guys want to raise your hands okay that's maybe five maybe some not okay who who thinks they have like have no clue what it is okay and rest in the middle so I want to show like I think this is the best example so I don't know if you guys heard of mist um this is like handdrawn digits um I could draw like digits like 9 one two whatever um there is an there's like a latent space of all ways that people
(14:31) draw the number one or two or n um and it it differs and the way that you draw different letters also our numbers has similarity between each other so this is one of my favorite like uh intuitions for what the embedding space of the number seven being drawn and all other numbers being drawn you can see there's actually this like similar continuity at the edges of sevens and nines where some sevens and nines look like similar you can see even there's some threes that end up looking like sevens um so this is what an embedding space looks
(15:05) like this is what it means to embed information um and what the features are of this what the characters are just literally if like if you like break down a region that you draw on into a little grids you just fill it in or don't fill it in that's what this data is uh so you can imagine instead is um are you outgoing or not that's a box filled in or not filled in and so we would make shapes with our personalities SAR to making a number or letter you can see there's like as you move around the space that things that
(15:38) are near each other are similar to each other and that going in a certain direction has a meaning that like for example in this case the eight goes from being slanted To the Left To slanted to the right it means something in this space geometrically to go from here to here it means something to be in this region and in the same way in personality Ed space things that are near each other people that are near each other are similar and going from this direction to that direction means this is a type of person becoming more
(16:07) this way so yeah that's an embedding space um in my opinion one of the most magical modern inventions can you hover over one of the most isolated clusters like one on the right uh yeah well I think that would be zeros or yeah Zer you can see there's some Nish zeros and even sixish zeros or well actually a zish six is really what it is um the label so this where SC so like label this stuff they would pay people to like label this stuff this is where you see some of the stories like Nigerians are like just like labeling AI
(16:43) Bros data that's like some guy Nigeria might have like said this is a six um probably more likely was like a un like an underpaid PhD student at like some University labeled data six um but yeah that's uh that's how it works you see the one just thing um and like really I don't think there's anything actually that far out of space Maybe This eight here it's kind of like in a pretty weird spot so yeah there are outliers too um boom okay personality space so one of the coolest things that happened uh with language space was this I want to
(17:17) convince you that there's geometric properties within inventing space so this is word Toc and one property is if you subtract um oh how do you how do you describe this it's like uh you this is the vector that goes from female to male and so what what is itus woman plus man oh yeah you do uh Kingus man plus woman this is man yeah minus man hold on this the usual rules tip minus tail it's not minus woman man plus woman it's like or equals man this is the Regal Vector um yeah like make woman Regal it goes that direction so it' be
(18:03) like you go this way King what is the man Vector it would be this direction yeah uh oh minus man so be going the opposite direction you guys get it it's a space with rules I always forget how exactly say big to biggest small to smallest so what does it mean to invent personalities yes that's what we're doing mean outgoing shy and you can see that there's an archetype combinations of traits so if you're nice and outgoing maybe you're this like meta trait Life of the Party um so yeah this is what personality embedding space would look
(18:36) like um all right what aspects of Personality should we embed someone said earlier like you just made OB white they didn't even like read what the thing was and I'm like that's not the point but anyway um what as has you bed as I said before everything if you want to know everything about if like everything that's going to be relevant you should embed all of it so here's what happens when you don't embed a personality at high enough resolution and then you ask a thing an AI model based on everything else it knows to render that thing based
(19:06) on how well you embedded it you get a lossy interpretation so imagine you're asking someone this is a person right it's one of you I embed you like this and then I ask some model do you think you know this guy would find you attractive um and then it says well this is what we think the a model thinks this guy looks like you know this person might find that person attractive you can see it could get lossy where like the information that you need to know that is not uh captured because it's not high resolution so this is why some
(19:35) people say Matt this won't work it's because they think that we're going to do this um we're going to capture only this amount of data about the human personality and they would be correct if we only capture a very low re resolution of someone's personality we are going to get very inaccurate answers to questions like uh do you do you like this guy or do you find him attractive um so we need a lot of data we need to have a very very dense uh personality embeding to do this correctly voice uh so actually like
(20:03) for example I don't know if you guys have ever had a relationship with someone like online like World of Warcraft lovers style um but like that's a real thing like people I had one friendship that was like almost entirely on Discord and every time we met in person it was like way different in like not a good way um this is a person I had a crush on too and like it was like Mutual until we met in person and then yeah was terrible anyway she I asked her one time like do you what if you knew someone was attracting me what would you
(20:31) what would you know about them and she said they must not uh have high standards they must not care much about looks so anyway pretty mean uh somehow friends uh so anyway Point down rough whatever okay yeah so you have to embed everything about them uh like for example Voice verse writing uh I think there's a lot of uh people out there maybe even especially if like women relate to this more where they actually like there's something about a guy's voice there's this meme on the internet that they get the ick when a guy leaves
(21:04) a voice message on like hind or something um so like they're typing to them and then like the guy talks and they're just like you know ick like the way he talks suddenly wrong we have to embed that if you want to give like if you want to set someone up on a date with where you get credibility that it's going to be right you need to embed The Voice you need to embed the voice because that's actually like a a vector that that correlates with relationship outcome you need to embed almost everything to not have this
(21:32) happen um modeling people uh these are some of my components of like how we need to model people we need to do so thoroughly um and that's what I've been discussing just like make sure we get all the information about them we need to do so efficiently um so there are two things that people prices people pay when we're modeling them uh they're either paying with their time or with their privacy um because one thing you could do is have them upload like they could do the click the button where you download all your Facebook data and then
(21:58) just click upload with like the zip pile uh but that's violating privacy so we do get a ton of data with like a little bit of time from them but they don't want that and so as we're thinking for like people who like want to make a startup about this about how do I collect all this data thoroughly is how do you do it with time and privacy costs being used efficiently um you need to increase uh confidence on important traits ratios um and need to extrapolate more accurately to get them to continue investing we
(22:26) need to do it communica so the thing is is that people don't want to just like throw a bunch of their data into a black box they want to like have you up the ante at each step where with very minimal information from them they want to see a magic trick that you prove to them our Tech is actually using this data to do the thing you want to do so one way you can do this there was a study that came out I don't have it up here but like they they showed people like about 20 different clips uh 30 seconds each of videos and they're just
(22:57) doing facial uh sentiment analys on the people watching it and they knew their Big Five personality trades um and there were certain videos where just the the reaction of a person in a very small number of seconds that video was massively correlated with certain personality constructs imagine if your website like you have a dating website and you open to the user with here's like a short little video clip and then like they watch it and then you just instantly tell them you're probably this extroverted or whatever and they know it
(23:25) about themselves to be true they like what the hell like this website I have to keep giving them more they actually like can map like a little bit info to know me deeply I think people want to be felt feel that they are known deeply because that is the that's like what a in-person Matchmaker does too they want to feel if you're like Consulting with a person they want to feel that this Matchmaker knows them like just from very little info they want to feel I'm known and understood because this gives them credibility um and we want them to
(23:53) understand also insights about themselves so that we can say we understand what Journey you're trying to go what goals you have in this relationship um and finally we want it to be done um usefully which is like maybe a hardto Define term but it's something like uh just because we've mapped your personality doesn't mean we're actually helping you um so we I don't I don't know this is like probably the hardest to Define but like we need to give you actually good outcomes um so yeah here's like an example of like how
(24:19) you can embed uh human like uh body movements for example so this actually done right up there at UC Berkeley um but like you can basically take just videos of couples dancing it's like almost like sci-fi gotman esque analysis where you can take videos of people dancing and you can see Synergy and synchronicity between couples um we are living in the we're living just on the precipice of this like the style or the this the style Gan moment for personality is maybe just like a year away and the stable diffusion moment for
(24:53) embedding human personalities is maybe only three years away and I think this kind of stuff is like the early papers that make this work okay so a little bit from Twitter um okay we'll just start with this one actually so uh this person uh this little piss baby um had one time I was really Manic and I paid a Psy kick on Etsy 30 bucks to grow my soulmate do any of my followers look like this um there will be a call back to this later but people are out there trying to do this people want this to happen um out of 8 billion people it's likely that
(25:30) at least one is scarily close to me in personality like a real doppelganger I'm trying to figure out how to a clausy coordinate with him so we can meet each other any ideas for shelling points um and basically this is like pick 20 variables of a person calculate the personalities um there's like uh this person's saying basically that like how do I there will be be another person like you they're saying that the personality in bending space is very high dimensional we can have people be unique in it this person's worry is that the
(26:07) Obama like low resolution phase is like we can't get more precise than that uh and they are incr or maybe I don't know okay so I'm going to start with a field that has already had its style Gan moment um and that is the field of human emotional Ed so this is Alan Cohen I tried to get him to come I think it's actually Cowen by the way tyen I don't think he's Jewish and it's Jewish name but uh it's actually C I think um so Allan C he lives in um SF or sorry he lives in New York City uh but he went to UC Berkeley here and he is like
(26:40) um like uh he's like the goat at this like embeding human personality or embedding human emotions um so he started with hypothesis that all human emotion research is in its pre-cp era right where we basically have a a prior on what the affordances are in Emotion space one of one uh this was like one of the earliest papers where we want to look for more plausible Alternatives because we have a bunch of data on human faces and we say uh this is anger this is disgust this is happiness oh this is also happiness oh this is fear and this
(27:18) is also a fear by the way this is anger and this also does this person look angry to you no no does this person look afraid does this person look happy okay um and yet we say um they reported oh sorry okay let me let me uh this person said that they felt happy but they look surprised this person said they felt fear but they look happy this person said they felt angry but they look disgusted and it's because we have insufficiently high-dimensional uh affordance space in Psych in psychological um literature to describe the emotion
(28:02) because the the was that everything is very primitive like we're in the dark ages of like this stuff we like we're just saying these are the only emotions that someone can feel those are the only labels we're giving if you look surprised and you're happy you're feeling awe if you feel fear and you look happy or embarrassed if you feel angered and you look disgusted you're feeling contempt this is what a higher dimensional embedding space looks like when you actually like say that all these priers people have
(28:28) put on Lang anguage of like how you can describe personality are insufficiently high dimensional so when people today here have asked me have you heard of the big five why isn't this enough this is the big five this is what I'm trying to create this is this is you know only five dimensions of Personality that we talk about this is add more Dimensions this is be more descriptive this is um life of the party you know instead of just if you're outgoing and nice you don't have to pick one you can just combine them
(29:00) semantically this is one of alen Cohen's earliest work um whatever uh he he this is the data he had a Target emotion so he' like give participants the word dismay and then he'd give them like 25 randomly chosen other words and you'd say rank these by how close you think they are to the word dismay to get human uh data on which emotions are similar to other emotions and that becomes like kind of a not exactly sure what the data set like what that data but this is how you uh can create an embedding space from that um from like that list of like
(29:34) here's the things closest to that emotion so you can see now we're starting to get a higher dimensional uh representation of what emotion looks like and you can see there's something over here that's like challenge or grit so earlier people just maybe say this is happy this is like sad and this is angry right hi and orgasm are pretty close where is that bottom left where is it bottom left oh yeah oh yeah pity and sympathy caring love orgasm lust um you can see like epiphanies and all and adoration go into like these romantic feelings um but you
(30:09) can see that like amusement and giddiness and Joy are not uh the same thing as those and you can see like romance down here you can see this like fear dread fatigue Detachment this is anger hate and irritation um so this is what it looks like when you uh turn like a simple thing into a high dimensional embedding space Oh yeah okay any call outs on this thoughts questions this is the embedding space of human emotions this sort of two dimensional but you're Ed into a very high dimensional space right yeah I wonder if
(31:04) there's some orientation where might see different kind of exactly I'm recording so this is uh the way you create this is something called dimensionality reduction you do something like or t uh I'm think I won't cover that too much in this because I have to have different slides that show that I actually might um this is actually pretty quick to explain um I have an intuitive idea what so here is a high dimensional thing this 3D coordinate of this uh Mammoth this thing like a a mammoth or MTH yeah mammoth fossil and here's a way that you
(31:46) can turn this into a two-dimensional representation that's what it means to make a high dimensional thing uh visible on a two-dimensional plane explain okay um this is like the state space of emotions um so there's like number of dimensions of emotional state space um this is the how much you feel the emotion and um I'm not sure the upper axis um do you think it's possible to create an objective map of emotions because don't emotions also get influenced by the culture that you grow up in uh we're going to get into that actually there is
(32:27) a paper on that we're discuss okay actually that is this paper um or it's the next one I think or actually yeah we'll get there um but you can make one that is like um high high resolution um yeah that is a good question um okay so here's um some more of his stuff um I actually wanted to show you guys one here so this is the the EV emotions evoked by video um so people said they looked at these clips and they said what emotions I think they were feeling or maybe there's even like a facial analysis you can see like um there this Sparks
(33:10) interest at least the emotion they saw is interest this is like adoration this is Amusement he's going to fall that's that one's a here's awkwardness here's Nostalgia yeah just going we're going to just get 90s 90s Nostalgia okay romance sexual desire yep disgust John Waters horror also anxiety went out oh oh that's cool anxiety yeah also censored um fear we can also find like in the inner Place between disgust and sexual desire I was not planning they nailed it they nailed it
(34:15) that is like you're not sure which okay um Nostalgia and something else that's I wonder entrancement yeah Hill's somewhat nostalgic I guess um and then here is like an unsure one between interest and ad uh Joy so you get it I think um there's there's such a thing as embedding spaces um audio to emotion eding I also want to show this one though I'm not sure if my speaker will be loud enough so this is sounds and uh Amusement we go to desire me tell you sadness distress here exactly there's such a thing as
(35:25) embedding spaces and it codes encodes a lot of um I think that's my main point okay quick question that's exactly yeah go ahead question um how how good is like hume's like voice voice recog Alan Cohen is the founder of Hume okay so not bad yeah he is they they invested uh he he's an AI grantee and they invested well having him try to figure out human emotion okay why is this so AI Grant is AI Grant is like a oh AI Grant yeah the AI has granted money from the future retro CA usar okay uh some emotion labels are
(36:06) okay to your question um some emotion labels are consistent cross culturism are not so this is actually similar this is from some of those embedding um things he did so this is uh emotions that where two people I think they mostly use people from India and the US um agreed versus disagreed on labels for certain um data sets on is this the right emotion so you see Indians and Americans I think this Indian americ we'll just go with that um think that sadness like they agree when the thing is sad like voice images a video
(36:36) whatever they agree with content anger they will not agree at all on what's embarrassing uh explains a lot of they will not agree much on what's shameful or what's disgusting uh what makes someone feel guilty what looks like ecstasy or arousal or sympathy so you can see in cultures uh there are some uh where in the space can be cross culturally shared and then somewhere there can and this would have to be accounted for so you could get more data to know which culture is a person from and then you would map them into the
(37:06) embedding space for their culture um things like that could be done here's also like Japanese responses to like labels they gave to things they saw you can see Japanese are much uh they'll give like uh veilance is like how much they like it and then arousal is like how excited I guess um and you could see like Japanese basically will not give uh low veilance thing low arousal so they won't be very aroused by bad things whereas American this is like maybe like some of our like you know Tik toks where people are fighting could be like low
(37:36) arousal and low veence it's like bad but we also are actually unaroused and bad yeah or unaroused I'm not sure it's like magnitude and direction veilance is Direction in this case the direction is a bit trivial of just positive or negative like left or right but still this is admiration over here for yeah I'm not really sure how to interpret this the the point is I think there's different distributions of emotions elicited in different cultures you're right like we have to actually include the culture and that's back to
(38:06) my Obama pick you need a high resolution understanding of someone you can't just say we' asked all these people who are weird if anyone knows like the we D and all caps like this group that only responds this way we have to actually map their culture too um here's some uh like other ways that you can look at like uh dominance versus affiliation like it's like negative dominant is hating someone having like weak but you know negative is like being morose um and this is like euphoric versus Like Loving um you can
(38:39) see also that like narcissistic traits are maybe um more up in this place hating and contempt um you can see that like depression is usually down here negative and like withdrawn bipol is actually like splits as advertised um autism is actually like low like kind of neutral on dominance and back and forth on affiliation uh so you can actually learn personality traits and I'm I'm showing you this to encourage you to understand that if you think there's something that this embedding won't capture you're probably wrong uh
(39:08) probably we can like know almost everything about a person just from even seeing the emotions that they elicit in certain data um we can know if they're bipolar if they are depressive you know if they have extraversion or not um all those things uh you can see like for example here's another way look at dominant filiation narcissism tathy Maan are up here there way you could like filter out users who maybe have no hope of like a good relationship with others um so yeah just you can know things about people through embedding
(39:36) space embedding human personality uh so if you recall um I was showing you how to embed emotions which is a field of research that's already been working now I'm going to show uh some stuff on this I'm not sure how much I have but um okay so this is a super cool chart um I think this is actually from the like Facebook data set that was uh the one that now is not allowed to be used because it was the thing that Cambridge analytica used um and the Andrew also used yeah Andrew what was the name of it my person what
(40:09) is it my personal my personality yeah I think this is maybe also used the my personality set um so picture this uh we have your Facebook we know all of your likes and I want to guess and I also know all of your Big Five personality scores like just one of you um I have to guess what your neuroticism conscientiousness extroversion agreeableness openness is is um to do this um I'm like either your work colleague or your friend or I'm an AI and this is log scale um I know 10 pages randomly chosen that you like on
(40:47) Facebook if I an AI that knows 10 Facebook pages you like I can guess your Big Five personality traits as accurately as a work po if I'm an AI that knows a hundred of your Facebook likes I can guess your Big Five personality traits as well as an average human um or family let's say I'm not sure exactly that um I think it's oh this is hum's average accuracy just in general for all people this is to prove that that computer get better um so then how close is your spouse to you there are only 275 Facebook likes close to you
(41:25) um which is to say that if you think even my mother can't help me figure out who I should dat I say she doesn't know all your Facebook liks um and like probably the computer can know you better than the best Matchmaker or family member ever could um what this is also just me demonstrating a scaling law as we know more about your personality we know more about all your other personality traits um I had one interesting thing about this which is I think like it's interesting how quickly people can figure out how open someone is I
(42:00) that inter um and it's really hard to figure out how agreeable someone is and actually this shows up in a few other things not a lot to inter things yeah you wouldn't expect it to show up a lot this would suggest decreasing marginal information to to knowledge about from liks which would kind of go cter to your hypothesis of knowing someone deeply better and better matches in terms of people well it's just a marginal it's just a marginal return decrease but you can still just know more yes but to say assuming that you go you know given
(42:38) assume a somewhat rapid marginal decrease that's saying there's assuming some rapid marginal this does not look like a rapid marginal decrease I would say it does but depends this is going up linearly it's on a l scale I know but uh also 300 Facebook likes is like not a lot of it's shockingly high already wait that's that because the X is marginal decrease because it looks I don't think it disproves the idea that you can just get way more data than 300 Facebook lives somehow it work same thing people said in 2018
(43:10) about uh imiss so wrong same argument I heard 5 years ago this is also just the five Factor model and something more sophisticated might have an entire long tail that is knowable and that no Matchmaker can read enough to find that out um I feel like if I was faceb and I trying to do dating what I would do is just like you have the whole social network just take people like on the border of every social network because then it's like they're similar enough they'll probably get along but they're not in the same Social Circle so like
(43:42) sakes are lower it's not as like weird I feel like you don't like this might be helpful but like just efficiency wise like all of us are gathered here because we're like things bringing us together but like you know social networks are a bit incestuous and like you just go a little bit outside that yeah um all right as data increases you can know a person's personality better a scaling law uh and yeah okay um what do friends and family know best about you that online data doesn't so this also um was talking about some of the
(44:26) stuff you guys were saying that like I think uh it was something like yeah we'll skip this one I I didn't looked at this in like a year um uh this one is very interesting this is how well um the difference between people self-reported and uh other persons report the how correlated someone reported their own Big Five personality traits versus how their peers reported it at different ages um so what you can see here is like the the less total uh Divergence uh or actually it would be the higher these values the more people in their
(45:02) peers agree on what their personality is like what a person personality is uh you can see actually extroversion like at a very young age people already agree with their peers on how extt they are and actually have a very big Divergence or at least it's a you know kind of a truncated y AIS but like less onc conscientiousness and you see as people get older they understand each other better on how open they are maybe also like people start to realize how not open open they are and they to agree with their peers but either way like the
(45:30) um people and also I found this one super interesting that people are not very good ages 14 to like 20 at agreeing with their peers on how emotionally stable they are and something something happens around here where they start to be like okay yeah that is how emotionally stable I am um real Myst yeah I thought and also like as you get very old around like 30 maybe people just finally start to figure out okay this is how agreeable this person is uh yeah I think maybe some people are like dis able and they just like yeah won't
(45:59) won't fess up to it and maybe that's where this increase come from um how hard is it for us to know ourselves uh there are some differences in gender um they actually not a lot to find here um get that okay so this is some uh research I did so there's something called open psychology it's a website with a lot of Open Source data um this is on Mak I traits so people are answering how moan they are this was lyer scored so 1 through five for each question um for example we can have here uh actually ra raise your hand um we're going to go
(46:39) through one through five actually okay we're going to just have like we're going to do zero through two most people are basically no you guys won't answer that one honestly maybe uh it is wise to flatter important people put no hands up if you say no put one hand up if you say it's kind of wise to f p people put two if you say it's very wise we're getting a distribution okay um keep your hands up double hands up let me see the double hands up okay two double hands up so these people would be red in here somewhere right um and who
(47:17) put No Hands Up raise your hand now one now after my own heart um does imply that it's not honest um essentially yeah yeah okay um and again this is also where I had in my initial thing the top left was not you answering personality questions just saying who I am but rather we just measure something about your reaction to a movie because there is a very valid critique of Personality uh inventory is that people can misinterpret the question even um and then people also just why would they be honest all the time or they don't even
(47:50) know themselves very well um so like we don't we it's not the perfect measures uh you just need a lot of data uh sometimes people self-report but also sometimes just data from like people ey tracking on like a video to really get to know a person um I'm not here that's just going to say we need more one through five personality tests and that's going to get us the relation sity I think that is a naive view I do think it would help it might even be efficient to get like the early gains but I think you need a lot of other data to really
(48:18) know a person well um so you and I would be this um idealistic uh little blue here or we might be like so I'll give it away this these people are pretty evil these people are not we also might be these evil people who are blue up here um because not everyone who says no to this is good um so why do I say these people are evil so this is analysis you can do um on so what I did is I took this High dimensional analysis where for each Dimension each question is dimension you can be in one of five coordinates here like one through five
(48:47) how much you agree um and then I upped it right remember umap like the mastedon gets flattened I umap that whole data into this uh so this is now a personality manifold acoss these questions about melanism um so for example um these people in red here believe that there is are excuses for lying to someone else these people in blue think there's never an excuse for lying to someone else so if I plotted one of you on this graph and you were right here I would say you're a kind of person who doesn't like believe it's okay to
(49:19) um to Li you not good lying you can see here like honesty is the best policy in all cases these people kind of believe that um these people have this view like the biggest difference between most criminals and other people is that criminals are stupid enough to get caught um these people don't believe that right so these people here think that like they're not okay with lying but they also think that like criminals are just ones that get caught whereas these people think they're generally is a split between like good and bad people
(49:45) but they are willing to lie to some people so you can see just like if you had a map of Europe and you have different cultural traits of each region in this map I've made of Personality each region has a culture associated with it different dogmas or views they have PT Barnum was wrong when he said there's a sucker born every minute um these people think he was right also this question's word away would don't know if people misread it um but these people think he's right that there's a sucker born every minute and some of
(50:12) these innocent people also do I think these people probably mellian where they do think there's a sucker born every minute and it's just and like they won't be suckered um so yeah you can see like the mellan um human embedding manifold and this is like a communic tool you can make where instead of having Myers Briggs uh you could have something like this where you are on this map of personality and you can plot people on there to have them understand in kind of like a buzzfeed ified sort of way like here's where I am you know get your
(50:41) results too okay um we're now getting into a section that is kind of like a mirror to Andrews um the Le deep Lex about this but first before I go on um any questions so far do you think that as people learn more like this type of information gets um kind of more popularly understood that it will skew like the actual values of like the self-reported that's what I said I think self-reporting is like uh yeah I I'll go Elon with it and say like same in the same way he said liar is a crutch self-reporting uh personality
(51:18) data is a crutch um we probably don't need self-reported personal do you guys all know that quote okay so so Elon making Tesla competing against weo where he has been proven wrong so far but I think he'll be right in the future um he said about wh with their like million dooll Hardware where they insisted you need lidar which is data that tells you how far things are um that liar is a crutch to solving South driving and that all you need is camera data because humans don't need a lar to tell how far things are uh Tesla CS probably
(51:49) shouldn't need them either um so in the same way I'm saying self-reported personality data is a crutch uh you probably don't need it to know human's personality you can probably get that through any way besides that and so any critique someone has of the accuracy of self-reported um personality inventories I will say like whatever we'll get it a different way I think that hypothesis of yours about self-driving kind of contradicts what youve been saying about more high resolution data being better um it's it uh that's contradict because
(52:22) it's not really a contradiction uh I think it's more just like I don't think you need that's just like yeah more data is always good but if it's falsely reported uh I'll just get it a different way I'll get the info in a different way and maybe the glider doesn't scale as well because you have to build these like model more information is always good and I think what the eventually a model would be able to know the kind of people who do lie and in fact this thing already does kind of tell you when someone's lying so I found
(52:47) in here that when people answer a question in a way that is too Divergent from the way that other people around them answer questions uh you usually find that there's almost no way that they could have a self-consistent ideology with the way they answered every other question and the way they answered that one so either they misinterpreted it or they weren't willing to be honest about their answer to that question so in the data when people lie you can still find it that's really pointing at you and to you saying
(53:12) I probably will just keep ingesting the liar crutch of self-reported data and I just don't have to take it seriously to get information about a personality in fact I'll only get more information that the other kind of person would lie about that specific question solid solid why is Li line like I I kind of VI it the same like it's like I it's just extra data yeah and he's saying you want more data get more resolution I'm saying um great then I don't have to throw it away like I said with the solution to him I can just process it in a way that
(53:41) doesn't make me get less information actually gives me more acors okay the cultural Trope hypothesis so this now we're on in my hypothetical thing we're now on the third section the bottom left actually the second before that we're going to get into both of those sections so he's talked about embedding human emotions embedding human personality um but these are still um less like are you a trait one personality trait no are you yeah exactly you're a combination of them um if you combine several traits you might
(54:14) get a Trope um do you have like a Trope that you think you relate to or like do you have a like a favorite like movie or TV character yes uh this is going to be crazy um what's that one movie with Kevin spy oh American Beauty amera he was he Hest not lying he's never going to lie on the person are you I wasn't planing on looking up the troes associate with each character anyway um Kevin spy has tropes associated with him um in the American Beauty and there's probably one of those tropes that he relates too um and that
(55:06) is like what's that I don't know if or you're drawn to it in some way I think it's a very interesting character yeah period and that's it yeah you're drawn you're drawn to it in some way um so that so that trait either could be one you have or it could be one you're drawn to potentially in a partner that you'd be trusted in but either way combination interesting yeah some people are into that like you know CNC so um anyway like not going to explain that one yeah you can whisper it between yourself or Google it um maybe it'll be another just
(55:43) imagine if you had a relationship with Kevin space from American Beauty and you liked it what you be into and that's what that is okay so the cultural trope hypoth is the idea that if you combine personality traits um you need to get to like a a higher dimensional representation of Personality um and that would be what a Trope is um so this is the Lex hypothesis this is actually what created the Big Five personality traits and the Dogma in it or the proposition in it postulate is those personality characteristics that are important to a
(56:22) group of people will eventually become a part of that group's language and my cultural Trope hypothesis my postula is all important traits will eventually become tropes in a group's stories um so if there's a Trope to know about in the distribution of WS amongst people they probably wrote a story about it and they made a character have that Trope they probably even have that Trope repeat across characters with different combinations of those tropes for different characters this is probably just a thing that happens same as
(56:50) language and it's actually a higher dimensional thing because it's not a Trope is not a single personality trait it's actually a unique an interesting combination of personality traits that's the cultural Trope hypothesis so where would we ever get data on tropes uh so if a person can have multiple tropes is this just sort of a you know different basis for the same Vector space of Personality whether you base it on traits or base it on tropes I don't think so what I think is is that tropes are combinations of traits and
(57:21) characters are combinations of tropes there's a lot of pro that you see around you but in movies are not very represented because the writers have giant spot there's never a movie character like aized girl but there's a million of them struggling brighter yeah nerds took over Hollywood and they made movies where bullies are the bad guys how do you map like the distribution yeah so the data set is imperfect and I think this hypothesis is probably not 100% correct um but it's a place to start and I think a way that we
(58:01) can start to fill in the manifold of possible characters with possible Trope combinations because what you just said was actually not a Trope necessarily It's a combination of tropes which is a character so you said it's like a a blonde scientist who's smart um though this might actually be I guarantee this is probably is a charact type of character Trope on TV Tropes but you're probably right that there's Spirit characters like this but each of those is a Trope and you just combine them into a single character the way that we
(58:31) can fill in the manifold of characters that are potentially under represented is synthetic data so we can already have uh Chach BT generate hypothetically what would the kind of things that would be said by this kind of character be if we need to get additional data to fill in the manifold of information about that kind of person we can create uh theoretical Hollywood writers to make these stories that have not been covered do that I think Marvel will set you up for life is this is this analogous to personas
(58:59) that marketers use what's that is this analogous to personas that marketers use um like are tropes personas that marketers use no I mean this this Notions that similar finding out a character let's say a combination of of traits into into tropes so marketers assemble personas out of different uh uh combinations of of trait from a marketing perspective I'm wondering if it's it's similar I don't know about like marketing well enough to know that but it sounds correct it's the same way that people I think probably most I
(59:33) think probably marketing people do it more explicitly uh because they have like kpis to meet uh whereas like I bet like Hollywood writers probably do it a little less explicitly I bet Marvel does it more explicitly um or at least to Disney probably is closer to marketing than pure Cinema uh in terms of like we need a character that we can turn into a toy and like one that like the kids will like and then one that people feel is like the big bad somehow they have one one thing that them I was saying how how would you
(1:00:00) define Trope so Trope is a combination of traits that's uh and so here we have one Trope um on TV Tropes there's the Wonka right and you can think that there's many characters who might be Wonka though uh this Trope is specifically named after L Wonka um I think the Talk's too long for me to give a whole talk on my TV Tropes research uh but there's a lot of good questions you can have about this but this like the Wonka is a genius nut who should biologically be fire but he won't because he runs the place and runs it
(1:00:27) well this character is typically male utterly confuses those around him and leaves them wondering where he is already there the na so this is also where we get into cool things about TV Tropes data that we have is that we and this is uh something I'll get to later but we not only have data about the tropes and characters that can exist we know from stories the relationships between characters with certain kinds of tropes so here we get to the naive new newcomer that's a Trope right uh which could also be like encompassing a
(1:01:00) character could be like na newcomer plus few other things will think he's just nuts until he proves himself by doing something genuinely amazing what if I had like a dating website where I had mapped one of you as a Wonka and mapped one of you as a naive newcomer and I told you in advance when I'm pitching you the relationship that some some outcome like this is likely to happen between you that I could predict in advance before your date these are the relationship Dynamics you two will disc discover because I know you have this
(1:01:28) Trope you have this Trope and in cinema movies whatever people with these two tropes when you draw an edge between those nodes often have these relationship outcomes this is a thing that repeats over and over in Stories the the Trope hypothesis that like this is actually we've already mapped a lot of this information the differ um not exactly sure actually those might be like similar enough that I don't I don't know where I draw the line yeah I think archetypes and tropes are actually very similar I mean this entire Paradigm
(1:02:02) is predicated on traits being um stagnant um do you think state state like traits are actually stagnant or are they like inquired over time no yeah they definitely can change and some have a lot more movement than others and some have a lot more affordances of like paths they can go through than others and I think the other traits determine what the paths are available to that and this is probably something I don't have enough images to discuss uh but good good concept and yes we have to like track the drift of traits we have
(1:02:30) to even predict maybe where it'll go based on what we know anything else all right you guys get what a Trope is uh maybe look through TV Tropes um if you get a chance it's really fun so TV Tropes is something called a bipartite graph a bipartite graph is a thing where there's two categories of things and in this case it's characters and tropes and you draw a line from a character to a Trope and like here here's all the tropes like one character will have like 20 tropes associated with them 20 lines 20 edges and there's this
(1:03:02) uh paper called node Toc um and this is a paper that uh this is a actually a GitHub repo hand paper that allows you to turn a graph into a vector space um the way it does that is it does random walks along this graph and it sees how often or how far is the distance for one node to reach another uh what that means practically is that if you share a lot of tropes with another character you'll often reach each other in short time uh and so characters that are similar to each other will be near each other in this aming space and then you you map it
(1:03:36) down to two dimensions and then you show pictures and boom this is the character tropes um embedded and mapped in Dimensions Sheldon Eli loer Abed Sheldon and Abed does everyone know who Sheldon Abed is okay raise your hand if you do know who these two people are that's enough okay yeah uh you guys get these um from Star Trek and Doctor Who are similar um you can see all these like prehistoric creatures are near each other CU they share tropes right there's tropes that these people share enough that they are similar characters and uh
(1:04:09) Galler from Harry Potter Count Dooku and Thanos are all near each other and in this room right now all of you who are similar to each other would be near each other if I figured out all the tropes that each of you associate with and I did no toc on you and I placed you in the room you would be standing next to people who Shar TR tropes with you and if I did these tropes along personality Dimensions versus like identity Dimensions that would change where I placed you but if I did them on all dimensions then you would just have a
(1:04:36) person who's the total sum of most similar to you and as you're doing analysis on labels of person relationship outcomes you can choose or even learn have the model learn which dimensions are highly predictive versus not of relationship outcomes any questions beautiful isn't it yeah a quick question about like opposites like um so who would the opposite of Thanos be in this like in this in this space I think when you get into high dimensional spaces for example in clip space uh when you type a prompt mid there's no
(1:05:08) opposite image there's things that are orthogonal in high dimensional space everything's orthogonal to everything uh so in high dimensional space there's not really opposites or even things that are like symmetrical across some axis because everything's orthogonal um so you kind of have to lose something like an opposite oh do you have a thought on that you're the map guy did I say it right yeah okay soard are right next to each other but I would I can imagine different visualization where it's Bard and Ern Shon on one end the Doctor Who Would for
(1:05:43) G that's what I was saying earlier yeah that that depending on the dimensions you choose certain people would or wouldn't be near each other and what a model would have to learn over all these Dimensions is which ones create geometric spaces that are highly predictive of relationship outcomes and so you might discard certain ones like probably people being dinosaurs probably is not very relevant to us um but people being villains are not probably is uh have you done like vector arithmetic on this model and checked
(1:06:10) that like you know character plus trait I have not and that is what I would like to spend the next few years doing is character arithmetic on this kind of data uh because I love it so anyway um so here is also kind of my earlier slides where I was showing the road map this is kind of this also rederived here are less complex um things and like more complex things that like basically it's like lower dimensional information like big five and there's like basic emotions have um you guys get the idea take picture you can have the feel yeah um
(1:06:50) broad attribute um friendliness but then tropes childhood friends tropes vitriolic best buds right so I'm showing you big five to what I'm talking about low dimensional versus high dimensional representations that's such there's such a thing that exists when you increase the dimensionality is that you get more information trustworthy only friend versus the reliable one curious the cute Bookworm versus the Watson helpful the broad attribute the caretaker versus we help the helpless you can get more information
(1:07:23) when you zoom in so if you say oh my friends they're both helpful people so I'm going to set them up cuz they like each other but one's this and one's this you might get a little bit something different than you expect the geometry only works the data only works if you get a high dimensional enough resolution um here are some male and female tropes um some some uh woman in southern France um did an analysis of male and female tropes um at some University and uh yeah Miss fan service does everyone know what fan service is
(1:07:55) you shouldn't if you do um there's fans of women in shows and you can imagine what you know giving them what they want is so um thisy secretary yeah these are all uh the female tropes like basically each character has either a female male gender or something else throw away there something else's and let's like find out which tropes are more corelated with male or female characters these are the most female ones these are the most male ones motivated Fe that's interesting um robot war cure for cancer Evil Genius Grand
(1:08:29) Finale we need like a like a woman robot war going for like a grand finale like three of the men one don't even sound like they're about humans at all yeah which one robot war cure for cancer and Grand ping maybe it's like the guy's curing cancer but phrasing it is cure for cancer yeah take it up with the editors at TV Tropes um they're they're like Wikipedia editors like they're just like adding data all the time um here's other information um oh here's one quote the top 50 Tres this is for books I think um or film and TV genres
(1:09:06) um so you can see how gendered oh wow um musicals are on TV is less so than they are in movies um you can see horror TV is more gendered than movies like how much it plays into gendered tropes um what is this Sci-Fi TV is actually not gender it's actually reverses gender roles um interesting anyway data exists um personality database.
(1:09:38) com so you can actually we have more labels about this even mapping to things like Myers Briggs mapping things like the anagram uh we know like characters and their traits we can map things to other things which one you notice oh anxieties and anog Stakes sorry yeah huh what is the six again the one with anxiety yeah yeah and and oh and like this Seven's like excited one yeah or like the wild one oh Minecraft is youu is inj in a thinker apparently that's weird um I think of him like an eight he's like killing people eating up it's a cool website more data you can
(1:10:23) always get more data so um on the gender Trope thing yeah it seemed like the genders were or the tropes and their gender wasn't really reflecting anything about people and it was more reflecting the space of uh the biases that the writers of fiction have disagree but anyway we can move on okay um so another thing we can learn um is through language and conversation I think I kind of this probably should have been before the tropes thing but I showed you a bunch of stuff about um yeah I I disagree that all uh gender
(1:11:03) biases are socially constructed I think some are reflections of reality I'm not saying that yeah yeah anyway um we could that that can be a a uh a political talk upstairs or something uh we can even have a 30 minute uh going over all of her work on gender bias and film she publishes a lot of stuff it's shockingly good um so yeah uh okay embedding through language and conversation so um we talked about like personality tests a little bit or like tropes but I think that the greatest source for some people especially the
(1:11:35) kind of people in this room is just like we have their chat history they've discorded they've Facebook messengered they just like blah blah blah blah blah blah talk all day and like we should be embedding most of our stuff through like kind of unsupervised learning over tokens that people generate one hypothesis I had for example of how to get a human personality embedding is if I had had for any one of you all of the tokens you've ever generated and I had a GPT model with a conditional embedding I don't know ra of you know what a
(1:12:04) conditional embedding is okay um a conditional embedding is like something you give to a model that helps that isn't its parameters learned over all data it is something that is specific to the thing it's doing for example uh when you type into mid Journey you actually generate an embedding a conditional embedding on the thing you typed and all of parameters are already Frozen it just knows how to generate any image period so if you do um conditional free um Mid Journey it'll just generate un image from anywhere it's like dreaming
(1:12:40) um if you give it conditional condition that will determine its path it'll be attracted to a region that is similar to the embedding of the words you put in um in the same way if I had a GPT model and then I gave a conditional embedding um of like and then like I I trained it where that conditional embedding was whoever was generating the tokens then you might imagine that what this GPT model has learned is number one how people talk and number two how you talk you specifically and so that embedding would become an embedding of your
(1:13:15) personality I think there's one thing you could say that is like if I can predict every single thing you would ever said like if I can look at what you've said and I can always predict the next token I have a pretty good model of your personality I know you pretty well if I can predict everything you're ever going to say I know you pretty well um now you have to do all sorts of ml stuff like make sure you've not just memorized the the training data set blah blah blah but there's an there's a theory that if I can predict everything you would say
(1:13:42) and I know it's you I have embedded your personality and so that's the idea so I think of this conditional embedding like if it's say language for example there's the unconditional thing about yeah English sent ver AR like this and the conditional part is would actually make conversation different than anyone else uh so like a girl text me yesterday saying hey do you want to come over how how should I interpret that well tell me about her that's the conditional embedding um you can't give me a great answer of what
(1:14:12) that means unless you know about her right this is not a real thing um but like that's the conditional that question that question tell me about her is the conditional Med because you are a general model that has mapped all human conversational pattern and you're being asked generate me a hypothesis for what this this language means and the person says give me a conditional embeding of who this person is tell me more data about them I want to do that with tokenization go what are some of the ways that the embedding would actually
(1:14:39) be used or Incorporated mechanistically like would you just add it to the other embedding vector or what I think um actually yeah I'm not really like a person who's like uh creating like the architectures for these things I'm not exactly sure what it would be but like however these models once you have two once you have the embeddings of every personality can't you just start doing you know a second round of learning on the differences between embeddings of people whose relationships you've observed and then you get a good
(1:15:07) prediction on you know the high dimensional Vector difference in two people's embeddings can predict their relationship Dynamics it sounds right it's too complicated for me to internalize right now yeah uh there's a lot of good things you can do with these embeddings and I'm just saying this is one way that you could use just tokens from a person to generate a personality inment okay um for example you can know male or female just by how often they use certain words oh wait let's actually I don't need to cover
(1:15:38) this um okay raise your hand if you know what Bushido is any women that want NIS I lived in Japan one woman what's B yeah you don't know yeah uh that's predicted by data by the way um okay Bushido is like Samurai uh style lifestyle of fighting okay now boys turn to get uh what is a bando raise your hand you know a bando is La guy and fashion guy okay what is a bando yeah sort of thing two best dress guys know what it is uh okay what about a Damas raise your hand what is it kind of clo traed from the East historically
(1:16:37) a very fine cotton okay apparently no guys know what this is what is a pum raise your hand if you know what a Peplum is one woman two woman three woman four wom no man know what a Peplum is including me how is it spelled p e p l okay and let's get no women again oh um okay raise your hand if you know what pzo electricity is we get a woman see data tells you okay what is electricity it's electricity generated by a movement all right what's that okay um author what what's the talk after M I'm doing it masc femin in this room no no
(1:17:28) what's the one in this room I think it's the um keeper no it's do doctor okay yeah I'll have to stop for him then um okay we're going to skip through this um we're get this this is like things that have been learned in the embedding space of language models right so Lauren was born in the year of blank El was born in the year of blank it can literally know the log likelihood probability of different years so this is literally just the next log likelihood probability difference between the likelihood of Elie being this year vers uh Lauren
(1:17:55) right we can learn a lot this is actually a thing from Google p pair um you can also see changes over time in language uh so you can see like how um in the year he she blah blah blah you get it I need to get to some other stuff okay language right does anyone know this paper raise your hand yeah one of the coolest papers ever generative agents interacted to macro human behavior they made a little town and they gave prompt engineered every single thing to have a personality who knows what this show is a million come on who
(1:18:25) knows what show this is black black mirr this is the one where they have they Sim get simulated doing dating oh yeah yeah this is them realizing they are little freaking dots in human embedding space this is what this is um they got embedded and this and like they got simulated great great why can't we simulate relationship outcomes using language conversations blah blah blah uh chat Bots doing this and we just simulate uh your dates over and over and what do we do with that so here's a simulation of this was a a game
(1:18:58) theoretic um simulation of people sharing resources and figuring out you have different policies different personalities what are the outcomes likely with different um starting conditions and what I want to do with relationships is predict you know if I match you with this person here's the Rel relationship outcomes you like to discover maybe you guys will get along but in the all universes this is where it tends towards but if I pair you this person this is your one in a million you're always going to get along with
(1:19:23) them even at a rocky start I'm going to tell tell you like this start keep going because in like most simulations you meant to be with them only language it was trained on was actual relationships of people and then the words they use yeah because if it's trained on archetypal characters then again we might have a completely the Obama picture we need more resolution for it to be accurate otherwise it's just wrong not the resolution but real life couples I the feedback oh for for saying what relationship outcomes happen like
(1:19:59) instead of these corny dialogues from Hollywood movies maybe 10,000 examples of actual couples that worked out for 30 years them great point I I wish I had more time I have to like rush through everything so maybe actually no more questions because I still have like probably 30 more slides okay Mees and political compasses of personality and relationship outcome man communication tools um this is if you're a guy with mommy issues and you are a guy who's the bottom not sure what that means this is the girl you go
(1:20:30) for yeah these are tou relationships don't cancel me or do whatever this is what they look like um I'm not going to explain any of these cuz they're all  up um but yeah that's the political chart with the kind of relationship you seek um okay um that was that was the slide that like we decided not have a reporter here like we don't one know here anyway um uh basically what I'm saying is that if you have these high dimensional embedding spaces what do you do next you need to make it communicable um this ASM tootes
(1:20:58) towards BuzzFeed ification of uh communication of personality but for a little bit more intellectual people it's just political compasses or sightless intellectual memes um have to rush through it here's more political compasses um here's like you know uh major morality mapped uh don't have time for all of it they're fun this is like EA tweets so all information can be mapped all Communications information we can like map People based on their tweeting right we know where they are in these two dimensions of many many many
(1:21:30) dimensions um people also move through the manifold I think you talked about that people move through the manifold um people personality changes so you can see this is the person talking about how they moved around politically and what activations made them move uh conservatism yikes Bernie 2016 government's kind of whack some capitalism is useful Ron Paul is Lord taxes don't seem fair taxation of Step m is B Li IND is pretty cool um we even have more this person in 2014 was up here and moved around to here oh
(1:22:00) politics are a thing go America and Jesus right Mom and Dad Capital made pitfalls but we just got rid of two party system reformed everything would be okay well there are really some UND unders people in our communities the CIA did what either we kill all social dominance hieres or they'll kill us first here's the pamphlet uh that gu um stories have shapes as people move through the political Compass chart as they have these Revelations about what their personality and what they believe it has a shape Kurt Von good already
(1:22:27) talked about There's the Man in the hole a guy the main character gets into trouble and then gets out of again and ends up better off for the experience man in the hole he pops out um the main character starts off poorly then gets continually worse Kafka with no hope for improvement yeah metamorphosis um Hamlet has lifelike ambiguity that keeps us from knowing if developments are good or bad Cinderella uh one like her life gets better blah blah blah blah blah it all comes crashing down she lost it all and then back up to Infinity she's happy
(1:22:55) forever um the Old Testament like we get the Garden of Eden and then someone ate an apple and now we're sad um this is it this these are the stories they have real shapes uh you can you can make this mathematical um here's a woman even explaining like really herself how she moved around and I think we could get people to explain their life stories to a chat bot or something too to know how their personality has changed throughout time um just another one of that this also implies that we can influence transition along
(1:23:28) personality Space by knowing what worked for different people to move them so you might say Matt why are you trying to make everyone in the region of the political compass that you want them to be and I say I'm not actually I just want people to not be in the region of Personality that makes them in unfit for relationships um so some people just literally are in if you imagine a political Compass of like relationship orientations and certain regions just always have bad relationships can we figure out what the transition paths are
(1:23:54) like a marov chain to move them into regions based on where they are into regions that make them better at relationships and I think the region they're in will determine what what paths actually can be moved through versus where there's Cliffs uh cognitively here is JRE like you can almost imagine this like a high dimensional thing plotted down into you that's why there's like these weird shapes this is type of guys like Tech Guy um like CEO like entrepreneur like in the B like Lon this is him labeling if he would or wouldn't he's like
(1:24:24) vaguely bisexual in his videos um so like I don't know he like yeah jre's a pretty wild guy um but he labeled this so this is human Lael data over a personality manifold that's what this is he's saying would he have sex with them wouldn't he said not even as a friend would I have sex with them so he wouldn't do it with like a killer sociopath he he would don't know what an istp is um he he e boy is ideal a soft boy is ideal e boy goes into like gamer goes into Normie goes into soft boy goes into a Dandy goes into like a preppy guy
(1:24:56) uh yeah he has a crazy video on all this but point is there is high dimensional personality space and people have preferences over this and you can label it and learn those features different types of labels so this is an outdated model of Twitter's cancellation policy um how risk of being cancel on Twitter due to your political views here's how risky you know it's like the traditional political gumpus chart uh so like yeah this is probably very outdated of like what's going to get you can on Twitter probably actually nothing will now um
(1:25:24) but yeah yeah this is like this is labels over a manifold right and so you can imagine that same so archetypes uh we talked about tropes archetypes also thing you know magician warrior king lover warrior king magician so like a lover at the extreme becomes an addict uh whereas a lover you know not loving enough becomes impotent a magician becomes aach too much of a magician becomes a detached manipulator too little becomes an innocent person a warrior is either a masochist or a satus a king can either be a tyrant or a
(1:25:52) weakling um there are like these these archetypes can be derived from high dimensional personality data and communicated to your users so as they go out in the D World or they just are in the D or they can know who they are so what I'm saying is basically don't use all the things that look like um data science and statistics when you're talking to your users or convincing people about high dimensional stuff use archetypes use memes to have them understand what you're talking about how what you've derived about who
(1:26:18) they are um Myers Briggs is a type of archetype clustering personalities along these axes personal possible logical present you get it people already do this stuff um people do it with emotions too uh you can even Buzz feedify everything so basically can we just make a real data BuzzFeed ification of all these like high dimensional personality edings um you know BuzzFeed is so dumb where is like you're a Michael Scott or whatever but like what if it was true what if I could actually make BuzzFeed personality quizzes and they were
(1:26:47) actually real information about reality um books can predict the location actually forgot to put the thing in here there's crazy crazy stuff about which book genres correlate with which Big Five personality tra distributions it is pretty wild but I don't have it up new friend Discovery I also these slides are not fully uh fleshed out but there's an idea that your current friends are similar to the people you're looking for and that you can like get these kind of social graphs this what I think Facebook should have
(1:27:13) been doing is finding people with personalities that are most similar to the ones you talk to most and like matching you to a new friend along the chain of social connection uh this was kind of the intu like I think Hing or Bumble or something like one of them was like all about finding Mutual Facebook friends that was like their prior on that and I think it's like somewhat true here's what a social network looks like you can see it's a real thing uh here's what another social graph looks like this is like Game of Thrones social
(1:27:40) graphs are real things and like we can we can use them as data to like do the cold star problem like moving people around um to like find them Partners um here's the structure of romantic and sexual relationships at Jefferson High School a famous graph uh these are people who've had sex with each other you can see there's just like these are like the like all the normies just having sex with each other um and then these are like people who like have their weird little side clicks why Jefferson in quotes I don't know I know
(1:28:05) Jefferson High School is just in quotes this might even be hypothetical I don't think it was a real it was a real high school but they gave it aous okay this guy has a lot of female Partners um yeah you can see this is like a social graph researchers develop algorithm to maximize friendship acceptance by strangers on social network so you can also just get probability of like people who they're friends with and their personalities can predict um who they might be friends with but will it always work perfectly
(1:28:33) no but it can offer improvements um and so it's the idea that this map of the world is not the actual territory you're in and I think that's always an important lesson to learn when you're actually like creating these embeddings and trying to learn stuff is that it's not meant to tell you that like everything is true all right um attractiveness we have to speed through this way faster or could we and could we maybe move the other guy to like a different room because I think I probably need another like 40 minutes or
(1:28:58) um Let me let me work on that okay yeah all right attractiveness um so a major component that mayor who raise your hand if you were thinking about this the whole time or at least it popped up in your mind about mutual attraction raise your hand really like do it honestly if you actually thought about it okay interesting yeah uh okay I show you your soulmate and you think they're ugly not good right yeah Matt you up that's not my soulmate I would never have sex with someone that ugly uh why' you why' you give me a soulmate that's
(1:29:25) so ugly uh sorry about that uh Matt I trust you with all my data sorry about that yeah uh canceled like removed like refunded like I'm done my my startup's over I I need to have them actually be attracted so uh can this be embedded too can we solve this too I have some uh slides to help you figure this out I want people to upload oh so here's like one weird idea I want people to have to upload several profile picks um including the ugliest one of themselves they can find so people always like puff up their looks so one way I'm going to
(1:29:56) reinforce this is that I'm going to use an aesthetic scoring model where I rank I have an AI objectively say how attractive each of the photos they upload is and the difference the discrepancy between the top rate one and the bottom rate one has to be a three-point discrepancy um which means people must show they like they must show a picture that is like three points lower in attractiveness than their top photo um and this is a way to have people actually be honest about like what they look like on a good and bad
(1:30:23) day and it's like going be fair for everyone where people aren't penalized by the network for doing so the customer support someone is just like too hot that they can't get it they're like I literally tried so we said like probably this would be a race to the bottom to find the the the face that the model thinks is ugliest that other people find the most attractive it's like you go but they're like oh so in a cute way so anyway um probably someone's just going to have like some ugly stuff around them and we probably just like it probably
(1:30:49) wouldn't work but it's an idea like their friends um cuz people are unespected attractiveness themselves um okay oh yeah I didn't fill this in this is from Christian rudder's thing um data CM where he said that in Okay Cupid they found that high variance people got the most matches um so here's an intuition like if me and Thomas uh R Jackson here each we give him like a seven um but then like Thomas gives him a three and I give him a 10 and Thomas gives him a 10 and I give him a three who are we going to message
(1:31:25) I'm going to message him he's going to message him Jackson gets no messages even though he has the highest average score you need to be high ear people message their top choices um so in the same way to be to be disliked by some is to be loved all the more by others that's one of the thoughts about attraction is that you need to be high variant um just a little intuition pump on that um the individuals get the most affection will be the polarizing one I said like an idea like maybe people's exes can tell us what's bad about them
(1:31:51) so we can see what unique things may make them likable um because maybe people's actually can just like help us understand their negative traits that they won't tell us that like other people actually might find them Dearing in general I thought we need a lot more like getting x's in the loop so this is if there's a really selfish guy or there's a guy named Kyle and he's like all these girls in this app are ugly or like I only Swip right on these hot ones and they never match me and we want to say Kyle buddy we did statistical analysis you're
(1:32:26) not swiping right on your looks matches okay like there is a distribution of attractiveness you fall in a certain area you're only swiping you're swiping right on women who we've statistically modeled to always swipe left on someone who looks like you here is a woman who is the highest um like our prediction of the highest amount that you are attracted to her where she has the highest mutual attraction to you how about we show you them instead um let's have you like like actually go for the people who are into you we are
(1:32:57) effectively trying to Moneyball mutual attractiveness uh does anyone know everyone know what the movie Moneyball is uh ra if you do not know Moneyball wow everyone knows the movie Moneyball yeah basically use data and statistics to find the guys who make your baseball team super good that costs the least money because everyone thinks they look weird or something um like this one guy just like looks Goofy and I don't want to like have him our pitcher even though he's really good as pitching I want a Money Ball mutual attraction I think
(1:33:22) that uh if you are taking like um relationships only from a purely sexual attraction standpoint there are so there's so much un um achieved Alpha in the world where there are two people that both find each other really attractive that like they most of the people they could date they don't find attractive and we just need to like pair those people up and I think we can do that through data science um one thing you need to understand is we can now generate the manifold of all possible looks we don't need real human photos we
(1:33:51) can show people interpolations ofr faces to really precisely hone in on exactly what they're attracted to um that should be a filter on dating s just what exactly are the regions that you are and aren't attracted to um we can see that you know just as you guys have learned throughout this whole talk manifolds are real and that being close to them means something and that moving in certain directions along them means something um there's like a bunch of ways to look um yeah faces eyes noses lips um and these will all be instead of one hot
(1:34:24) encoded embedded in Vector space we could even use this model's embedding space as the vector space for all of our looks so if we just project it into that embedding space we already have an embedding space of human face we don't even need to invent it it already exists so if you give me a label where I like how this guy looks that's a label in embedding space you've already given it to me um okay after we try an aesthetic model on a person's preferences we could then get the average embedding of their highest rated
(1:34:50) things so here's the interpretability part right remember I keep saying that like if we're going to know something about we have to prove to them we have before we get them to invest more clip interrogator is a thing where you have an embedding space and you tell the person back here's an image embedded what is the caption and in the same way I could have a thing where I embed the photos that you like the regions you like and I could have it turn it into language unilateral or I could have it turn into language to say here's the
(1:35:18) traits you like wouldn't it be so nice to like swipe right and left on faces and that get projected in embedding space and then I have a model that can map from image embedding space to language through clip and it says um dolma right domma you seem to like guys who have like this shape nose but only if they have medium length hair or you like blonde guys but they have to have a man bun like I'm starting to describe the regions of embedding space that you prefer and here's like by the way in that inan space I have the
(1:35:49) average ratings of all people like you or all people in general and I say here's a region that Everyone likes that you gave a very low score to and here's what that is it's guys like this and this is the kind of magic where you can prove to your users we have actually learned something about reality we have figured out um what you like and don't like that differs from everyone else and they'll say wow this is money balling mutual attraction um the aesthetic scoring model that made St Fusion beautiful let me show you this this is actually a
(1:36:18) project um from uh I did like some work uh with this group and this is a thing where this is like a subset of lion 5D and they train a model on some labels of images and then they put all images into this space and this is like the model's rating and these are like the ugliest images apparently and as we scroll through we're going to see that models can learn which images are beautyful or not these labels are only about beauty so we're getting higher uh we're getting into like 3 to.5 3.
(1:36:53) 75 we're just going to scroll through quickly you can see it gets better some are also un hot LED I ignore those um you can see it gets prettier and prettier these models can learn what people find beautiful it did have some weird things at the Tails but again just higher resolution more data um so yeah we can actually learn preferences over images even when we've never had a label for them before just by having them in embedding space uh here's Beauty scores uh we also have like different races like asian females Asian males here's like them getting
(1:37:23) more attractive according to a model that is like Based on data uh probably hurt some people's feelings anyway like this is just we can do this um here's like also different ethnicities even more with like looks um what's that yeah here you go actually technically yeah these labels are also one hot and coded we probably could just have all ethnicity be in embedding space you know or like now it's just down to the family level um anyway point is uh we can label this um here is a model that specifically again I told you about
(1:38:00) conditional embeddings here's a diffusion model with conditional embeddings for be a Target Beauty score and so all these embeddings are kept consistent except for the igen vector that would maximize Beauty with while changing the fewest other things so apparently a f boy is hot this guy has to become Fen yeah that seems to be uh so this person probably should transition the restroom okay they are I wonder if you could even discover in like embedding space some sort of like gender dysphoria maybe that's even like
(1:38:35) a thing that could be discovered through human and personality embeddings I'm telling you guys human personality embeddings can change the world um interactively assessing disen disentanglement in Gans this guy has not know his famous as he should be I reached out to him but he was all the way in like Indiana um um Korean guy in Indiana I think he's that like Purdue or something anyway this thing is amazing so we can have a walk through the latent space of faes and you can say exactly when you stop becoming attracted and we
(1:39:05) know the exact traits that are associated with that walk in eding space which like big lips are changing blonde hair blushy eyebrows what is co-varying in the latent space so I'm telling you instead of showing you real users photos I can just show you things that are optimally selected to most of efficiently figure out exactly what you're attracted to so we can Moneyball things where you can get matched with people that you're attracted to that are attracted to you um in a way that uses as few images as possible to figure out
(1:39:35) exactly what you would like and this can be done in the latent space of things like style gam and so this is how we can do the data analysis um the question is where in lat and space does attraction begin and end here's like more what we just know like um her traits and how they change where does attraction change begin end whatever there are regions and embedding space of looks that you like and don't like certain amounts back to this guy what if I actually generated this for you you know instead of being a psychic on Etsy for $30 what if I am a
(1:40:11) dating website that ran you through this process and literally did generate you know here's this person sharing on Twitter here is my exact regions of looks that I like that see that are known to maybe like me that other people don't like as much anyone on Twitter look like this and so you don't even need to match them we're making tools for people to go out and actually do this as a real thing any questions about this I think that how you people part to meet but it's not Maxim yeah and then reverse search for
(1:40:48) like my IDE the tra are just not going to L out like might just be the wrong approach I it's probably number seven on the list yeah also probably more important for men than women I think I think it's actually the opposite or depending what you meant by like I find that in my experience a lot of guys they'll go out with people who are conventionally attractive no matter what but a lot of women they highly index for a specific type or couple specific typ um yeah whatever prior you guys want to put on what you how you think this tool
(1:41:21) will be used is is a tool that be used and we will learn how to use it for our benefit but it is a tool that can actually help us optimize in this space in this whole talk I have views and priors but really not that strong my only prior is that we need to make these tools to allow us to take every one of these directions whatever you think or whatever you want to have a machine model learn um on data we can do if we actually take this approach to solving these problems um here's another one just more beautification of faes
(1:41:55) yeah same moving along the vector of beautification but what if it was along your personal Vector of beautification anyway um we're money bong in right might be enlightening for people to see who they value the most that is the least value by others so they can strategize on how to get traits uniquely valuable with them and the partner Moneyball we Moneyball uh attraction um we can let someone when searching decide how much they want to factor in Universal appeal versus ones they've been specifically measured to like um
(1:42:21) yeah uh there's this one website this kind of aside but like there's this one company that literally tells you you so you could also put changes you want to make in clip interrogator to figure out what the words are l so she thinks that her nose is too big we find her nose is normal it's her weak chin that makes it look big um so you could have these models also tell you interpret exactly what's going on um so that you like don't misunderstand what you think about yourself or what you think about what you're attracted to um this is another
(1:42:50) funny one an uh can I say oh sure yeah one of the founders of photo fer um and co-host of the Symposium uh helped make this site um and it tells you how like on the distribution of traits these are labels on this guy's so also he would pass the test right of like three-o difference between his best and worst um this is like you know in the threes and fours this is in like the eights and nines it's like we could literally tell like tell this guy on Bumble or whatever you have to upload both of these photos and we could like algorithmically
(1:43:21) enforce that um but then there's also different so and took the algorithmic proach and there's this random guy called wheat waffles you guys wouldn't believe I I also paid wheat waffles on Fiverr to rate my face um it was pretty brutal um I think he gave me like maybe a five or six that's where he gives most people as he advertised in the distribution um so yeah not a Chad not even a Chad lit but like almost um pre- Chad lit I think it's Chad light by the way yeah almost a Chad light I think chadl actually is like it's like man you
(1:43:53) know yeah yeah so anyway um he makes these videos point is attractiveness statistics distributions whatever you guys get it uh this could be a useful tool geography um what if you had an interest by sub region for you what if we told you which regions people are most attracted to someone who looks like you so we don't have to even show you to these people we can simply guess in embedding space where you lie in terms of your personality and looks and where people in those regions lie and what they would think of Someone Like You why
(1:44:24) is this not a thing I guess morons want to talk whoever that is we also know that there's different ratings distributions in different communities um and we know that certain communities give more or fewer like detailed ratings this can be incorporated into our data but you see good reads like the average rating on Amazon for books is just like poor uh Goodreads actually has like opinions um cultures are different I'm not going to explain this uh but like Vietnam's reative apparently uh Germany's yeah I'm not going to explain
(1:44:53) this I don't even understand it I did understand it at one time but I forgot linear yeah it's an interesting if you take a picture if you want to figure it out later action it's an interesting model of culture okay um here's where more women and men are why aren't we doing Geographic like strategizing figure out where you should go actually everyone in the bay already knows this is the move it's the whole like fly from SF to the B actually it's you have to be two artillery distances away get out you listen to something
(1:45:24) okay I have to finish so yeah you guys get that there's different distributions of men and women might be a where is that Sacramento might be a thing Taho tripto Tahoe anyway you guys get that there's like Geographic differences right you can optimize over geography not just attraction money balling attraction you Moneyball geography I think the pink one is Sacramento I think the other two the SF and San Jose here's a website not sure why she's Asian I think like Chinese people made this website so she's just like know those like uh or
(1:45:54) what's the website that's like Amazon for China you like Alibaba model is like here's my factory like I think she's that's but like for this website like I think it's the same thing so anyway you can literally sort by like women how what the ratio is of like women to man um of a certain age and BMI and height and actually it goes way down there's like a lot of ones for this website why is there not more Geographic optimization uh when I'm on Bumble I like change like regions based on where where I think people would be more
(1:46:24) interested in me why is this not an automatic process why are they not shopping you around to like places where you can know where you're actually optimally desired should be a thing um cultures are different cultur different cultures have different preferences there's this is like you know these Cuisines are liked most like Phil this how Filipinos feel about these Foods they love Filipino Cuisine no one else does um uh this is Italian cuisine is apparently the most popular across all except China they're haters uh though
(1:46:55) actually you could also say that Italian cuisine was just taken from China by Marco Polo in the 1200s so maybe they're just like they have a reason don't appropriate our spaghetti or whatever um point is like cultures are different so like we can optimize over which culture someone from to know that you might have a better shot in this region or you might be more valued in this region optimizing and moneyb human attraction and relationship outcomes um gender differences Rob Henderson had a really good thing uh I'm not really sure like how to fit
(1:47:25) this in the whole theming except just more data about the fact that for example women now think a major reason that they're they're saying a major reason they are not in a relationship is they're not meeting meets their expectations um he's pointing to like this also I have to do the fact that they're college educated men aren't and that women want to have someone who at least is educated blah blah blah like men want someone who's not much taller woman want someone who's not much shorter there's like things about people
(1:47:52) that you can optimize over data on what they're looking for we can Moneyball mutual attraction and relationship outcomes keep that in mind too you know coming off of the like attraction thing this can also be optimized for interpersonal relationship outcomes what people like in another person's personality um here's things like women and men rate higher yeah whatever um okay uh something about actually I just included this even though I'm not sure what the threat is except like uh men's genetic apparently are consistent
(1:48:24) between but like what women want changes um point is it's a complicated High dimensional problem let's get a lot of data and labels to figure it out um I think this is like a yeah relationship Readiness okay any questions before we go into this section I think we're in the final like 25 minutes of the talk so any any questions so far so we've talked about embedding human personality how we're going to do that um and like I think we still haven't gotten to the actual relationships part um because there's details in there so
(1:48:56) one thing you should consider is that in a market where there's supply and demand um matching people who are demanding and supplying more efficiently is a way to improve the market but also improving the goods in the market themselves is a way to improve the market I have uh actually not as much to say about this as I think um some other people at this talk might have but I think that using these same techniques these same data science techniques same embedding techniques we we can make people better Partners better suited to having good
(1:49:27) relationships uh and I think you know a lot of people say I'm already perfect as I am or no one says that like some do I don't know point is some might say I'm already perfect as I am and what I want to do is uh I just need to like have the person who's perfect for me and I'll be happily ever after and I think a lot of us know the truth is is that you need to change too I don't if I give you from God a sorted list of the most compatible people that still might not fix all your problems you might need to improve as a
(1:49:56) person in order to have that relationship this this is about relationship Readiness I really don't have a ton to say except just some ways to like do data science also like type of single guys like some of these guys are in personality space right and some of them just don't have hope so if I find someone in a certain region of Personality space I the problem is not finding them the right person the problem is making them not the wrong person um and I think that's the thing that you have to consider in all this is
(1:50:21) that like uh some people just are more or less hopeless in the region that they're coming from in terms of character personality space uh so like in the CL I didn't even know going be there aggressively online dating not believe married yet colist guy who Peak too early this is yeah um I guess I only had one thing on I didn't even have enough date on that I think I would love to have some people give talks on what they think it would make people more ready and more and be better Partners I think this is the thing I have the least expertise on I
(1:50:52) think a lot of other people are much better at talking about this but I think it's a very real and important thing and my tools are not a pan like P to solve all the problems uh with this you actually need to become a better person too what's your so we have to stop um at 20 after the hour so you have 10 more minutes okay yep all good okay.
(1:51:13) com I think this is what like the label on the manifold would look like where the higher one is people you like more um people already tried to sh comes with like Myers Briggs for example uh so this is just a lower dimensional attempt to like label relationship outcomes I'll post I'll post this uh thing later there were friendship subset hypothesis um so like basically okay yeah this was like yeah I don't have time for that um pitching the match um so there's a thing I was listening to uh the guy who's in charge of Netflix recommendation systems
(1:51:44) and he said our biggest problem is actually not necessarily finding the right video for you to watch it's convincing you to watch that video because all we get is a thumbnail and you've noticed like Netflix actually tried doing things where they change the picture based on who you are that is them embedding your personality trying to figure out what how to pitch it I think if I had a list of the optimal people for you to like attempt to date pitching them to you figuring out how to convince you that that is the person you
(1:52:09) should invest time in is a very important problem to solve if you're having a dating app um there's like also like interpretability explanations for like you can turn a an embedding space into like a language explanation for that embedding space is that one from the blue one brown yeah um people want to be known but what if the app can help you show that you know them get to know each other um things like that um relationship outcome Improvement statistic can't tell us what happened they can tell us what might happen um
(1:52:41) Can personal relation bets help us systematically identify and relationship conflict areas and direct them regions of better outcomes um types of couples this is language analysis uh we'll skip that but just know that this is more ways that we can analyze language and understand personalities um personal Improvement Readiness people just need to get better um to be better oi okay I think we're g to just do app ideas and then be done so this is the major thing I want to discuss I wish we had still stage actually for this because I think this
(1:53:10) is the most important thing to think about when you're making a new dating app is is your app that is social optimizing for fame or love is it about worship or is it about reciprocity is it about creator or communities is it about relationship being Intense or being casual it's about advertising or about direct pay forget that part um some apps are more about love and some are more about fame um here's how couples met everyone's meeting online now uh here's raise your hand if you you know the joke that's related to yeah bad
(1:53:44) bad Minds dir anyway relation dating apps are not working why because they're not making efficient filtering tool we're not doing efficient search we're not doing the embedding space stuff all the stuff I'm talking about don't have time sorry um you ask me after right now what is your ideal sexual relip people are looking for different things in in dating apps um so what if so here's like people wanted e-harmony versus Okay Cupid e-harmony was like this Ultra Christian thing that like trying to get people to have low divorce rates and
(1:54:09) have kids together and we see people want different things in different regions we can optimize over geography um here's hinge Bumble SE seeking.com most popular in Palm train California by the way uh and then Gainesville for to interpret that how you want um seeking.com is like finding sugar daddies um dating have to not appear on nowh they I hate the baring um how how are you when most likely work if we match people we need to have the we need to like give them a thing to go off of in order to make the thing work
(1:54:39) this is like pitching but like Mutual pitching where like we make sure that the conversation is actually going the way it needs to go we need to like monitor their current mood and attitude to make sure that they're not using the app in a lazy or disingenuous way you need to like make sure they're not like just like blinding through the app experience um privacy tradeoff like as people want more privacy we get less information so you need to get someone to tr get them to trust you with at least enough give them enough privac or
(1:55:05) give up enough privacy to actually be known how you do that convince them that you know things that you can help them um this is s for Satoshi uh one of my favorite tweeters but this is an idea that will chat TBT just become the ultimate dating app because as um the number of program or data relationships happen inting use the word relationships but just like as the complexity of data increases software complexity gets so high the number of permutation and combinations gets massive to where chat gbt there might not be any apps anymore
(1:55:32) it's just one app and that generates apps for you based on the exact circumstance you need solved so maybe the ultimate thing looks like a chat GPT M do becoming this God that matches all of us um and that's my final SL I think we have uh we have like three minutes so any final questions what was a joke oh uh about yeah the graph nope if you know you know you might find out tonight any there's a manful market on Bo that's great uh any final thoughts what's the next step in your master plan I really don't know I think
(1:56:17) um I'd love to find people to work with on these ideas who believe in it um I'd love to find people who want to engineer it who have like the engineering skill I'd love to find people who could even be like more like um CEO types to actually handle the stuff that I'm not really that good at um and let me kind of you know be more of like a researcher on the team I'm also like good at the back end but um yeah I'd love to like be working with a startup that saw all of this and said oh we can actually do this and um like this is how it should be
(1:56:47) done and we actually will achieve amazing results with this and uh yeah people actually believe in the vision all right I wasn't here in the beginning so maybe you adjust this but why expect that uh the problem is not good enough uh personality emitting because we had um Okay Cupid and that seemed pretty pretty good not nearly good enough at the beginning I explained cool a long one but I have this theory that even if we had algi that showed you the top mat plan that really work for you I think our minds would have a hard
(1:57:27) time comprehending that F soulmates are all great now I I think a lot of success I see is like in like love is blinding these crazy TV shows where they put 10 people there there's a there's a deadline there's an expectation that in like three weeks you'll be married it's not perfect I don't third of couples actually end up being married which is a way better R than anything so I'm thinking like maybe framing I absolutely agree basically like the frame that the next person who meets the requirement whoever's above a seven
(1:58:05) absolutely agree that a list from God of the most compatible is not the the solution to every problem you also need to change other stuff to set it up right so if everybody have a everybody Money Ball this at the same time what it would Converge on there's a stable there's a thing called stable marriage problem Gil shape and it would yeah I don't know it question I I don't know the answer but like that's a place to start looking at what algorithm ask there's properties to that yeah still probably like 30% of my
(1:58:35) thoughts that I didn't get time to get up or present but Gil shape algorithm is very interesting one notably the stable marriage algorithm does not correspond perfectly at all with the intuition of well happy marriage or what people think when they imagin marriage it just means St proper Gil shapely if it's male P male optimal would Beal it's like it's a zero summ game yeah the way that it's R hope you guys all enjoyed and yeah feel free to ask me questions if I if I can just say that my my talk on masculine feminine uh sexual
(1:59:10) Clarity and intimacy has moved to 420 or whatever time it is now and it's going to be in the Attic not where it was origin yeah