THE SYNTHETIC RUBICON: CROSSING THE UNCANNY VALLEY

E93 THE SYNTHETIC RUBICON_ CROSSING THE UNCANNY VALLEY

Transcript

Dalton Anderson (00:00.556) Welcome to Venture Step Podcasts where we discuss entrepreneurship, industry trends, and the occasional book review. We have officially crossed the uncanny valley. It is almost impossible now to discern between synthetically created images and those that are real. This is the first step of the synthetic Rubicon.

the next couple steps are close. Those would be imitating voice completely and video.

In this episode, we're going to be discussing what needs to happen to cross in the synthetic Rubicon. And then what does that look like when you do? And I have previously touched on this topic many times and four separate episodes. So if you find this episode interesting, please check these out. So these episodes are episode 23.

The Filter Bubble Blues, The Personalized Reality, Episode 63, The Imperfect Echo, AI Voice Cloning and Its Current Limits, Episode 82, Google's Nano Banana, The Viral Trend That Dethrown ChatGBT, Episode 80, Zero Click, SEO for AI Search, Google Rankings, and Agentic AI. The first episode I talked about, The Filter Bubble Blues,

was a

Dalton Anderson (01:39.982) on Google's algorithm and social media algorithms. And then also talked about AI influencers that are becoming more more popular. think estimates say by 2030, it's supposed to be an $8.5 billion industry, which is nothing crazy, but the fact that they're already projecting what this AI

influencer industry is going to be and seemingly that they've already ran all the tests and that it's socially acceptable. Like people are accepting that influencers are their valid peer in society, which I think is odd and I don't agree with, but I made how I feel pretty clear. I kept saying I'm disgusted in the episode.

The Imperfect Echo was an episode where I cloned my voice. There are limitations, but the first eight minutes of that episode are...

or is a synthetic clone of my voice and me speaking. It can't imitate the video just yet. That'd be cool. That's the whole reason why I do video podcasts because, well, eventually it's going to be easy to turn out podcast episodes that have no life, no opinion. They're not real people and there's no character behind it. And one way

that you can discern whether or not someone is real or not is if they're on video and they're a person and they share parts of their lives and they're personable. All of that can't be done without video. mean, you can do it, but is it real? So video just adds that little extra touch and also video podcasts are more engaging. So there's a whole bunch of things, but

Dalton Anderson (03:40.303) One of the main reasons why I decided to video all my podcast episodes was because eventually people are just going to make these AI podcasts and that's going to be the new trend. Nano Banana, the Google's Nano Banana in episode 82 was a cool episode to where I was creating images and then I was editing images in real time, like asking models to change their pose or the background or

the garments that they're wearing or the color of the garment. All of that, you couldn't tell that they were like, I did that. Like it was, it was really good. It was like a professional Photoshop job that I did with a simple prompt, like change this shirt from brown to blue. Okay, now make that shirt a collar shirt. It was very interesting. And then episode 80, and I'm kind of sad how the episodes weren't the same. Like they're supposed to be in chronological order, but.

whatever, zero click was talking about the future of what it's going to mean to search the internet. And increasingly the internet is going to be built for.

non-humans. It's going to be built for...

agentic search and it's going to be optimized for AI. And for that to work, you really need to provide more context. It's always talked about these AI systems that providing more context adds more value and adding more value to the AI system. You're going to be prioritized and then

Dalton Anderson (05:28.012) that allows them to provide a better output to whoever they're speaking with. And so I haven't thought that the internet is gonna become more technical because everyone's gonna be wanting to be on the AI watch list for whether or not they're searching for new gloves or new bed frames or whatever it may be.

as long as there's enough context provided that allows the AI system to navigate it in a good way, then you're going to be used. And if you're not being used and people aren't going to your website, because people don't go to websites anymore. They ask the AI to do that and the AI comes back with what they need. also very interesting. But in this episode, we're breaking down a couple of things. It's a multi top or I guess a topical episode about

The image where we're at, I have some examples that are incredible. Video and audio, what's holding it back, agentic web. And then this C-T-P-A, which is...

Content. Content.

sorry, coalition content, prevalence, authenticity. And it's basically, I would explain it as a nutrition label for synthetic images and videos. Like where did it come from? What was it made of? Like how much of it is AI? How much is real?

Dalton Anderson (07:13.422) But of course, I'm your host, Dalton Anderson. I like to read, talk about interesting topics on this podcast. My background is technical. I worked, you know, data science-ish type of role. And then I work in insurance. I enjoy reading books, running and building my side biz. And love doing these episodes currently on episode 93. And huge announcement, by the way.

I appreciate getting to 100 subscribers. That's really cool. 100 subscribers on YouTube. I'm still getting dogged by Spotify. They're telling me I have 18 followers on Spotify and with my listener count I should be at 60. And so I'm gonna keep asking, hey, can you get Spotify off my back and just give me a follow? It's not even a me thing. It's like I don't even ask for it. Spotify is like, you need to get your stuff together, man. So I'm like, okay, okay, I'm trying my best.

Okay, so the Rubicon of static imagery has 100 % the uncanny valley has been crossed to where the image used to have little subtle things like a couple years ago it was like the hands you'd get all these weird, weird hand things so then people would would like make sure that the AI model wouldn't wouldn't print the hands and then there was a weird skin tone thing going on and then one of the biggest tells was the eyes the eyes just

They just looked dead. Like they didn't have this liveliness to the eyes. And then the skin tone looked a little weird. Like there was weird things going on with the skin tone a lot of times. All of that has been solved for. And right now I cannot easily discern whether or not this was AI or a real image. And that's a problem. And one way that these models have

combated this was they inject this metadata into the pixel and basically they create the image but within the image there's like this secret.

Dalton Anderson (09:30.992) that's within the image, like hidden in the pixels when the pixels get generated, because it's all synthetic. And it basically generates a synthetic key that ties that image to the model. And so you would be able to tell whether or not it is AI. The problem with a lot of these safety systems is if I go and take that image and then I edit it and say I upscale it or downscale it,

and then I change the colors, then the pixels need to get regenerated. And when the pixels are re-rendered well, then the key is destroyed. So it's not really a safety. I think maybe...

the safety could be that they're working on it. But it's really easy to circumvent the safety, when I say safety, the identification. So it's easy to circumvent and then it's harder to discern. So it's kind of a problem in itself. But let's look into some examples. This is really focused on influencers and

the ability to control what's going on. So I think I talked about it earlier in the episode. I said, I think like, know I did in 2030, the AI influencer market is supposed to be $8.5 billion.

Dalton Anderson (11:04.587) AI influencers don't come with...

brand, these random brand hits with controversy with these influencers doing nonsense or they don't have to sleep, they're completely controlled, they have zero autonomy, it's autocratic, and they never age.

And so you just have this complete control over your image, which brands like. And you don't have to pay them anymore. Once it's created, it's done. There's no logistics involved with an AI image. Maybe put that thing on a flash drive and you plug it in somewhere. But it's all code, so you can edit everything.

Dalton Anderson (12:00.167) and maybe it just gets better as the models get better. scalability, it's it's frictionalist, it's scalable, there's, as I said, no aging, no sleep, no travel costs. It's just like you could be in Paris fashion show, then you could be at the New York Comic Con. You could be everywhere, anywhere all at the same time. And so there is no issue. And then another thing is you're able to collect data

and talk to your user base in a more personalized manner than a human could ever do. A human will never be able to scale to the level of an AI influencer. I personally don't like AI influencers. I think it's disgusting because there's not like a clear label of like these things are AI or not. And so if you're seeing, if you see these images and I really talked about it in

that filter bubble episode talking about just the lack of clarity of what you're looking at and where you are. And I really talked about its influence on like little girls, like on social media, seeing these things. like at a certain point, its prevalence might out, its prevalence might

have influences on how you feel about yourself. People are already insecure on social media, and then you add in these perfect AI influencers that don't have any problems, that always look phenomenal. Yeah, it's a problem. okay. That is where we're at. I wanna show some examples before I move on to the audio and video.

So I'm gonna show this example. This is a video. Oh, well, I'll show the video last, because I think that's a good transition. So let's do that. So let me share this tab.

Dalton Anderson (14:12.687) And a lot of these are Nano Banana Pro that just released with the recent release of Google's new models. Screen. Okay.

So this was the first viral comparison and it was by this guy named Sid. It says I'm a Sid X. Okay Sid, but it says Nano Banana versus Nano Banana Pro, we're cooked. And so the original image was using the old model and then the.

The new image or the comparison image uses the new model and it just looks so real, man. Like it looks insane how real it is. you can, I mean her skin tone looks realistic. Like she's not wearing any makeup. She just woke up. Her hair is a little messy-ish, you know? And the shirt with the.

the wrinkles and like the way it creases over her shoulder and her elbow.

And it's crazy the detail of the cup and the flower with the water and how the water is reflecting the lights from the ceiling. The guy pouring the drink in the back.

Dalton Anderson (15:41.372) And then if you look at the original image, like the older one, it has this weird skin tone thing going on. the skin tone just looks like, it doesn't look natural. It looks like Barbie-ish. It looks plasticky and doesn't look natural. And then the water is not reflecting the light from the ceiling.

So if you look at the water, the water doesn't have that same reflection where the, where, man, I'm blanking on what the term is, but when the, when the light bends the water, some of these episodes are tough when I do them live, cause like you kind of just blank on stuff sometimes. But anyways, so the reflections in the water and then the cup, like the details of the cup, they just look better. Like everything just looks

Way better.

Dalton Anderson (16:44.927) and then the table doesn't look as real.

Dalton Anderson (16:52.783) Overall, great image. This is NanoBanana. And then Alibaba has their model that's open source, which is really good. Let me share this. Alibaba's model is like very strong. Theirs is Z-image. And so both of these, this is NanoBanana on the left and then Z-image on the right. It's hard to say whether or not one of the others better or not.

They're both really good. It's difficult because not every generation of the image is going to be perfect. And so it's never a really apples to apples comparison. You would need to have the same prompt and maybe you generate it twice and then you compare both of those images or maybe do three times and you have four prompts.

I guess that would be enough, three or four, yeah. And so that gives you those examples and then you can use those examples in the three generations to see how consistently it looks and what the comparison is. This is another example of the Z image model.

Dalton Anderson (18:12.345) Well, actually, this is Nano Banana.

So this is Nandibandana Polaro. It created this Japanese influencer in Japan that has pink hair and is wearing this dress. And then they crop the images to make these cute little insidiary things that are trendy right now. And it looks super real. I don't know. I can't tell. It looks as real as it can be.

Dalton Anderson (18:51.501) This one is Nano Banana and this is by Sid. I'm a Sid X again. Our friend Sid, the original model is this woman is eating spaghetti, the refrigerator's open and she's got a hat saying, a girl's best friend. And then.

The hat has some weird stuff going on with the lettering. The spaghetti looks decent. The hands look good. They look a little spread out. The bottom hand looks very spread. Like the finger distance. Like she's got a big old hand on the bottom.

But then...

Dalton Anderson (19:42.148) Just how real this image looks is crazy, crazy. I don't know. It's so hard to tell that this is fake. I would never be able to tell me this is fake. And I'm saying me when I'm talking to myself. Just the light reflecting off of the microwave and the little blemishes that are on the stove.

the blemishes that are on in the reflection on the pot that's on the stove. It looks really good. And all of the

All of the-

brands that are in.

Dalton Anderson (20:30.735) that are in the refrigerator are like legit brands like Heinz, French's. She's a lots of mustard. I don't know what's a little mustard and she's got the beers, crazy stuff.

Dalton Anderson (20:47.055) Okay, so this is another image of...

I don't know, it gives like valley, valley, like, I don't know, forest, like European vibe, I don't know. But it's this woman who's got black hair, white skin and like goldish eyes.

And then the updated image, same type of shirt, like a Tommy Hilfiger shirt, same type of bag, just looks better. And her hair looks really good. And then her eyes look like, not that the other ones looked real, they just like, this just looks more realistic. Like it looks so real. The sun is on the right side of her face. It looks to be like 530 or something where the light's a little softer.

near the end of the day, the way that light is hitting. And man, crazy stuff.

And then this is the last one that I'll share and I will generate the image ourselves. We'll do one. I'll use this prompt.

Dalton Anderson (22:04.554) Okay, so this one is the before and after, same type of vibe, crazy real, hard to tell. And then let's move over here, share this tab instead. Let's do a image creation. So I use the prompt that that person gave me in their post. I really like when they include the prompts, it makes it easy.

Dalton Anderson (22:40.036) So wireless is generating, I think that just puts it in perspective how crazy these recent models are and where we're at on a synthetic.

the synthetic crossing, the synthetic Rubicon. We're definitely over the hump when it comes to.

the lack of the ability to discern synthetic created images versus real images. And I think this puts it in perspective, like she looks real, like she really does look real. She looks...

She looks as real as she could be.

It's so hard to tell. So hard to tell. I can't tell this person is not a real person.

Dalton Anderson (23:34.256) Crazy stuff.

Okay, so then the last thing that I'm going to share is a video. So this is the last thing and then we'll move on to the video topic. So I'm going to share my screen, make this full screen. It's one minute long. Let's just watch it. The voice, I'll say that the voice is not as good, but the visual, the expression and the vibe and the aesthetic of, I think her name is Marcy, this AI video that the character is really, really good. But the

the voice needs work.

Dalton Anderson (25:08.816) you heard it in her voice, but her laugh sounded authentic and her expression in her face was lively. so we are finding ourselves very close to crossing the so-called synthetic Rubicon. And then I talked about the first hump. The second couple of humps is audio and video. And I think we're closer in the video regard where video...

has some issues and I think the issues are more or less like hallucinations, not really on the physics sides as much, but I think where it currently still has some glitches is that the models are built on basically predicting the pixels that are gonna be in the next frame. And so if something disrupts the view,

then what happens is it might forget that there's a subject in that video that's supposed to look the same or have this kind of vibe to them. And like a good example would be like if you're walking and you walked behind the house and then you're walking to the other side and the camera view is from the street or something, it might forget that you have some kind of hat on that's a certain color or certain brand.

And so on. It's referred to as object pertinence failure, permanence failure. And that's close to being solved for. I don't think it happens nearly as often as it used to. You can have consistent characters in videos nowadays. The previous issue was like physics based where if you dropped glass, would

would melt instead of shatter and the way it shatter might be weird or if you jumped on the paddleboard, the paddleboard won't understand or the image model or the video model won't understand the shifting of your weight and the water should ripple and all these things like that. That stuff has been solved for. And so it's very close to being able to generate how an object would interact in the real world.

Dalton Anderson (27:35.665) So those are the things it's kind of working on. It needs to be less predictive of the models and actually create the predictive of the images. The models need to be less predictive of the images and the pixels that are gonna be next and more or less create what should it be, like the actual story.

Dalton Anderson (27:57.659) But we're coming along very, very decently. And then the last thing that I really think is a problem when I was trying to clone my voice and create the video of me was I try to animate myself and make like an animated version of myself.

Dalton Anderson (28:17.661) That was a problem because the length of video that you can create is so low. You can't create large videos. When I say large, like when I'm talking about time, the length of the video is too short for it to be realistic. Like I could only create 30 to 45 second videos. I wasn't able to then piece those separate videos together and then I just scrapped the whole thing because I thought it was a waste of time.

So if they could figure out, all right, if they could figure out the length of time that you can create videos, the other stuff I think is pretty close. It's very close.

So it leaves us to the last, I say the last frontier and there's been less emphasis on this than there has been on the video. Just recently, models started including audio generation within their videos that are created or sound effects and stuff. it's very, it's not a problem that has been worked on as long or with as much emphasis.

and is just starting to be included by default in these models.

So, audio. Audio has this weird thing where it sounds really good and then it just lacks that tone, that consistent tone where the tone changes, it has a flex to it, where it's noticeable, where it goes in and out, or longer where it might beep out and become more robotic.

Dalton Anderson (30:05.154) When I cloned my voice...

Dalton Anderson (30:09.647) a while back, it sounded good for majority of the time and people couldn't tell that it was robotic. I could tell because I knew and then there were subtle tells on the audio in itself. But in a general sense, it's really good. It's really good.

But if we're get to long form videos with audio included, there needs to be more consistency on the tone of the pronunciation of words. The words need to have emphasis and emotion and tonality to them where you can go up, you can go down, can whisper, you can go get mad, you can be stern. All of these different.

types of speaking or I guess it's really all these different types of tones.

add to

Add to the emphasis that you're human and that you are engaging and that people wanna listen to you and that there's something there besides just some synthetic code and prompts.

Dalton Anderson (31:25.873) I think much closer than people think on this, people have a really hard time. Untrained ear can only identify AI created audio, I think 50 % of the time, and then a trained ear about 70 % of the time. So if a trained ear can get down to 50, that means untrained ear theoretically would be at 30 % if the number still holds true. So then if you can only identify it,

three out of the 10 times, you're just listening to a lot of AI stuff. And once again, this is one reason of many that I record these podcast episodes in video, video, the emphasis on the video, because video is more engaging, but also.

It protects people from listening to AI slop.

Okay, so then there's agentics, these agentic web. And I talked about it in that episode that I referenced in earlier, but basically we're gonna be getting to the point here coming soon, like a couple of years, where people are not gonna be searching the internet as much. They're just gonna be using AI search. I talked about the numbers. I can't quote them right now, because I, one, I didn't research it before this episode again, but.

the numbers are quite compelling and are definitely trending into the route that people are going to be switching how they search and interact with the internet. And if you're going to switch and switch how you interact with the internet, then you're changing who interacts with the internet. And if you are replacing the users queries with AI queries, AI query values different things than a user does. A user values simplicity, quickness to answer, and

Dalton Anderson (33:19.577) just tell me what I need to do, how I need to do it and move on. And you want to be quick, brief, direct. Whereas AI doesn't like that. AI doesn't want you to just tell them the answer. AI wants to know why it is the answer and how you got to that answer and what's the background behind the answer and what's your thought process, all these things. AI needs to be able to navigate the website and understand the hierarchical structure of how things

how things are pieced together and look through the metadata and be able to navigate the index of the website for it to be queryable. All that stuff is way, way over what a human wants. so in my thoughts, many websites that aren't human-centric, like I would say LinkedIn would be human-centric, but a business-answery type of vibe.

But a lot of these other ones are going to be shifting towards more technical content where the information isn't simple. The information is as detailed as you can be. You want to create that authenticity and authority on that topic that you're releasing to the internet for these AI agents to take in with the model context protocol and

discern that you are valid to site and include in their output. And then if the user's interested, hopefully they click on your website, but there's gonna be a lot less clicking by humans and a lot more clicking by AI agents. And that's always been the case, like a lot of the internet traffic are bots, but now with these changes,

it's becoming increasingly apparent that we'll be shifting from strictly this SEO to AO or geo. It kind of depends on who you ask. This is like an ongoing change. So I don't think that there's a set nomenclature that's put in place on what to call it, but basically creating or optimizing your website for agentic search versus human search.

Dalton Anderson (35:44.498) which is a whole different ball game. And if frankly, if I went on a website like the one that I have created for a venture step and my website's Dalton Anderson.net, and then I'm going to create a sub domain and have ventures step.daltonanderson.net. But in a basis, a lot of these episodes, I transcribe them, but I make that like a smart transcription to where it's the episode, but then it has anecdotal information, all this stuff going on.

where it adds a lot of value to AI, but not as much to human, but it doesn't lose the context. It has everything. Everything I talked about in the episode, it's included, but it's just in a smart formatted way to where it's easily understandable by AI. It has hierarchical headers and such on different topics. And then for humans, if they had a question, they could just search it and see it. And that's the hope where

It's not really made for humans to search, it's made for AI. And if humans want to search it, then they have plenty of context and can understand. Also, my content is technically more, more AI centric and technical. So maybe if you are searching and looking at those websites, then you're already technical and you want the extra context. But a lot of people just want it to be direct. Tell me what I need to do and I'll do it. I've got other things to do. I don't want to waste.

an hour reading this nonsense.

Okay, so the last section of this episode is about trust. And where do we lie in that regard? And then we have that CP or C2PA, the coalition of content.

Dalton Anderson (37:39.855) prevalence and authenticity. I'm sure they could have came up with a better name.

It's like horrific, but...

This is a whole push to make a

Dalton Anderson (38:02.642) cryptographic proof of

Dalton Anderson (38:09.336) of an image came from a camera sensor.

slash newsroom screen or screen to where there is no gap on identifying that this came from a legit source. And so they're taking the approach that, hey, we can't trust the safeguards that are put together by these AI models. So you know what we'll do? Let's create a key for things that were authentically created. And if it doesn't have a key, then it's inauthentic. That makes sense.

I think that's a great, great thing. As I talked about earlier, it's easy to circumvent what is real and what's not real. And so if you have a key on what's not real and then you just slightly alter this synthetically created image or video, then it's technically the key is broken. So you can't identify it. So the only way to combat this is to create proof of what is authentic and original.

I don't know how they're structuring it and like the technical details of how it's going to work, but it is a proposal. How do we combat the lack of trust that we're going to have in the future? And this is one thing that they're thinking about. And maybe this deserves an episode on its own, like talking through the CPPA and where they're at, where they're at with their coalition and where are they getting support?

who's backing it, who's sponsoring it, all that stuff, because I think it's important to understand the background behind what's going on. But in general sense, I do think that's a good approach. I didn't necessarily research this thing for many, many hours. It was more of like, hey, I like this suggestion of we can't combat, or we consistently can't,

Dalton Anderson (40:12.412) have an easily identifiable way, or I'm getting my words all jumbled, but. Jam-bled, I didn't even say jumbled. Having a way to identify synthetically created information consistently. The only way to do that is to identify what is authentic. And so if you're identifying what is authentic, then you can tell what is synthetic.

And so I like the approach of thinking about it in reverse where, hey, let's restructure how we're creating our cameras, how we are using video cameras and all this stuff. Let's create a authenticity key. And so we know that this is authentic and it's real. And then if it doesn't have the key, then it's not real. Perfect.

I think it's great. It's hard to fake that if they build it right. I just don't know how it would work with a edit. Like if you edit your image, like if you shoot it in raw and you edit it, or if you edit the video and you cut it, you change the colors, you cut the video, slice and dice, and it's multiple videos all combined. How do you tell that that's authentic or not? I'm not sure.

But we've officially crossed one of the first humps of the synthetic Rubicon and where we are in a post Rubicon society, synthetic Rubicon society, I don't really know. It's gonna be a mess. It's gonna be a mess. Everything needs to get restructured. And maybe the shortfall of that is people feel less engaged on the internet. When I say the internet, it's really like social media.

if you can't really tell what is real and what's not real, then, and if that becomes a whole thing, oh, this cool thing I saw on the internet, and then the first thing is like, okay, well, is it real or not real? And then you've gotta go on this little discovery session of figuring out whether it's real or not. And I think that's going to turn a lot of people off, and the people that it doesn't turn off and they don't care whether it's real or not, I don't think they were the target audience to begin with.

Dalton Anderson (42:36.851) If you're listening to this episode, you care what is real or not real. And if you're not, that's fine if you're not listening to this episode. I do think that majority of society would care whether or not something is real or not. And then maybe 20 % don't.

So maybe there's less engagement on the internet and so with less engagement, people just, I don't know, go outside, touch grass more often. Who knows? Maybe they lean into authentic creators that have been around for a while and they go back to their childhood creators. There's definitely gonna be a shift. It's hard to predict what the shift is gonna be, right? How are you supposed to know that? I don't even know when this is gonna all happen.

How people react is very difficult to understand, but I do know that there's gonna be some changes. And then the other part is when the internet is created for agentic search versus for humans.

Dalton Anderson (43:48.445) Do people become overwhelmed with the information that they're constantly getting or do they not? I would lean towards they feel overwhelmed and I think that drives them closer to using AI to search. If you build your website for AI and agentic search, then when humans go there, they're like, man, this is a lot.

and then they go to their next website and they're like, man, this is too much information. And so then you optimizing for agentic search, I think would then increase the user base or consistency of users using agentic search, simple because there's too much information to understand what to do. And so the user asks AI to discern what

is important, what's authentic, and then tell the user what to do, and then cite where they got that.

That's what I think. I could be wrong, 100%. This is all philosophical. And I'd love to get people on the show to talk about these things because it's huge. Just these couple topics in themselves or episodes on their own. And it's just overall a very interesting time we find ourselves in. The society has to adapt and improve and alter their ways so quickly. It's never been like this.

I'm excited for the next frontier for sure, the challenge. But of course, wherever you are in this,