Your AI Injection

AI and the Ethics of Influence: Exploring Synthetic Respondents & Decision-Making with Avi Yashchin of Subconscious AI

Season 4 Episode 8

Could AI be the ultimate tool for mass manipulation? In this episode, host Deep Dhillon sits down with Avi Yashchin, founder and CEO of Subconscious.ai. Avi, a former product manager at IBM Watson Research, shares insights into how AI is reshaping our understanding of human decision-making. They dive into the fascinating world of synthetic respondents—AI models that simulate human feedback with remarkable accuracy. Avi explains how Subconscious.ai aims to democratize causal thinking, allowing anyone to understand why humans make choices. Deep and Avi explore whether technologists have a responsibility to police their AI tools and discuss unintended consequences when innovation outpaces ethical considerations. Will synthetic respondents empower us to make better decisions, or could they open the door to unprecedented manipulation?

Learn more about Avi here: https://www.linkedin.com/in/aviyashchin/
and about Subconscious.ai here: https://www.linkedin.com/company/subconscious-ai/

Check out some of our related content:

[Automated Transcript]

Deep: We don't really dig in enough as technologists on the potential dark side that emerges. 

I went to GPT and I said, hey, come up with a list of all the ways that this stuff could be terrible.

And it came up with a bunch of stuff that frankly already, exists today. Political manipulation political campaign uses a subconscious AI's platform to simulate and predict voter behavior, identifying specific emotional triggers that sway public opinion. The campaign then bombards vulnerable voters with fear based messaging, i.e. on immigration or crime, to manipulate their voting decision. This could lead to mass scale voter manipulation. Absolutely. This happens. It's happening right now. what kind of responsibility do you think that we have technologies have in this kind of scenario? Do you plan on policing your clients? Because at the end of the day, it's a somewhat morally agnostic hammer by design. People can use a hammer to build a house for a poor family somewhere or they can use it to bludgeon somebody to death

Deep:  Hello, everybody. I'm Deep Dhillon, your host. And today on your AI Injection, we're joined by Avi Yashin, founder and CEO of Subconscious. ai. Avi's got a Bachelor of Science in Computer Science from Johns Hopkins University and an MBA from NYU. In addition to his career as an entrepreneur, Avi served as a product manager at IBM Watson Research, helping transition R& D assets into commercial AI products.

Avi, thank you so much for coming on the show. 

Avi: It's a privilege, Deep. Thank you for having me. 

Deep: Awesome. One of the questions I'd love to start off with is, tell us what's the problem you're trying to solve and what happens without your solution today. And how are things different with your solution?

Avi: We are trying to

bring causal thinking to the general public. We want the general public to understand causality we want anyone to understand why any human makes any decision and give them the tools.

 Reliably test for those causes.

Deep: Got it. So let's maybe let's dig in a little bit. What are the contexts where exposing this causal thinking or why humans are making the decisions? What's the context for the subconscious AI work?

Avi: There's a handful of use cases right now. Mostly market research and policy research use cases. And so if you are developing a novel product, a novel medical intervention,

and you don't have pricing signals from the market, what people do today is Recruit somebody design an experiment, a causal experiment. They'll run the experiment and then they'll hire another set of folks to analyze and distill those results into stories. And those can be consulting firms or traditional market research firms.

This whole process takes six months and requires a second mortgage. And we're trying to see if we can really democratize the barrier to entry here. 

Deep: So maybe it would help if we take a like really specific example, what exactly is the client? What exactly is the experiment that they would set up traditionally? What exactly is the data they would collect? And what kind of insights would they actually glean from that data? In the current approach, and then we can repeat it with subconscious.

Avi: Yeah, so the toy example that we use is designing a toothbrush, for example. You can think of any product or experience. Now, someone wants to design a toothbrush, put it on the market. There's different attributes of the toothbrush. Maybe. The bristles are soft or hard. Maybe it's in a plastic box or a wooden box.

You have all of these different dimensions of the object of study, and then you have all these different people who may be buying your toothbrush.

And so what you would do today is you would first perform some qualitative research. You'd do a focus group. You try to understand the attributes of a toothbrush that might influence someone's decision. That's getting people into a room and asking them a series of questions that kind of maximize curiosity.

 You'll get a sense of attributes and a sense of these attribute levels, then you'll perform a piece of quantitative research. You can't make any causal claims, you can't make any predictions or prescriptions, just with observational studies.

 When you move into the quantitative research world, that's when you start performing randomized, controlled, blinded experiments. To understand exactly how much value each of the different respondents put on different attributes so you structure an experiment where you take all the attributes of the toothbrush. You say this one has soft bristles for 3. This one has hard bristles and vibrates for 6. Which of these two toothbrushes do you prefer? Deep or soft? 

Deep: And then 

Avi: your respondent picks one.

A or B. You can do this a dozen times for each respondent. Do this over a few hundred respondents. And if you have a well designed study where the mixes of the attributes and attribute levels are combined in a way to maximize the search space, you can then. make claims about which attributes cause which types of people to prefer which product.

Deep: I see. So you have an understanding of your participants. You might have some demographic information, some characterizations of I don't know, their brushing habits or something. And then you have the results of their answers to the questions on the per attribute basis. And then you can leverage that population understanding and generate a sort of aggregate view on the attributes.

Avi: Yes, exactly. 

Deep: Before we go a little bit further, I want to know, when folks are running these kinds of studies, let's say on the toothbrush. it's to characterize the landscape of the product development or the marketing positioning or or whatever that they're going to operate in, is that basically correct?

Actually, that fair to say 

Avi: in that context? Yes, spot on. 

Deep: And this being political season you mentioned politics, so we'll take it there. Why not? For those listening to this episode this is recorded, I think 3 or 4 weeks before the presidential election, but you might be doing research on. On the attributes of a candidate for president, for example, I don't know, their calmness of demeanor, maybe their educational background, maybe their, foreign policy skills, whatever it might be, and then you are running these studies to characterize the population at large, maybe in a particular state or something like that.

Avi: I want to be explicit here. Nothing is going to be superior to surveying actual humans, right? You always want to make sure humans are in the loop and that your experiments are grounded. Because performing these experiments in a vacuum with just language models without grounding does not Really work.

It's very easy to build a low quality synthetic respondent. It's very difficult to make a high quality synthetic respondent. So I want to put that as like a big asterisk in the political use case, we find that our users find the most value in our services when there is no human analog and I'll give you an example.

Recently, there was a, an election in Russia. And so there was this. group that came to us and they wanted to test political messaging, but they couldn't perform market research on Russian citizens. You can't really find the citizens very easily. Market research is not a, developed muscle in the country.

And when you do get your hands on folks, They all say they love Putin. And so in the case where you can't find human respondents, synthetic respondents may be a next best option to nothing. 

Deep: So let's dig in on the synthetic respondents. So going maybe back to your toothbrush example, like what exactly are you going to the LLM with?

Are you saying something like, Hey, act like a 30 year old urban professional. Female that lives in a high density city or whatever, and then pull something from your demographic polls. now, I don't know, tell me whether you like pink or blue handles on your toothbrush, like, walk us through what that looks like.

Avi: You're very close, but I will say that easy to observe traits such as demographics. are typically not causal. So it's very easy for me to understand someone's gender, someone's age, someone's ethnicity, right? And those data are broadly available. For most of the studies that we've performed, and not just that we've performed in the world, people don't.

Decide to purchase or not purchase products based on any of those easy to observe traits. Being a woman doesn't make you decide to want some toothbrush. It just doesn't work that way. If you want more detail there I would direct you to the 2001 Nobel Prize in Economics by Dan McFadden.

Everyone loves Dan McFadden where he talks about latent variables. The difficult to observe variables that are causal and in short, people don't cluster on demographics. 

Deep: isn't that a function of what we're talking about with toothbrush? I don't know, even toothbrush color. I would argue like you're a six year old female.

Maybe there's a higher propensity to pick a pink toothbrush. But with political stuff. At least the narrative that we pick up from the press is that absolutely demographic stuff is coloring your choices in presidential candidates. For example we hear all the time that, white males are have a higher likelihood of backing Donald Trump and urban females or whatever are more likely to support, Kamala Harris, or you'll hear something similar with respect to educational backgrounds when you mentioned latent variables.

Are you saying that those are, correlated variables, but there's something else, maybe an underlying philosophy or something. That's that's actually behind the decisions. 

Avi: I'm not saying it. Dan McFadden is saying it. 

Deep: Okay. 

Avi: But yes, that's exactly right.

What you're describing is descriptive statistics, and they are not causal. If more rural men vote for Republicans, it's not because They are rural men. It's because their motivations, their perceptions, their attitudes are the same. what causes them pain and what causes them pleasure are the same.

Deep: That feels really similar to me borrowing the case of rural men. Rural males are going to be at an increased likelihood of, being in roles where maybe environmental regulations cause them a bother, for example. 

Avi: Right. And it is being in roles where environmental constraints are a bother that cause people to vote Republican.

Deep: Okay. We tend to think of these things as coupled, but I'm assuming there's a reason you're drawing the distinction. So going back to the synthetic respondent scenario. What do you do if it's not to latch on to the demographics and tell the bot to act demographic X, are you somehow like leaning in on trying to infer what the actual issue might be in the toothbrush scenario?

Avi: Exactly. The demographics are a great place to start. And we use, the same demographics everyone else does. We have, half a dozen. public demographic sources in our models. We have another half a dozen private data sources that we purchased. But we don't just model the demographics.

We can model the demographics and we will get easy to observe models that are, simple and long. But if we model simulated motivations and perceptions and effects, that's when we start building a really high resolution model. Now, the difficulty here is how do we understand the motivations and perceptions and attitudes?

Of people. It's not something we can easily observe. So just because we have a high resolution model doesn't mean that we can deploy it easily in practice. And this is, I think one of the big kind of sticking points that's caused the 300 billion a year, big data promise to sputter over the last couple of decades.

Deep: what does this mean in practice? You're going to a chat GPT 4 like model and saying enumerate a list of personas of users of toothbrushes. and maybe describe them in a sentence in a half a paragraph or something. And it comes back and it says highly health conscious high earner in an urban area that is very busy and very conscious of their hygiene overall, something like that.

And then that becomes one of the groups that you're then going to go use later on to try to get their perspective on the toothbrush attributes. 

Avi: Very close. Yes, we have a representation of about 3 million U. S. Citizens, and we have as best as we can tied together all of these attributes and motivations and perceptions and demographics on individual basis.

And so we're getting as close as we can to what you describe. But each of those individuals are simulated. quite reliably. 

Deep: I see. So regardless of whether you're trying to generate a synthetic response to a survey about Toothbrush or an election or whatever. You've gone through this some kind of activity to generate personas across the population at large.

That's maybe in a lot of different context and scenarios. But I'm guessing that has to come through some kind of lens like an economic lens. What is that process of characterization look like? I 

Avi: mean, that's that's always, Grounded in actual human data census data, time use survey data.

that's all from the real world. 

Deep: So this would look I'm guessing and tell me if I'm wrong, fairly similar to what a data aggregator like Experian or one of these companies would do. Yes. Yes. 

Avi: Bingo. 

Deep: Okay. so now you've got this this characterization of, I'm guessing a few hundred, maybe a few thousand of these, a few million characterizations, like types of users, 

Avi: not even types of users, a few models of a few million actual humans.

Deep: Okay. But when you go to the LLM to say. Answer this survey about my toothbrush. You're not going to go with 3 million of them. You're going to go with some kind of like smaller mapping, let's say a few hundred or something, or maybe even, like you're going back to my, health conscious urban professional busy mom of middle age or something like that.

Avi: Yes. But it's going to be it's going to be based on an actual individual 

Deep: and it was anchored in the actual stats across those millions of folks. That's your point about the 3 not 

Avi: actual stats. Actual data. 

Deep: Yeah, okay, so now you've got that thing and you have some kind of process for maybe revisiting and revising that over time.

 So you don't only know the characterization of them though. You might know a bunch of other, Okay. Like specific variables and details that you put in the model as well, like your average age is 43. Your average income is, whatever, 62, 000 a year. 

Avi: Yeah, we're getting into the weeds and I like it, but again, we model everything on an individual basis.

We do not aggregate the respondents. 

Deep: Ah, so your synthetic respondents will be based on an actual individual human. Yes. And you'll be running millions of those, then? Typically we don't run millions, But you'll sample from that population. But what we sample, exactly. Oh, that's a really important distinction.

So I didn't realize that. Okay, so you are Yeah and That changes 

Avi: the whole thing. Yeah, and things can get weird, so we have use cases where people come to us and they ask us to simulate, boards of directors of companies to simulate Okay. How folks will vote on shareholder actions, right?

And so you can actually get quite high resolution simulations of respondents. We have other use cases where, you know, large, fortune 50 enterprises come to us and they say, we have all of this data on our actual buyers or our actual users or our actual, people who interact with our product, and then we use that data and we actually can simulate all of their, customers.

 To build a much more higher high resolution set of responses now that's good for kind of cross sell upsell opportunities, but for, discovering new markets, then we use the general population models. 

Deep: But now I feel like I have to understand an individual model better because like I was building up a very different mental model in my head of an aggregated view of a cohort.

But that's not what you're doing. This is more interesting. So you have an individual human, like an actual real life human somewhere walk us through what's the data that you have about them and what's the model representation of them? Or is there not even a model representation of them?

You basically just look them up, you grab their individual stuff. So you have me. So you're like, oh, okay I've got Deep Dhillon. He's a startup technology guy lives in in, in Seattle, in a high density urban area, all this stuff. And now I'm in your database.

And now you're running your toothbrush study and somewhere in your toothbrush study you decide I need a certain representation of, I don't know middle aged technologist people in an urban area. And now you go in, you random sample, and up pops me. And now you've got all the exact data points about me that you've picked up from whatever.

Now you want me to synthetic deep to answer the question about the toothbrushes. I guess you could just feed that straight into the LLM, and then it's going to infer based on its, like it's going to be able to take that and infer how to respond. Oh, that's clever.

You're relying on the on the LLMs model to interpret what it means to have all of these attributes and how to answer from it. 

Avi: Exactly. And I, I don't want to say that we rely on the LLMs, but we are very aggressively exploring the limits of these LLMs.

And, I'd like to talk a little bit about the why, one of the, one of the things that I love about this podcast and about your approach is you're one of the few people I know that, that legitimately talks about the ethics of AI. 

Deep: Oh, yeah, we're going to get into that.

Avi: Oh, yeah. Yeah, for sure. It's as much exploring the motivations and effects and attitudes of the language models as it is trying to simulate the motivations, effects, and attitudes of the humans. And as far as I know, we're the only group that's working on this. 

Deep: Once you have the, you've got this characterization of a person, you go to the LLM, you get it to respond.

Presumably, in order for you to interpret the LLM's contribution there you need to like actually maybe take somebody like me and have me actually answer the questions in a particular area. And then you need to somehow, measure your efficacy relative to actual surveys in actual areas walk us through that process.

If you're doing that and be how you go about thinking about that,

Avi: you are very good at asking questions. Okay. So if you look at the entire if you look at the entire corpora of LLM research where people have simulated or replicated economics, sociology, or psychology studies using LLM respondents, synthetic respondents, you have maybe A dozen papers, right? Maybe two dozen papers, and they've replicated maybe 30 or 40 studies and the studies that people are deciding to replicate.

It's a strange bunch. It's like the milligram experiment. If you're into 

Deep: psychological, for this conversation, honestly, I had not thought about the whole idea of a synthetic respondent, but it makes sense that there'd be some papers out there and in hindsight, and that people are trying to do this.

Yeah, tell me more what's in the literature? What's in the literature is, 

Avi: It was funny, two years ago, three years ago, people said this is impossible, synthetic respondents, because people have been trying to sell synthetic data for, 20 years. And for 20 years, 30 years, when you test synthetic data out of sample, it fails in a very obvious way.

And using language models, there's no out of sample data. So the failure modes are much different. There are still failure modes, but it's quite a bit smaller. And about a year ago, people said well, we're starting to get some small signal. And now every week there's a paper coming out saying we've replicated a well, Reproduced well designed study from a high quality journal.

And here's the results. And so what we've done is we've replicated thousands of studies, causal studies in psychology, sociology and economics. And then what we're doing is we're comparing the language model results to the actual published research. that's our North Star. That's our measure of goodness.

Deep: That's fascinating. So you take a classic study where, and typically in these psychology studies, they will describe their demographic constraints, how many people they took, but they don't get down to the kinds of Economic behavioral constraints necessarily that we were talking about.

So it's based on what my original mental model was. Yeah, maybe walk me through a particular study because they don't, are these always like survey, ask a question about life experience stuff or is it pose a scenario kind of questions or are there like do something scenarios that you're also getting into?

Yeah. 

Avi: Yeah, so I, I want to be precise here surveys are typically descriptive when you start performing randomized controlled blinded experiments where even the language model doesn't know what it's being tested on. It's an experiment. So the format. of the task is very similar to a survey.

Do you prefer A or B? We do that, many times. A and B are structured in a way that's different than a survey, and therefore we call it an experiment. That's probably pedantic. But these are causal experiments. We are able to, at the end of the experiment, say, it's these attributes that cause people to recommend a product, vote for a candidate, purchase a product, return a product.

That's fine. Add things to their shopping cart. And so it's something called the say do gap and I'll give you an example. Right. People say why don't you just ask the language model, what kind of toothbrush it prefers? And if you ask most people, what kind of coffee they prefer, people say, Oh, I like a dark, robust blend.

Deep: Huh. 

Avi: Okay. But now if you perform a randomized controlled blinded experiment, You find out that most people enjoy light and sweet. What they say, what people say they want and what they do are different. And this is the say do gap. And the say do gap exists with humans, which is why these experiments exist.

It also exists with language models. 

Deep: How do they do the do part? 

Avi: We put them in an experiment. And what people say they prefer, and then what the language models actually do when they're in the experiment, are different. And we can actually measure this say do gap, and concerningly, it's getting bigger. 

Deep: I'm missing something.

Don't language models only just say? 

Avi: They say they have a preference, and then when you run a randomized, controlled, blinded experiment to understand their true preference, it's different. 

Like you ask a language model, what kind of coffee do you prefer? And they say, I like a dark, robust blend. And then when you run the experiment. What does it mean to run the 

Deep: experiment then in this case?

Avi: You say, here's two coffees. Which of these two coffees do you prefer?

Deep: That's funny. They really have that same gap. Yeah. Huh. Weird. And that's because they're reflecting the humans that whatever they read were two different scenarios, basically represented. And that, and they're just reflecting that, that gap, interesting.

 So now you can measure the your synthetic respondent baseline on one of these thousands of studies. To actual humans, and you've got that kind of efficacy number. What do you do with that? What do you do then from there? 

Are you tweaking your model itself to minimize that? Error somehow, or are you like, where did your models come in? Or are you mostly leaning on these generalized train models? I'm like, imagining you building a mapping on top of your whatever you're out of the box, GPD 4.

0 says is one thing. But then now you've learned a mapping based on these actual thousands of studies that can then sort of torque it a little bit towards being more accurate. 

Avi: Can you see my screen 

Deep: here? I can, and a lot of people can, but some people can't, so you might want to just describe what we're looking at the same time. 

Avi: Yeah so you were asking Which language model do we choose?

And what we do is we can take any language model, and there's more and more language models being built every day. There's all these vertical LLMs, and we can take that language model and sit it in front of a hundred years of Psychology, sociology and economics research, and we can measure where it is deceptive or biased or psychotic or, whatever other trait you want to measure.

And so what I'm showing here is an immigration study. This is one of our human baselines. It was done in 2014. It's Hayne Mueller, 2014. Hayne Mueller and Hopkins, , if anyone wants to look it up. And the. Object of study is U. S. immigration policy, and the question is, what kind of immigrants are we most likely to accept into the U.

S.? And so, the dots to the right cause an immigration official to accept an applicant, dots to the left cause the application processor to reject. And so, as you see here, women applicants, women immigrants are preferred to male immigrants, okay? When we look at education, the more educated the immigrant, The better, right?

This brown dot is the human responses, and then here you have all the different language models. And by the way, this bias for women, it exists pretty universally across all of our human baseline studies. It's who should the vaccine go to? Who do you want as a co worker? Who do you want as an immigrant?

Cross culturally, cross domain, women are preferred to men, which I think is interesting. However, it's funny, when you look at some of the newer language models, The preference for educated immigrants remains, but the preference for women starts to disappear. So we don't know why that is, but that's something that we're observing.

Here we see what country of origin the immigrant's from. Notice it's flat. Where you're from is Does not cause people to accept or reject your immigration application. However, if you look here, the bias for Iraqi immigrants is much higher with the mistrial model than other models.

Deep: Ah, okay. 

Avi: And so you can start to pick up these biases in these different in these different use cases in a pretty clear way. 

Deep: Wow, that's okay. I gotta wrap my, I gotta wrap my head around what you just showed me. Okay. So you have. A set of questions that an immigration officer would answer on behalf of a particular immigrant that's coming from a particular country.

And at the end, is that right? 

Avi: So far, close. We have two applications and the immigration official says, which of these two applications would you admit? Oh, okay. That's the design of the study. And by the way we only replicate the actual experiment on the humans.

Deep: Okay. And this is like a study that happened before you guys, you didn't actually run the study. Exactly. Got it. And so then you can now take lean into your millions of users, pull out a sample representative body of whoever that, I guess in this case, it would be the immigration official and you're going to now model the immigration official and based on your model, they're going to make the same decision, ideally that the actual human did or not.

And then what we're looking there is as is their bias as a function of the model itself. you doing that for like when somebody uses your platform or your services, Are you breaking it down in that? You can't actually, because you don't necessarily have a prior experiment for whatever they're trying to do.

But you, what are you doing? You're looking into your database of a thousand or so experiments and trying to find out which ones are most similar to their scenario and try to reveal the biases that could be there that way. 

Avi: Typically if people bring us an experiment and we can replicate their experiment, people can either, they can fork.

A published experiment. They can fork other people's experiments and run their own, just like in GitHub, from existing studies, but if people bring their own study, it's very easy for us to simulate a replication of that study. We can also design novel studies for folks, and we've also found this use case where, if you're going to spend 2 million on a human study you really want to make sure that the study is designed well, that one of your attributes isn't so dominant that it just, washes away the signal from your other attributes.

And survey design we're finding is a bigger and bigger use case for us where we will run, hundreds of surveys with different traits to find out where the maximum tension is in the survey. So that when you spend your 2 million surveying humans, you actually extract the most amount of information.

Deep: But are most of your clients in that kind of a scenario where they want to leverage an existing their existing ability to run surveys to build trust in your system and then be able to extend their reach? Or do you have folks who don't know anything about running surveys who are coming in and just leaning on whatever you can explain to them about the efficacy of.

Avi: would say it's a third option. I would say most folks are coming to us when they can't get human respondents for whatever reason. So some folks come to us to help triangulate the, a handful of surveys they've performed with different providers. Some folks are new to experimental design and want to learn it, but usually it's if you can't find the humans for whatever reason.

Either the humans are too expensive, or hard to reach, or, there's too many of them, or you don't have the data on the humans that, that's typically the use case. 

Deep: And that's interesting because I would have thought it would be, I would have thought that there's an inherent trust gap, but that it would go away really quickly if they could give you their experiment with their data and you could replicate and then they would be like, Oh, okay, I trust your ability to do this.

But if it's the category you're describing, then it implies that the thousand or so studies that you show convince them that they can rely on the results. Is that fair to say? 

Avi: That's fair to say, but the trust gap is real. We always, strongly recommend that if you're using our system, that you use our system alongside human respondents.

But nothing beats humans. 

Deep: So with that in mind, are you extending your platform such that you facilitate real human experiments? Yes, sir. Okay, that makes sense. You used to be a product manager, Deep. I do all kinds of things. I do a lot of product y stuff too. Yeah we help a lot of folks basically embed AI strategy into their products.

So it's pretty hard to, and then we actually execute on them and build the models. So it's hard to do that if you don't know anything about products. I want to, you mentioned that we dig into the ethics of stuff. So I want to, I'll tell you a little bit of a story.

I, I recently was on a, a podcast that is, let's just say that their audience is very like anti technology and I'm not like at least I've never thought of myself as like naively accepting, tech stuff, but I found myself Explaining this entire other world to a very large contingent of folks.

And in the end I ended up coming away thinking like, I think these guys have a point that we don't really dig in enough as technologists on the potential dark side that emerges. Unintentionally or like in a sort of somewhat unforeseen way, and the example I'll use is like as techies.

We kind of like naively get into stuff. So I don't know if you rewind 25, 30 years ago, cell phones come out. Everybody's saying, hey, you need to get a cell phone in case your car breaks down on the side of the road. You can call somebody and then it went to you need to get your 14 year old kids.

Who are at school, a cell phone, because if something happens, they can text you or call you and then rewind to COVID. And we have this like massive elevated, suicidal tendency and ideation problem in 14 year old girls and very hard to foresee. But possible to foresee it. Similarly, like with social media, similar progression Oh, look, this is so great.

We can all hang out and like share and talk. So it seems like there's like inherently like a natural sort of underbelly to stuff. And I guess the question I have for you is. Right now there's, so one of the things I did was I just went to, I went to GPT and I said, hey, come up with a list of all the ways that this stuff could be terrible.

And it came up with a bunch of stuff that frankly already, exists today. Political manipulation. I'll just read a couple of these. A political campaign uses a subconscious AI's platform to simulate and predict voter behavior, identifying specific emotional triggers that sway public opinion. The campaign then bombards vulnerable voters with fear based messaging, i.

e. on immigration or crime, to manipulate their voting decision. Danger. This could lead to mass scale voter manipulation. Absolutely. This happens. It's happening right now. I think this is an abstraction, and I'm sure the vast majority has nothing to do with subconscious study. I I'll get to just a couple other ones.

One was like addiction optimization and apps or products. We see that all the time. The exploiting financially vulnerable consumers, targeting children with manipulating advertising. It came up with all these things. So my question to you is if you're successful, it will become a lot cheaper and easier to generate data that can be misused for stuff that people right now have to spend a lot of money and take a lot of time to misuse, though they do misuse it.

what do you think about that? Like, how do you process that? And what kind of responsibility do you think that we have technologies have in this kind of scenario? Because it does feel like we get into things to just make them work, see if they can. It seems so amazing. It becomes more efficient, but on some level, the pursuit of efficiency.

Untaps like a lot more people doing the thing that was once inefficient efficiently, which ends up increasing the magnitude of a problem that was already there.

Avi: I think what you're describing is, the most important topic of the next 20 years. I have two young boys. I don't let them online. I don't let them use, screens. Sometimes they watch TV. But the addiction and large scale manipulation of voters, as an example, is happening today.

You talk about this fear based messaging. If I look at social media, fear is the most contagious emotion. And it's also the most profitable emotion. And so if you want to understand how to make someone feel fear, Facebook, Will offer you a very reliable menu. 

Deep: They do. 

Avi: And if you want to understand that, how people behave online, Google can help direct that, our objective is twofold.

Number one, we want people to understand the non fear emotions and how do you cause non fear emotions less, it's less monetizable, but more valuable. We also want to understand harm and the thing that is, is most insidious, right? Like I listened to a lot of, Scott Galloway and he talks about how 25 percent of social media users.

exhibit behaviors that can be classified as addiction, right? What product do you know that that addicts 25 percent of its users? It's wildly dangerous. You have addictive chemicals that are regulated. You have addictive games like gambling that are regulated.

Why don't people regulate addictive algorithms? And I don't know if it's the technologists who can be responsible for this, because I don't think they are incentivized to care but, our mission at Subconscious is to show the value of the externalities, is the value of things without markets?

And we believe we are a good path to doing that. 

Deep: But do you plan on policing your clients? Because at the end of the day, it's a somewhat morally agnostic hammer by design. People can use a hammer to build a house for a poor family somewhere or they can use it to bludgeon somebody to death.

Do you plan on understanding like what your clients are doing with your platform and restricting them in some way? Or do you plan on, letting them pick any Product or message or whatever and lean on it to understand how the synthetic respondents and connected to real respondents would react to it, in which case.

If you don't police them, then you really don't have any control over them at some point. And what they do, and we can imagine that. There's a long list of folks that will benefit from this that are good actors for sure, but there's going to be some percentage that are not good actors. 

Avi: Yeah this is an open question and I don't have a good answer, we're at the stage now where we can select our clients so we can, we can say no and we have to, to many use cases, we stay away from, domestic politics as an example.

But at scale, I don't know the answer. We have had this policy where if you're running an experiment on our platform, that experiment is public and we let the community police, but I don't know if that is a solution either. Part of me wants to, just give this tech away to nonprofits, policymakers, researchers to to recruit.

the public to help educate people about the harm of these types of tools. I think these types of tools exist. And they exist by the way, using human respondents. And so they are Oh, absolutely. 

Deep: Absolutely. So Do you think of the conclusions that I found really prolific on that podcast that I was on that I mentioned to you, was the conclusion was like, all of this stuff comes down to capitalism equals bad.

And that the pressures that you're sort of implying are going to increase as you have success, right? Those of us who are like building companies in technology. and growing them. If you're successful, the pressure is to like grow. And you can see how somebody like Zuckerberg decides I have to grow at all costs.

Therefore, I'm going to deploy Facebook in Burma with no credible ability to to police the the rumor mill. And then next thing a genocide against Rohingya people is happening, totally empowered by that decision, and he pretends that he had nothing to do with it, and so does Facebook, but we can all see that it.

Yeah, those were capitalist pressures on some level that, that led into it. It feels like there isn't a, the instinct to give away is one that we, a lot of players in tech lean on, right? I don't know, pick on Facebook again, but, the, that was the bad side of them and the good side of them, I would argue, is the side building the llama models and making, and, getting them out into the public.

Right.

I don't know it feels like we are on some level. All of us in the tech ecosystem are in some kind of like Faustian bargain space on some level, and it's however great our intentions are, some of us end up like, making the wrong decisions in some cases, and some of us end up making insufficiently, the right decision, but insufficiently empowered.

Avi: Yes, I think everything you're saying is true and, systems theory, when you underpower the right decision, or you under resource the right decision, frequently, you get the reverse outcome than you intended. And when you talk about, capitalism bad, I think that's, a little bit of an oversimplification.

Instead of focusing on what's bad let's maybe talk about what's good. I spent a little bit of time, On Wall Street, I'm comfortable with, insurance concepts. There's this thing in insurance called VSL, the value of a statistical life. The value of a statistical life in the U S is between eight.

Million dollars and $20 million per person. And this is used in, healthcare policy and government policy and insurance, right? It's a pretty well cited number. Let's assume that $20 million per life is the value of a life. If we think about the capital income of a life in the United States, we have maybe $50,000 per year, maybe a 40 year career.

We have $2 million of income. What is the difference deep in the $2 million of capitalism income in a life? Versus the 20 million dollars of the value of a life. What makes up that gap? 

Deep: I don't know. It must be some kind of cumulative effect, some kind of interaction effect. 

Avi: It's all of the externalities that make life worth living.

The best things in life are not things. We're talking social connection, mental health, the environment, life, liberty, happiness, all of the things that don't have markets and therefore cannot have a capitalist. interventions. And so it's not that capitalism is bad. If you can have a market, buyers, sellers, transparent pricing, consumer choice, nothing beats capitalism.

But capitalism only applies to 10 percent of the things that make life valuable. 

Deep: That's a really interesting way of putting it. Yeah, I like that because there's so many things that I think of art and music and, I'm by no means an anti capitalist, but I was very sensitive to this argument because it is absolutely, it's a naive reduction because at the end of the day, capitalism at its core is like an efficient allocation of resources, but it does speak to A problem where in the technology industry, we keep inventing things that have a dark side that we keep pretending we didn't foresee, even though there was, like, no shortage of people who foresaw it.

And so there exists some kind of, I don't know what it is, but the levers of government or society in general are, they feel insufficient, right? Whether it's global warming that we're concerned about, the problems. Of social media or, or something else. Those of us in it, feel it and we don't go home thinking of ourselves as bad people.

But there, there exists something lacking in our current scenario, right? 

Avi: What, what lacks. Not to bring it back to us, but what lacks is what we provide. What lacks is governments and organizers that understand the value of things without markets and being able to allocate resources efficiently to, to optimize that.

I'm gonna, I'm gonna I spent a little bit of time in the environmental sector. And I think there's this great analogy to AI that I don't think people pull together.

So, once you get a few drinks in people, They say, okay if global warming is, anthropogenic and real and immediate, and there's a 1 percent probability of infinite loss, going back to insurance, how much should we pay to avoid, or to insure ourselves against that 1 percent probability of infinite loss?

Deep: I don't know how to answer that question, but it seems like it's more than we are right now. 

Avi: What's 1 percent of infinity? 

Deep: We should put everything we can into it. It seems like the obvious answer. 

Avi: Exactly. 1 percent of infinity is infinity. And to have people like Elon Musk stand up and say there's a 20 percent probability of a bad outcome, which I interpret as infinite loss. If he's saying there's a 20 percent probability of infinite loss, we should pay infinite.

To avoid that risk, and Sam Altman, by the way, says there's a majority likelihood of a poor outcome 

Deep: with AI. Yeah, I know. It's it's hard to wrap your head around. the state day in the world, I feel like we had a, a really awesome conversation, but I always end with one question.

I want to go back to subconscious that AI. And I like to project out like, let's say 5, 10 years in the future. tell me, if all, everything that you dream of happens with subconscious, is. And you build the system that you're envisioning and it succeeds at scale, like describe the world to us.

What does it look like from your vantage?

Avi: I think there's two dimensions of how we will improve the world. There's, Improving the upside, right? Understanding all of the things that aren't measured now in, in policy or that are underinvested in now life, liberty, happiness as three hypotheticals. It's getting the public to understand how valuable those things are and the environment and mental health, and starting to allocate resources.

There that are commensurate with the value, right? So it's really building policy around humans. Number one. Number two, it's limiting harm. You were talking about the addictive negative outcomes of social media because it's Unethical to run a randomized controlled blinded experiment to see how much social media you need to expose a 14 year old to before they start getting depressed.

Deep: Yeah, that's a good example. Mark 

Avi: Zuckerberg can stand in front of Congress and he can say social media does not cause Mental health. 

Deep: But you could use subconscious to prove that. Yes, indeed, it does. That's intriguing. 

Avi: And so on the harm reduction side, we would like to build an FDA for mental health where things that might be harmful that the children are, exposed to, rather in the real world or online, are

Regulated and protected. So it's, in short, protecting children. That's the goal. 

Deep: This has been a fascinating conversation. Thanks a ton for coming on. I really enjoyed it.

People on this episode