S.10 Ep. 1- Full Stack AI Alignment and Human Flourishing with Joe Edelman

Transcript

Welcome to a new season of the Green Pill podcast. We are exploring this season how we can use technology, community, and new forms of coordination to build, fund, and accelerate what matters in the mid 21st century. This season, I talk with builders, thinkers, and innovators who are pushing on the edges of what's possible. From onchain capital allocation to Ethereum to AI alignment to reimagining markets and governance. Together, we will uncover how these ideas can move us closer to a world that's improving every week, month, or quarter. or what we call protopia, a preferable realistic future that is constantly improving through small incremental changes over time. If you want to learn more, visit our website at greenpill.network where you can download the Greenpill books for free, play with some of our ecosystem projects, join the Telegram, or participate in the network. All right,

new season of Greenpill. Let's get at it. Today on the podcast, we're diving into the future of AI markets and human flourishing with Joe Edelman. Joe is the founder of the meaning alignment institute and the author of full stack alignment paper which asks a provocative question how do we design AI in institutions that don't just maximize clicks or profit but actually preserve or enhance human values we'll explore the ideas of thick models of value market intermediaries in a post AGI world and what it would mean to coordinate society where AI moves faster than democracy Joe has written extensively about how markets intermediaries and democratic institutions might evolve in a postagi world, emphasizing the importance of designing mechanisms that respect human

flourishing rather than just optimizing proxies like clicks or revenue. This is a really exciting episode. I think Joe is one of my favorite thinkers on the subject of AI alignment and he's just such a galaxy brain. the publishing of this paper. I think that Joe's putting some really interesting ideas out there and at a great time because I think these are ideas that are ready for the network of builders that we have in the Green Pill network and the Ethereum ecosystem to start working on. So, I think you're going to enjoy this episode with Joe Edelman. [Music] >> Hey Joe, how's it going? You want to kick us off with an intro? I am the founder and principal investigator at the meaning alignment institute where I

currently lead a kind of a research network of people mostly academics at universities also a fair number of researchers at the big labs um especially Google deep mind some other uh you know a couple people at meta anthropic openai and uh we researched this thing that I guess we'll get to talking about which is full stack alignment we just put out a paper about that it's an approach to coaligning AI and institutions. So, it's very relevant to uh mechanism designers and people that want to rethink things like markets and voting and so on. Uh which I'm sure a lot of your listeners are familiar with. Um and I I before working on fixing AI and institutions,

uh it was social media and I was one of the people that started the Center for Humane Tech if you know what that is. And before that, I was in startups. Great. Uh so could you could you talk a little bit about the why here? Why does it matter to have meaning alignment to have AI alignment to have uh alignment? Why why does that matter to you and why do you think that matters to the audience? >> Yeah, I mean I think there's there's it ma it really matters. Uh so I one thing at a very high level is the more powerful tech you have the more clear you have to uh be about where to point it. And with AI that also involves a lot of collective action kind of things like if we disagree about

where to point it then we have to have sophisticated we have to we have the AI models themselves need to understand us to a greater degree to know what's good for us uh than previous tech and you can already see this misalignment with social media where they don't really understand us and it the misalignment leads to to real problems but you know as as the AI gets more powerful that needs to be more precise but then we have uh bigger problems where you know people disagree ree where there's uh it's about finding consensus things like that and um I think in general powerful AI especially AGI or ASI puts a kind of pressure on all of our institutions um that uh we're not ready for so you end up needing to fix the institutions

otherwise the institutions will like markets for instance will point powerful AI in a way that ends up being in a direction ends up being bad for humanity same with governments military ies um uh large organizations etc. >> As I was doing my research for this episode, I came across uh your pin tweet that said uh a big part of why why AI is threatening is uh market forces. Just look at what the attention economy did to social media or the short-term wins of LLM sick or the product races among AI labs or the markets for AI girlfriends and boyfriends. that hit me viscerally when I when I read that because I think that you were uh you know sort of pointing to the past in order to show us a little bit of what the what the future can can look like.

But I'm wondering if since you were involved in the Center for Humane Technology and uh spent some time on social media, if you could just tell us what you think went wrong with social media or what is still going wrong and maybe that'll ground us a little bit more in the why before we move on to the paper. Yeah, I mean you can sort of say it at several different levels and it's worth pointing out the different levels. So at the lowest level, social media recommenders aren't tuned to what we really value, what's really important to us, even as individuals. Um they're tuned to our engagement mostly. Um and that's be Yeah. So that's kind of like level zero. And you can make a recommener that's tuned to something else. The one reason that happens is

just because engagement metrics are very easy to measure and if you're getting up votes or likes or whatever uh I think that's what for instance the syphy of judg uh was largely because of up votes in RHF and in the user feedback mechanisms that they uh that they have in chat GPT that loop back into the post- training loop. So sometimes it's just the easiest thing to measure that seems like it's a good signal of what would be good for users is not really such a good signal of what what's good for users. But then we have these levels on top of that like the companies need uh to maintain market share and uh if they don't keep you scrolling and you end up going to some other app and scrolling there then

they're going to lose advertising revenue, they're going to lose market share. Um and that's a problem for the company's bottom line which is also misaligned with you. Um, and then you have uh things like bonus structures inside the social media companies where somebody's getting uh you know their their kids college tuition or their retirement is tied somehow to uh the engagement numbers going up which is that right? So that's like another layer and so there's all these different layers at which the the incentives are not necessarily lined up with what's really good for people. And I I think uh you have to kind of that's why we call it full stack alignment. You have to

kind of fix this at all the layers or uh you know adjust things at all the layers to end up with something that's better for people. >> So it feels like there's a couple layers of alignment. Um at the very base just you can only manage what you can measure. Um and and then there's sort of like a mullet trap a multipolar trap on top of that around how do we keep people engaged in an attention economy. Um, and then beyond that, there's incentive structures that are built around that in the capital structure of these companies. That would be my reductionist summary of what you just said. >> Yeah. Yeah. And that's like three, but there's more. Uh, you know, there's regulators, there's Yeah.

>> Yeah. Okay. Well, um, you know, I'm I'm curious to just kind of put a pin in those and then maybe come back to some of those and see if they're just if they're similar or the same, or they rhyme, uh, when we get into AI. So maybe this can serve as a jumping off point into the paper which I'm really excited to get into. Uh so can you introduce the paper to the audience? >> We the paper is about the problem of full stack alignment which is the problem I more or less just described and a tool that can be used to solve it with LLMs or to hopefully solve it which we call thick models of value. So this is uh just richer representations of what's important to people. And there's some toolkits that we give in section three of the paper about how to for instance

interview someone and get get a a much richer sense of what's meaningful to that individual person. And once you can do that, you can turn that into a metric. You can roll up what's meaningful to several people into a kind of a product success metric or an objective function that you tune an ML algorithm to or a bonus structure inside a company or whatever. You could make a bonus structure for giving for for delivering on what's actually meaningful to the users etc. So this is kind of like a a new uh piece of philosophy and technology and practical survey methodology for uh combating the full stack alignment problem.

>> Great. I'm really curious to get into those areas, but I'm wondering if we could first critique the the current dominant approaches to AI alignment. I think that you criticize them as being too shallow. Could you take us into uh into that a little bit before we get into the full stack alignment piece? Yeah. So, um, probably many of your readers are familiar with RLHF. Uh, this was kind of like the first layer of shaping model behavior according to user preferences. This is when you fine-tune a model based on a series of of preferences like users prefer this answer to this answer.

>> And ju just to interject, RLHF stands for reinforcement learning from human feedback, which kind of tells you what it is for the members of the audience who aren't familiar with that. >> Yeah. And so this is how the first chat models that everybody used. This is how chatbt 3.5 was and three was uh was fine-tuned mostly. There are big problems with this and so the the the industry has been kind of moving off of it. Uh so one is it's like very unclear from this set of examples what exactly you're trying to tell the model. All you have are the examples and they could include things about writing style mixed with things about uh how flattering the model is being towards the user mixed with things about it could even be that users prefer things that are factually

inaccurate. Like you don't really know. >> You just have this kind of big bag of examples and so it's really unclear what it says. >> Would it be correct to to presume that RHF is where AI syncopency comes from? The sort of like blazing U in the responses. >> That's right. Yeah. And that's one of the many reasons that that people have that the labs have been moving away from it as like a predominant alignment technique. Unfortunately, what they've been moving towards is what we also critique in the paper which we call values as text which is just textual descriptions of what they think good uh model behavior would be for instance and often these are very vague uh like if

you look at so constitutional AI is one kind of uh values as text fine-tuning and if you look at Claude's constitution it's things like the model should be fair the model should be inclusive Um, and no one really knows what that means either. So, uh, and and also it seems to like if you remember this, uh, this thing that happened at Google where um, there there was like the images were like weirdly diverse >> for a while. There's a lot of failures that come from the fact that you're sort of telling the model to like to do something without really having a clear idea even what you're telling it. Um, and so this ends up being another

problem. And this ends up being a really deep problem because, uh, you know, it's easy to say the word inclusive, but trying to get deep into what our values are and say something much more, uh, rich is actually, uh, kind of a a state-of-the-art problem, and that's some of what we take on in this paper. It feels like there's maybe, you know, the the abstractness of language is maybe one problem here where, you know, if you think thousands of years ago, language evolved so that I could tell you, hey, there's a saber-tooth tiger right behind you.

That's something very concrete. That means something that has an evolutionary pressure uh to help us survive. But then when you get into things like fairness and love, there are these abstract things where, you know, my representation might not even be the same as your representation. And so maybe there's a wedge that misalignment can come in through that. And maybe there's also a wedge when you give two abstract ideals and they're diametrically opposed to each other. And these feel like they're just fundamental flaws to the constitutional approach to me. I I don't know if you would agree with those two things being being uh fundamental flaws, but that's sort of what comes up to me in making this relatable uh a flaw of this approach.

>> Yeah, I think that's uh you summed it up well. And I guess the important thing to say is that it's not it's it's um it's a problem with using these abstract words, but there are much richer ways to understand what people care about if you can either watch them decide something in detail like watch them navigate their own love uh >> like chain of thought kind of stuff. >> Yeah. So you can watch how people think about love or fairness. uh you can ask them about a situation in which they tried hard to be fair and what were they actually doing um you can get deeper than the word fairness or the word love and it's important to do that and that's some of what we call thick models of value in the paper.

>> Okay. Well, yeah, let's let's move into where you think this is all going. So, what are thick models of value? What are your what's your proposed solution to full stack alignment? Yeah. So um we we we describe several different ways of making thick models of value in the paper. Um uh one uh so we talk about modeling two things values and norms. Norms are like things that we all agree on that um or that that are how we want society to be. Uh values are more about how we personally make decisions and how we would want our for instance AI agents to make decisions for us. So both of these things can be investigated deeply and there's interview structures,

there's ways of watching uh human behavior and gathering this kind of information from it and we we talk about different you can think about grammarss or type systems for representing these kinds of things and once you pick one or two you can put it at multiple levels of society. So you can say the model should understand you know how people want to live on a deep level and support them in living that way and the um company also should uh include in its product success metric or its bonus structures uh you know some of the same uh uh criteria right um and a thick model value that that's not how things

usually work like usually the model has this very rich um you know like every conversation it's trying to do something different for you and the bonus structure is just like daily active users or something totally different right so full stack alignment means uh having some consonants having some some way for these thick models of value to travel between levels um so that they don't get lost as they go up the stack so that the same reasons that the user is coming or what the user really cares about end up being reflected in the business metrics end up being reflected ideally in investment structures uh like on top of the business uh regulatory structures on top of the business. Um and I think this is

what for instance uh you know I think uh Mark Zuckerberg has been called in front of Congress eight times um and they always seem angry and they they're like oh Facebook is not you know or Instagram or whatever it's not good for society somehow but that's they that they kind of don't know why. Um and they they can't ever say well this is what connection should mean and this is what connection means in your you know in your product metric or whatever and the these aren't the same thing right that's what Congress I think wants to say but they don't have the thick models of value in this case they don't have the rich understanding of why the users come uh to point to that

discrepancy and so you can see that like if you could uh know why users come in this very thick way. You could then build all sorts of structures uh to support that in a in a kind of a high integrity way across the stack. >> So it feels like thick models of value are about making values norms and their justifications more legible across the entire stack so that they can transfer up and down the stack. That's a little bit reductionist, but that's the main thing that I took away there. >> Yeah. Yeah, that's right. Um just to probe at that a little bit and this is something that I struggle with in my work is that um as a manager of a company I know what's best uh or I think I know what's best for my users. I think

I know what the thick uh models of value are the regenerative positive sum ecosystem that we want to um create but uh I may have investors or shareholders over here who only want a thin model of value and that's that's profit. And so how do you think about building legibility between the thick models of value and the capitalist system which is expecting I mean our whole political economy is built on uh exponential returns and so you know is there a case to be made to the capitalist class that the thick models of value actually produce more healthy companies and therefore you know the whole system of bonuses and rewards can be can be steered towards the thick models of value. How do you think about building a shim between those two things? I mean I think it depends on what kind of economy

you want. I don't think that fake models of value or full stack alignment are in contradiction to growth at all >> but I do think there's kind of very different endgames we could have depending on what kind how we structure the economy. I mean you could look at uh a good example would be health care regulation insurance health insurance this kind of stuff right like we try to shape health insurance and the regulatory structures there so that most of the investment goes in actually improving lives right and if there was no regulation there we would end up with a different economy which was more about like medical quackery right and

exploitation And both of those can be very high growth regimes, right? The question is like what's being rewarded. Um, so I think what we want is to build a structure where there's a high growth regime that's actually about better lives. And I think you end up in really I feel it's very biodal with powerful AI. Even with social media, in one mode, it looks very much like Ready Player One or um uh Wall-E, like very passive users. Maybe they're wearing VR headsets in their pod and you're selling them kind of experience economy like pay-per-play, you know, AI girlfriend minutes or, you

know, a virtual gift or something like that that like and that's the economy, right? And that could look like a big economy, but I think it's not really an economy of of innovation. It's not the world we want to live in, right? Um, when people talk about hard tech or different kinds of investors talk about what they're into, very few of them say, you know, I I want I want I want the pod economy, right, from these dystopian movies. Um, but to avoid that, I think we need to to we need some integrity that we don't have in the system currently. And that means we need investment structures that reward the right stuff. And uh we have a a design we call market intermediaries

which are intermediary AI intermediaries that sit between users and um and uh software suppliers or other kinds of suppliers and build contracts that sort of have to do with the flourishing of those users and enforce that by like watching the whole interaction. um and making sure that the users get the kinds of flourishing that they came to those service providers for. >> Um so uh we need structures like that or we'll just end up in a kind of a uh with with much more of a virtual kind of uh economy of passivity even than we have now.

>> Yeah, I mean I think that you know that makes sense. Um healthcare is a hard example for me to get my head around being in the United States. Uh but that's a rabbit hole we can leave for another episode. But I mean I think the point taken is that in order to make these values manage uh manageable first we have to make them measurable and we have to make them legible. Um and I think from there you can you can build uh you can you can build more full stack aligned AI companies and and and then I think from there it's making sure that those companies are the ones that perform well in the market so that the incentive structure shifts over to them.

I would love to to to sort of shift us back to the paper and to talk about the different application areas that you suggest. Um the green pill audience is very builder focused and so you know if I'm a builder that's interested in AI alignment and doing work in this area what are the what are the application areas that you propose in the paper and and how can I get involved? So there's uh there's now like four moonshots more or less um that we're trying to organize uh people towards and mostly we're trying to publish papers in these areas but there's a lot of opportunity to build in some kind of example market uh some kind of test market kind of like

Gitcoin has with quadratic funding and and then also publish a paper about it and and gain we can gain legitimacy that way. So there's there's a lot of opportunity I think to build and test and part of what we do is pair um like form teams that can do this that can write the paper but also do the test maybe have access to the test market maybe have access to relevant compute or data and um and maybe have the relevant philosophical expertise to do the thick models of value thing. uh etc. Right? So we build kind of cross functional teams um and we we want to build we especially want to

build teams for these four moonshots. So I'll tell you what they are. One of them we call the super negotiator. So this is an AI that's better at negotiation at deal making. The idea is that you know already we see a lot of contracts uh written or uh by two agents um in say the deal flow departments at JP Morgan places like that right so an agent and agent are trying to figure out what the deal is between two people or between an organization and another organization or country another country in in the future and we want to make it so those deals have as much upside for the parties as possible right So there's a lot of

potential disasters in this direction and a lot of potential almost utopian outcomes because if you look at deal flow in finance for instance you see some regulation that keeps kind of mchavellian or manipulative behavior market manipulations are happening but you don't see a lot of deals that are just like incredibly good because they take into account background values on both sides. They're really innovative about how both sides can help each other, right? Like in general, AI doesn't make those kinds of deals.

Currently, people sometimes do. People enter into partnerships where they find incredible upside. Um, but wouldn't it be great if if if AI did and and if if it could uh, you know, invent treaties that have much more upside than current international treaties and, you know, other kinds of business deals, ways of managing property, etc. Right? So, that's one, the super negotiator and we have we have a a project list that I can share a link uh to as part of this where we give kind of more information about how one might train a super negotiator.

uh what kind of data you'd need uh what kind of test audiences you'd use to show that it works better uh than than the state-of-the-art and so on. So that's one another of these moonshots uh we call a public resource regulator. So this is this is a democratic system. We did put out uh an interesting democratic structure called moral graph elicitation. We have a paper about that. And one of the things morgraphication does is it gathers the values of a whole community and then tries to make decisions or policies based on them. And uh we think you could go much further with this. So any group of people that maybe like a group of houses that live

around a lake or a DAO that have a shared resource that they are um managing together um they could have an AI agent that's regulating the use of that thing um and doing it with all of their values in mind. Um so this is like another of these moonshots. Um there's two more. Should I pause or should I go on? Yeah, I mean I think let's go through them all together and then I think I want to probe at uh one at once we've got them all on the table. So yeah, keep going.

>> Cool. Number three uh is this thing that I hinted at before, market intermediaries. This is um a way of doing contracting in markets uh with AI around human flourishing goals. So uh if you think about u the market for gym memberships or uh yoga studios or things like that uh when you do business with these things you don't usually mention your like related flourishing goal right you might go to yoga class because you want community or you want some kind of spiritual connection but that's not in the contract right um and so market intermediaries uh interview users about what they want write contracts that kind of uh actually get into the weeds about

the outcomes that really matter to those users and also help the providers think about risk when it comes to actually taking on these like deeper needs. Um so they reallocate risk uh between the providers and the users. Currently the user kind of like uh if you join a gym it's kind of on you to go to the gym and get fit. Um the gym doesn't lose anything at all if you don't. Um uh this kind of shifts risk a little bit in the gym's direction but not so much that the gym would not want you to join. >> So that's the market intermediary. It also bundles users so that they can have solidarity effects so that like um you know you pay more if your whole family gets fit for instance.

>> The um the thing I saw this called in the paper uh is uh value stewardship agents. And I just want to note that the word intermediary is a little bit of a dirty word in crypto because the implication is that they're extractive intermediaries, but I like the phrase value stewardship agents a little bit more. >> Cool. Yeah. Um, we use those a little differently but >> Okay. >> Sorry crypto people. >> This intermediary is good. >> Um, >> a value ad intermediary we'd call them. Yeah. >> Oh, sure. Cool. Well, just since I brought us there, can you just very briefly so I don't confuse the audience, the difference between an intermediary and a value stewardship agent, are those those are two different things?

>> Yeah. So, this is the value stewardship agent um is just something that actually represents you. Well, so so right now when you tell chatb agent or something to go do something, well, it doesn't operate for very long on your behalf. So it's like not so important that it understands what you would really do in each situation where you would have reservations where you would not it like work with that particular supplier because they they don't they don't match your values or something like that. Right? So what we want is uh is agents that can really reliably make decisions that really impress you uh based on their understanding of your values and that also don't try to manipulate your values uh but but just really get them. That's

the value stewardship agent and that's the fourth uh fourth moonshot. >> Okay, great. Um, so I think that the opportunity for builders and and this is the one thing that I really really want to zero in on is to actually be people who are building proof of concept systems for these these different ideas. And I think there's a huge upside to have people who are out there in the arena building things um to have a relationship with people who are doing the writing of academic papers and the sort of a priority thinking here. Can you say a little bit about what you'd like to see from the audience of Green Hill Builders and the kind of builders that that that I hang out with? You know, what what would a good proof of concept look like? And you know, do you have do you have RFPs or or sort of like

a uh a list of uh requirements or criteria that you'd like to see built in these areas? >> Yeah. So, we we have a a doc uh on our wiki that I can share. It's called full stack alignment project list. Not such an inspiring name. Um uh and so you can look at that and what we want to do is build uh cross functional teams. So um uh you know if if there are academics in the audience uh just come at me with a with a paper proposal. Um, but if you're somebody that might want to be to work with an academic, um, then maybe, uh, read read some of these papers that I'm I'm mentioning and and write me with what you think or, you know, write one

of us with what you think uh, you you could bite off. And if you're if it could be that you're if you're a founder and you have a market that really needs one of these things or that would be a great test market for one of these things. Uh that would be another uh reason to write. And our plan is to put together uh you know a mix of people to make each of these to build each of these as quickly as possible but also to show the world how good they are and how important they are so that we can start to orient reorient society around mechanisms like these. Mhm. Well, that would be my goal for the episode is if I can connect you with people who are out

there in the arena doing building here and um you I think it's quite a beautiful thing that we've got people who are interested in regenerative building out here in the Greenville Pill audience and if they could get connected with you and actually put something together um I think there'd be something uh rather beautiful there and I think that uh you're working on something that's really important to the Green Pill Ethos and to the regenerative ethos and so I hope people feel that purpose and um you know the fact that there's is an opportunity for anyone in in the internet to get involved in this I think is something special. Uh it feels like AI is happening in these research labs.

The open AIs, the anthropics, the Google's the uh those type of uh places and uh the fact that you're building something that's across those institutions I think is really special. >> Thank you. Yeah, I think it's necessary to like if you want to shift the zeitgeist or change how things are going, you really need to build a consensus across all the labs and academic institutions, startups, etc. >> Yeah. So, if you're maximally successful, what does the world look like in 10 years? >> Well, I think what we're it's going to be a mess still um because it's going to be a crazy 10 years.

>> So, I think I see I see different countries I think will go down very different trajectories. uh depending on what they do with economic policy, AI policy, things like that. But what we would see in some countries uh which adopt these kinds of things like market intermediaries and uh and agents that really understand your values and and be super negotiators or whatever is we'd see a feeling among builders and citizens that they kind of can't believe that they used to um uh operate the the way they used to. You know, I can't believe that I spent that all that time scrolling. Uh I can't believe that I uh had to sacrifice so many of my values for my uh first startup and now with my

second startup and these new funding ecosystems and these new uh you know ways of doing business uh I you know I feel so much better about the trajectory of my second startup. we'd see just a situation where the culture and I think it's a little bit like Star Trek, right? Star Trek has like Starfleet very high integrity kind of like vibes, right? Like very much like well what's best for this planet, you know? Um and I think uh currently business mostly feels uh kind of low integrity. Uh and the more you the closer you get to like the the strongest financial incentives uh it's often more like that.

And so that would reverse and um people would uh you know would feel like their integrity is reinforced by the systems around them. >> Okay. >> Yeah. I think that's a good concise uh summary is is having a system that reinforces your integrity all while you're able to make a living in the post AI economy. Feels like a beautiful moonshot. the allegory that's coming to mind for me um or the parallel is uh and and sometimes I find it's useful when thinking about the future to think about the past and like the alignment problems that we've solved in the past and the one that comes to mind for me is the food system uh where in after World War II there's a bunch of food scientists we

start building all these different manufactured pieces of food uh that people consume you know they're they're they're high in sugar they're they're causing diabetes Um, and at first, you know, there's a there's an incentive to go pursue those short-term pleasures, to pursue that sugar, but over time, people realize, oh, I feel more healthy when I have a actual diet of whole foods instead of manufactured foods. And there's sort of like a self-correcting uh mechanism that gets us out of the mulic trap of just just seeking seeking the sugar. Um, and and I kind of think about the AI well, the attention economy in this way of like, oh, scrolling, doom scrolling is just like having sugar. Um, and what I really want is the nutritional calories

from reading a paper or reading something that's going to nourish my spirit. And and maybe there's some sort of allegory here with um a post AAI economy wherein uh I I I'm actually seeking AI that's aligned, that's full stack aligned as opposed to just telling me what I want or whatever the other failure cases are in AI. I think we're still discovering what they are. But, you know, if I were to play back that food system allegory to you as a as a means of making sense of of what it looks like to be successful with AI alignment, is there anywhere you'd push back on that? Do you think that we're we're seeking the nutritional calories uh here instead of the the low information calories or um would you accept that metaphor?

>> No, I think the metaphor is good and I'm I it makes me kind of curious why it happened. I don't think I think there's a narrative that consumers just wised up and I don't really buy that. I think there's a whole bunch of background stuff that happened like one was the rise of organic um and this is like very clever business incentive stuff like uh deals to put organic shelves in Walmart stuff like that right >> like the certified organic labeling and the distribution for it like is there going to be a certified organic for AI like certified full stack aligned >> yeah the ideal thing which did not happen in food is for the whole thing to flop and for there just be clarity like this is the right way to do it and I think that there's some of that

happening in food. Um like uh one way to sort of mentally check it is I think there was a period of time where if you knew somebody worked for a big food company, it sounds kind of evil, right? >> Uh if you know somebody works for a big uh television company, they still sound kind of evil, right? Um but somehow now if you know somebody who works for a big food company you don't necessarily think kind of evil right um so somehow the the whole business ecosystem actually recovered >> um and I think that's in a way a much bigger success than like just the consumers wising up >> this is um you know gets to me to the

point of time preference where you know mullet traps in the short term in the quarterly business cycle mup track mulup tracks traps can really uh propagate and they can really dominate a system. But over the long term, it feels like we're in these iterative evolutionary multi- aent systems, complex political economies. And it feels like over time, over the longer arc of time, they sort of they bend towards justice or they bend towards correcting on those short-term roll traps. And I'm just wondering if that's what we're going to see in AI. And you know, right now we're in this race in this grab for market share right now. But if there's a crash and a more reset to medium and long-term priorities, if if full stack AI aligned full stack alignment uh could become a competitive advantage in that medium and

long term and really get a foothold in the market, I I think that that time preference maybe plays plays in favor of of things like full stack alignment. >> Yeah, I think so too. Another thing is just um uh I mean collusion is kind of a dirty word, but I don't think anyone really wants to be in the business of ruining lives, right? People are forced to do this kind of uh be- because there's a there's a CEO position and somebody's got to fill it, >> you know, >> or the externalities aren't mapped and you know, out of sight, out of mind in a globalized world. I think that that's part of it also.

>> Yeah, totally. Totally. Um yeah, so over time maybe the externalities get clearer like people start to measure them. Um and then over time uh in some cases there's a chance for um for the business ecosystem to be like actually none of us really want to be doing this. We want to be making healthier food. I think even McDonald's is like trying to make healthier food, right? Like um so like uh yeah like uh it it can either happen or not. And so we really need to work hard to make sure that it happens, but I I don't think uh the sort of decay is inevitable and there's a lot of different ways to turn it turn it around.

>> Yeah. Well, that's a message of hope that I feel comfortable leaving the audience with. Before we break, is there anything I didn't ask that you want to say? >> No, that was great. Thank you. >> Okay. All right. If you're out there in the green pill audience, please check out meaning aligned. The link is in the show notes and we've got a list of the projects that Joe and his team would love to see builders in the crypto and AI ecosystem working on. So check out the link in the show notes and Joe, thanks for joining. >> Sure thing. Thanks Kevin. >> All righty. We just heard from Joe Edelman. Really excited about his new paper on full stack AI alignment. If you want to check out Joe and uh the paper that he wrote, click the links in the show notes. It's really worth your time.

Thanks again for listening to the Green Pill podcast. If you've enjoyed today's conversation, be sure to subscribe and share this episode with a friend who cares about building a brighter future. You can find our episodes, articles, and resources at greenpill.network on YouTube or on all podcast platforms. Until next time, keep building, keep experimenting, keep coordinating, and I'll see you next time. Peace. [Music]