Plural Publics

Transcript

hi everyone I'm shrey I'm an applied scientist at Microsoft research under the special projects Division and the title of my talk today is plural Publix um a very real question that I'm sure many here in the audience have started to think about is whether or not it is possible to live in a pluralistic society in a world where humans coexist with large Foundation models and by a pluralistic society I mean a society that is free a society that is self-governing and a society that flourishes when there is participation in the context of diversity and this question is very real and top of mind for policy makers for AI

pioneers and citizens alike and so in this talk I'm going to be focusing on how we address that problem and how we can preserve Democratic systems a framing that I like to use wait for this slide to go but is known as a narrow Corridor framing the narrow Corridor framing is a framing that balances both between centralization and Chaos the way you could think about centralization is concentration of power and illegitimate checks on authoritarianism and chaos is when you have social fragmentation and what we're starting to see today is immense amounts of centralization emerging in the top AI Labs we've started to see how a very small subset

of Engineers are capable of creating some of the most powerful transformational Tech that Humanity has ever seen and they've been able to do so in a way with very limited Democratic accountability in that process in a way that has led to the emergence of immense amounts of concentration of power information and processing speed when we start to relax the safety measures we start to see how there's can their chaos can emerge and this is more so to do in the open source Foundation model space and the open source Foundation models we started to see how models like llama alpaca have started to lead to immense amount of deception and both the speed and ease at which these models can cause deception has started

to cause a tremendous amount of issues not really sure where my slides are so I'm going to keep just talking over it for now uh okay there yeah and modern liberal democracies rely on three key principles there's a sense of identity there's a sense of property and a sense of voting and they take shape in many different forms in some Modern liberal democracies you see how identity maps from a birth certificate all the way up to a health card to a passport in the land sense you see how it takes shape and the physical land context or the data context in modern sense and finally there's a sense of collective decision making whether it take place in Civic elections or Internet polling today all three of these

um set all three of these are Under threat and we're starting to see how different AI leaders around the world are actually trying to critique that problem and instead of me who maybe conveniently trying to be an adversary to this problem we can start to see how even the people who have the most to gain are critiquing this problem as an example Sam Altman who's the founder and CEO of openai has echoed the problems that large Foundation models can have to strengthening authoritarian regimes and that we need to start building these models in a more democratic way we've also seen other uh policy leaders talking about the fall of democracy and Gathering some of the most senior CEOs of all the big tech companies to start

to put together um uh the a policy towards this uh okay technology and democracy have seemingly been in in tension with one another for the past couple of decades in the past we saw how Democratic systems were some of the best developers of Novel technology whether it be the internet lasers or GPS but now we're starting to see some of the best technology emerge from private institutions and that Corridor between centralization and chaos is narrowing and narrowing in a way that's getting concerning and so what I'm going to be presenting today is

this vision of plurality that our lab focuses on which is this vision of finding consensus across difference and widening that Vision between centralization and chaos plurality is current is grounded on a lot of different social theory whether it be the secrets uh literature from gear ximl the writings from jcr lichlider who is the founder of arpanet the precursor to the modern internet as well as John Dewey who's American philosopher that wrote a lot about how we can collectively deliberate to find solutions to Collective problems and uh these leaders of today's plurality vision is being led by both Audrey Tong the digital minister of Taiwan as well as Glenn Weil who's the research director of the lab that I

belong to at Microsoft research along with a variety of different academic Labs across the world including Harvard getting plurality Network Berkeley Central Center for responsible intelligence and the other names listed here um an example of the plurality work that you may be familiar with in the ethereum ecosystem or the web3 ecosystem is is quadratic voting quadratic voting was developed by Glenn and others uh where the goal of this was to find as much consensus or representation across differences and reward those differences in a way that is proportional to that difference and we've seen examples of the power that this has had in the git coin community um right now though what was and as we've started to learn is that we know that

Democratic systems rely on participation but we know that it relies on having voice and incentivizing voice in many ways but as we're starting to see participation is at risk we no longer know how how we can trust public commentary we don't know whether emails or letters were generated by the people who are actually writing them so Senators don't know whether or not they're actually putting forth policy that's representative of its own citizens in town halls in internet-based polling mechanisms are no longer sufficient or credible if we're unable to verify that those people are in fact the people we say that they are and so the focus of our work is to do two things one how can we find new ways to enhance participation because large

Foundation models actually have a lot of opportunity here to reshape the way we think about democracy they have the ability to potentially have long Global instantaneous participation in a way that Democratic theorists or reformers can only imagine but they're only important if we can verify that that voice and input is actually what they say it is to be and if we can do that then maybe we can get to this vision of plurality that we've been envisioning and so the the core focus is to look at how has this changed in the emergence of the internet uh in the past 20 years the internet to many is currently viewed as one of the most promising Revolutions in human communication it's allowed people to communicate

across the globe at speeds we could only once imagine and allow people to do so so long as they have an internet connection but the internet has actually hindered our ability to communicate with one another effectively the reason being is that it doesn't preserve context when we communicate with one another context is the critical part of human communication and context is under threat now because we no longer know what's real or what's fake and so when the Bitcoin Community started to talk about don't trust just verify being this very Niche culture quote in this culture it's now becoming mainstream because we can no longer trust and we have to verify information if we want to be able to participate in Democratic systems I mean so when I said context is

everything what exactly is context well context is precisely the symbols and language we use to communicate with one another the only reason you are all understanding what I'm saying is because we all have some consensus around the English language we all have some assumptions we made when you came in to listen to this talk around what the style of speaking or the framing of it is going to be under something to do with ethereum or edcon or web 3 of the sort and we can see how context is critical to human communication because without it we are unable to converse with one another and so if we can't preserve context we can't communicate with one another and context is under threat so much so um that we've had so many different authors in the social science space

write about it and talk about the impacts that it's had on networked teens along with the way we shape privacy as well as our secrets Dana Boyd who's a co-worker of mine at Microsoft research wrote the book it's complicated and talks a lot about the impact that the internet and social media platforms have had on the inability to preserve context for teens when someone posts something on Instagram or Twitter they may post it with a given assumption around what they wanted it to to mean to someone but when someone else views it their expectations their dispositions and the way that they view that content is very different Helen nissenbaum has tried to reshape the way we Define privacy privacy is not this binary thing between I do share or I don't share privacy is contextual what

might be private to you may not be private to someone else but it may be private to another person so our privacy should be views and viewed in this very contextual way and Gary zimmel has talked about the power of secrets on strengthening our relationships with one another which will become important at the later parts of this talk context is under threat um in many ways sorry um and as we've started to see here where um we can no longer trust the voice of our loved ones anymore if I have a picture of your face in three seconds of your voice I can create a pretty credible version of you in a way that we've seen in scams can persuade someone's grandmother to give their supposed granddaughter twenty one thousand dollars and these scams are only starting to emerge in much more persuasive ways so if your

face and your voice is no longer meaningful as a way to preserve your identity online how are we going to start to communicate with one another and trust one another so recently our team published this paper in the Harvard ethics Journal titled plural publics which offers a framework for how we believe we should be thinking about both the Privacy Frameworks along with how we can build communication channels online to be able to have integrity with one another we actually think that AI might be the precise attack we've always needed to communicate with one another more effectively for the better part of the 2000s privacy has always been this nice to have the reality is none of us change our passwords as frequently as we should and we don't care enough to do tell our

parents or our friends to do that as well but privacy goes from this nice to have to this necessity because if you lose the key parts of your identity you may no longer be able to authenticate yourself in the digital world you may no longer be able to access your bank account communicate with your friends or communicate with your parents in the digital world and so now privacy goes from this not like yeah nice to have to a much more of a necessity and so this plural public paper tries to offer a framework for defining what exactly an incredibly High information Integrity space could look like and we take the framing from John Dewey who's a 1920s American philosopher that had this book titled the public and its problems and in this book he defines this idea of

emergent publics emergent publics are spaces where people come together to deliberate and talk about how to solve problems for issues that are expressing against a very specific set of people jcr lichlider who's the oh sorry that slide was missing jcr licklider who is the father of the arpanet which is also known as the precursor to the modern internet um had this idea of building upon the initial vision of Dewey's emerging publics and jcr look later was unable to do so in a way that we wanted to now talk about and so there's five key tools on that slide previous that we started working on to build tours and so Jaren Lanier who's a researcher at Microsoft as well had this

uh post titled how to fix Twitter and all of social media and a lot of it was focused on how we need to start communicating in more group chats instead of these broad-based social platforms like Twitter or Instagram group chats come with this very interesting idea of reputation and social cost to communication when I enter a group chat with someone on iMessage signal telegram whatever it may be it's mapped to This Global identifier with my phone number in this context potentially that people can penalize if I operate in a very harmful way and what is harmful harmful is contextual so if I show up into my group chat with my family and start communicating in a way that is that is very different from my friends my family may look at that

username in a very different way but what's important here is that the reason why we think group chats are incredibly important is because it's really easy to delineate whether or not someone's using an artificial version of themselves or not in a way that respects social norms so as an example in this room here if we had some inside joke about the blue chairs in this room and I started using an AI to communicate with one of you that didn't pick up on that joke then you would know that it's not picking up on some contextual cue or symbol that our relationship had with one another um another really important primitive that we released earlier this year is this idea of what's known as a designated verifier signature a designated verifier signature says if I send you a message you should be the

only person who can verify that message and no one else in the world should be able to verify it the reason why that's really important is because it's for two reasons one of which is you should be the only person because you are the contextual individual that I've assumed has all of the assumptions to interpret that information meaning if I post something on Twitter maybe I should post it to the people who were at adcon instead of all of my Twitter followers because they didn't share that same experience of being in this room when something may have happened the second reason is that AI labs are actually incentivized to now verify the data that they train on not for any purpose of like great creating better models but purely because today gpt4 Llama Or all these other tools are built on this

distribution of human language that is actually human for now but as these models start to pollute the internet more and more we're no longer going to know what is representation of the human Corpus and an artificially generated Corpus but if we can verify through these types of signature schemes that we believe are quite powerful that something is actually human generated then AI Labs will be incentivized to pay in for that for that information in a way that also preserves information Integrity for the communicator um for time purposes we'll skip over the circuit um there's also a really important power that blockchains have in being able to create what's known as common knowledge common knowledge is this idea of I don't know some I don't just know something

you don't just know something but I know that you know that I know something and that is really powerful when it comes to contextual communication I don't want to just know that you know English I want you to know that I know that you know English so that I will actually communicate in English and there's this sense of common knowledge that is critical for coordination purposes that is really important in supply chain military sense but also even just coordinating our everyday actions that we believe a blockchain can be very powerful in enhancing finally we did a lot of work last year on creating identity certificates because we don't know the state today of what Primitives are useful for varying verifying someone's identity there is a centralized view that maybe scanning someone's Iris or having oauth depending on one platform and there's this more

chaotic View and having attestations across various different types of Platforms in the web through ecosystem we don't know what the answer is today but we know that we need to be exploring it in a way that enhances privacy in a way that preserves context in a way that we can feel is credible and so we did some work on that here as well um we put together a lab today it's one of the largest research Labs that's working on this specific problem at Microsoft research it's called the plural Tech research collaboratory with over 100 researchers focusing on this very specific problem of preserving information and integrity along with very various other sets of research problems in the web 3 and AI adjacent space um some of the ideas that we've been thinking about now are specifically in

my research has been focusing on is how can we preserve the Integrity of collectives identities so we know that these large Foundation models are incredibly powerful what if each Dow could have their own AI of themselves and monetize that Dao's Ai and protect the Integrity of that dial so what does that even mean it means if I took all the conversations all of the metrics all of the data that our Dow created in our Discord chats or telegram chats or slack chats whatever it may be and use that data in a way to strengthen the power of an lfm and open source lfm through this process known as reinforcement learning with human feedback maybe and just maybe we could actually create an AI for each dial in a way that those Dows actually

own and for those of you who are unfamiliar with reinforcement learning with human feedback it's what actually took these large Foundation models from being able to predict next tokens to being able to mimic human-like performance and a lot of that process Focus focuses on fine-tuning on human-like data or having human annotators label the outputs of these models so instead of having arbitrary open AI humans label that data why don't we have our own dials label that that data and own that data and own the AI that emerges from that data um the other part of this is identity data we need to be thinking more about how are the identity data that we have can be used to both authenticate ourselves

in a way that doesn't take on this very centralized view as I talked about before there's this idea of what's known as intersectional social identity which says that instead of taking just one attestation and whether it be a government credential or a biometric scan let's take the intersection of many attestations and the intersection of those attestations is what will actually have a much higher probabilistic estimate with more credibility for you to be a unique Human Than otherwise hmm secondly we need to think more about contextual communication flows today when I post something on Twitter I really only have two options I can post to everyone or I can post to my close friends list you see similar style of private versus public Forums on Instagram with your private Story versus

your public we need to be thinking about more contextual ways of posting in a way that people can verify and preserve the Integrity of that information maybe I want to post to my close friends and everyone at edcon or close friends and get coin contributors and it should be done in a way that is designated verifiers so only those people can verify that message and no one else will have any assumptions or judgments towards that information when I post about it and so what is the future of our research look like the future of our research is focusing on building context preserving Technologies it's focused on Guiding policy and finally running experiments in communities and dials that want to start working with these context tools the final thought I want to leave people with is this idea that our secrets are

our Turing tests today when you if you ever heard of captcha captcha is this way of looking at someone and determining whether or not they're a human or a bot and a captcha it takes on this very frequentist approach of saying in this very moment can you solve this puzzle and if so then you are probably not a robot those puzzles are now mundane and no longer sufficient for determining whether or not someone's a human your face is not even sufficient for determining whether or not you're a human and nor is your voice even sufficient for determining whether or not you're a human we may need to start looking at our prior informations our secrets to start determining whether or not someone is a human because the only thing we can determine about one another is the shared memories we share with each other those shared memories are the

things that we can start to use to authenticate the fact that I shared a meal with you and the wines built on the floor and that like these types of memories are the things that we're going to start to emerge and so if this work is very interesting to you and you feel called to contribute to it please do follow me on Twitter or substack or reach out to our team we have a lot of Open Source projects both in the web 3 community and AI spaces that we're looking for more people to contribute to and thank you thank you [Applause] [Music]