DAPPCON 2024: Why Our AI Future Needs Data Provenance - Joel Thorstensen

Transcript

I'm by no means an AI expert um but I've had a lot of experience with data Providence because I've been building decentralized data protocols and so today I'll basically be talking about sort of my observations over the years and sort of how I think these two things intersect so I wanted to start talking about sort of the problem of AI like the risks ahead and it seems like there's we're in this path and there's basically two paths that both are kind of bad so um in one scenario we go into a world there's centralized controls someone controls all the AIS there's no privacy there's censorship um and I can sort of do what

they want and and the risky thing here I guess like also is like a winner takes all scenario but this is not really like a future we want and the other side which I think the side that like many of us here probably are thinking is the right side is like the decentralized AI like everything all the capabilities of AIS at the edges and this is sort of nice if we think about it from like a self- soverign perspective but it actually brings a lot of risks um so deep fakes is one thing but even beyond that like if you can just use an AI to create um sort of drugs or like very dangerous weapons it's a real problem and so uh we need to actually have some way of having

alignment which is not only aligning the AIS to humans but actually aligning humans all across the world to like not everyone compete for this Winner Takes all centralized control scenario um so I wanted to start talking about trust uh so if you think about this deep fake problem uh there's there's a few different approaches there being approached um uh the main one is being by a Consortium companies called content authenticity initiative this is basically like a few companies coming together uh to sort of verify content online and basically what it is is signatures hashing and time stamping and funnily enough is stuff that like the

crypto Community is actually really good at but what this looks like in practice is that you snap a photo with your camera so maybe you have a Canon camera um that camera has a key on it that signs the content and that key has a certificate from Canon the company then you edit this photo in Photoshop and Photoshop sort of signs that edit and so you can have this like uh flow of all the edits that happen to a photo and this is nice right we can if the website supports it we can actually look at all the things that happened in in the edits and we can know like oh this is definitely not a deep fake there's sort of like two problems with this the first one is that someone might not just use this and just upload a random picture

without any metadata and we have no way of knowing if it's like real or if it's it's if it's a fake one the other problem is that now we have these like uh big companies that all control what is like true right um and we can imagine some entity maybe wanting to put pressure on all these companies to like censor some particular happening in the world so like all the photos coming from that particular place actually is like now considered uh false uh so there's like two problems with that the other approach to this problem is something called perceptual hashing which is basically taking a neural net um an AI tool and looking at an image generating

sort of an hash that represents the picture and you can have two different pictures that is like slightly edited and you can get the same hash uh so this this is nice and I think there's more to be explored here um but it sort of has problems that sort of neuronet things have in general which is that uh they can be easy to fake um and so there there are some like problems like actually making this really robust but even Beyond these two approaches that I talked about uh they really only apply to like images audio video maybe like written content but it doesn't really work when it comes to

like the stuff we get with jailbreaking model and like generating like really harmful content so imagine um someone who's using a model to generate uh like a like a like a recipe for how to make some drug or to how to make some kind of harmful weapon um that we really don't have a problem or we don't really have a solution for this problem except like the centralized control where someone sort of surveils everything and looks at what everyone is doing um and like the one key thing to to realize here is that like if there is someone sort of that is not sort of mentally well they might actually get access to a

model like this and do something like this and so we actually need in order to solve this problem in a way which is not like authoritarian and surveillance State we need some sort of web of trust we need to have communities and prer accountability in like actually small groups of people and that needs to sort of scale up to multiple layers up until like the entire planet and so this is really Big Challenge and I think the web3 community actually has the tools to be able to get at solving this problem um I think also if we have like a we of trust like this it can serve as like um a basis for something like quadratic voting and funding which which is really good way of like thinking about more uh digital democracy um but how can we make this

web of trust web of trust has been this like term that's been around for a very long time uh and it never really like worked never really took off uh I'm sure a lot of people here since it's it's sort of a web 3 conference you use signal but how many of us actually like go and verify all of our contacts that like we actually verify each other's Keys it's just like a social interaction that is like has a lot of friction so in order to make web of trust work I think we need to make web of trust implicit in everything we do all the actions that we do online and in sort of digital environments um so so ideally we can end up in a place where we can go to like an ether scan of data and we can just look up any action that was taken and see

that like yeah it's valid and it was taken by this person that's like one hop basically my friend knows them um but there's a problem with this of course it's like if if it's a down vote should that be like treated the same way in terms of trust as a very good positive review well obviously not right so how do we deal with that we basically need some way for the computers to understand what's going on and so this actually been a long-standing effort to do this called the semantic web for those those who know it it's kind of funny the semantic web was the thing that was called Web 3 before web 3 existed so this effort by

w3c to create the standard called jsld to make the web machine readable but Json LD not really took off it's like burdensome for developers to implement it but the nice thing now is that we have llms and llms sort of solve this problem of making the web interpretable by machines and we can create something called vector embeddings and Vector embeddings is basically pointing your model at a piece of content like a website or a PDF or something like that generating something called a vector uh that basically represents uh this content in a way that the AI model understands and then we can create an index of this so if I search the web I can get um results based on content and if we can do this in a way where like

all the content that individuals creates have embeddings uh we can actually search sort of more uh in a more like human way with content so I can say like is this person over here um you know reputable when it comes to biochemistry and I can get back uh s written human language result but I also have references to cryptographically signed sort of at the Stations from my friends and like this hops to like make me trust that statement and so this is pretty cool can we just like embed everything on the internet uh and have every combining

this with like the signatures we talked about before uh we kind of can but the main problem is is like how do we trust the Victor embeddings if I have a piece of content how do I know that this embedding is this Vector actually is accurate someone could create like a fake Vector for a piece of content to put it in the wrong place um basically the way you would solve this would be to have um these embeddings signed by people who actually run the model so people who are running these like big closed models like open AI Google and so on actually can you sign the outputs that they give us and assign the embeddings that we generate using these models uh so that's sort of solves the problem unfortunately like it puts a lot

of trust in these like big AI companies ideally we could do something like making a a knowledge proof of an embedding that is using one of these like open models like mol or the smaller llama model um I think in theory this possible uh it's probably like really hard in practice right now um okay so great we have all the internet in as embeddings everyone has signed everything this sounds pretty terrible for privacy right like we're not really that much better than the world where this a centralized entity that is like monitoring everything uh so what do we do about that well I think first of all we should just realize that in general it seems

like right now ai is pretty bad for privacy um everything we do basically involves talking to some company that gets all of our data and uh in order to run these like inferences and give us like this really useful uh um you know interactions that we get from AI so in the web three space there's been a bunch of talk about something called Data backpack which is basically this idea that you put on a backpack you put all your data in it and you go around to different services and essentially like give those Services pieces of your data and selectively disclose them this seems like reasonable but unfortunately in in the world of like an AI where we want as

much relevance as possible we're likely going to end up walking around with an open backpack where like all the services that we use just get access to all of our data so it's it's not really like a great solution and so I think what we really need to do instead is instead of users taking their data to the applications that they use um we need applications to come to the users and use the data locally with the user so we need to return to a world where we download the application and we run it locally on device um otherwise there's no way we will have privacy in the future um but we can do this and we can also have social spaces between users

where we have privacy globally but we have like sort of sharing data and sharing our Vector EMB betes between people uh and I think one of the cool things about llms is that we can actually use these tools to generate uh the applications that we need for the moment so if I need a very specific way to like edit an image or edit a spreadsheet I can just ask the AI model to generate that tool and now I have it locally maybe no one else have it but I can share it with them if they need it and we can collaborate on that piece of data all right so um I wanted to talk about a little bit of sort of few promising projects that are moving towards this direction uh so first of all getcoin

passport um they're basically building um the fundamental building block of the web of trust uh they allow you to uh basically verify your identity against like a bunch of different services and then use that as a basis for uh civil resistance score and you can uh use this in gitcoin of course to do quadratic funding um but I think it can be used for so much more especially if they start doing more sort of privacy preserving parts of your passport that you can just selectively disclose to different people um index network is another project um that is taking it upon themselves to

sort of start building This Global vctor index of of data on the internet and started sort of doing that indexing but I think what's really cool about index network is uh they have this this this feature that allows you to not subscribe to a channel not follow people but you can subscribe to a semantic search um so basically what that means is like I can go to to to like the application and say hey I want to know about all new hattricks that happen in football anywhere globally and I'll get sort of a data feed of like all that that happens when those sort of games happen right so that's that's pretty cool uh coordination Network and these I laabs are building sort of tools for

decentralized science uh so I think science can be like a really cool first use case for uh this type of web of trust where in some cases uh D labs for example is doing a um decentralized peer review system and so that's like really tangible of this like web of trust thing Powerhouse is doing something similar but mostly for Dows and so on and coordination network is cool because it actually starts to allow you to customize prompts for different things and share those between users so we can have this like on the-fly generation of uh applications they're sort of getting moving in that

direction um and so where am I talking about these projects in particular well the thing they have in common is that they're all built on the project that I'm working on which is ceramic so the main thing that ceramic provides that makes all this sort of easy is that um is is mutable objects and so these objects are version controlled um they're temper resistance so time stamped on chain and they're signed by cryptographically by end users so with cryptographic access control and the nice thing is this objects are independently verifiable so I don't need to synchronize the state of the entire network to know the state of one object I can download one object keep it locally on my computer and verify that independently without necessarily even

revealing it to other people and this is all in a global name space so all data on ceramic is addressed with a cryptographic identifier and each node can sort of Select which part of the network that they share data to okay so what about privacy well right now all data on ceramic is written in public and other nodes can synchronize it one of our key focuses right now is building a physical Access Control layer which means that I can select who is allowed to synchronize data from my node and data doesn't actually physically move around the network unless I allow it to do so and with this we want to push data out

to towards the edges so that users can hold their data without sending it around and any type of like company or other entity can just like hold the data and share it with whom they would please um all right so based on sort of what we talked about so far like what what is really missing in in the space of like data Providence and and and sort of the intersection with AI I think we need better tools for Comm Comm Community formation and what this means is like this web of trust only really works if we can have some way of like uh determining what are the different communities and how do they interact uh this is also like really

important for things like quadratic funding and voting and so on um and so we need just like better tooling for creating communities I think one thing that's really cool that comes out of the complex systems science uh realm is something called Community detection algorithms so this is actually like a branch of like research so basically algorithms that you run over social networks to determine like what are the different communities in these uh social networks and and sort of what are their relationship um more tools that use like quadratic voting and signaling I think in in particular it would be interesting to see sort of using quadratic curation to curate like an rhf data set that we can use to sort of have a more

democratic uh democratically trained AI finally we need better ways of sort of having accountability between peers and in communities and this needs to be at all scales like small groups of people and Friends larger communities uh nation state size and like ultimately globally and one key takeaway that has like uh is is really impactful to think about at least for me is that in this world of like AI That's just like like more and more advanced the the health of local communities is actually a critical requirement for us to have security

globally um and so what do I mean by that well basically if there is like a community that has like a bunch of uh unhealthy like mentally unhealthy people they get access to the massively powerful AI tools and can you see like oh I just want to destroy the world and they just like go and figure out how to create some sort of biow or something like this that actually is like Global risk and so it's actually critical that our local communities are healthy so that we can have security globally and um I hope that with these tools that we've been building for years in the web three space we can actually help to address these issues all right thank

you thank you Joel We have time for just one question over here yeah I'm sorry it's not really a question but I just wanted to congratulate on your last sentence I think you're absolutely right the only way the only thing that uh a rogue AI could not attack is actually small communities because that's where we can filter for AI all other stuff we cannot filter for AI and that's where they where we cannot defend so yeah that's probably the key to AI safety yeah it's it's really scary but it's also like

sort of I want to live in the world where we solve that problem because that seems really nice thank you so much Joel for your talk Round of Applause for Joel