Decentralized Trust and AI | Golda Velez - 3box Labs

and everybody welcome we got gold is speaking and and we are good to go so everybody enjoy stage is yours all right thank you guys and um yeah actually I'm really excited to be following Tracy from Pluto who was just talking about AIS and fraud and Trust because I'm also going to be talking about AIS fraud trust and uh those kind of problems I think it's kind of interesting I just actually was conference hopping I was here at eth Denver and I went to two AI conferences and back to East Denver and uh you know yesterday in San Francisco it's like everybody's like oh oh AI is everything everything in the world is going to be AI it's going to be so great and it was like so much hype and it's actually very refreshing to come back here and also get the anchoring pardon my pun uh to Identity and who actually is talking and uh you know what what is true what is not so AI actually causes as as Tracy was saying a greater need for Trust Systems because it's so much easier um let me go back here uh it's you know risk I used to do risk work for risk work for Postmates so I was actually fighting uh credit card fraudsters at scale and when you think of fraud it's not always just like Joe saying stupid stuff on the internet that's misinformation using something it may be a very organized crime ring that's actually doing fraud for profit and when you're fighting risk Rings you're actually fighting another team team of frauders uh that's that's attacking your system and it it's kind of that spy versus spy thing and so there's various types of risk and fraud that you can think of um some of them that we would run into uh is account takeovers where maybe there's a guy's account that was a good account but then someone got the password or private key uh they do a lot of massive you know uh laundering so all those stolen credit cards they don't just get used by the guy who got them they sell them they bulk transfer them and then there are people who uh you know launder them at scale um you know fishing attacks uh social hacks is actually a big one we used to always tell all of the drivers you know we never ask for your password and then of course somebody calls them was like oh we're Postmates can we have your password and they were like yes so a social hacking is actually a big thing that's that's how actually a lot of chains have been compromised and of course ransomware so you really have these kind of high stakes games that are being being played so it's not just a matter of like is this a human but like there's actually attackers that are systematically trying to subvert whatever system you might make um and generative AI makes detecting fraud harder because llms can simulate human behavior much better than anything we've seen before so telling whether something is a human based on its Behavior based on its timing based on its click patterns um all those things that we normally were able to do a lot of those are not going to work so really I'm seeing that we're going to need to transition to more trusted and gated spaces um for a lot of us and that's kind of what you're saying as far as do I want to interact with all the Bots do I want to interact with all the noise so there's a lot of opt-in spaces nowadays there's like allow lists you know your mail is probably being sorted to people that you have previously talked to you know my phone is blocked unless I have people in my contacts essentially we're creating ourselves little walled Gardens that are safer spaces for ourselves and it's interesting you know coming at it like from a biology perspective there's a reason that cell walls evolved inside a trusted space you can be much more open and it's actually more efficient um and I can tell you this from you know looking at what is the bottlenecks on for example an ipfs if you're reading all the traffic from ipfs the bottleneck is going to be the signature validation so signature validation all of those types of validations are expensive even without going to the blockchain even without paying the gas blockchain is expensive to use signature validations are expensive if you have a trusted space in which you don't have to validate you can go and scale much more I mean we would we would even not use https on the internal networks because you want to speed things up as fast as you can once you're getting to massive scale you don't want to have to validate every packet um so how do you create trusted spaces supposing that you want to um you need really two things you need long-term identifiers and you have to be able to share data about the behavior of those identifiers I think a lot of people here are probably familiar with the prisoners dilemma and what that is is it's simply it's a it's a game theory that if there's a repeated game they're going to behave better you know the classic example of splitting the check in a restaurant if you know the people you're going to split it more fairly than if it's like a onetime thing and you're never going to see them again and if it's Anonymous of course you're going to possibly behave worse so if there's long-term identifiers it creates the potential for accountability the potential for consequences for Behavior so how do you get long-term identifiers in this world where an llm could like pretend to be anybody at any moment and then how do you gather the information about them um one of the big things actually from working at Postmates when I was doing risk and fraud is I would do all this work and identify some kind of identifier we called them assets that were hard to necessarily replace so it might have been a device ID it wasn't really a did with a signature it was just a device ID but these things still would have reputations in our system and if they had a bad reputation we wouldn't let them in now I might find a device ID being used by hackers like 5,000 logins just got associated with this device ID but I would not tell the people at Lyft and the same guy is probably also laundering credit cards through Lyft and laundering them through Uber and laundering them through every other place that takes credit cards I mean what they do is I'll just tell you a funny story they would actually advertise Postmates Pizza on WeChat and Postmates didn't sell on WeChat but if you ordered pizza on WeChat they would actually fulfill the order by taking a stolen credit card making an order at Postmates the guy gets his pizza but he would pay the WeChat dudes so it was this you know he they could do massive fraud at scale with pizza because it's being orchestrated so these are the oh yeah let me keep this going um let me make sure okay so anyway I I would like to be able to share this behavior and that's one of the things I'm going to propose here is that ability to have long-term identifiers as kind of a cultural expectation and then a way to sh share that data in data pools that's the two things I want just two things um so here's a did if you haven't seen a did before this is a did um dids are nice because they're a standard they have private keys that can rotate depending on the did methods if like lose your keys you might be able to have a rotation method where you could restore your keys and they're standard across different things so like blue sky uses dids every message on Blue Sky is signed with a did Noster uses dids every message on Noster is signed with a did did ceramic uses dids every message on ceramic or orbus which is built on ceramic uses dids and those are completely separate protocols but they're all signed with dids and did is the thing like in your metamask wallet when you sign something you're signing with your did your decentralized identifier so they can be cross a protocol cross chain because there's a did method that's actually a w33 standard um so that provides a powerful identifier that can cross all these different things and one of the reasons you want to cross things is if you have graph information if you have cross domain information you can protect against fraud better because you have more knowledge about the guy when a guy just shows up you don't know anything about him you have no idea if he's a fraudster or not but if you see him in a little ring with other fraudsters probably he is if he's connected to them so there's if he shares like an IP with them he probably is if he shares a device ID with them so so you want to have this cross doain information though you also want to preserve privacy so I just want to also give a shout out to separation of concerns old-fashioned programming stuff okay please um so signal and models are two separate things the fact that I sign a message doesn't make it true the fact that I sign a message this guy's a human doesn't make him human maybe I got fooled it's just a piece of signal so the way we would do it old on web to is you have a Kafka stream ton of events coming at you boom boom boom you process them you store the significant ones in a database and then you roll up features and you feed them to models and believe me I used to be a proponent of heuristics and rules models are smarter I'm sorry at Scale Models can catch the fraudsters better than my rules no matter how hard I try to like guess at the best rules so if you have a source of Truth for the model where you can tell it hey I found out this was fraud and we used to do that by allowing some of the stuff to go to charge back and the fraudsters would win a little bit um but then we would find them so if you have a model it can roll up all these signals and make a prediction and then if you have a source of truth of knowing when those predictions are true you train the model and the model then is trying to outwit the fraudsters as they constantly change their behavior so I think from an engineering standpoint it's better to separate these ideas of a signed thing or a thing that's on blockchain or even a thing that you proved with ZK trust or whatever is just a signal and it may be a strong signal it may be a weak signal but it's not an absolute it's it's a signal that goes into the model as a feature the feat and then the model makes a prediction fraud not fraud and then you have maybe in some percentage of cases you have some humans looking at it you have bad outcomes depending what outcome you're concerned about you'll know if it was a credit card fraud because it'll be a charge back two months later so anyway you want to have signals you want to have models that read the signals and make predictions um so yeah web three and now comes the ceramic part I'm going to put my ceramic hat on I work for ceramic I like ceramic for this because it is composable it's all about composable data it's very low cost it's cheaper than an L2 um we have basically schemas that you can register other people can consume them it's just like mongodb except that it's on it's on like um well right now it's on ipfs but if you run a ceramic node it'll actually become a little better than ipfs soon um and they talk to each other they share the data on streams that you can publish to and consume and they're cheaper than an L2 but they are anchored to ethereum so the idea is that it's very scalable so a lot of people can publish and consume on composed EB at scale um and the hashes are being anchored to ethereum but the whole thing isn't being published to ethereum because that'd be really expensive so I want to do a live demo and I figured there would be no good um uh Wi-Fi here so I did it locally we'll see if it works and this is just a really hokey demo but it's like I would like there to be a culture that every entity should sign with a did now normally of course it's not going to explicitly like talk to you and sign with a did it's going to be something you would call or you would use some code to do but I just made a fun demo um I'm running Neo GPT here um and I'll just show you that it is live if it ever runs and so this is like a really stupid model it only has 250 million parameters so somebody give me something or I'll uh I'll say hey I'm Tracy from Pluto and let's see okay so Neo GPT is thinking and he says um here to give you more information about the command config he's just making up because this is a really dumb model but I'm going to ask it who are you ooh and my little rapper answered because I programmed it to that if you ask it who are you now it signs a message with a did now obviously you're not normally going to talk to a model and be like who are you let me see your did but if you if you did that programmatically or directly I think there should be a cultural expectation that everybody everywhere should have a did that's going to talk to me on the internet if I don't already you know if it's not over some kind of secure Channel that I already know because also you can't really tell AI from people unless you make every human get kyc and I don't want to make every human get kyc so I have no idea if I'm talking to a human or a model but I just want the damn thing to have a did so I can talk about it so for example going back to the demo um I did this earlier and I did the same thing and I made it give me the did and then I put the did into uh this is a a company I I sometimes see but you you could program lots of different places that would do this but this is a group that's doing decentralized trust and linked claims and then I'm going to tell uh this interface hey this did doesn't make any sense it's probably not Tracy really it says it is but it's not and so then I'm going to sign that attestation with my metamask and I'm just going to say something about it and I can do that whether this bot gave me permission to or not because it has a did so I can say stuff about the did and then the stuff I say can get published um on the ceramic Network and then other people can see it like cers scan is not my thing I don't run cers scan but as soon as I said something about this did it shows up on cers scan because it's part of the ceramic Network and now other people could consume that data and they can run their models so all the signal can go out on the network and we're using the ceramic Network as an example because it's nice and scalable but you could do it other ways I can consume it and make a model over it you can consume it make a model over it if you don't like my predictions make your predic itions use your anchors of truth but if we could share this data then we can all be more powerful against the spammers and this is an example of that little guy link trust he makes a because he doesn't have enough data right now um those guys don't have enough data to build models they're just showing you a graph of like who said what about who that you could explore so that might also be useful but but the main thing is to share the data and then I can decide which people's data I'm going to consume like I only trust these guys I'm only going to consume their data I'm not going to trust those guys I don't know them and then it still is a game of spy versus spy but the good guys you know if you ever played Avalon you know the good guys are the ones who don't have to lie if you play the very naive version of Avalon the the guys who can tell each other the truth and be open those are the good guys um can win um so I'm actually ending kind of early I'm going to give some chance for questions um but here are some places you can look at more stuff share data at scale we have the stuff composed DB where you can like make a model make any schema you want publish it consume it make an app and then totally separately not related to ceramic is like if you want to play with AI go get some models from hugging safe mistol and you can start playing with the AI models which is just like they're doing more and more stuff faster and faster um so yeah I want to take a few minutes for questions I don't know if anybody has any any that'ss about scale okay um all right well if you're interested in anything about oh wait there's a question oh good I like questions questions are good um hi uh so everything you talked about is underpinned by using dids uh and uh using dids very pervasively but the the the amount of stuff in the world is so huge isn't there like isn't there a risk of did explosion and at that scale what does a did mean if there are billions if not trillions of dids being generated all the time okay I I'm not sure I heard everything you said but you said basically will dids scale and like what does a did mean at scale is that yeah I mean an explosion of dids could be a problem yeah I mean obviously yeah you do have scalability problems it depends on what you know State you're in I think the dids are really the core thing I mean we were able to store data about millions of you know card fingerprints millions of devices millions of logins at you know without any problem just like in postgress and it fits very nicely into a model um I think dids you know there's less of them than there are messages because like there's n billion people and then maybe another n billion AIS and maybe another 9 billion other things but you know it's it's on the scale of countable billions of dids which a model can handle that a big model can um so yeah somebody has to be consuming and then rolling up that data and yeah that person who's rolling it up is going to have to have some resources and probably charge for their models predictions at some point you know maybe they'll give some premium or something but yeah actually rolling up the billions and billions of dids and the data is going to be expensive somebody has to do that and be funded to do that that um but just the fact that they're being transmitted um you know you don't have to store all of them they're going to be flying by they're just kind of like traffic that's flying by so everybody's not necessarily keeping all of them all the time but yeah somebody has to in order to do effective risk you have to aggregate a lot of data question uh the ceramic model that you mentioned is are there any use cases in in the world right now that is getting traction on um wait I'm sorry can you say that one more time I definitely want to hear that are there actually used cases where this uh model that you described is uh gaining traction on I would like to learn more about the real life use cases what are the real world use cases right now for ceramic models correct yeah yeah so like orbus is using them so orbus has uh a social network on top of ceramic models there's um yeah thank you for that question and thank you for the other one um there's a some decentralized science uh use cases where they're sharing scientific data um I'd like to see people actually sharing the parameters from Federated models I don't think people are doing that yet but like for predictive models you can actually do Federated predictions and just share the parameters somebody should do that um of people who are actually doing stuff uh gitcoin is the biggest example example right now um all of if you go on open Collective and you're you're you what get getcoin passport stamps if you've ever done that like your validate your Twitter account that's all on compos DB that's all on ceramic um then there's uh o is doing some stuff um I should know this because I actually measure these guys um oh I guess my time is just about up uh gosh I shouldn't leave people out but there there are a number of people doing stuff uh from decentralized science to social packed um so so there's a number of people a lot there are reputation cases and other kinds of things in social networks and game points yeah all right oh my time's up