The Privacy paradox in AI by Daniel Luca [Eden Block] x Ethereum Cypherpunk Congress

today I want to talk about um the Privacy Paradox in Ai and I'm going to go through a few ideas and try to explain exactly why I think there's a paradox especially with AI and what that means to us as individuals or as companies or as um anything else which is related to this technology and also with privacy at the same time so I'm Daniel Luca um also known as clean unicorn I'm like everywhere clean unicorn so in in in one sense I'm not super private uh I'm an entrepreneur and to be honest mostly a developer like the more the CTO type um and I'm also a security auditor so I was a security auditor at consensus diligence um and also Air Tech which was my own own security practice um there's it's not part of the uh part of this but also co-founded the stable coin so I had the opportunity of working pretty closely with uh investors and building a a product and after I did that I thought maybe there's a few ways I can improve it so um now I do take at Edom block which is a crypt VC so we invest a lot in uh web 3 projects and AI so I think privacy is a basic necessity is part like if you're a human it's an innate need to keep some facts to ourselves uh private and only share them with people you trust and you care about and you value their judgment and who understand you um or even someone of those facts you don't want to share with anyone um I do think that you might have heard this but some people say if you haven't done anything bad there's nothing to worry about um and I would love to see in the audience if if people believe that so if you can put your hand up so I we can kick you out um I think this is a good explanation of like it's it's not just that it's not about doing bad things um so everybody has their own limits when it comes to how comfortable they are with u privacy and so everybody has some kind of privacy need they have to accomplish and this is I think a good example I mean maybe you want to say like I don't need a lot of privacy but yeah I'm not sure so there there is a spectrum of like I'm extremely private so I don't even have a SmartPhone I only use a Nokia 3310 which is even like worse than you know having a graphin or anything like that and uh on the other part of the spectrum is like I use password as my password um and I honestly was like that when I was really young I used password but you know spelled backwards so I thought that was smart um it was it was pretty cool back then but I was maybe 12 so I think privacy has uh one of very interesting Parts is that privacy is safety for people I want to talk about like really understanding why this is the case and how we should use it so this is fny um but this is my dog I named him after uh haly you might have heard of him one of the uh Bitcoin developers and the fact that I'm sharing this with you says something about me and how I relate to you right now so I am open to share this fact with you with um a circle of friends with people that I trust and I feel like I have something in uh in common um thing is I have to walk him every day because he's a dog and he has to um do some kind of exercise but also he has to poop so when he poops because I'm the one responsible and I have him on a leash I kind of watch him I kind of have to watch him him doing his business so he doesn't have a lot of privacy and um to be honest I'm trying to give him more privacy but he has to earn it over time cuz I'm the one responsible I know he doesn't uh he's not super he wants to be free but he knows that freedom comes with a price um now the the interesting thing is I really don't know if he wants more freedom or not but like I assume he does uh this is uh Dr Sydney jurard he's a Canadian psychologist who in the 60s and 70s he studied people in different instances of having um levels of privacy and he created a self-disclosure theory and um this theory is interesting because it emphasizes the role of self-disclosure openly sharing personal thoughts feelings and experiences in developing relationships self- understanding and even personal growth he argued that uh self-disclosure is essential for mental health and allows individuals to gain insight into their own identities and understand others better um this creates stronger bonds and um jurard saw this act of disclosure as Central to authentic connection and he was suggesting that withholding thoughts and emotions could lead to alienation or poor psychological Health in in some extreme cases it actually led to uh physical pain so I I'm I'm presenting all of this because people in this community have a hard time finding each other online you're like this weird Avatar with a madeup name and a madeup username and you go to these conferences and it's so difficult to find the people you actually interacted with before um so I don't think this is very beneficial for the whole Community as a whole because you do need more connections and stronger connections with the people in order to also organize yourself but you know being an anarchist kind of destroys the idea of being organized so there's there's a duality there which I find interesting um so what happens in the uh extreme case this is the uh the panopticon this is um for a really long time this was a concept created by an English philosopher and social theorist uh his name is jemy bom so he created this ideum if you look at the image you see on the outside this is a prison so like just to be clear is the design for a prison and the idea was that if you create such a an environment where you always see the inmates they are going to behave better and on the outside are the um uh their cells and on the inside you will find guards who always watch them the idea is that you don't know exactly if you watched right now but you have the feeling of being watched this whole design is built in such a way that you're not allowed to have a lot of privacy and when individuals are constantly monitored they do alter their behavior uh this forces them to conform socially adopt common norms and kind of change their beliefs or preferences in essence constant surveillance creates self-censorship and people avoid actions that could bring unwanted attention this does create multiple problems or a few things is like you lose self-expression you lose autonomy and it's very difficult to create genuine relationships now I think that we kind of live in a digital panopticon um in this digital space it's maybe less obvious that we're being watched all the time but all of our data is tracked even if you use a Nokia 3310 you you're still tracked the um the Telecom company knows where you are so it still has your location um in the digital world being highly literate and knowing your data is collected knowing your data is processed and used um makes the experience kind of knowing that you are under surveillance all the time it kind of makes it objectively worse because you know you can't stop it and you know you're being watched so you kind of alter your behavior you're not yourself so much so I think that in this case uh knowledge is anxiety it's usually knowledge is power but having this kind of knowledge creates a sense of paranoia uh when you realize how much is being collected and the more you know the worse it is so I'm coming with a few examples right now this is um uh slack teams uh Discord I'm not sure if you know but your uh slack admin can read your DMs if they want to this is kind of scary when you figure it out uh maybe Discord and other like you you don't read the DMS but you still have a lot of power over what people um what you know about the people about your members there and even if the admin doesn't have that the Discord uh has access to that data the Discord team um if you look at other Industries uh okay so this I'm not sure if anyone recognizes this uh I'm going to give you a hint it's uh part of Blizzard uh blizzard is this game company uh created Starcraft and and a few other great games um this is deer Kane um it's kind of interesting that this character from Diablo was providing the uh the player advice gossip and it would even identify items for you um he was he kind of was the the knowledge owner the knowledge master so ironically or maybe non ironically he's presenting this uh end user license agreement which um is kind of tough to look at so this is how um items would look like unidentified for example and if you go through the uh and use their license agreement you will see that blizzard can transfer your data to other servers so that means that different rules apply once that data is in different geographies and you do consent to be monitored all data and Communications are collected but you know at least you get to own the games that you really love um however in fact you don't even own the games they're not sold to you they own the games they can take it from your account they can do whatever they want so you're not even getting that um a bit about open AI so open AI has this uh and use the license agreement and you have um different geographies so that means that they try to take the most and use the most out of data based on each geography so like it's not like they're trying to be private they're trying to collect as much as possible as much as the law lets them um this is cursor cursor is a a pretty good ide to write code and um if you enable privacy mode and if you keep privacy mode off you have different settings of doing it but there's um they still send data to other services and then those terms apply and if you go to them like it's it's not clear that they're not connecting data and I actually wanted to uh enable privacy mode in cursor and it's very difficult because it's not called privacy mode or in all their help is is you you call it privacy mode but then they design it they named it as trust and you have to edit the Json uh config configuration file so it's not easy to do it um I use Jud GPT honestly and this is the memory that it automatically creates about you so I spent a week in Japan I was trying to learn some phrases and like now it's it thinks I'm spending a week in Japan um it knows what I drink it knows I'm Romanian and I'm trying to improve my English and whatever like it just adds this kind of information about you and you might say Okay so maybe I'm not spending the week in Japan right now but this software will improve over time and it will know what information to make obsolete but it still has that information about me so yeah the more you know the the worse it is so okay where do we go from here we have this um privacy a paradox because you need a lot of data to get good results so with AI it's always like garbage in garbage out but it's also when you're a user the more context you provide the more information you provide about you and your actual problem the better results you'll get but that's against being private so how do you solve this okay we have some options so um there's lots of technologies that if you look at them like you have Federated learning which is spreading the AI training over multiple machines and you can also add some cryptography on top of that you have homomorphic encryption so that means that you encrypt information and you can work with it without knowing what's there so you can share that data with other Serv but they don't know what they're working with so it's still private for you you have you know differential privacy secure NPC like all these all these Technologies which are good but they're pretty difficult to set up they require more resources and we already know that AI is using a lot of resources so it's going to be difficult to put this into practice just because like a lot of information actually makes the service better you can can also um anonymize the data so like one way to do it is to create multiple accounts like multiple identities and I I used to do that when I was uh part of a hacking group and I kind of killed that identity now but I had to do it because I had to connect with the other hackers I had you know my own handle but you do the most to keep that private and keep that separate from everything else there are other techniques like you know data partitioning so you split your data around you um obscure personal likees or you tokenize this is like removing some names and adding a a token instead of that so instead of having a a clear name you just add a unique ID to it um but it yeah it's it's a lot what I actually do is that I run a lot of AI models locally um so I need a pretty strong machine to do that but I'm then I'm able to really talk to it uh freely um cuz otherwise I feel watched when I use cloud when I use open AI when I use anything else but it does require you know resources time and I see my uh battery in my laptop just going down when I when I run an inference and you need like at least 24 gab of ram so there's a lot of uh you have some spefic specific Hardware requests so in a way privacy kind of becomes a burden for me because I I want to have privacy but it creates all these difficulties for me and if I um the biggest problem that I see and I think is going to be very difficult is that some people who use the open Services the want that collect data for for you from you are going to those people are going to get better results those people are going to be more productive and if you're private you kind of have to hope that the people who do that are going to have a bad event which is not something good to wish to another person just to be right because you're you're doing all this work it's a burden for you but and they don't do it so they're going to get burned at a later date in the future but I also think that um privacy should be Freedom you have to think about privacy as freedom because if you think about it surveillance influences Behavior so how much of your authentic self are you willing to trade for convenience since surveillance kind of reshapes behavior it influences us to conform or withhold information and not show our authentic selves with us and with our own technology technology that we should own so I I'm going to leave you with these ideas because it's it's not easy to find the right balance between the two but I think both both are very important and both are antithetic so like I'm I'm glad you're here uh this really shows that you care about technology you care about privacy you care about Ai and I I salute you for that and I want to leave you with with one really important thing if we can't protect ourselves we can't truly be ourselves thank you [Applause] I'm not sure do we have time for questions or hey thanks great talk question about like being able to do private like inference ourselves what are you seeing emerging in terms of Technology that's you know bringing the cost of essentially you know access to tokens per second down for the average individual um I know that there are some solutions that are out there but you know typically they're ranging 50 60 70,000 or more um so are you seeing anything coming out that actually drives that cost down for retail I I don't think we've explored completely exactly what we can do so there's a lot of innovation in AI in terms of uh learning how to do better models and how to use models better but you know rags and all kinds of new approaches to add stuff to AI but also um model architectures or distilling models which is a way of technique of you know extracting exactly what you want from a larger model and using that locally um I think Macs are a pretty good thing CU they give you a lot of shared memory and you can use them um but I want what I want to see more is that in in crypto specifically you have this good coordination mechanism that crypto is really good at so you could create uh networks that do the inference for you you interact with them and they provide you the results but you also have to add some kind of Technologies on top of that so it's it remains private um and you also have to trust the technology um so I don't think we're there yet I think this part is just getting started and we'll see more and more interesting things happening in in the near future like very very soon one followup um do you think then that you know perhaps Innovation or exploring more things around like small language models that are more focused on specific topics that you know are more cost effective in terms of raw compute maybe like a viable path to actually enabling more private inference yeah so like distilled models small language models which have uh your own database somewhere accessible uh they make inference much cheaper and very useful to you uh and also very private um so like I think those work however in practice right now um to be honest like for me chat GPT is one of the best ones but I try to use it as little as possible so I also run llama on on my laptop and it's um we're going to see more different architectures in the future cuz uh you know back propagation and and Transformers have got us up to here but we're going to see more interesting things um pretty soon I think great thank you thank you Daniel again let's give him a war Applause people