Building Consumer Private AI Apps with MPC & Tee | Tom Terado, Nillion | ETHTaipei 2025

Hello everyone. Uh thank you very much for being here on the last talk. Uh I know it's been a rainy rainy and big day for for you guys. Um so my name is Tom. Uh I work at a company called Nillian.

Um and to be honest, whoops. I will not be talking about Nian as much today. Um I want to talk around consumer AI AI products and privacy. Um so just as an introduction uh I do developer relations. So I came from a marketing background um in social media and then moved into software engineering and then here I am.

Uh previously I uh so what Nian does in layman's term is orchestration for cryptography primitives. So pets I mean pets MPC and TEES and previous to that uh I worked at a company called wallet connect which is a very big layer in connecting DAPs and wallets. Um so a lot of uh apps ma major apps that we were able to interact with. So that was a lot of fun. Um and in terms of my personal journey just to share and talk about why we're kind of going down the privacy route and it seems like a lot of people here in Taiwan and traveling around the world really care about this.

Um I think I initially wanted to start with decentralization financial independence and you know transactions that are you know on the open ledger and things and then I moved more into kind of like wallet connect and the kind of user journey of oh let's do account abstraction and do all these things which I think are awesome and I think there are great things to actually get from that. Um but I think kind of where the world is heading is a lot more concern for what privacy is um which I'll touch on shortly after. Um so today as an agenda once again I won't be talking so much around a company but more so asking a question and trying to understand why we need to build with privacy um as opposed to just you know going with current conditions. So I'll cover the current state of privacy the pet primitives so privacy enhancing technologies MPC uh TE FHE ZK um I believe in the last talk uh TE was quite covered so I don't think I'll cover too much on that. Uh, one sec.

And yeah, and then so use cases of kind of in the real world and in our web 3 world and then more so around kind of the trend that we're seeing outside of crypto, but also can combine what we have here today. So, uh, I hate to be the guy to want to ask and get involved the audience. Uh, I know this two-part question. Uh, how many of you guys actually care about privacy? Please raise a hand.

Okay, that's 80 90%. And then the last question I really want to filter out on is how many of your friends, family that are not in the space care about privacy and cryptographic primitives? One, one, two. Okay. I think that's the challenge.

Um, and I think we're all in this room because we understand ZK stuff, TES, and financial transactions and things like that. So like I think you know on the like the curve meme we might be on more on the right curve thing kind of situation. Maybe we're mid-curving with our knowledge. Um but the majority of people here and outside actually I don't know I don't think they really care about privacy as such and I think that's the barrier that we as founders developers people in uh marketing and community have to care about. So left side is that we kind of a lot of us here 80 90% of us or 100% hopefully care about privacy and you know retaining who we are and our identity and things like that.

But the other side is and I was talking with my best friend who works in uh travel travel and tourism. He thinks uh privacy is a private Instagram account. So, it's a very that's kind of I'm not saying that's just the standard norm, but that's kind of the mindset what people outside of this industry perceive sometimes. Um, and I think it was a good kind of step back and I asked, "Hey, do you care about signin with Google? Do you care about signin with this?"

Or, for example, do you care about uploading your family photo onto chat gpt 40 image generate gibli style? He doesn't care. So, that's kind of the spectrum we're on. uh left side is us maybe and then no I think we're on the right side of the curve and then left side of the curve is just the general population and how they think about privacy. Uh next so uh just as a kind of recap of kind of what's happening in privacy.

So the left hand side uh is around 23 and me. uh they're a DNA company um and more so they've been in the media and the news so very much so that you know they're maybe not doing well in a business and financial operation kind of thing and I believe the court was like hey you can kind of sell your data if you want if you want to and try and figure out how to run your business from there which for us in the room is pretty bad right like it's it's not really something that we stand for and something that we'd want to see happening of this being sold to big farmer and trying to do something and harvest that data. Um, and then on the right hand side we have uh I just Googled uh data breach and we have some data breaches in Australia or in the US and different places everything there's always a data breach here and there whether it's enterprise level in terms of their security and encryption and things like that. Um, and yeah, personally for me, uh, not to disclose, but yeah, I've had my healthcare and my teleico and one more my healthcare provider being breached with data and what they do is they send me an email and say, "Hi, we're sorry. Full stop."

And so ultimately that kind of sucks. Um, they don't really care. Maybe they get a slap on the wrist, but that's kind of the state of privacy in the outside world of web 3. And uh so some of you might be familiar with what this was said. Um so yesterday I attended DAK by uh Vitalik and the rest of the organization that organized that.

Um and I hope I'm not paraphrasing or kind of changing the ways that he was trying to say things. But basically the three factors there is kind of what's happening in our world right now. Good for the for the best for the better and also for the worse. So mass communication is say how we're so hypervigilant and global citizens and we're all connected. But then AI, I love using AI, you know, but also it's kind of scary and I have existential crisis every single day.

And then we also have biotechnology. So things like COVID really showed us that with hyper hyper pandemics um and the kind of situation that we live in. And then the kind of the underlined quote here is centralized security is a source of risk. And I believe he had a really good diagram to kind of all uh assist this kind of quote and statement. Um and that is because I'm not saying privacy is going to help you know mitigate the risks that come with those three factors below.

But it there are many elements of having secure databases and authentication and different ways to approach our identity that really matter with things like AI previously mentioned around tees, LLMs and things like that around mass communication. My normie friends I shouldn't say normie my normal friend is uh you know concerned around social you know how does you know how does his messages get leaked out and things like that. So that's kind of like just the trailing thoughts and I think a theme that people have been seeing with the theme of privacy and especially at this conference and I feel like I I really like this conference and u people I met here because it is very cryptographic and privacy friendly and what we have here is kind of an opportunity to kind of talk about that um which I really appreciate. Um, so this is something I will come back to and this is pretty much the only takeaway I want from my presentation today of I think about this iPhone or Android toggle and a lot of the things that we do is pretty much on the left side of public and we just have everything public facing you know data breaches and things like that and then you know all of our socials are kind of on the public and things like that but I really want us uh you know for the people in the audience I assume quite a lot of developers um is that I really want us to uh kind of toggling the private mode and the thing is it's actually not that hard right now. Uh I think if you go back 5 years around ZK and all these other things that these privacy preserving technologies provide is that it was a kind of bit harder even go back 20 years ago it was much harder.

Um so I really want this to be the only takeaway if you do remember what I talked about. Uh, and to kind of go on the the theme of what privacy enhancing technology, it's a great acronym. I was going to put some other dogs, but this is the first thing that came up of pets anime. Um, and then there is no relation towards any of these characters and the cryptographic primitive, but I just did it for convenience. So hopefully that makes things a bit more memorable.

Uh, so this I I linked it below. Uh it was a mirror article I wrote uh a few months ago around pets. Um this is currently the current ecosystem. You know, I'm sure a lot has changed and uh I definitely met a lot of people in the conference that should be added onto here, but this was a quick ecosystem map and it's kind of what I like to do whenever I enter a new not industry but different niche especially in the space whether it's an AI or something like that. And we have the MPC, FHE, TE, ZK, and we have a subsection for MPC wallets.

And a lot of us here um on this map and many more extended uh are really doing awesome things and which I will tie into AI. But uh yeah, everyone here on that map and and furthermore are doing amazing things. Um and this would have been a much emptier map 5 years ago. So a lot of progress is starting to be made. And uh for those that know of him, David Schom uh is a godfather, an OG um in the cryptography face.

uh and can you can anyone or uh maybe I won't subject you to the audience but this was written 40 years ago um in one of his papers and the highlighted part of tapping into data gathered in everyday consumer transactions can secretly conduct mass surveillance interfering individual lifestyle activities and associations. This was written 40 years ago. I think that's pretty wild and you kind of look into that in the current state of things. Um if we talk about and relate it to the things that uh was in the centralized risk um slide but once again to iterate it was done in 40 years ago which really represents kind of the challenges that he foresaw within the community but have kind of fruition which is a bit sad but I hope it's not too late. I I'm not sure of the context and maybe you guys had a lot more with TE in the previous slide.

So I won't touch too much on these primitives but NPC if you think about it um it's just written on the screen but the easiest way to understand the millionaire is with a millionaire problem. So uh we want to see who is uh richer uh Alice say hypothetically has a million Bob has $2 million. How do we reveal it and say one is richer than the other without revealing the amounts even though I just did? And this is pretty much the easiest way to kind of uh visualize it. It's spread amongst nodes and then put together uh te uh I think the previous person did an excellent job on it so I will not get too much into it but easier to remember face ID and other things allow it to do programming and run it in a secure I guess garden is kind of like the way I like to think about it.

uh zero knowledge. I think a lot of you probably have a lot of context. These are the conditions that you would probably need to meet zero knowledge technologies and to prove without revealing any additional information. And um computation is performed on encrypted data without decryting it first. So currently the way that we do it is that when you have to transport data there has to be a decryption but this allows for different things.

Um, FHE kind of took me a while to get my head around. So, there are various parts of it. There's like partial homophobic uh encryption and different things. Um, but the article that I linked uh has a whole primer if you would like to learn a bit more about it. And I'm I hopefully will not continue to keep using so many anime things to kind of visualize this, but uh this is kind of how I like to envision and understand the pet stage as three tier.

Some people think we're on the the stage one. Some people think we're on stage three. I think we're kind of stage 2.5. A lot of the crypto primitives in the history of cryptography were uh created in the '90s and 2000s in research and discovery and even earlier.

I would say that was probably like the egg phase. Um but right now I feel like we're in the the present of production ready pets for usage. um and composability. And I think this is my last anime thing, so apologies. Uh so the fusion of pets uh we just saw them all in isolation of different tiers and this and that.

But what we can actually do is uh think about pets as uh kind of like think about like having a dog and a cat and another pet and another pet all in one garden in your house and they're able to help you, right? Um there's different composibility that that we need to think about um with these different primitives. Um, so TE, ZKP, uh, ZK proof, uh, TE MPC. Um, the one I like to spend more time thinking about is TE and MPC. So they are composable and that's kind of the stage we're at right now.

So hopefully this has given you a good primer um into crypto cryptography and the primitives of what what's involved um so that I can tie it into the conversation towards the end around AI and consumer applications and uh kind of in the use cases I'm sure there are many more that I'd like to put together um but in the things that came to my mind um having come from a wallet and UX background in the previous uh job is that a lot of NPC wallets are doing a Great job. Um just the three um three kind of like shares and shards and then reconstructing it to assign transactions in the ZK side. There's rail gun email and even country based um identification to prove who you are and also in NPC databases. So um this is probably a kind of a more uh cultural aspect within our space that I really want to nail in. I uh maybe it's more on a hot hot take or something, but uh I'm kind of sick of going around the crypto conferences and reading the slogans.

We are an L1 powered by XXX powered by this and we can do 2 million 3 million TPS. I think that's the wrong language. Um it just it doesn't make sense to if we think about the people that there were two gentlemen at the front who were like, "Okay, my friends care about privacy." those people will not understand these pet primitives that we just talked about. Um, which is why I want to go back towards this of I really want to kind of engineer part of our language and maybe the way that we build things to be a bit simpler in the terms of public and private um and kind of taking back the values behind it as opposed to selling it as um zk this uh te this.

Um, and I think we're getting there. Um, and that's what I really want to hone into it is that I think the better that we can do the developer experience and of course with the marketing and the language of how we talk about it is that we can get into the more freedom of what privacy stands for. Um, so uh I hope to tie this now in um I hope it's been good enough primer. Uh so I this is me thinking around uh the stack of what it is to build a consumerf facing business at the moment. Um think about a I don't know an apps like Tinder or something.

If you think about it from a technical perspective this is all you kind of need. I'm I'm sure I'm missing a lot more but this is hypothetically if I'm at ETH global hackathon and I need to think of how do I build something. This is how I think as a developer. Um, and right now on the right hand side is uh some graphs. There is a very popular database called Superbase which is heavily used by many developers.

And on the right uh I I didn't put the label but it's a company called Replet. And the reason why their hockey growth in the last quarter, say from November, December 24 is the AI hype, LLMs, and the readability that it has provided to really hockey chart and get people to use things. And that's kind of what I I want to kind of tie into. So, a lot of these people, a lot of the people building on AI stuff, there's a uh I watched uh if you guys follow a a channel called Fire Ship, he calls it the vibe coding virus. Um, and a lot of us has started to vibe code.

And this is to be honest why the LLM side and also the vibe coding element has really pushed forward our ability to be able to make consumer applications. And I I will retract the word consumer. I think I'd like to just call it applications. I think in this industry with trying to be like we're going to build the next consumer application that does this just call them apps. Um but this is basically what the AI world has kind of allowed us to do um in the past 6 months and it's it's really impressive.

I encourage you guys to play around with it and see what you can build um even if you are from a nontechnical background. Um, so this is kind of the stack that I think of and this is kind of what is being reused and people are being able to really make cool things and cool product jigs and what I want to think is this is the private mode of uh right now uh we can really use the APIs the primary APIs that people are using in the AI world are these clawed and uh open AI and all these other things and what we can do um and I won't touch too much on it is the the model side of things. Um, we can run private models in TEES and really kind of keep the prompts and as well as like the security of it quite quite tight. Um, we can do databases. Um, so uh the database that we run is with MPC and it's just a MongoDB.

Just think of it as like private MongoDB. So like these these are the kind of things that are happening within the stack that is being used every day. authentication with ZK is you know possible. Um analytics I haven't seen as much but I have seen private analytics. So this is the stack that's been unbundled.

Um and I want to bring about just a thought exercise about this um of this phenomenon that we had the other day. So if you think about if I wanted to build this jib Jibli uh style image anime generator, you needed primarily three things. You needed an authentication to maybe Google or something. You would need the OpenAI 40 image generation API uh which isn't available. So I think people use like Gemini or something else.

And then you needed a database to store the files and retrieve them if you wanted to. Pretty simple stuff as part of the stack, right? So we go back to here. Superbase would be the database. Uh authentication would probably be sign in with Google.

And then the UI would be spot out by v 0ero lovable or bolt. Um and the API is / chat completion with an uh open AAI or Gemini or whatever. But then I want us to think of this toggle right private or private AI private databases same product. So we didn't really compromise. We don't have to compromise on what it is.

I will give a caveat. Speed is a bit of an issue right now if I want to put on this toggle but it is there. It is available. And I think that's the thing. So with private orth you can use zk private AI you can use um uh t uh you can run the model in te or offline and things like that.

And then we have private databases which are possible with MPC. And yeah, so uh I think I have a little bit more but once again I just really want to hone in that I think building consumer applications right now and I I really want to encourage like people in the space and and the community in general is that uh let let's not just build the next unis swap clone with ZK and this and that. I I think it's better that we start thinking of okay what is it that our friends want to use and how can we just abstract it in public and private mode with the primitives that we have because then people will care because think about it like sign in with Facebook I don't think I cared about it the very first time it came out right so we've gone into a very different stage of technology right now where we have to consider the need for the user if they want public and private and I really want to encourage that. Uh so this is just a a very simple diagram. Uh we also do LLM inside at TE.

Um but I I think uh this was previously mentioned so I won't spend too much time on it. But if we tie it into the whole Jibli uh example we had, the normal flow of the Jibli would be you would say, I want a cat uh jumping uh in my house uh in Gibli style LLM server and then response. On the private side, it's a bit different and all of that is within a secure enclave. Um, so in summary, uh, I hope, uh, you guys learned something and were able to take away something around the crypto primitives that we have readily available. Um, they're pets, they're very cute, some weird acronyms.

Um, but privacy is here to stay and moving in a positive direction. I think it was already here when Monero came out and things like that, but in terms of consumer and developer and marketing and community, I really think it's a place where you can just hit the button and start doing it. But let's just not say we're a ZK protocol that does XYZ. Let's try and frame it in a better better way. Um the future of privacy lies not in a single solution of just using one of the primitives but maybe thinking about combination.

And I still think that's a bit of an experimentation phase. Um and really I want us to think about the uh toggle. I really hope that's the main takeaway of that we can really build privacy consumer apps within with the same features and same ability of web 2. So I think that's it. Uh that's just my Twitter if you want to stay connected.

Um but thank you very much and I hope you have a good evening. Cheers. [applause] Thanks Tom and a huge shout out to all our speakers. [music]