Miguel de Vega, Ravital Solomon, Remi Gai & Pascal Paillier I Pragma Denver 2024

[Music] welcome back everybody hope you had a great lunch break and you get to chat chat with a lot of people who are here from different parts of the world we are ready to resume the rest of the day and that means we got about six more talks to go talks interviews panel discussions all the above and let's Dive Right In so for our next panel we're going to bring on four amazing people we're going to be covering privacy and anic Technologies shortterm pet and we're going to talk about how we can actually make cryptography more applied this means we're going to cover what the use cases are what some open problems are how they're getting address what scale what time frame they're getting addressed on and how you can think about using them in real world applications so for this I'd like to bring on our four awesome Founders I'll bring on Miguel from Nilan Robbie T from sunscreen Remy from inco and pascala from Zama please welcome all four of them on stage there's a lot of new ground to cover on this one because this might be the most sort of with an Asis technical panel will do which means we have to just introduce a lot of terminology but um what can to do is I'll have all of you start with a quick intro uh starting with you Miguel just let us know who you are and uh what you do and uh and I guess maybe why that's getting too philosophical just take the first two okay just the first one uh hi everyone my name is Miguel I'm the chief scientific officer at nilon and essentially my background is at the intersection between engineering and and math so I have a PhD in math but also a master's degree in Telecom engineering and over the last um 20 years I've been working on um developing and implementing algorithms and protocols on distributed networks hi I'm Rital I'm a cryptographer by har started working in lattice cryptography while I was an undergrad at Berkley and math have stayed in this space since then and I'm happy to be here I'm Remy I'm the founder of inco uh so previously I one of the found member of a D5 protocol on on uh polka do so we built a par chain Al also a software engineer Google and Microsoft yeah and I'm Pascal hi I'm the C at Zama and my background is in cryptography been working on FH uh well on H uh for uh what year is this 2024 for 25 years yeah amazing well since we opened up that can of worms already Let's uh let's kind of talk about what H is and what FH is um so maybe I'll kind of go back to you bascal uh tell us in a hopefully easier to understand way what is homomorphic encryption and fully homomorphic encryption yeah so um so initially if it she was a concept that were even that was invented like nearly 50 years ago but it was just that it was like a concept for many years and the concept was if you look at standard encryption what can you do with an encrypted data and the answer was nothing you can just store it or transmit it along to somebody else but it's like sending the to the Future right you're encrypting you get something that is going to be decrypted later in time right but you can do anything with the cyer text it's just like date data it's it's not that really you don't want to get rid of it but you're sending the data to the Future right so the concept of FH at the time was what if we could compute over this encrypted data without having to decrypt it and so at the time but you know it was like the' 70s it was the time that modern cryptography uh suddenly appeared and people had like these new Concepts about publicy encryption and obfuscation and and and plenty of uh you know ideas and it was just one more concept yeah and it took some years before people discovered actually H homomorphic encryption wasn't fully morphic encryption where you can multiply two encrypted values you do an operation of the encrypted values and it it gives you an encryption of the product of these two plant texts without having to decrypt anything and in the irrespective of that there was there were also like crypto systems where you could do addition but there was never a solution that would combine both addition and multiplication and that was the dream to achieve that that is exactly F morphic encryption and it wasn't until uh 200 n that J Tree found the first uh crypto system that actually fulfills enjoys these these two properties simultaneously and there was an ey opener for a lot of people including myself because we've been with you like colleagues cryptographers we were like trying to prove that it was impossible to be fully encryption because it was so hard it was so hard so were figuring out okay it's not normal that it's so hard probably it's impossible let's prove that you cannot be fully mic encryption but it was also very hard to prove that it was impossible now we know why and so gent tre's crypto system when it appeared was stupidly inefficient but it had the huge Merit of answering the question of the existence right and since then I mean it was like 15 years ago we've seen a tremendous evolution in the field and suddenly a lot of people were um you know attracted by this question how to improve it make it efficient and now we're at a time where we're very proud that to see that it's it's finding its way to production this is how far we've come you know over the course of these 50 years it's now a reality and it's it's very exciting to see that that was an amazing overview and I think that kind of really sets the stage of like how far we've kind of come from just research papers to just mathematically knowing it's possible to not having the computer actually do all these things um so kind of just jumping on that same theme I'll kind of go to you which is okay so we know from that paper it's possible but it's just not as efficient and now the game is make it efficient and fast and this is something that you're doing so what is it that you get to optimize and what does that look like long term in terms of how crazy of a scale and efficiency are we thinking we're talking about and just what goes into making those decisions yeah so the challenge really comes that there is a very large gap between what an average engineer who's not a cryptography expert can do versus somebody who is an FHA expert what performance they can get out of it there are all kinds of parameters you can play with you can think through things like security do you actually need 128 bits of security for your particular application especially if it's short load you can think through things like air rate the depth of the computation and you can really pull out a lot of performance from that so from some of our work on FHA compilers we found there can be a 100x THX performance difference between what a nonexpert can get versus what an expert can get out of fhe there is only so much we can do on the algorithmic side I would say the next Frontier really has to do with hardware and we we'll get into that too I want to make sure I cover cover the in on on most of the panel so FHA we kind of cover including H uh you're working on NPC tell us or tell the audience how does that kind of differ from how people think about cryptography and just overall kind of privacy enhancements that go with it yeah so FH we've been talking about FH um it's its own thing when you think about a client server architecture in that the client has the the cryption key and sends the cipher text to the server the server performs some pre-processing or some processing and then returns back the cipher text for decryption but FH can also be threshold IED it can be used so that the private key is uh shared amongst a number of parties and it is in that context that F now becomes in a sense flavor for building uh an NPC protocol an NPC protocol is just a way for a multiple for a set of parties to perform computations on a function while keeping the inputs uh secret and so FHA in that context can be considered as one flavor and there are more flavors than that for example it can be built from carbal circuits where the idea is that you encode the input but uh as a difference to FH you also encode the circuit itself and then what you obtain when you run the protocol is an encoding of the of the outputs which you collectively decode and it can also be built for example from linear secret sharing but instead of applying the secret sharing to the to an encryption key you actually apply it to the secret directly and so the parties perform computations for additions and linear operations it's easy because those operations are local but for nonlinear operations it requires some communication so those three kind of flavors hit different spots on a three-dimensional um performance Cube where we could look at the number of rounds or communic round complexity the amount of bandwidth that you need to send which is communication complexity and then the computational complexity each one of those three flavors is hitting a different spot each one has like good things but also an achillus heel so to speak and also I mean none of these are like these are all complimentary to each other or you you can have multiple ways to communicate and this is about information sharing not about neily proving some information so there is not FH versus NPC here is want to make sure we we don't confuse that audience on all this um and lastly can now what I want to go to is you Remy which is Robbie talin and Pascal just talked about compiler optimizations for FH and you are and making you more practical with inco with and evm FH equivalent uh actual developer environment so tell us more about like what is possible with inco right now and kind of how thinking about it and what sort of made you make this work with the evm yeah um maybe just to start like why we're excited about bringing FG to blockchains um you know I think if we really want to use blockchain as this value later um there needs to be a minimum confidentiality and and this is because you know you don't want your neighbors or your colleagues to know how much you're making you don't want them to know that you have a lot of crypto which could give you a lot no security risk we really think that you know if we were to zoom out the reason why public blockchains are transparent is because you need to be able to verify these transactions right uh but it's really more of a bug or just more like a short-term limitation in my opinion uh because the idea is if you can do um computation over in cryto data in a verifiable manner I think this should be the kind like the Natural Evolution of this blockchain layer because like I said earlier I think not having any confidential is actually itself quite difficult for Mass adoption in a very practical manner um now you know when you apply FH to you know blockchain or to evbm um the major benefit that you get is now you can store encrypted data on chain but not only that you can also compute on top of it uh so if we think about this today uh we can only build use cases that make sense on a transparent database today on ethereum or polygon or any public blets this is really not how Web Two Works what if you want to build like let's say a poker game where you have to hide values what if you want to build um like a top that has some kind of secret key that verifies against uh time for example you need to have something confidential on chain to check against uh and that itself doesn't necessarily mean it's about privacy it's just you know the way certain applications are built you need to enable confidentiality on Chade and so this is what we do at inco with uh you know evm uh that has FH enabled is we think that we can all kind like the next Frontier applications that weren't possible today and so we do it in two ways one is people can build on inco as a modular blockchain and we also plan on extending confidential to existing blockchains which will have hard time supporting F given certain you know Hardware requirements there's also add again NPC component for decryption so this this is where we see kind of like the the future or kind like the next Frontier uh applications awesome well maybe this good time to open up for for all of you can we think of other examples to get a better intuition on like why privacy enhancement confidentality is great on chain and and applications that will be definitely way more possible that are either hacky or not possible right now well I think there's some low hanging Tru uh like for example um you have Dows you can you can do voting you have like on defi you have um um encrypted uh order books or you can have also auctions private auctions or you can have dark pools or you can have private techs um you can have as well like the policies Associated to transactions you can have them also in privacy presing mode beyond that uh you can have also in the in in like a supply chains applications and and so on but there's one we we'll probably Explore More lwh hanging fruit but there's one that I would like to bring detention to which is um it's not a low hanging fruit but I think it's really important and I think it's it's AI so the way I think it's going is that AI is of course going to become smarter and smarter but it's also going to become personalized so it's going to learn everything about you or everything that makes you you and that's really dangerous we we've seen that already with uh open AI with this feature that they have unlocked which is memory now they can remember things about you and of course when you read the legal text there is this pinky promise that kind of says oh if you if you let us if you tell us we will delete it and we will never share it but I think what's at stake here is really sensitive because we're talking about all the emails you've sent and received we're talking about you know the lies you've told to people the private messages You' shared with your partner um like what you really think about your boss or um you know a lot of things that you may know and things even that you may not know like for example you may not know that you're not as smart as you think or that you're not as funny as you think or you may not know that you actually the way you talk you're developing already symptoms for Parkinson or some You Know M mental sickness so all of this information can destroy a person really and so you a pinky a pinky promise is not going to be enough we really need privacy preserving Technologies I think this this situation is putting privacy preserving Technologies to the test um so the society actually needs uh what we are working on here I think all of the projects are working on on private Ai and related topics because this I think is fundamental for our society is a is a really important challenge yeah maybe uh I guess we spend a lot of time thinking about use cases kind like the no use cases um personally I'm very excit about payments uh generally so as we all know stable coin is one of the Cure use case of crypto in the the context of payroll if eror is transparent actually it's quite inconvenient because everyone within the company knows how much every's making this sort of a bad practice non bad practice in web 2 already so know just adding later confidentiality to me is a fairly practical use case um the other thing is you know often time we think of privacy kind of like against the institution in terms of like you know the brush in setting but the other part that a lot of people don't think about it's really against each other too I wouldn't want my friends my family to know how much I'm making for example right this is the ler confidential that I think maybe a lot of people don't spend enough time thinking about which you know including the the payroll use case same with priv voting right today if you do a down voting it's completely transparent and there are repercussions especially if you're from the qut team member or you're somewhat influential and you know in the web two world I know families that you know don't talk to each other because they have different political beliefs right so uh you know just having some later of confidentiality on priv voting is going to be extremely important for for this Bing system to work uh and then lastly I think you know I'm pretty excited about gaming uh you know I grew playing video games you know today crypto games are very limited to the ones that make sense on a transparent database so you know something like Tic Tac Toe where it's okay to reveal kind of like your move but most games um that actually fun requires some hidden mechanics right so if you want a bluff right like a a poker game or kind of like the mafia game that we built um previously it's a fully launch Mafia game built on Ino where you know you could be a doctor Detective or killer that's a hidden role that's gener onchain also store onchain but now you can change the game uh the Dynamics completely right so because you can Bluff can strategize uh in other games like you can cons consal resources uh you can spawn people so these are the more novelty that we think we can bring to the crypto space that just aren't really feasible today or very difficult to to achieve and uh and yeah so yeah now that's that's just great because right now anybody trying the same things ends up going offchain for that compute and then say I did it but that still relies on some assumption that they were honest and why why I have that barrier and and this kind of helps us do this yeah right and this is really like why I think FH is very interesting because it's uh it's deterministic right so I think in the context of blockchain I think it's kind like a very natural suit for Rel this purpose yeah maybe the other thing to dig into here will be U just kind of actually talking about for ref just is compiler optimization so you you set anywhere from 10 to 1,000x Improvement could be out there and possibly more or with the hellow gpus that scale or more um what what does that actually mean in kind of practical terms like is it that the hardware is just not there yet that we're spending minutes on generating proofs or is it that look we can do it a couple minutes just the weekend we saw the ZK proof Revolution last five six years and it's going of G to like now stop second or stop 100 seconds um or milliseconds and and is it is it that phase we're part of or or we have to sort of invent new ways to think about compiling these circuits and and that's kind of the phase so what's a good way to think about those benchmarks or the open problems or just overall like can we put some numbers on understanding what the scale is yeah so that's a great question when protocols approach us about integrating FHA into their application what we really ask them is what is the minimum viable performance you need so there are many cases where we can get the performance today on something that's workable with no gpus on a 64 core machine people want an idea of actual numbers we can run a first price sealed bid auction on about 64 bids with 16bit precision and under a second time on a 64 core machine we can compare numbers together in around 30 milliseconds as well um the main challenge I would say is that while those numbers are okay as of today for certain applications like dark pools if we're willing to throw pretty beefy machines at it it is not acceptable for anything Meb related so some of the applications we've seen that are quite interesting have to do with block space auctions and collaborative searching and there the performance isn't quite what it needs to be even on like the theory or algorithmic side you really need gpus potentially even fpga to get us to the final bit I would add to that that there is um so performance- wise there are plenty of different uh you know subjects that that could be optimized so ultimately for certain use cases where you need low lency of course Hardware acceleration is like the big next thing for f um but there are also like things uh like for instance circuit generation when you compile you actually want to generate a circuit made of atomic homomorphic operations right now um and I mean circuit generation is a hard problem in itself I mean there are plenty of NP NP hard problems when you want to have a prably optimal circuit uh but right now it it's like we have just a strategy for a number of identified pre-identified functions so we have a library of strategies we're just plugging that in and maybe we have a little bit of optimization but I think there's a lot of room for Improvement there with badass compilation techniques that actually generate optimal ciruits there's a lot of slack in there um but yeah um as rabit was saying we we're like missing uh some you know speed up factors for some use cases for some other use cases it's it's already usable you can you can deploy I mean it no no wor is there but if you want to do like encrypted AI uh a lot of gpus first inference I'm not even talking about training uh even though recently we've made achievements in the field we can now train not neur networks but like decision trees we we kind of cracked it so we have a first solution to that but um even I mean the smallest llm would have like already like millions and or even like billions yeah and so uh yeah so it it's a lot of workload and uh we're no nowhere near approaching what you can do in terms of performances in plant text with what we have today so we really need Hardware acceleration to achieve kind of the same what we coin with the term uh you know the um amorphic Supremacy amorphic Supremacy being that moment in the future where with a hardware exploration you can you can of have the same kind of performances that non Hardware accelerate plantex equivalent computations so by default it's going to be F encrypted unless there is some I mean you will have as open AI if you're open AI you will have to justify why it's still being done in the clear it will be suspicious because probably you want to access the data the https moment of working on data pretty good expanding on the idea of of this library of optimize comp compilations of some functions I completely agree with um with that approach and if you think about the plain text World um it just came to my mind um this this book at MIT introduction to algorithms and if you go through it like there's um Decades of work uh collected in that book with the best algorithms for sorting for searching for I don't know graphs for databases like binary trees how do you insert how do you delete how do you find um for greedy algorithms linear programming and whatnot um I think in in general in NPC and/ FHA uh we're lacking something like that I think we are we have the knowledge to start writing some of the chapters of the equivalent book because the the techniques that yet they use there in plain text don't R directly apply to kind of secret shared or or you know privacy um enhanced uh computations unfortunately so we need in a sense to reinvent all of that body of material a lot of that knowledge and although we can write some chapters I think we are still a long way to being able to write a complete thick book like that which covers like the best way to optimize and build all those different uh solutions to those those different projects and until we do we don't do that we won't be able to write efficient compilers as in the plain text uh World we're doing a great work uh great job of course um what with what we have but I think there's plenty of room for optimization on that front so maybe one thing I'm kind of curious about it is does this kind of imply that the only way you can kind of get this to be more practical is just throw more Hardware at it or or is it well one one is it is it a bad assumption like is there I don't like should there's something be wrong with that statement um or is this the only way because we kind of saw the Pentium one to i7 era for the last 20 years but this is just what it means now you have to think about gpus getting more commoditizing faster and better and this is the only way we can do this thing or is it kind of considered a bad thing that we have to already consider as6 or fpgas or more gpus so that means it's impractical we're trying to make it more practical what what's a good way to think about that sort of way to introduce a technology that is very promising but is fundamentally limited on Capital here to to make it more applicable so I guess for us you know yeah through is we do need this Hardware to rescale know 100 th000 x mud higher um but in the meantime there's also another idea which you know you should VIs s leverage ZK for example combin ZK and F and one example I can give is for example if you need to in a game where you need to validate let's say a h a move against a 100 existing moves then you can also do this off chain using ZK and just do the state transition using FH on tra right so there depending on the use case there are also ways you could potentially leverage ZK as a way to just like offload a lot the heavy commutation validation logic um at least this is how at least how we think we yeah we want to be pragma on what are the US cases are feasible today before har resolation and I do think ZK uh could provide a sweet spot for you know compensating for for that as well yeah yeah and I think there's um I mean zik is also one of these other pets right but there I mean in the realm of pets there are plenty of citizens right so we've been many we've been talking about FH at MPC ZK is was obviously well not one one other technique but like a pool of different techniques doing different things uh there are also plenty of others that are like uh you know Anonymous signatures uh even like group signatures or thing like that that are not really deployed and that are like low hanging fruits that web 2 is not like using in any way um and that web 3 is like could could adopt like overnight if there was a use case for that there there are plenty of different solutions I I can think also of um white box crypto for instance this one I mean in pets there are problems that have been cracked and we have like reliable Solutions uh with security proofs and the like and there are still like fantastic dreams like if it she was for a while that haven't been cracked yet but the day somebody cracks somebody cracks that that problem like white box crypto it can be completely disruptive and transformative white box crypto meaning um I can embed I can artcode private information into a an executable so I take an as key I put that in a in a software that just does an as decryption it's kind of obfuscation if you want but it should be probably secure so so I can give you the executable and you will be able to perform like a decryption or encryption but you're not capable of finding the key that is embedded in there and so somehow with that concept you can achieve public encryption with a yes for instance you can you can do so so much more that we can do today so it would be another way to kind of replace secure hardware and we wouldn't need te anymore and so would so that would be another way of like good future to look forward to but uh yeah there are plenty of ideas in the in this space and uh you know but also very very hard problems to tackle to get there early that is a great point it's not it's not just a couple of problems in in in kind of the whole area of PS there is a whole set of other tools that you have and you also can combine them and you should because and on one technique you kind of Y optimiz is not the only way to get there right um kind of speaking of all of that again this kind of applies to all four of you um how do you think about the developer experience of course a lot of this introduces new technology and just kind of new ways to think about it do you want people to kind of be thinking about everything as multiplication addition probably not uh you already chose the r of making evm compatible code just be kind of uh bated in that sense but kind I'm curious about you how do you think about working with NPCs and kind of doing key uh generation and and verification and as you write your compilers what would you like the ideal developer experience to be as well yeah I think it's it's a good topic and and a challenging one as well uh because of what I was saying before that there's um a big difference between programming in plain text so to speak and programming on this type of technology so developers need training developers need like a new way of thinking and simple things like what happens in Brun time is completely different here like for example in a developer experience from from PL text you typically run a program it crashes it gives you feedback as a developer that you have something wrong so you have a bug and uh to the user if if it's happening in you know to the user then they know that they cannot trust that program for whatever reason but what do you do if you have like a division by zero or if you have an overflow um you could add checks that kind of um look for that type of situation that would have an impact on performance but the problem is not just that the problem is that you would be leaking information because if we're if we're running a multiparty computation and let's say I'm injecting some some of the variables um I could just choose my variables to force this type of situation and when it happens I know some information about your variables because they were also involved in this overflow so something as simple as sprun time which is something we take for granted completely changes when you when you face it from from this perspective I think I can build on that too so I would agree with pretty much everything said the experience right now with fhe is really really far from what it should be for developers the ideal experience would really involve a developer just writing a program and python or rust and us being able to effectively transform it into its private equivalent the reality though is that if we take that approach the performance is so abysmally bad it's not usable for really anything and what we have to do is ask the developer for help so we have to ask them well which of these inputs are encrypted what kind of security do you need and in many cases we can transform the programs pretty well with very few changes but in more complex cases say for dark pools or auctions you really have to rethink the entire application design to get something performant yeah maybe just so after that so you know granted we are integrating FH to an evm which makes it a lot easier for Builders to build they can of also use existing tools like manam mask Hard remix on the application layer yeah I think there's another design space or like design principle that people have to be aware of is that for example you cannot use eile statements if you have Inc data use eile statements that would essentially leak information right so there are circuits like CX multiplexer that allows you to you know um do operations BAS based on encrypted bullet right if if it's if this encrypted data is true then you run this logic or this path if it's false it runs this other you know logic or path or separ effect right and this in a way is something a little bit little bit different and a little bit newer also in the crypto space where thinking about designing confidential defi or games you really have to be careful about not leaking information uh in in a way that's not intentional basically yeah yeah uh to explore that more a little bit more in detail so the so as you were saying if you do a conditional branching in a program everybody's aware I mean Anyone Who develops I mean uses an if IFL statement like uh you know maybe 30% of of instructions probably more probably more in a given program is actually a a a you know a test and fully morphic encryption is contaminates everything it touch is right so if that predicate that spe of information depends on an encrypted input then it's going to be encrypted as well so you can uh decide what to do with that control flow at that at that moment in time where you're supposed to jump in the then body or the else body uh you cannot decide so fortunately enough we can you know if if the then body is like I assign three to the variable a and the L's body is I assign 7 to a fortunately you can kind of regularize that that uh portion of code into a cmax when you are uh saying I can do multiplexing like I take the bit that is encrypted so it's 01 but encrypted and I assign to a uh 3 * the bits + 7us uh * 1us the bit and that I can do a morphic way and there there you go kind of transform so in in the compilation at some point we should also consider regularizing PR knife programs that are not that are not aware of of the of that issue so that every you know if else statement is converted automatically into linear you know branchless code but fortunately enough we we kind of know how to do that it's just not in production yet but it's coming that was actually a really good explanation um but and it means like now you have to give the choice that do you want to do this thing as a compiler optimization or tell them that this concept needs to be replaced with an NLS because this is what happens and the side effects yeah exactly but what we want ultimately is the developer experience should be you know um oblivious of what's Happening next in the compilation process you know so the developer should have like no change in the develop in the developing experience itself or at least we're trying to get as close as possible to this Vision where anyone can actually use the technology without having to know anything about it and from from what you know now do you believe that's possible or does that still mean that we will have to still introduce a handful of other kind of new Concepts on kind of new paradigms or rather just new conditionals or Loops or just Primitives that requires them to know specifically how that environment is is we're we're kind of already there I mean outside of the FM um we also have like concrete python which is a python uh front end to our compiler concrete and actually you can write down like python code and this front end is going to to convert that uh directly and and and kind of apply this topological transformation uh is just not as powerful as we would like it to be it's it's getting there but it also like sometimes very hard you know to tackle all these problems at once it's possible or we just need time to get there the but the experience right now is yeah the developer just knows fan and kind of use like encrypt data types and and tags you know which which variables should be encrypted and that's it and and the rest you leave to the compiler to figure out it already works today not a yeah well um as a last question everybody um what can people look forward to with what you're doing with your protocols or your products and uh what's kind of something you want to share with with them that uh might be interesting to the audience on on PTS so we're currently very interested in in the problem that I mentioned before about Ai and it's a it's a moonshot it's a very difficult problem to tackle right now with uh with existing technology as as Pascal was saying we're talking about billion parameter models um but running um private inference of such large language models would be very beneficial because they would not only protect the model weights but most importantly they prompts um and although it's a it's a moonshot there is like lower hanging fruits along that path that takes to to that position and that is for example uh retrieval augmented generation or the fine-tuning of some smaller version of of that model uh those are actually the the ones that will probably start manipulating personal information or private information so it's it's a good way to go and we're trying to tackle these problems from the standpoint of using a a theory is called wavelet theory is from digital single Pro processing and um applying it to to this problem and so far we are still at you know a preliminary stage but we're getting some exciting results and we hope to be able to push uh the limit a little bit forward to make this um you know this technology a little bit more feasible we're most excited about bringing privacy to both CFI and defi we think bringing privacy there will really allow institutions and sophisticated actors to join the space on mass so that they can hide their strategy in terms of some exciting announcements for us that are upcoming we have been working pretty hard on the theory side of FHA so we've managed to make some advancements there I can't say it's what Hardware is necessarily going to get us but it certainly makes a significant difference yeah uh for us uh we are quite excited to see what people can build on top of inco uh we really think this is going to be the next Frontier for level apps NOA experiences and we'll be hosting a few hackathon this year with e globo but also other folks and uh I think that's going to be the best way for developers to re get first experience is to do a hackathon so come talk to us we also have a lot of ideas can give you feedback um and yeah just looking forward to see what people are going to build this year well for zma zma is a technology enabler right so we're not a blockchain company we're not an AI company what we want to provide is you know a um state-of-the-art f tools that anyone can use and then people have to figure out their own use case what they want to do it but people are naturally very creative people are Builders so yeah for us what's coming is like improving the tooling everything should be simpler everything should be smoother more efficient so we're like partnering with uh uh Hardware houses for Hardware exploration um and yeah and and we'll probably make new announcements during the ethereum conference this summer aming stay tuned a lot of cool announcements to look forward to well Miguel Rital Remy and Pascal thank you so much for joining us for this amazing discussions and hopefully the audience has now a better understanding of PS thanks