NEAR AI Office Hours #29 w Nillion and Gateway

Captions as YouTube publishes them, usually machine-generated, so product and people names can be misheard.

1. 0:01

   in inen hi everyone uh the next episode of office hours is here today we have Nuna from Gateway laabs and Miguel from nean um let's begin with the quick introductions and uh some background on the project Nuna you want to start yeah absolutely happy to be here um nice to be meeting you guys as well and thanks to yeah for inviting me so I'm Nuno one of the co-founders and I leave everything engineering related here at Gateway what Gateway is doing I'll save like the

2. 0:40

   sweet spot for like more for the end but what we are trying to do is to bring privacy and unified state to um blockchains in order to actually bring more use cases more composability to yeah normal to the blockchain spaces is everyone yeah my name is Miguel I'm co-founder of of nilon and essentially what we're building is um a new type of decentralized infrastructure that can handle private information uh so that you can put that sitting alongside a blockchain and you can run and store all

3. 1:19

   the private information all the private data that you have on this infrastructure it is decentralized so it has the same uh you know look and feel as as you would expect from a decentralized network but it it's specializes on handling these typee of uh computations and storage operations and then it seamlessly interconnects as well with blockchain technology okay so so when it comes to

4. 1:42

   privacy I think there are several active um areas of research or I guess some some of them are much closer to production than others uh that people pursue so I would be interested to hear how how you guys think about it like what tradeoff do you see what uh what approaches you pursue I guess let's go in the opposite order Miguel you want to start yeah so essentially we think that you know privacy should be tackled from the standpoint of privacy enhancing

5. 2:13

   Technologies which is the you know it has been you know they been along for 40 years or so but now is the first time that the're really production ready and that's kind of the core Baseline of our infrastructure but we also at the same time believe that there's no single Silver Bullet there is no you know P that will elbow out the other ones and so that's why our approach is rather to build a sort of orchestration layer that

6. 2:39

   is able to interconnect the different PS so that we can offer our developers uh the best uh possible point in the trade-off space regarding the requirements you know around security and performance yeah I would I would say something similar in the sense that like privacy has been a concept since like since forever really and now that we are seeing like uh computers and Technology evolving to have several different ways of tackling the issue I think it's not a

7. 3:12

   matter of finding one Silver Bullet but rather like tackle the actual problem how can we keep stuff private and how we can use that or how we can put that in the public eye so to speak while preserving the Privacy so we are doing something quite similar in the sense that we want to use technologies that actually scale for well millions of users and for various use cases but we are not like saying that this is the way and all the others like are terrible

8. 3:45

   Technologies on the contrary we think that uh M there there is space in the industry to attack privacy from multiple angles um and I think like there will be space for everybody to win a little bit and on that front I would add that I don't see like you know Gateway or or ourselves like competitors I think um privacy is competing with other narratives in the in the space like um so for example there are other narratives that it's you know they're all about making it faster making it

9. 4:18

   more scalable but is the existing technology so privacy opens up a new dimension in terms of use cases you can tackle in terms of you know growth for the industry it's more like a disruptive approach compared to the more you know uh incremental approach that other narratives offer and in that sense I completely agree with you Nuno that uh we're kind of you know on the same page uh we're trying to push forward privacy

10. 4:41

    because we I think all believe that it's an important ingredient that is missing in web3 so when it comes to privacy I think uh the way I I think a lot about privacy From perspective of uh using it with the emergent large language models right so uh they require for some for some other use cases I think privacy is easier like anything that just requires exchange of messages or files you know I encrypt it on my end you decrypt it on your end and for quite a few use cases you can you

11. 5:16

    can achieve you know uh complete privacy with relatively little effort you know like messaging uh document sharing but when it comes to using things like large language models it's uh it's more problematic because given majority of people cannot host a large language model they have to effectively send the prompt to the machine that host the large language model right so you either send it to um

12. 5:47

    like not even touching aspects of whether we should trust sending our data to entropic and open AI more relevant to our use cases if we have decentralized networks of uh uh of like inflence providers we definitely don't want to send our proms to those guys right because that's completely random people living on the internet and when I think about it I primarily uh uh think From perspective of using tees because

13. 6:21

    tees uh they they don't require any uh any advanced techn like you there's no cryptography involved that is bleeding edge uh they work relatively fast but they have an obvious drawbacks such as a you still have to trust a big player and B in the past like you know previous generations of he technology have been in different ways shown to be broken right so with every new technology they they may might not be shown to be broken yet but the TR

14. 7:00

    um and and maybe you know maybe they are fundamentally unbreakable but the trust has somewhat eroded right and there are many people who would not be comfortable with the KE is so so so I would like to hear your thoughts on where ke is fit in the picture where we should not be using ke is uh if if any of you guys use fhe how far F has come uh and if you guys use MPC uh you know what what MPC works for what it doesn't work for so you know

15. 7:30

    PR pretty pretty open-ended question and pretty generic but I would be interested to hear your opinions here who start I I I I I can start um I think like these have their own merits to what they want to tackle but as Miguel mentioned like it's not a silver bullet and I think like as as you mentioned Alex as well like it comes with a lot of drawbacks especially because like if you have control of the te in the physical environment you can

16. 8:02

    easily like attack it uh via S CH attacks so it's not the standard that you want to have when dealing with such sensitive data and especially for AI and God knows what you end up putting on the models right so for that like I think tees might be complementary for other pieces of technology that might evolve you mentioned FHA uh you mentioned MPC uh in particular MP is something that we are strongly building here at Gateway in the sense that we believe that MPC with

17. 8:36

    more recent Innovations more recent research has shown to be scalable for the type of work go that we want to Target especially when it comes to AI especially when it comes to model training especially when it comes to private inferences on sensitive data so tees could be simply like a piece that enables like these multiparty computation servers to to run H while not being like the core piece of the

18. 9:08

    entire technology they can still uh play a role uh while the big piece is instead like the MPC technology which will allow you to put the data in a privacy preserving way uh via linear secret sharing or Garbo circuits or whatever scheme that you want to use there um and then you would you perform the computation in a completely privacy private way completely preserving the users's Privacy applying that to AI you

19. 9:39

    have not now because it's not yet there but in the future you might have like large language models running in a completely decentralized and sharded way so to speak with data that is also sharded across multiple computers and to like close the circle you have tees that can just powering up everything to add an extra layer of privacy to ensure that everything and to end is completely or almost unbreakable yeah I completely agree uh

20. 10:10

    wct is another tool um it's it's the right tool depending on on the degree of you know sensitivity of the data that you that you want to be using in your application for some some applications it's fine for some others it's not and we see the you know the distinction between uh tees and other software-based privacy enhancing Technologies like FH and and NPC not as a binary decision um I think it's it goes along the lines that you were mentioning Nuno I think um

21. 10:40

    both can be combined uh even in the same solution in the sense that you can for example run some layers of the llm inside of a te some other layers inside of NPC or FH even you can make distinctions and run linear layers inside of MPC or FH because it's more performant and then the nonlinearities like activation functions or normalization layers you can run them on the te there's there's various

22. 11:02

    combinations and approaches to to these that being said um there is of course a performance gap between NPC and FH that needs to be bridged and so we've done a lot of work on that front uh what we did is to take um technology to kind of borrow it from signal processing um in in particular the discrete wavelet transform which is used to compress signals for example is used in the jpeg

23. 11:27

    2000 standard uh for image comp impression and we ported it into cryptography both to for FH for tfh in particular and for MPC and we saw great Improvement so significant Improvement in the evaluation of the nonlinearities because you can represent them as a lookup table and then you can compress that using these techniques so and when you look backwards you see that in the last five years of development we've

24. 11:51

    gone all the way from evaluating your models with thousand parameters to now models of 1.6 billion parameters so there's been a lot of improvement and will continue so I believe and so the the bridge uh the Gap that exists between tees and and other software based pets I think we'll get closer and closer um and also we have an experience like we've tested both FH and MPC and we kind of concur with you Nuno that uh NPC seems to be more performant uh particularly when you have to when

25. 12:21

    you're working with a few parties um um than FH but it does have on the other hand the disadvantage that you have to pre-process information ahead in time something that you don't have with FH so again it really really really depends on your use case and application yeah and just to add like a small note on that uh I completely agree with you like it's a matter of you have all of these tools on your tool belt and it's a matter of like deciding which ones you want to use for particular use

26. 12:51

    cases we are big Believers in like mixing and matching all of these different Primitives so for example as as you mentioned as well like if you want to use MPC for a particular piece of the puzzle and then tees or both or and then like the outputs you want to take them and you want to share those outputs securely with another party via I don't know something like proxy re encryption like the fact that we are walking towards uh a future where we have all of these Primitives that before

27. 13:24

    were like super uh kept by academic so to speak now that we are walking in a Direction Where We have all of these at the distance or at the quick of a button or at the distance of an SDK it really enables like all these different companies of these different new use cases to emerge and it's just a matter of mixing and matching and building whatever you want really and I think that's a key ingredient that you mentioned right there which is the fact that this no longer is uh you know

28. 13:55

    pertains to the domain of pure academic research um there's a lot of commercial institutions for example this work that I was mentioning uh before we did it in collaboration with meta I think we've reached a point where um we're we're kind of um the the the protagonist of of this era of pets of this stage is the engineer is the kind of Applied cryptographer that is able to take that

29. 14:19

    theoretical work and actually U you know implement it at scale uh this wasn't the case before like five years ago we were still in the stage of you know developing the the theory and kind of testing and and coming up with new protocols right now is more about taking those protocols and implementing them at scale yeah what is h so the thing you mentioned with Matt it's called Neil AI

30. 14:42

    right what is a what's the exact semantics of the so our platform is kind of divided in three main what we call blind modules um so one is the general purpose compute um module under currently under NPC that is called nil VM then you have nil DB which is the general purpose um module for you know the storage at scale of information in a in a privacy preserving way and then the Neil AI is the one that encompasses all these innovations that I've been talking about is the one that where you have the

31. 15:16

    tees where you have the the MPC and and and all that um um you know all that kind of soup of technologies that that can be combined in different ways and forms and then on top of those is where you build what we call sdks for example you can build private racks by combining the database product that you use to store the embeddings the vector edings and then NE AI which you use to compute the distances and to compute the shortest one and so on so that's that's you know in the combining those three products is

32. 15:47

    when you build the sdks that provide them inference provide private rack applications and stuff like that so um we have this completely unrelated to private um product called the not product but the um component called Fain signatures where you have a multiparty computation that signs transactions on other blockchains uh and we um we we plan to to introduce their tees so I think Tes and NPC work very well together in the sense that if all the comput if all the

33. 16:28

    computations every party in the MPC happens inside the te then if you if in fact people execute side Channel attacks uh you get you still get the whole security of NPC uh and in the absence of such Chanel attacks someone who corrupted post factum a large percentage of the of the parties will not be able to do anything at this point because once the computation has completed the side Channel attack I guess is meaningless right so I think it's

34. 17:02

    interesting I think one thing that would be interesting to to to say explicitly is what are the security considerations of each of the approaches right so my understanding is that with t i mean with t we already sort of mentioned them there are side channels attacks presumably they have been shown to work pretty well with this GX I I don't know if side channels attacks have been actually shown to be effective against DDX they have some

35. 17:27

    mitigations uh then fhe if if my understanding of f is correct then as long as cryptography is sound uh then there is no no external event can result in me losing uh you know you know my data being exposed right with MPC uh depending on the configuration if the sufficient sufficient number of pets collude my data will be revealed right yeah correct here it's a bit more Nuance that that I would say um because for example there is two two forms of security mainly in in cryptography one is semi-honest security which means that

36. 18:11

    the parties cannot or or the nodes running the computation cannot access the secrets but they can tamper with the result from the computation and then you have uh malicious security which is when you're even able to you know account for that and you're able able to counteract that type of attack with FH it is more expensive to do the latter because it essentially means that you have to repeat the computation across many

37. 18:34

    servers and then compare the results or you have to to do a zero knowledge proof of execution which is also very expensive to run on top of f so it is true that you don't have the collusion problem but you do have that other problem with that with MPC can be accounted for because you do have malicious Security in MPC I see okay so with fhe if I delegate to someone else to do the computer they I know for sure they will not see my data but without redoing the exactly the same computation or having some

38. 19:06

    other cryptographic proof I have no idea if they actually did the computation correctly exactly they could be running another computation makes sense yeah yeah and then note on on that as well like as it pertains to MPC um the fact that you send off your data in a shared like Shar rather way it means that like the data won't be accessible by by other nodes as long as they like are in the honest majority scheme um but on that note as well like if they are in honest majority scheme if

39. 19:46

    the majority of the nodes go good then you can temperate the results of the data not necessarily with the inputs per se with the malicious uh you can like counteract that so um because it's malicious the nodes are configured in such a way that um they cannot or I wouldn't say cannot but like the system ensures the confidentiality ensures the correct the correctness of the computation um and obviously like the sancity of the inputs so it's more secure than just normal um

40. 20:29

    normal like honest majority schemes but in both cases like the inputs granted that like they are properly sharder than no part in the mle like is able to extract the shards or the mask or anything like that everything is kep right but but it is the case like I think this is where the difference is with fhd and MPC is that if I'm like if I don't trust anyone at all so so I I I I think it's important it's similar to uh let's say uh zcash

41. 21:07

    right so with zcash there are two um two goals that the protocol pursues right one is that when I send money to someone I need to be sure that nobody ever knows that I did it right and the second thing is that I would prefer uh the money not to appear from nowhere right and in my majority of cases the former is stricter is more important than the latter specifically in the Privacy preserving uh cryptocurrencies in the sense that yeah if uh If

42. 21:40

    U like effectively ZK snarks can have two different or like not specifically but technology that provides privacy on top of which you might build a cryptocurrency like Z cache it provides you two guarantees one is that uh data that went in like ZK part that went in is is not visible and the second one the argument of knowledge part is that it is in fact sound and uh you know the statement that

43. 22:06

    test to is correct right and I think the failures in the former is quite unlikely the failure in the latter is you know the were bugs in the past which would allow you to create a snark for something that is not true right but the failures that you get are different right in the in the former case your data leaks and that could be way more disasters for for some entities than losing money that were sent right so so I think here this is similar right so when I'm sending data for

44. 22:40

    inference uh it's one thing it's different levels of risk for me for that data to become visible to someone else right like imagine I'm sending a data that can be used to to Blackmail me eventually right and it's another risk if if the response from the from the model is tempered with right so my understanding is that with fhe the former assuming the cryp cryptographically sound and there is no bugs the former is impossible right no

45. 23:05

    matter how many even if the whole world colludes against me right uh it's impossible to to get my data but it is very you know it's possible to to give me garbage out right to give me a response which is completely incorrect right so against which I can in some way or another try to you know like again send it to multiple people of course again if the whole world colludes all bad in case of NPC if my understanding is correct that then no matter what you do uh if

46. 23:38

    everybody colludes against you if every single node in the multiparty computation is malicious both are the case right both the data will be revealed and the output will be tempered with right so so so from this perspective I find F to be stronger I I I don't know if that's how how other people look at those Technologies like if this is a meaningful I would say for most cases you're right um I would I would use as a counter example uh threshold signatures

47. 24:10

    which you mentioned before which typically run in a dishonest majority setting whereby if you as a customer are one of the nodes then even if all of the other nodes collude in NPC we're talking about they wouldn't be able to reconstruct your private key so there there are settings uh the honest majority say things where that doesn't happen is not the case even in NPC but you're right that in FH um you are

48. 24:35

    protected by the heartness of a cryptographic problem which is learning with errors which is thought to be Quantum um you know resistant and so which means that essentially if you gather and harness all the power in on on Earth you wouldn't be able to to crack that in in MPC you usually get that guarantee or even stronger you get even information theoretic security for the share that you're getting as an individual as a party but you don't get

49. 24:58

    that guarantee when you when you kind of you know look at the shares together if you're able to corrupt multiple multiple parties then you enter this other uh discussion of you know what is the what is the threshold that you're using in the in the scheme yeah and to add on top of that as well like um as the example that you gave Miguel about you running a node and you are part now of like this committee

50. 25:22

    that executes the computation that is a way in like dishonest majority in the malicious setting of like almost guaranteeing that everything is going to happen your way because like you are the client you are the person like executing or asking for the computation to be executed so it's in your best interest to well not C with with the others I would argue um and to also like do a full circle here with what you said Alex like I think for

51. 25:51

    different uh things different Primitives might be used uh I think like especially for example with NPC MPC is interactive in the sense that like you need to be here to like send your data Shard your data and then like execute the the circuit or the computation based on that data that's more aligned with the entire ethos of blockchains as they are things don't really happen automatically like without the click of a button or with a Chown job happening on chain everything

52. 26:24

    happens or all the triggers that might be automated they happen on chain but you need to still trigger on chain actions to happen which is more aligned with the entire ethos of being interactive therefore like implementing an NPC scheme uh to do computations like that FH you could like look at that as a non-interactive um standard where for example you can encrypt Healthcare data keep it there encrypted run computation just keep running computations and like

53. 26:55

    because it's non- interactive um you don't really need to be there 100% of the time let's put it like that so yeah to to this point it really depends on what you want to do what type of data is where where you are using this technology is it blockchain is this outside is this AI we want to sh this across multiple people we want to put this on a server and call it a day I yeah I mean there's no single silver yeah so

54. 27:25

    I go ahead yeah thank you I would add to add that actually actually those Technologies are combined right uh so in in some setups of FH you threshold dies the private keys so they're they're kind of secret shared amongst many parties and that's a technique typical of NPC and also in other techniques both in NPC and and FH you use zkp to to make sure that the inputs satisfy certain

55. 27:50

    conditions that there's no over overload or you know that the inputs are within certain range so that's I guess what we were saying before nun and I that all these Technologies are are not like binary or trinary they they can be combined really at a deep level of interaction and you get a result that is actually better than any single one of them separately but but for me to fully

56. 28:12

    understand right and this is purely thetical uh but let's say that we have a technology uh that allows you to run so I generate a private key locally right so the key pair I I I have full control of and then we have fhe that is sufficiently performant to be used right so like you know to do the inference and then we have sufficiently performance scheme to compute a zero knowledge proof that the computation is

57. 28:37

    correct right would that have any uh so so so if we have that construction which is performant and usable what advantages if any MPC would have on top over it well you would have to add the you would have to threshold IE the private keys so that there is no single element or entity in your whole equation that has the power to decrypt the information in that server so that what if only I have the private private Key Well if you the the

58. 29:10

    user yeah yeah I have the user yeah you could have that as well but again you have the problem of you know owning that piece of information the client is typically not the best place to store that sensitive information so it rests better in in kind of server infrastructure controlled and and maintained by professionals um but you you can see that in your example you've already kind

59. 29:34

    of mixed the three of them because we've concluded kind of that you know threshold iing would be a good idea um um then you used Z zkp as a proof of execution and F for for for running I would say the problem with with your with that with that scenario is that you know if you mix FH with um zkp on llm inference it's not going to be performed yeah of course that was more of a theoretical experiment right if they were performant yeah but I guess then my

60. 30:07

    my next question is so so so now let's let's let's for a second concentrate on NPC um how how performant is it today can can we use it for llms uh or do we primarily use it for other forms of machine learning and uh what is the uh well let's begin with how performant they are and then I guess my follow-up questions will depend on the answer to that question yeah I think I can answer the last question can we use that for llms and the answer is no um I I mean at

61. 30:40

    least not at at scale at least not with the same speed that you see on chat GPT uh you could in theory build like an NPC circuit that would compute like that llm for you but that would take a an enormous amount of time um point is like how fast it is today uh we've been running some tests and like we've been implementing some of the most recent optimizations that you see in Academia and for uh like normal arithmetic bullan

62. 31:15

    circuitry you can uh get really interesting results compared directly with TF to give you an example we can see like results that are 16 to 25 if my numbers are not mistaken 16 to 25 times faster so at scale for something like for example a blockchain where you are not going to run an llm right now at least you can put in theory uh thousands of TPS uh compared to at least the current iteration of fully homomorphic encryption as for machine learning and

63. 31:57

    AI the way that we see the Technologies evolving is that all of these can become the the plumbing or the pipelines for providing better and faster machine learning uh if you think about like today running an nlm with or run any sort of like model decision Tre like anything that comes to mind with private data the only realistic option that you have are like tees or like uh enclaves

64. 32:28

    running on gpus I believe Nvidia has like some product offerings in that regard which which are heavily controlled if we evolve towards a world where all of these can be properly sharded like the data that you pass and you can have like execution times that guarantee that a model can be converted into an NPC scheme and be properly computed in whatever scheme like honest majority malicious although malicious is

65. 32:59

    a little bit harder but let's say honest majority you can in fact bring like complete decentralization to the entire execution complete privacy to the entire execution because like now everything is properly shed no one has access to the data you can pass that to models and now you can train models you can run inferences and you can enable a lot of use cases for example medical use cases um for example example um collateralized

66. 33:30

    loans with real world assets that you might not want to make public the sky is really the limit yeah we're we're seeing similar numbers I think the state-ofthe-art um for for LMS under NPC for example is that around 1.6 billion uh parameters uh parameter model which is small gpt2 uh it runs around a second per token which is very slow if you compare it to kind of plain text it scales up linearly though with the size which is good but

67. 34:02

    it's is sufficient I mean it's clearly not enough if you compare that to tees where you can run Lama 7B up to 70b uh maybe with a penalty in performance between 1 and 10% that's kind of a different ball game uh so that's why it really depends on on your security requirements if they're more stringent than what you can get from tees then um the I think the solution is to combine uh so combine both approaches as I said before so some layers can run in NPCs

68. 34:32

    some layers can run in in tees and but still you know given those constraints you can also do other things around uh llms for example you can build a private rack uh solution um that you know can then fit into the te execution like like inference part of the of the solution you can do synthetic um you know data generation you can do even on the training side you can do uh um um the combination of weights from from Federated learning under MPC so that no

69. 35:04

    one sees those weights coming from the different um um computers training training the models um and and you can do for example you I think you were referring to that Nuno you can you can actually work on the data sets you can you can normalize them you can you know um kind of aggregate them under NPC as well um in essence it becomes like a tool set that you can where where you can start chunk by chunk you know

70. 35:29

    providing more and more privacy into the equation and of course there's going to be always this push between what can be done and what needs to be done but as I said before this is a a moving this is a fluid limit that is moving very fast in in in in our favor and I would also say very briefly that I think also there's applications coming into the future where you can wait a little bit you can see that already with some agents where

71. 35:52

    you hand over the task and you wait for a you know a few hours for the agent to do research you have more patience because the llm has more autonom um autonomy um so that plays in favor again of privacy proing Solutions even though they're small slow because um the llm is becoming more intelligent so you're able you're willing to give it more time to come up back with a with a meaningful

72. 36:16

    result yeah I'll just add like a small note on that as well like just because we are not at the stage right now where we can realistically run like large language models with a lot of parameters it doesn't really invalidate that we can't do other things on the data as you were just saying Miguel I think that like if you consider that tees can pre-process postprocessed data as well as NPC as well as FHA as well as for example creating data wakes where

73. 36:47

    you can have for example the data or your users's data in crypted sharded uh just the data sets there you can retrieve them like using threshold cryptography as well and then the combination of all of that you can create actual prompts or data sets or whatever you need to insert into like the the models for influences or for training so we are at the stage right now where the technology already exists

74. 37:18

    to facilitate and to bring a certain degree of privacy to AI but we want to be in a position where all of the pieces of this equation Gathering data processing actually running the inferences and then getting the results everything happens in a completely private way whether that's with NPC that which is what we are working towards or any other technology we'll be here to see but we are already at a stage where you can do some really cool things with the tech we

75. 37:49

    have and I'm curious to hear what are the sort of in in a short term like let's say the next half a year two year what are the biggest use cases that they enabled but by MPC that you know you guys expect to see being deployed in production yeah so I can briefly touch on that I think that like if you can bring fast private execution with unified State on chain you can like actually have private data as a default in the way that you interact with chains so you don't need to I don't know use a

76. 38:28

    co-processor mechanism to send your data process whatever you want and then send it back or do encryption on like off chain send on on chain and do some weird things what you can actually do is like make both private as well as public data Cod exist in the same space be treated as one and that enables like a lot of use cases let's say for example we can prevent Mev to begin with we can do like private defi uh that that's just one that comes to mind as well as like

77. 39:01

    confidential transactions uh the sky is really the limit like everything that you can do on chain today using like normal uh solidity whatever code whatever like uh programming language you use if you bring that to a space where everything is by default encrypted or secret shared you can like put more sensitive pieces into the chain and do as you wish with that that's on our end we we're seeing a lot of traction um from developers building data analytic

78. 39:35

    solution where you where you can kind of um send your data over to this uh decentralized database and then you allow for other third parties to perform computations to get insights from from your data probably possibly mixed with with other customers data so this Falls within the data to earn kind of um um you know um business model or so even data to access um and also the interesting thing is that this nicely connects with AI because some of those um computations that people want to run on your data have to do with AI so there

79. 40:08

    are actually AI models that that run with various degrees of complexity we're also seeing people building uh uh marketplaces for private agents and we're also seeing people build um you know defi Solutions like like you were saying Nuno for example encrypted order books uh private auctions you know private limit orders or private private liquidation auctions or private trading policies and also there's people

80. 40:31

    building onchain games whereby you want to keep the private state of the game um secure or secret and so you Outsource that over to um you know privacy Network like like uh the blank computer from nilon um and then the game continues on chain and it you know kind of gets that Randomness and makes use of it but this time you have the guarantee that no one is tempering with that generation of of the the randomness so that you know you

81. 40:56

    cannot gain Advantage by designing the game uh you can protect your users against yourself cool okay I think we we're getting close to the end of the office hours uh as as a closing thoughts I would be curious guys to hear your thoughts about what is the um well well I guess one question is uh will NPCs and FH C up to to run larg engage models if not what will be the future of those uh like are are we forever stuck with tees

82. 41:29

    or will there be another solution uh and generally how do you see the world like the Privacy landscape shaping up longer term like let's say five years from now yeah that's that's a really interesting question I think like if we don't catch up with tees or normal execution then we fail as an industry like just to summarize but in a more serious note like I think uh because of the resources because of the interest that that all of these technologies have been getting over the past few years the

83. 42:02

    natural course of events is that like everything ends up being improved such that running like large language models in a decentralized setting will be the norm in a few years um at least that's what what I want to work towards and I believe like Miguel would be of the same opinion um and as for getting stuck with tea I mean again tees are good for some use cases I don't want tees to be the norm

84. 42:35

    like I don't want just NPC to be the norm or FHA or anything really we want to have like all of these Technologies at the stage where they are um fast and performatic enough such that they can run really anything that we want including large language models AI agents like process the data do everything that we been talking over the past hour um and then like you just Peck and choose based on your criteria whichever you want to go with that's the

85. 43:06

    goal yeah I think we will see deeper Integrations between tees and MPC for example MPC can help out also with key rotations inside of the te to make them more secure um and as I said before we will see those hybrid Solutions whereby you can mix some some of the layers inside of the te some others inside of NPC or FH um so it won't be that much of a switch uh onoff switch between te to MPC MPC and and F I think it's going to be closing the Gap um but I I don't

86. 43:37

    think it will ever be kind of completely bridged uh but we will reach hopefully a point where it really pays off to run the the you know the the largest part of your load inside of MPC and yes then that maybe those nodes those MPC nodes can run also inside of a te for additional protection but te wouldn't be like the main source of security and then privacy in general I think it plays a major role in opening up the space for

87. 44:02

    decentralized AI I don't think decentralized AI can exist without privacy without that you just get the verifiability part of things which is okay I get the guarantee that this model actually run and not a cheaper one but you don't get privacy and and things are going to get really really sensitive when those models start you know interacting with you in a more personal way and you start sharing your you know

88. 44:26

    your tax declaration and your finances and your health records and and you know you you start treating those models as your as your personal therapist so that's a a big red line um for you know decentralized to go forward without privacy and at the same time it's a great opportunity to for for privacy to actually uh grow and allow decentralized AI grow Beyond its current limitations

89. 44:49

    and in general also that applies to the larger web three space awesome okay thank you y Nuno for joining us today uh and yeah see you everybody at the next episode of office hours