Decentralized AI w/ Illia Polosukhin, Anand Iyer, Patrick Bush, Niraj Pant

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

1. 0:01

   all right IIA NJ and Patrick should we uh should we kick this off I know we're a couple minutes in we don't want to keep everyone waiting especially you two uh why don't we why don't we kick this off um thank you so much first of all in advance for you folks uh for joining us we really really appreciate it um obviously there's been uh a lot happening in the space um and uh you know I think you folks luminaries Builders and have a lot of thoughts about what's happening here so

2. 0:34

   uh maybe we could do quick rout you know rapid fire introductions and I know Patrick kind of looped you in last minute here but uh when we start with Ilia you know I don't I don't think you need any introduction but I think it'd be great for uh just a level set in terms of how uh you know how you got into the space what you've been thinking about in this in the vein of decentralized AI and we'll kick it off there and then hand it off to nage for some background sounds good yeah I think I I

3. 1:00

   saw a really funny comment from uh my podcast at Lighthouse which was how how's Ilia has background in large like five years so whatever background in language models when they're only two years old and yeah I mean the background is obviously language models existed for like uh 60 years or 50 years large language models uh inside bigger companies have been around probably four seven eight and so yeah was at research

4. 1:31

   working on kind of natural language understanding and the co-author of the paper attention is all you need which introduced Transformer technology powering uh a lot of the large language model Innovation that's happening now and uh for me I think the important piece and part why I was very excited getting into blockchain space as well is I really believe in open source I was uh

5. 2:00

   back in Google one of the major contributors to tensor flow because I saw it as a kind of really important bridge between Google's kind of internal Technology Innovation and open like I'm really getting a lot more developers in open source being able to build um important kind of machine learning applications and uh at the same time I think From perspective of you know if we think and we take an open source mind said uh what we have right now is um

6. 2:34

   pretty much more and more big companies uh kind of driving and like open sourcing pieces that they they think is like strategic for them but less and less in the uh kind of General space because you need to have resources you need to have uh kind of access and you know you need to pay bills and so uh so big companies are the only one who can employ the people doing this and so I

7. 3:06

   see the central the and kind of in general leveraging blockchain space in crypto as a way to contact that create a communities and incentives for individual contributors and small teams and startups to really Build Together uh in the open continue maintaining kind of this technology to be accessible to everyone one and at the same time you know create enough incentive and and uh value for people who are actually contributing thanks IIA uh NJ uh would

8. 3:42

   love to love to hear some of your background and and what got you excited about uh about ritual I know there's a I know some of the backstory but I think everyone would would love to hear this from you yeah thanks for uh thanks for putting this together I'm excited to to have the discussion so my background uh originally is in cryptography um mainly did a lot of work with uh privacy systems uh around payments and smart contracts in 2017 and where I've spent

9. 4:15

   the majority of my career was as a VC at polychain where I focused across a bunch of different Industries within crypto but but largely spent a lot of my time time excuse me on on infra as I was looking through you know what got me originally excited in investing in this space um I I think one of the biggest things was crypto provided new tools for developers that were previously unavailable um you now have these kind of trustless platforms that allowed you to easily create

10. 4:56

    Innovations on top of other Innovations build new products products a whole new incentive mechanism and all of this great you know cryptography uh underlying all of it um as I started exploring AI over the the last couple years um I I think it really was the the creation of llms and seeing how powerful they were in I think transforming me into seeing how AI could be uh could be

11. 5:23

    very valuable and useful to the world um ai's gone through many Winters and Summers over the the past you know dozens of decades and you know we've seen a bunch of AI crypto companies starting from 2017 but none of them really kind of took off at the time and I I think there's a a bunch of different reasons for that which we can go into later um but seeing the opportunity set of AI but also the potential centralization risks which we've seen in kind of traditional

12. 5:56

    Tech platforms that brought a you know kind of an opportunity in my head of something that would be really valuable to build uh and so started ritual last year with uh you know we're now a team of 15 all around the world the goal of ritual is to make it really easy to connect smart contracts to AI models and we really want to help grow the opportunity set of crypto adjacent and sort of AI enabled uh crypto applications in the ecosystem and um you

13. 6:30

    know we're happy to have Anand as as an investor and and Ilia is one of our one of our advisors so again thanks for setting this up and I'm excited to to chat thanks araj uh we have Patrick as well from Vanek um I think Patrick might be having some technical difficulties um oh Patrick can you uh I think you might now be a speaker and if you are um Let me let me give you some let me give folks some background because he was a late ad but uh uh his team put out a report yesterday at you know focus on

14. 7:01

    this intersection which I thought was very timely and topical and uh we thought we'd have them on just to talk about the opportunity here so maybe Patrick a little bit of background like how you started to put this report together and any highlights that you want to discuss but I you know I think we will get into the meat of it during this chat but uh maybe we can start with uh sort of what pequ your interest in wanting to explore this intersection yeah hey you can hear me

15. 7:25

    correct yes great um to to me like there's been so much discussion about the intersection of crypto and AI for quite some time now and lots of it seem like fluff and so like here my job is to kind of figure out like which part of the markets and which tokens will go up like that's my role as a research analyst and it seemed that there might be some sort of narrative formation around that like and so that's that's kind of where it started so I started really diving deep in December which of course does not compare to some of the

16. 7:59

    more distinguished speakers here in terms of time spend the topic but you know I've covered a lot of things and I've thought about it pretty deeply about like the ways in which you can you can pair and it kind it kind of goes back to the basic principles of of what crypto does um it just open transparent immutable and gives clear ownership properties I think those things apply to

17. 8:22

    AI the models the data themselves and then other kinds of faucets to provide the ability to create those models and provide inference use them whatever um I think just makes a ton of sense that that's kind of how I I went about this from an investment standpoint and then I just um thought a lot about the idea of what comes next awesome Patrick thanks for joining us um you know as you know but NJ and IIA have been looking into the space either from an academic standpoint or

18. 8:54

    obviously professionally for quite some time and and NJ has got a great lens both as an investor and as a builder um I want to dive in the thick of it our hope is that we can do this on an ongoing basis there's a lot to unpack here there's some really cool people building some really cool and I'm hoping that we can do this on an ongoing basis and and really start to unpack more about what's exciting um but uh maybe for starters NJ I'd love to get your thoughts I know you and I have chatted about this it's coming up in

19. 9:18

    maybe almost a year now on uh on the on the thesis behind decentralized Ai and and what you saw as sort of the opportunity set and there were some things that you said and I I don't want to I don't want to take the thunder away from you but that that really stood out to me in terms of like what timing and you know how you you were sort of uniquely positioned in some ways given your uh you know your Vantage Point as an investor but also having done

20. 9:42

    research in this area before that made you want to pursue this it was like a a Now or Never thing that that you sort of alluded to there was this interesting time pressure but I guess more sort of from a from an ethos perspective right like what what does what does this mean to you what does decentralized AI mean to you why is it it important why should we all care why for all those people listening and you know what were the things that clicked for you when you

21. 10:05

    were like this is where I want to invest my career because you could be doing anything rer you're so talented like why why did you decide to do this yeah absolutely um the really the ethos of I think what this space provides maybe to just zoom out and cover the I think the properties of crypto you know what crypto brings to the world and then what AI separately brings to the world crypto to me largely is a tool for self- sovereignty so as

22. 10:38

    we've seen you know everything in the world tends towards centralization that's just how it happens Capital um you know sort of social things Etc and crypto is really the the tool in our toolbox to allow for self- sovereignty again and so we get all these great properties from that uh privacy computational integrity censorship resistance and then to kind of take where where AI is I think especially post um post

23. 11:09

    llms you know an analogy I like to use is for the last you know 50 60 years that we've had computers you're probably a little longer um we've had to you know talk to computers with code that's how we interacted with computers and and now computers can interact back to us in human language which is really really powerful mechanism that is is sort of a recent Innovation I remember you know learning about llms and and you know it

24. 11:39

    kind of had a moment or I kind of had a moment that was like when I first saw Bitcoin it felt like almost this Divine creation of of something like unique and interesting that was um that was getting created so where do these things intersect and and you know why is this interesting and you know why um did I decide to go do something here this to me um is one of the biggest opportunity sets in the crypto space uh broadly um in the ways that crypto can

25. 12:09

    help AI as I mentioned before AI is starting to centralized in a in a big way as as things tend to do there's a couple of big model providers there's a couple of cloud storage Prov or you know Cloud uh Cloud providers whether for storage or for for compute and that is a problem once these llms are starting to be used in everything around the world all of the products and um and devices and we saw with the open AI

26. 12:41

    Fiasco a few months ago you know just a few people are already kind of controlling the fate of one of the biggest and most important private companies in the world so where I think crypto can be really valuable to AI is giving these self- sovereignty guarantees to AI bringing privacy back bringing control and and sort of open governance back uh building uh or bringing sort of computational Integrity back and then to you know quickly bring up the other side how does crypto or how

27. 13:14

    does AI help crypto given these properties of AI being this tool for improving the way that we can do you know we can interface with computers and providing this this kind of new statistical computer you could call it there are a number of different areas within crypto that are really exciting that can be further improved with AI whether that's governance and improving the governance processes of defi or of dows doing media

28. 13:47

    Generation Um whether it's audio video really whatever modality you may want all the way to things like gaming you know being able to create more custom experiences to the user that are are actually using them so what fascinated me most about getting into the space and why I wanted to is part of myo at at polychain was trying to figure out what developers and crypto would be building in over the next few years and crypto AI feels like that asymmetric upside

29. 14:19

    opportunity to build in and you know I ultimately am very passionate about bringing tons of users to crypto and we've seen cha GPT get hundreds of millions of users in a matter of a few months um we can bring that to crypto you know that is there I think we really just have to build the right the right tooling and and also show developers how to build it so that's kind of my my

30. 14:43

    general kind of high level thesis on you know the space and where I see some some interesting opportunities awesome thanks Raj um a lot uh you know lot unpack there but IIA would you know I think you you've been spending time here more than most and yeah I would love to hear sort of having spent time at Google coming at this from uh a vantage point of being in a company that's you know putting out products that are necessarily centralized nature from an AI perspective curious like what

31. 15:12

    what motivates you what's driving you today um and uh and maybe one other sort of sub question here which is nuraj was talking about some utility and obviously the ability to bring hundreds of millions of users I saw you nodding your head as he said that would love to hear sort of your uh thoughts as well yeah how do you know what do you how do we make that happen with um with with the focus on this intersection because it it

32. 15:34

    is definitely promising yeah I mean there's a lot to impact for sure and and Nash gave a good like motivation here um in general I mean for me I mentioned right I do believe I mean there's like a core set of the lease right so sovereignty is probably one that we all share being in in blockchain space and actually put out a kind of more detailed blog around like kind of self sovereignty fully full sovereignty which includes kind of AI

33. 16:07

    like your data which in a sense you know related to AI related to experiences you build and so really interconnecting that and kind of within that I think that capabilities like this you know thinking of this more as a tool right this capability should accessible to everyone and should be in service of you of each individual right and so this is like applying self sovereignty mindset to Ai and to achieve that you

34. 16:42

    cannot have you know a single company that has kind of uh due to network effects and data mode you know controls pretty much everybody's access to intelligence right because they will naturally however good the people are they will naturally need to censor because they need to comply with different regulations across different uh areas they will naturally will need to kind of uh limit things but also it opens up you know if you want to lower

35. 17:12

    the prices you can I started talking about this I I have this sometimes feeling where like maybe I should not tell people like bad ideas of how to like manipulate people um but I'm assuming you know anybody uh can figure out those things but if you want to run right now this AI you know chat startup the what would make sense is to offer it for free and then run on every single token a auction an add auction which sells it to

36. 17:45

    he bidder what this allows you to do now is you know like mostly it will be generated based on the top token but sometimes it will change it you know when it's like a mentioning a product it will change it when it's mentioning you know which presentent you should vote for it maybe shifts it right based on the auctions and so this creates like a you know best opportunity to manipulate people but for for a business that

37. 18:10

    builds it it's highly like it's very much something on their like clear business line because it creates a ton of Revenue and they can offer now their product for free uh and like I mean that's kind of how Google was built right but like in this case you know you wouldn't even say that you ran the ad because you just run like nobody knows which model you ran um and so I think that's an extremely kind of scary reality that we're going to get into because I mean these things are

38. 18:40

    expensive to run and so um anything that will you know reduce the cost or potentially makes your revenue is something worth doing for a startup and in turn that means like that we cannot believe the tools right they're not actually on our side they're not uh producing the output that you know kind of statistical model gave you it actually produces a highly uh biased output based on a bunch of criteria so that is what motivated me that's why I

39. 19:09

    was excited about ritual uh because like we need provable compute we need way to provably verify that the output I'm getting is indeed from this model not just from whatever you know somebody like else can write an if statement and you know if you're from vanike then output that you guys should buy all this tokens right so like you don't know like you don't know what the output you're getting so that's a really

40. 19:34

    big motivation for me and so kind of turning this to the user side um I I think important part is kind of the S like we want to get billions of people s sovereignty right I mean which means like they should have a blockchain account Etc but like at the end they should have self s and I think so far we haven't had really good examples of value self value that people can benefit from it's mostly being you know a into this

41. 20:04

    token or you know swab this token or really good uh you know uh gift picture that you can put on your profile and I think like private data you know being able to uh kind of use it on your like use it kind of this language models on your data without this data leaving your devices without needing to uh kind of give up all your privacy to some company and right now there's like a bunch of startups actually that offering this which like hey you know give me all your

42. 20:37

    data and I'll build you personal LM and that's like even scarier because like at least with Google you assume some level of you know private data privacy protection all those things with a startup I mean I bu a startup right you're just hacking things together initially and so like you have no idea how it's going to get hacked tomorrow to get all the private data out so so we need this kind of offering that is way more robust but

43. 21:01

    also it creates new opportunity right now I can sell maybe some of my data on the data market for people to train better models right so I I can make more money I can have a model that like votes on my behalf in B elections and you know uh onchain governance so I don't need to like waste waste time on that I just need to tell my model you know answer a few questions and now have enough context to do this so for me like that's

44. 21:27

    it's a new opportunity a new tool set but also like wearing selfserv hat is like really important to deliver something um that is not manipulating that is really on your side side and that provides you new new economic opportunities thanks IIA very helpful uh as you was speaking I was thinking you know when uh when you started talking I was like when do we pull the Google out of Ilia and then you landed the plane so

45. 21:55

    I'm glad happened uh Patrick I saw uh we're all in a zoom together at the same time Patrick I saw you might be raising your eyebrows at some of these did you have anything to add there what do you think no no I I I I was just working through like what's the so what for me and like I come into this process as as somewhat of a child who wanders into a movie halfway through right that's a big lowski quote for you um and I typically

46. 22:19

    come in with a really simple approach a very left C of understanding of what's going on and why and then unfortunately I start gathering enough information and become mid curve informed and then typically I use those learnings to go back to the left curve and then justify them using mid curve explanations but like the key things to me about ai ai models is that you want to know like who built the model like is it good and

47. 22:45

    where is that data from and to to me like the the ideal environment to have those ownership properties and those proving properties and a place to host those proofs would be blockchain and that that's kind of like my a really simple way of looking at this and why I think it's a really good Nexus crypto and AI yep I I think we're all we're we aligned um so for sure that that makes

48. 23:08

    sense um N I wanted to pick up on you know the the high level goal right your motivation getting you know seeing the the spike and the rate at which chat GPT grew and the number of users they were able to mass and and your desire to want to do something similar for uh which which can Propel utility for the blockchain as well what are some areas do that you think are sort of ripe for

49. 23:31

    the picking today obviously there's you know IIA was talking about the concept of like you know data marketplaces data lockers and that starts at the lowest layers um and then obviously there's incumbent Solutions on chain defi nfts that could leverage from that could leverage AI but uh you know what does that path look like for you and ritual like where how how do you start to get

50. 23:54

    to that point where it feels like you know a user is using some of these Primitives uh Without Really maybe knowing that they're using blockchain Primitives here and where AI is serving purpose for them so I'm just curious like where do you see the lowest hanging fruit today but at steady state you know where do you see uh where do you see things happening so a good a good frame of

51. 24:18

    reference I like to use I I think I might have mentioned earlier is if we take what and we're just talking I mean let's just talk about LMS for for a second um what do those provide the world it provides this new human interface into using you know using a a computer uh it's able to summarize information in in a bunch of different ways and it's able to understand our intent um and and what we're trying to do with certain things and so I I think

52. 24:50

    one of the challenges this space will face and we've seen it a little bit historically is that the number of crypto people that also understand AI even at a kind of 101 level is low and I think we need to we need to push that further because I think once you understand a basic bit of both then there starts to be some some interesting intersections so thinking through you know what has worked in the web 2 AI world and and even just naively copying those things over um so for in inah's

53. 25:24

    example imagine you had your own sort of self- sovereign ident identity with your data that's portable and and encrypted and you know stored on a on a blockchain or offchain imagine you could build uh you know build a personal agent or personal assistant that is able to now plug into your different daps and execute things on your behalf it can start suggesting you things hey you typically claim you know this token from this network around this time or hey your tokens are unlocking here or hey there's you know

54. 25:57

    this asset that you really like you should go do this trade basically being able to supercharge the what today is a very manual experience and I think thinking through being able to provide recommendations being able to act on the behalf of users uh is a very very big opportunity set um to be able to do and then this is not to mention diffusion models um now being able to do media generation imagine that you could build you an nft marketplace where all the

55. 26:30

    content is is sort of autonomously generated you build nfts around it and then you can you know offer those for sale it it really brings down the barrier to entry of creativity and I think thinking in in this sort of vein I think helps conjure up a lot of really interesting ideas um and and this is not to mention you know other ideas that that other folks have mentioned on the call so far you know like like defi for example there's tons of sort of immediate use cases with you know fairly

56. 27:01

    low lift um and not super complex models that that you could build so my Mo with ritual and I think with the space broadly is um let's get people really educated about ML and crypto and then the intersection and once I think you have those Frameworks in place people will start coming up with with weird all the time you know I remember it's kind of like 2017 um ethereum had smart contracts you know the only thing

57. 27:31

    people were using smart contracts for was uh to do icos and then over the next coming years we saw compound we saw dydx we saw Unis Swap and um those were just independent Creations in many ways so I'm really looking forward to that moment in this space when people just start coming up with stuff independently but you know it obviously requires a little bit of starter energy to show

58. 27:54

    what is possible on chain super well said yeah I think the the talent piece is is fascinating and I think Ilia with your work and and the your team and ritual it feels like there's going to be a beautiful sandbox to kind of build on top of and uh yeah I've I've held the the idea or notion that I think what defi did for blockchains around the Circ at the time that you were talking about youraj 2017 2018 I feel like AI could be that Catalyst to boost our Tam for crypto as

59. 28:26

    we look forward so uh excited about this and excited for more talent to come in and build as well um to to keep the Train on the tracks I'd love maybe to seg a little bit like to unpack this a little bit more you're right like you're you're creating this uh this sandbox this playground um can you tell us a little bit more about like how ritual might be facilitating that like I know I know you've got some some opinions on this but if a builder wants to get

60. 28:49

    started and wants to start tinkering like where do they go what do they do today with ritual so the um right the the ultimate goal of what we're trying to build is is act as a conduit um between between AI models and uh and and and you know crypto systems so the first thing that we've released that people can play around with and use to start building applications uh is called infranet it's kind of like Internet with with an F so the best place to learn more about this

61. 29:21

    is um ritual. netproduct and it has the this kind of crazy sort of skitso diagram we put together of all the the the different pieces of you know the eventual system um but if you scroll down there is a an explanation on on what infernet is and the the basic idea of infernet is we're coming from the frame of how do we start getting people to use AI on chain today and we've taken some steps where you know maybe it's not

62. 29:54

    the most secure perhaps or perhaps you know it's kind of only so you know so far complete but not fully complete and um what infranet does is it's an offchain compute Network um you could think of it kind of like a like a chain link and what it does is from say A you know from an ethereum contract there is a solidity SDK that we have put together called uh infranet SDK that is able to

63. 30:25

    go and dispatch calls out to the infnet network which um you know anyone can go and and start playing around with you can set up a node fairly fairly easily um and that system is called it's in infernet node or the infernet system and so a smart contract developer can say hey I want to build a chat app maybe it's a mostly offchain chat app but maybe there are some pieces that you want to you know interact with onchain so it can say hey I want llama 2 these

64. 30:56

    are the Fe features I want and um perhaps you want to fine-tune it and and here are the inputs that I want um from the user it's able to dispatch it the network takes care of all of the you know compute fulfillment tasks that that you typically would need to do and then Returns the result back so this is kind of what we've built so far that people can start playing around with and um we actually about two or three weeks ago also released a whole Library

65. 31:27

    potential or I should say um you know implementations uh on infranet that you can play around with and and you know build you said this is this is a this is like a trust system correct that you're trusting like a nodes or consensus amongst those nodes is that is that the way to understand this yes so this this network is not fully decentralized yet but but it will be uh in the future um

66. 31:51

    the things that you can do um I'll go into an potential application and then I'll I'll explain where the proofs come in there one of the applications we built is called uh friend rug it was a sample application we released about two months ago the basic idea with friend rug is it's a friend tech room that this bot is the one that's responding to you know back and forth from requests it's a you know it's a popular open source LM that we've you know put a bunch of examples

67. 32:25

    into to kind of you know have it be a little bit harder to to um answer queries and basically the goal of what you're trying to do with friend rug is you're trying to convince the bot to buy your frch keys so you can come up with different different ways different methods of doing that and the app you know the benefit or I guess the the purpose for doing that was to show how you can use these non-deterministic llms to um you know come up with a little bit of stochasticity you could say feeds it

68. 32:57

    into a classifier uh that's on chain with uh zkl proof and then you're able to go and um and you actually use the application so we'd like to see a bunch more things like that um I think there's a lot of excitement about agents and um that's one application we put out to sort of show that so that's um that's one way that you know developers can build with with AI crypto today inside

69. 33:22

    of their applications awesome thank you uh by the way for folks listening if you have questions for IIA NJ for Patrick uh please just you know uh at message me you know we'll make sure that I don't you know my voice is uh is uh is not the most Pleasant to listen to this whole time so please feel free to hit me up and I'll I'll make sure to unmute you and maybe you can ask questions director to the speakers as well um you know just kind of piggybacking off of what said I curious like when you think about I

70. 33:52

    guess you know in my head there's like this 2 by two like this decentralized centralized as open source and not open source and you know what what do you what do you really care about and how far does you know where does decentralization need to go in your opinion in at steady state is Everything decentralized is Everything open source for you uh in your mind or um you know do you feel like there's some compromise here at varying points of the stack and I'm just kind of still riffing off of

71. 34:16

    what you were saying as far as like data for example goes and sovereignty um and in know providing infrastructure at those layers versus something that's further upstack and curious how you might sort of uh parse that yeah I mean I think there's like different use cases and different kind of Parts um that require different trade-off here um so like for example you wouldn't want your private data to be open source um at the same time you want to

72. 34:49

    have like a you know decentralized application for your private data that's encrypted such that you know like Amazon cannot just say like tomorrow you boot it out a Dropbox can like boot you out right from your Dropbox also they don't not I mean sometimes encrypt sometimes it don't right so that's kind of the you know the framework there is like what makes sense I think the kind of for decentralization

73. 35:13

    in AI right I think where it does make sense is when you want smart contract to be an actor right you want either like you know have some money on Smart contract have multiple people interacting with it have some form of uh uh activity that you know will be onchain and for that you want to have a decentralized model with some way of proving that indeed it uh I mean decentralized inference for a model with

74. 35:45

    some way that uh proves that the output came to the blockchain smart contract such that you can act on behalf of this uh on the flip side if I'm running something like on my own data I might as well just run it in my device right and this is where um kind of we've been experimenting with some Edge models and kind of been doing this Edge intelligence uh Summit last year and so like a lot a lot I think a lot of it will also bifurcate there because

75. 36:17

    already you can run like I have a mistal model 7B on my laptop and I kind of use it from time to time the Mixr I I could run it on my M2 I wouldn't say it's like performance the the mixture of experts but like as compute becames better right the models also as it become smaller are like more intelligent for the same parameters uh we'll see kind of improvement there so so I think like there's kind of the you know like two by

76. 36:47

    two there as well which is like is this data private um is there some shared state that you interacting with or not right data private or not so those kind of the decision where it's like I want it local uh but open source I want it decentralized um that's kind of how you make those decisions and then probably the other important piece is um right now at least Right decentralized comes with a lot of overhead and so understanding where when this overhead

77. 37:21

    is reasonable versus when not is also important um and also you know some people right now for example use aash which is you know a compute Network you can go and just like get you know they have few A1 100s on aach and so you can kind of uh use it from there but you don't come with any proofs right and and so you don't know if they actually run it the model you asked or it can be run by something else CL run and again it maybe matters and

78. 37:53

    maybe not right that's kind of really depends on on the U use case as well so that those are kind of really decision criteria that you need to be operating under for anything that's like Financial right if you want a smart contract making decisions you want you know provable you want it to be like verifiable you want to make sure that the model came from a trusted source as well I think going down the pipeline and this is where you know I think in blockchain we kind of care about this like inference part in AI actually

79. 38:25

    people care about the data model training regime like those pieces as well and this where it becomes really complicated because right now actually people compete on how clean the data set they have what are what types of copyright the data that went into it and who they want and what and don't want to tell about it right like there's a whole like slew of problems there that goes

80. 38:49

    into like like if some of the startups right now that have pretty good performing models potentially they open source their data set they may get in trouble right so like that's uh gets into like complicated question of well is copyright even make sense in the uh in the new world of generative um kind of language and images right and I would actually argue it doesn't because you know we can just generate a bunch of

81. 39:13

    possible text and then say everybody else is violating our copyright sotic data so I think there there's like pieces there that are interesting that like we also need to keep track because again to not not give somebody an idea but I'm pretty sure somebody's already doing it if you want to change how people will think in a year from now the way to do this is to start infiltrating Forums on internet with content just

82. 39:40

    like by posting stuff because this will then be you scraped put into training data and train AI models uh in a year from now and this AI models will be highly biased based on whatever data you put in right now hopefully again I'm not opening a America here but this is like the best way to start influencing hundreds of millions of people right now that's like very very hard to navigate

83. 40:04

    very hard to remove right and like literally just post like you know go to all the forums and post whatever like types of text you want the ways people should be thinking and you can use actually current chat GPT to like prepare the text for that and make them you know uh discrep so like that is really important part because like whatever goes into training data creates biases and all those models that then comes out when

84. 40:29

    you're actually asking stuff and so uh that's where like I would say importance ac across especially Security in alignment will be and again blockchain can be a helpful tool here but there's some pieces that really really need to figure out for that to happen yep yeah I think you just described the reviews section of every Amazon product listing or comments on YouTube or Reddit um uh sadly for better for us um you know one

85. 40:59

    thing that NJ mentioned which you is talking about Talent um and the second thing is you know the the rapid Pace with which a centralized company like open AI was able to move and we're at this point today like you know I I I'm going to date myself here I took my first AI related programming class back in 1999 uh we we were taught this thing called lisp uh which was supposed to be the thing that unlocked AI but obviously

86. 41:23

    it was way too early right there was not enough data for AI to make sense and then to your point earlier you know we were able to scrape large amounts of data and get this far really fast and um so Maring a couple of things Talent data and the fact that today when we think about decentralization you know inherently there's a few issues that come up right there's some trade-offs that we made along the way so for example when it comes to blockchains we

87. 41:47

    talk about the trama um and so when I think about speed and I think about talent I feel like data is now to your point IIA is like it feels like it's it's REM modable problem we've kind of almost were there to the point where either it's synthetic or it's real and we've got this data to work off of but what are to your in your mind when you think about what could impede decentralized AI uh or what are some trade-offs that you know Builders might

88. 42:13

    have to consider when they're entering the space what is that in your opinion is it latency is it um is it too much de to your point like you know you don't want to open source your data right like that's a you know you know is so are there any such trade-offs come to mind and open to any of you by there Patrick n or IIA feel free to jump in but I'm curious what should we be considering and what are some pitfalls that we're not sort of thinking about

89. 42:35

    yet well I would just start with the obvious one which is compute I mean the reason open AI is so ahead of the game is because they got their hands on a ton of compute I mean this is saying from somebody who was doing very much the same things as a startup which had like 700k funding uh you know we could not afford building out clusters of a gpus that didn't even exist back then um and so I think the general story and right

90. 43:06

    now is that kind of in open source world and even in like smaller startups there's pretty much lack of access to Compu and so they need to rely on the apis that are provided by um Google and open AI or interestingly kind of meta became this new hero which open sourced like you know a model that they trained on massive compute and that everybody can now F tune and play around but if you

91. 43:35

    are you know talking about kind of lowering barrier right the compute being the probably most complex things to get hands off hands on and and and when we're talking about also like production usage it's just expensive right this some of this queries you know they literally cost cents per query at the end and so um you know if you have a user who sends like 10 20 30 queries per day that's like a you know not significant

92. 44:08

    price to actually for interaction so so I think like that that's kind of generally like that's not a problem of decentralization it's just general problem with current language models this is where at least from my perspective The Edge compute is comes in really interestingly because well if you can use your own hardware and everybody just uses their own Hardware then well great that's um you

93. 44:32

    know removes both centralization points and and kind of doesn't require the companies that are building products to be uh responsible and have huge costs and so we just really need to like make the model smaller but making model smaller itself requires you know a lot of experimentation and Innovation and that requires a lot of compute so so I think there's kind of that that interesting tradeoff but you know hopefully as time progresses like I for example uh testing

94. 45:03

    out actually new service that uses like custom chips like there's there's a a lot of cool stuff coming out around that uh to lift at least partially this this challenge all right you teasing a little bit therea but would love to hear more about this uh the hardware and chips maybe a later time uh over to you NJ uh curious if you you know any any trade-offs that you feel uh that you've started to encounter as your building ritual um and then uh Patrick I have

95. 45:35

    some questions for you as well yeah there's there's definitely a few um there's there's one is the the limitations of the cryptography that's being used to Garner some of these guarantees so the Holy Grail is still privacy which um you can't really use a lot of the existing techniques that you have available to you um because you know you're trying to compute on private data which is sort of one of the the um you know hardest problems in in privacy

96. 46:08

    um so that's to me that's you know more of a time thing than anything now that we have a bunch of companies that are excited about crypto aai I imagine we'll start to see more investment into this into this area um very similarly to what we saw from you know call it 2016 to you know now 2024 in in the zkp space um you know generating a zcash transaction used to take you know minutes and minutes and minutes and uh and now you're able to put you know whole VMS into snarks and

97. 46:42

    prove them in a relatively efficient time period so the these things have accelerated quite a lot and um you know Ilia sort of mentioned Hardware um a lot of the math that's being done in you know AI compute is is very similar to the math that's being done in say a ekp um to to actually do the proof generation so we'll also be able to mirror some of the techniques um that that we have from there there's certain techniques that we need to bring into

98. 47:14

    the crypto world you know namely you know going from from floating Point math to fixo MTH that it's cryptography efficient so that's one whole big area that I'd love to see some more development in and you know we're we're trying our hardest to uh to try to accelerate this um I'd say the other limitations uh you know largely lie around the semantics of what the AI is actually doing so I I I think vitalic

99. 47:43

    summed up the different parts of where crypto and AI can intersect in a really interesting way he took it in in sort of a more of a highlevel kind of philosophical manner in relation to you know the game calls it and I think thinking through the different semantics of how these things interact is is a big Challenge and something we have to think through more um there are pieces of how much does computational integrity actually matter for various use cases um

100. 48:12

     maybe for having you know doing uh recommendation systems for social you know you want weak computational Integrity but it doesn't need to be you know zkl or something like that but maybe if you're if you have an ml controller inside of a defi protocol you may want that to be zkl you don't want any you know any kind of lack of finality or or or potential ability to uh to dispute so I think we're we're starting to grow up in the sense that we've gone from like crypto AI is really

101. 48:43

     interesting now we're starting to encounter the initial technical problems and also what are the most interesting use cases and a big question I think about a lot is you know what will be be the first crypto AI thing that gets to you know 100 thousand users and then a million users and um some of them may be more in the background like you might just be playing a game and it happens to

102. 49:05

     use AI or it might be something you know right in your face and uh you know maybe it's more of a chat like experience so these are you know this is not all of the things that I'm thinking about but these are just a few that um I think are important uh as we're in this development of the space thanks much I think uh you know with infernet you've kind of created the sandbox and I think uh I I I recall Chris Dixon's blog post from 2010 when it's when he says you know the next big thing will start out looking like a toy

103. 49:35

     and so I think you start to experiment and create things which will eventually find utility or become mainstream uh and so uh yeah you know excited to to get this uh for people to start tinkering and Building Things I think that's uh that's incomp upon us um we have about 8 minutes left uh I want to be mindful of your time uh again folks if you have any questions for these uh fun Gentlemen please hit us up on Twitter uh we'll be

104. 50:00

     sure to answer that but I thought for the next few minutes we could just talk about the report that Patrick and his colleagues put out um and any summary you can provide Patrick any background would be super helpful um but uh I think one of my take ways and I'm curious to hear everyone's opinion on This was um the forecast the AI crypto projects which will do roughly you know 10 billion in Revenue by 2030 and and what that really means and you know I I think we we all have different ways of interpreting that

105. 50:26

     and I I think someone messaged me said hey what do you think of this and my instinctive reaction was I I think these forecasts are utility-driven which is actually promising the fact that we have a forecast is uh is really promising although I think it might be an older magnitude lower than where we might be in 2030 but uh I don't want to influence anyone's opinion here but Patrick I'll

106. 50:42

     turn it over to you maybe to share a little bit about the report and then uh would love for IIA and Nash to to chime in and add their thoughts yeah you know the like the the first idea was just like ballparking with that Target might Market might be and you know working through some cide analyst stuff and reading through information myself became pretty clear that like the clearest Pathway to

107. 51:04

     understanding what monetization might look like what that market might look like was just saying okay like what are the productivity gains like or the losses and costs right and so that's kind of difficult and so there's some pretty good analysis of precisely which jobs will be impacted and why and how they will be and there's some just really broad estimates and so like it's a really hard to take that and try say

108. 51:27

     anything super meaningful so take those numbers of the grain of salt um as a is mentioned and so the idea is like okay like once we understand like okay what are the productivity gains The Next Step sort of thing okay like how much of that productivity could be captured by enterprise software right sorry so what how much that productivity gains could be captured amongst the different

109. 51:49

     branches of AI so you think everything from building the hardware to writing the software to all the different faucets of supporting it and like AI infrastructure to service the idea was basically like okay like let's assume that software is going to be the most value added it's going to be the most interesting it'll capture the most of that that stack and then so working from there the next step is say okay like of that like what can software companies

110. 52:14

     extract from Enterprises so there's a bunch of different examples like that you have right now for everything from Microsoft Office to accounting software you say okay like what amount of productivity G s or costs do these things save and what are they taking and so General estimates would range between like 3 to 30% um and so ballparking that that's kind of like how we came up with this like software Tam um with the other

111. 52:38

     ones like there's there's a sense that there's a little bit of arbitrariness to like the numbers we chose but it was just kind of like to get an idea out there like hey there's like this this Nexus and so Nexus was kind of explain okay like there's this Market size that you could take how is it going to take it and then so you see like the current projects that are doing it there's GPU provision um there's building AI models having marketplaces for those models and

112. 53:02

     then also um data marketplaces as well and then finally having the ability to train and supposedly fine-tune those models something like like bit tensor I think is like the the one that's pretty unique in that capacity um and so that's just kind of like gist of it's like okay like here's what we have today here's the potential the future I think the most interesting thing um to take away from this report is is again like the basic principles like hey I want to

113. 53:25

     prove something happened right so it's like in blockchain and crypto what are the ways in which you do that like what are which the ways you take data that's not on chain and bring in on onchain so you have maybe consensus of offchain nodes maybe you have a like trust Network like a chain link maybe you have like economic value backing truth like maybe like an igen layer or a Babylon something along those lines and finally

114. 53:47

     you have like cryptography um and in this case it'd be like you know ZK ml or ZK or whatever that might be to prove your your your um your mathematics and so like thinking through those like the most obvious clear use case to me comes from this proliferation of onchain dexes that run offchain order books and so like the most obvious one for that is the idx the idx right each validator instance runs its own order book offchain and so you have to say to yourself okay how do you

115. 54:19

     know that there's like fairness in the in the way in which they're structuring fills in which the way they're ordering trans transactions and so there's an entity called skip that's supposed to monitor them and there social consensus for slashing them but there's no way to prove like okay what they saw was correct accurate fair and there was no coordination between validators so like logically you could you could run some sort of like AI compute that says like

116. 54:42

     hey this this mathematical computation was fair and on top of it maybe you can even like look for nefar actions by those validator sets like ways of extracting me in order books that only some people have like talked about in theory and not seeing practice obviously like those those AI models would have to see those things and would probably be open to novel attacks that haven't seen before but the premise is there like when I go to like the Chicago work

117. 55:05

     exchange I trust that the Phil algorithm that it says Works accordingly and the whole structure of how I get filled I summ my order to the time it gets filled and then sent back to me like works as promise and so you have that and you just trust that but for like a dydx um and some of these off order books you don't have that you don't have a clear understanding of how the fil algorithm works you don't understand if it was fair and then having the ability to to

118. 55:30

     prove that that fairness I think would be a really powerful immediate property uh that like a AI compared like combined with like zkl which I'm sort of like Pro system can can give to you awesome thank you wow that's uh uh very practical utility rent all the way from like you know computes portion of the stack which you cover in your report all the way to like practical use cases

119. 55:55

     appreciate that uh I know we're coming up on the last minute here uh I wanted to uh give the floor back to Ilia and NJ I know they're um Ilia has has been working on some things uh on the edge and you know there's a summit coming up at e Denver among other things Naraj is going to be at e Denver as well so I just want to turn over to you folks any call to actions anything that you recommend uh our listeners pay attention to or other areas which we can tune in

120. 56:19

     uh and very quickly you know um whether ier and nage like this or not I've kind of recruited them to do this every couple of weeks uh I don't know if they'll sign up for it but my hope is just unpack this category a lot more and uh and I think to n's Point get more Talent building into the space so but uh yeah so tune in maybe every every other Thursday at 9:00 a.m. Pacific is the

121. 56:41

     goal but IIA um Edge Summit or other things where people can find you or learn more in the next uh few weeks months yeah so we have a hackathon in Denver uh it's called autonomous agent it's you know attracting folks to really dive in and um like hack some things as as n pointed out just like let's get some people hacking let's get some use cases going and then see what really

122. 57:10

     works so we have this opportunity at Denver um February 26 27 I'll be speaking there as well as folks from polygon and other others coming in as well and then there's AGI day uh and kind of decentralized a day 28th where neaj I also speaking um so kind of inviting uh people to join or tune in uh and yeah I mean we'll we'll be continue doing this one way or another as well as there's like a project I'm contributing which I'll be rolling out talking about a little bit

123. 57:51

     later in a week or two into space lovely andaj will find you in Denver as well yeah at East Denver yeah get get used to all of our voices because you're goingon to be hearing them a lot more in Denver um yes we'll be we'll be doing a couple of of talks and panels and workshops and a bunch of other stuff so excited for for that um and yeah if if there are any questions on on what we're building um we're on Twitter at ritual

124. 58:23

     net um or .n net is our our website and uh yeah excited to see what what questions people have and uh what people build thank you all so much uh I want to be mindful of your time we we budgeted an hour and we're slightly over uh really really appreciate this and uh yeah let's let's keep doing more of this and and unpack this category more but more than anything else super appreciative to both of you who've been thinking about this and um you know I love nage your comment about you know

125. 58:53

     how you know thinking about the next hundreds of millions of users who might come on chain because you know there's a an alternative to centralized Ai and how that could you know boost crypto Tam across the board and but also provide super you know awesome utility that's uh that's comparable to what we'd find in a centralized environment so thank you all again so much Patrick I know this is a last minute ad for you but thank you for making it and uh yeah we'll we'll catch

126. 59:18

     you all again soon bye everybody bye bye thank you