DevHub Live 39 with ASIMOV Protocol (@asimov_protocol) - discussing data, AI, and more

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

1. 0:20

   hello hello welcome back to another episode of devhub Live I am your host Joe I am super excited as always but today we have a really special guest and I uh I want to tell you I always take time to prepare for my guests and this is one of the times where uh as I Was preparing and I had literally pages of notes and questions to talk about like um asthma Pro protocol but also the

2. 0:51

   founder the team so much stuff here and I I really don't want to delay talking about it but this is going to be really I think fun and exciting episode so I encourage everyone to stick all the way through till the end you can always watch replays on our uh YouTube everything's always there we watch it on X wherever it might be couple just really quick things if you haven't subscribed to the devhub newsletter I encourage you to check that out and

3. 1:15

   subscribe all the details are always there about what's happening developer ecosystem uh super exciting stuff that's happening uh there as well but without any further Ado uh Asma protocol you may have heard uh about just kind of through the ecosystem it is a newer startup that is publicly uh they've been building for a little while though and uh the team is all people who come from I want say all but the founders come from uh near in Aurora so I'm so excited to have arto

4. 1:47

   here to come and talk about this and talk about everything that he's been building and everything else so without any further Ado uh let's just jump right into all of this AR thank you so much for joining me today Hey Joe uh good to meet you and U thanks for the invite Happ thank you so uh there so much to get into but let's start at the really high level here of just what is Asimov protocol sure uh so we we are building on top of near um with a focus on uh

5. 2:19

   improving the world's um quality of information so you you know that the the situation right now is a is a little bit interesting um You probably um following news yesterday and and the like there's a lot of misinformation out there um there's a there's a lot of um um uncertainty about how are we going to in this Century uh actually be able to believe anything at all right um any everything we see whether it's um text image or video is now subject to

6. 2:51

   question thanks to generative AI uh but also also just more fundamentally it's actually always been the case uh due to misinformation disinformation and and the like so um the only way that we are going to have a good future is if we um improve the quality of our information and our information hygiene you might say um we we understand where um the information we consume is coming from uh we we understand to what extent it has been vetted and and and verified we

7. 3:24

   understand who has shared it to us um all of these things um track throughout sort of the supply chain of information uh will turn out to be extremely important uh for this century and and this is the problem that we are working on and specifically building on top of NE protocol of course there's some path dependance there yeah definitely very relevant to obviously what's happening

8. 3:47

   just in general with information for you is this a data project or is this more for an AI project or is it kind of both uh well it's it's both so I I have a dual background I um I've been working in allish representation since 2006 so that's um almost um 20 years and um I did my first startup uh in 2010 in uh knowledge crafts which is very much the same technology that we are building on on now uh so so that was a bit ahead of

9. 4:17

   its time uh so to speak um I think we go ahead of the market in five six years Amazon came out Amazon web service came out with something similar to what we were doing in 2016 and uh we were a little bit ahead of the market term but so so I have this I have this background um and and now this background has become let's say pressing given just the need uh to first of all um break open the data silos so that our AI assistants

10. 4:47

    or AI agents can get that information in a reasonable way I mean consider on your on your phone you have most likely you have if if you're like others you have more than 200 data silos right they they don't really um talk to each other with some notable exceptions like Google and apple Health Data perhaps um workspace data but be beyond that those are siloed well and good um and how how is an AI assistant going to be able to to help you effectively when your personal critical

11. 5:20

    data is is is being siloed and held hostage in this way so that's that's one aspect of the problem that needs to be solved um and and I'm happy to talk about why crafts are particularly good way to solve that um and and the the other aspect is that um well I I call it the garbage in garbage out problem right so as we as we train um AI as we put um AI agents to work um it starts to matter what is the data that they consumed what

12. 5:49

    that they were trained on MH um and if that wasn't high quality well then it's it's not that surprising that you have failure modes where the the output also is is garbage because it Go fed garbage in right right so for something like this is Asma protocol is really the infrastructure though for this you then obviously need to have other stakeholders come in and provide this data or are you doing like your own

13. 6:16

    verific like how is that process working with the actual data that's coming through yeah uh yeah so we we we are not really in the data publishing business uh but of course for the for the sort of initial seating of the protocol uh we we have taken a curatorial role as well um necessarily uh but uh there's a there's any any number of useful data sets that um initially will make sense to provide

14. 6:41

    on top of the protocol for example even as we speak um we are uploading um open street maps the the largest um open mapping data set in the world uh onto the near testet and U that will be a generally useful um data set to to draw up yeah but uh but we would would certainly anticipate that um we'd be although we might see such a data set um we'd be handing over maintainership of those data sets to to the parties actually uh creating the data sets once once the protocol gains enough

15. 7:15

    prominance that the um the team over there would uh be interested of course so for this protocol and I just want to clarify for people maybe unsure of it this is not like IP verification this is really more about data Integrity like what you are seeing is accurate information it is not about like necessarily partially who owns it but it's not really just focused on simply the ownership or is it both uh it's it's it's both so I I think

16. 7:48

    that the future is inevitably one where as I mentioned data needs to travel with a bunch of metadata about it meaning uh in particular um where did it come from who produced it um which which um chain of events brought it to you um it it matters if uh information came to you from your body uh versus you know some some toxic Reddit comment right right and um or some random download uh so it means that uh generally speaking we going to to want to have this this um

17. 8:20

    record proof of origin attached to any information we consume and that's certainly something we are enabling through the protocol um but the other aspect of it is is what I what I mentioned about the data silos that right now we have a world where um all data is is pretty much siloed and this this is first of all unnecessary because the ways to counteract this have been known for decades um but there maybe hasn't been a pressing enough need outside of something like an Enterprise

18. 8:51

    context or or perhaps a military context or something like that sure uh to to to unify it uh but but also it just not going going to to be able to build uh the future that uh so many imagine in case we keep this data silos and and don't don't break them open so so this is where the um this where the data model comes in that um if you if you want to be able to to map um any data you want to able to intersect any data any data sets you need a universal data

19. 9:21

    model and uh knowledge crafts are a universal data model meaning that they subsume all other data models you can represent anything knowledge grafts right whether it's financial data you know a CSV an Excel whether it's a relational table anything that it might be you can represent as a Knowledge Graph and then you can bring these things together and you get one large graph to rule them all uh this is the this is the kind of let's say Enterprise data integration book that um I did um a

20. 9:52

    lot in the previous decades uh for for large companies meaning um in in those companies you don't only have two data silos you may have a thousand plus right information systems and and they are generally not really talking to each other and data integration is a is a huge consultancy business because there there's so much of it to do and and all of all of that work is is fundamentally um in in a sense useless as in if only we architected systems in a better way

21. 10:22

    um with regards to um how we utilize information then uh that work that the cons is doing over there would not really even be needed but for the moment we are not in that mod sure so I I I think uh I feel like I fairly have a good understanding of why you would choose near in blockchain to do this but can you me just talk for the audience a little bit more about like why not just why blockchain because I think the decentralization piece that all makes sense like the data has to be open and

22. 10:54

    available but also then like why something like near well um so honestly the answer is a lot to do with path dependence right so um we we have u 10 10 plus X near people uh not only from near in Pagoda but from the foundation and and several other ecosystem entities and U so we we all know near we we love near and we help to build a lot of the underly tech and in in my case of course um I I was hired in 2012 2020 to lead evm development at

23. 11:33

    near Incorporated and um we sped out into aora a little bit later uh once we once we had uh uh the the bridge and evm teams um come together and and and form our Labs so um we we we've been around the block um we we know near extremely well including all of its um dirty Dark Secrets um so for example it's not always been the easiest thing to run um you know a lot of near core nodes but um

24. 12:06

    we we have people who have the experience in in running those by the hundreds and um yeah that's that's the fundamental reason but but of course there's also a very good match so both in terms of the the let's say the new positioning um of the chain as the chain for AI well that's that's a good positioning um and and I'm I'm happy that finally some clarity has been achieved on the positioning that's a

25. 12:30

    very short statement right the chain for AI I I like that yeah um and probably good idea to to stick to that um and uh on on the other hand also the technical capabilities so so things like chain abstraction the account model uh storage staking there there's so many um so many good technical capabilities that enable uh us to build this on near in a in a reasonable way that said we we are certainly not um uh we are certainly not going to be limited to to only near uh

26. 13:02

    this is just we are leveraging uh what we know to and and we we are looking at deployments elsewhere as well including ICP and and so is that the the the reasoning for going to other chains is it more or less to uh hedge your data so that it's now in other places or is it a technological choice to uh leverage other chains for different reasons well if you if you think about it uh so I'm a I'm a database guy because of the first startup I mentioned

27. 13:33

    where we we wrote a um distributed database from scratch um so if you think about what a blockchain is it's it's fundamentally an app and only immutable transaction log so that is coincidentally the the same thing that underlies any database any any modern database anyway and and what that means is that um as a coordination Point as a name space the blockchain is tremendously useful right so of course you could do um all kinds of uh things without a blockchain but uh in in this

28. 14:05

    particular scenario that we are building um the blockchain is not a solution um in search of a problem this is actually a very very good fit uh to to a blockchain that makes sense um I want to dive a little bit deeper into some of the technology here but I want to and I want to bridge the gap here from the blockchain piece to the AI piece so in some of the documentations Stu and uh press reles you guys have put up before you talk about combining the knowledge

29. 14:36

    graphs with llms for this setup how is this different than like a a normal data pipeline in structuring AI data U so on on this point um I have some uh slides um if you if you wouldn't mind yeah yeah let's let's dive into it for sure uh yeah so I I went to nine research conferences last year here around the world including uh Japan and um presented on on exactly uh this this question uh so um this is an adapted set of slides from our presentation at um um a a European knowledge engineering

30. 15:16

    conference in December awesome um so basically what what we are pursuing what we've been pursuing in Asimo protocol and and um um prior to it for the last several years is something called the neuros symbolic synthesis and as you said this is exactly the the combination of um large language models with knowledge craft so building hybrid AI systems and I'll I'll note that um in in

31. 15:40

    sort of previous um AI um Summers because just like crypto AI has had uh Cycles just you know longer over a period of decades instead of just a few years um in in previous Summers um it has always been the case that you've had um people get excited about something single technology so for example right now uh Transformers based um neural networks uh but then um the the expectations have somewhat uh um

32. 16:09

    exceeded the reality of what uh how how reliable these systems can be in production absolutely yeah so so then um in practice that has led to always hybrid systems being the outcome of of that kind of Boom periods and I think no less uh here so um there's multiple in in any any field of thought there's multiple schools of thought and and different Traditions often also geographically segregated um in in in this case I'm going to talk about just uh two uh so I

33. 16:39

    come myself from um the who are known as the symbolists or the knowledge Engineers uh so also known as the semantic web folk if you if you heard about that in previous years um generally symbolic AI um and for example Ilia pukin comes from the other tribe which is currently of course the dominant tribe neural neural network people machine Learners connectionists uh so this is sort of the

34. 17:04

    uh the story of these two two tribes condensed into a a very very short uh deck so this of course what I'm showing is a is a Knowledge Graph and U uh we have um uh we have a great name for the protocol Asimov based on the of course the science fiction author Isaac Asimov one of the three greats from the last century and so to to give a nice example here of a knowledge graph we mapped out a little bit of the uh Apple TV Foundation uh TV series which is based on asimus books somewhat Loosely based

35. 17:35

    but in any case um and that represents one of the one of these tribes the other one of course is the familiar one uh where you can for example um categorize um cats and and and dog pictures and and find out what they are classifiers um I like to tell this story in terms of uh the personas involved um so I think many many of the many of our Watchers today will recognize the gentleman on the right that's of course Yan Lun Professor Yan Lun who is heading

36. 18:05

    up the um AI fors at meta which means he's the guy ultimately responsible for the open source llama goodies and and all of these things uh so one of the Godfathers of deep learning as as is known uh on the on the left we have um no less a pioneer of AI but much much less known now um he's douglin nut who sadly passed away just about one and a half years ago uh so uh Doug here represents the symbolic AI side of

37. 18:35

    things and I've chosen Yan to represent the neural network connectionists um this story um is recounted um on some level in in the great book by Pedro Domingos called the master algorithm which talks about the five tribes of uh AI uh so I would recommend that I I've given away literal dozens of copies of this book uh so first on the Doug lard side um he is um um he he he was one of the

38. 19:08

    major Pioneers in symbolic AI um there's absolute path dependence to to to what I've been doing for the last uh 20 or so years to to his work in in previous decades and uh he um headed up a project called psych cyc uh by score the company uh founded in 1984 as as I've highlighted for the info boox there um psych was the most ambitious project in AI perhaps prior to um current times such as building the gr supercomputer and all of that uh the most ambitious

39. 19:44

    project in Ai and also considered its biggest failure uh so psych dealt with um the specific problem of how do re IMB computers with common sense and you can find for example D giving talker shortly before his death um so if you just think about this problem um it's it's actually a surprisingly difficult problem um for for example this problem stat statement over here this seems like a tremendously trivial thing but um when you when you actually think about what

40. 20:18

    would be involved in being being able to solve this problem um it's not enough to for example pass the the sentence into a syntax Tre and and label its parts as nouns and and verbs and and the like no you actually need a fair bit of knowledge about the world so Tom was mad at Joe because he stole his lunch so this is a a credit assignment problem meaning who is he and who is his and

41. 20:42

    when you flip the meaning uh by flipping because to so again who is he and who is his um so if you think about what you actually need to answer this U problem well you need to know for example about uh shed refrigerators in workplaces you need to know that uh people eat lunch you need to know that people might make lunch at home and then bring it to work uh you need to know about property

42. 21:06

    rights perhaps what is stealing why would somebody get upset if somebody took their lunch and and so on and so forth so there's a there's a hell of a lot of information about the world that you need to answer this question properly and that means that uh you know such such problems were quite sticky for a long time uh and uh Psych is one of the um Endeavors at attempts to to actually answer uh this problem the common sense problem um you might have heard about expert systems in the 1980s

43. 21:36

    these were you know part of the AI boom back then uh so expert systems um are considered a failure because of the Common Sense problem meaning that expert systems worked relatively well in in narrow verticals like medicine so whenever youve gone to the doctor they probably use something like a expert system for the last decades um they input some of your symptoms they get some diagnosis and out of that they can uh give you some treatment options but that's a clear why it would work well in

44. 22:08

    medicine but it it doesn't work so great for example in in law which is very open-ended and ambiguous and you need to understand a lot about the world in the common sense way that I just mentioned so so what this means is that um generally these are considered failures from from that previous AI boom uh so uh as as a result uh the the psych project which ran for 40 years from 1984 until

45. 22:32

    CA the death of dlut its founder um is is used as an example of the biggest failure in AI in Prior times um but perhaps perhaps it just means that the uh the scope was actually uh too ambitious it it wasn't really um something that humans it wasn't the project humans will ever complete so what didy Tred to doy tried to encode all common sense knowledge in the world into symbolic form in the computer and

46. 23:02

    uh this this seemed like it would solve the common sense problem but it was too big of a task um so then that that brings us to uh the current generation of AI so just a little bit before uh Doug's Doug passed away half a year before uh he will have seen chat jpt he will have um um used it and uh and understood that there was something something uh interesting now in terms of the possibilities but he will not have been um able to see it through because he

47. 23:34

    passed away just shortly after uh so if we now take any uh modern llm and we try to query things like the uh examples I gave um it's no longer a problem we we actually do get the answers we would expect at least from the larger models in the in the smaller cases maybe maybe not right and and we can even insist on getting um just the actual answer so so this problem is is solved by llm

48. 23:59

    um that brings us then to to the other side to Yan Lun um so again um Godfather of AI Godfather of group learning winner of the touring Award with Hinton and and benum in 2018 and uh he's um known for his foundational work on convolutional neural network CNS um so this was a year ago exactly a year ago actually um there was the world world government Summit in in Dubai an event that competes with uh uh Davos U with the world economic forum and they

49. 24:31

    had invited Yan Lun um and when he when he got there they plastered his face on the bush Khalifa the tallest building in the world uh so so as I understood the story um they took him out to dinner with a nice view um and um they didn't warn him beforehand they just U surprised him by blasing his face on the bush khif and as you can see he was rather pleased about it uh as as perhaps anybody would be and um um I was

50. 25:02

    thinking to myself when I saw this at the time that this must be for him a bit of um or sh I arrived kind of moment because see the thing is that while he's now a superstar uh it's easy to forget the decades prior to Superstar do when neural networks were considered a failed technology right neural networks um were not able to get investment neural network smelled bad and um still despite all of this he and and other die hards

51. 25:30

    pursued them uh so so that's um maybe good good with reference to a timeline to to think about that so as I mentioned Cycles in AI just like um in crypto but somewhat longer uh and now of course we are in the third Golden Age by this this accounting anyway and and this Golden Age begins in 2012 so those like you Joe you and me uh who've been around for a while uh we probably remember the emergence of deep learning in in 2012 2013 sure absolutely but prior to that

52. 26:04

    um there was decades of obscurity for neural networks because they um they were considered to have been have failed in the 1990s uh I still remember in 1990s for example Microsoft putting out some press release about um speech recognition based on six neurons not 6 million not six billion not 6 million not 6,000 six Mor right yeah so we comp Limited in the '90s and things didn't really work out um so but if you go a little bit further back just to kind of give the larger

53. 26:37

    perspective on this um cycles that our Industries go through uh so this is uh the New York Times 1958 July new Navy device learns by doing um and I've highlighted here from 1958 uh some of the predictions about the future uh so this is the um the perceptron um one of the first uh very very first neuron Network systems single neuron MH and um all the predictions about the future highlighted H they've all come through though it just took uh

54. 27:11

    six decades instead of uh the maybe expected five to 10 years that right right yeah people expect everything too fast yeah yeah U so um in in 2010 this problem this classification problem C doog uh this was a research project to assign PhD students to it um and in 2013 this was a homework assignment so that's in between we have the arrival of the deep learning Revolution and what was that based on well this is where this ties together into Asimo um well it was

55. 27:45

    based on two things it was based on the availability of compute uh so gpus were available and iask who later became one of the co-founders of um open AI um he uh utilized uh gpus to create something called alexnet but uh equally important was the availability of a very large labeled data set called imag net more than a million hand labeled images uh so that the network could actually be trained and in the uh annual image net

56. 28:17

    challenge um he wiped uh the floor with the competition so Alex net just obliterated everything else out there uh it's a it's a phenomenal results mhm and that got us started on the current exponential so my my head of research tells me that there's more than um 200 machine learning papers published every day now on archive and that's up from 100 plus a year ago so we're on quite an exponential there um 13 13 years in on on that one

57. 28:49

    and of course that's how a convolutional neural network works um that's Alex specifically uh quite a small Network by today's standards but uh huge at the time by 2016 this picture is from 2016 computer vision was substantially sold so even even a might could be recognized and in in that time frame I I worked on computer vision on drones so um really um very satisfying years of uh finally

58. 29:18

    getting uh good computer vision and uh being able to do a lot of stuff for sure um and and then of course by by 2017 better than human on the imet classification challenge uh so turns out imet had some bugs too in terms of the labeling um and and 2017 of course is when um the story ties into near because in 20 2017 we have uh Ilia pukin um one of the co-authors of the seminal attention is all you need paper um

59. 29:48

    exiting uh Google um not having funding for AI available ironically I would say and and pivoting to to crypto right and and now we we've seen recently kind of pivot back to the combination right yeah uh and and so everything I just showed you was still prior to 2017 prior to the arrival of Transformers which then uh improved those results even further and of course great videos available for how those work so that brings us then to um the GPT era so gpt3

60. 30:22

    and and eventually the release of Jud GP when everybody um sat up and paid attention U and a couple lessons in marketing um when when you are ahead of your time in terms of Technologies and um it it fails in some way Rebrand and keep moving that's a lesson from this history yeah yeah uh so the reason we all learned about deep learning 13 years ago is that neural networks um was not a invoked term by now neural networks have been

61. 30:55

    rehabilitated of course yeah so it's okay to talk about them again but at the time suddenly deep learning deep learning deep learning a brilliant marketing move um and on my side of the table in the other tribe um the symbolic AI tribe uh also many many rebrandings over the years uh so uh this this technology is always seem to live a little bit in the future uh but now now now it seems that

62. 31:18

    we have a bit of a golden opportunity uh to to exactly combine these two two tribes in a in a substantial way for sure and that brings me then to the conclusion here that um what uh what what I've been pursuing for for a couple of years now ever since I um exited um Aurora um is is this synthesis of llms and knowledge graphs um which is known as the neuros symbolic synthesis um and uh maybe the the thing

63. 31:49

    that I didn't explain yet is that um although it's fairly clear um why the common sense problem is an issue for the symbolic um represent of knowledge and reasoning about it automatically um it might not be clear on on why LMS need knowledge crft why they need explicit knowledge representation uh so just very briefly as I'm already uh hogging most of the time here um very briefly um so llms of course have a number of deficiencies that we all have become

64. 32:20

    familiar with over the last few years including that um hallucinations are not a not really a bug it's a feature uh and you can't really tell when it hallucinates and it can't even tell itself when it hallucinates uh similarly justifiability and introspection is is quite poor uh so if you ask J GPT why did you give me that answer well it'll it'll give you an answer about the answer but it's going to be a confabulation and um it has no particular access to its uh internals uh so um these are some of the

65. 32:52

    some of the problems that uh motivate an explicit knowledge representation uh instead of just using uh abusing that llm itself for po little llm um and um um similarly on the knowledge graph side then why do we need uh to plug Elance into it well because of this common common sense shaped hole meaning that my three-year-old son actually knows more about the world how the world Works um compared to any computer until

66. 33:23

    about 3 years ago so nice coincidence there um and um this would include exactly the kind of things I mentioned about lunch at an office but it would also mean you know why why is the arrow gravity pointing downwards why do humans live in in uh dwellings why do we organize into social groups why do we wear clothes you know multiple reasons why we wear clothes right computers um had trouble with

67. 33:47

    these kind of basic things about the world until recently um and the on the other hand the knowledge soup encoded into llms through the massive pre training um inputs uh is not particularly accessible it's not particularly introspect it's not particularly amenable to mutation and update so all of these are problems with just relying on that encyclopedic knowledge which is which is nice but um it it it is going to be a lot better of an idea to have smaller models that

68. 34:20

    don't necessarily carry so much of the whole Encyclopedia of the of the human um world in them uh other than the really Baseline stuff I mentioned as common sense and then have an explicit knowledge base that they access as one more tool in the Arsenal and That explicit knowledge base can be introspected it can be verified um it can be um mutated as it can be updated as needed um all of these things um are possible when you have a database and um

69. 34:50

    the the model then can shrink in size um and and focus on on the core propositions that it provides including you know a great understanding of human language including the common sense knowledge swop that it carries within it and and including of course um the reasoning abilities that are emerging so so this is um this is the synthesis and uh pursuing pursuing this from multiple different angles the as Asimo protocol is about um fueling the the data needed uh for for making

70. 35:23

    systems like this and why isn't everybody doing this I'm going to U briefly mention that um in Academia this is very well known as one approach it's not my invention um for a couple of years academics have have in fact pursued this approach but in Industry industry is very driven driven by Fats and and sort of um Fashions and and focuses on particular Tech we had the we had the vecto stor rag fad now we have the agents Focus uh and soon enough it will

71. 35:57

    be something else and um the larger perspective is that now that the uh Foundation models are becoming Commodities um it it means that you let's say you have a a foundation model on your phone on your laptop um when you when you need an upgrade to it you know it's going to be a software upgrade similar like anything else so overnight um multi gigabyte file will be downloaded and uh in the morning you'll have a slightly better AI right so you won't really care that you have

72. 36:30

    llama 3.1 3.2 3.4 or something like that uh you'll just care that you have a slightly better AI with a software upgrade and once you take that perspective what happens of course is that um if if the llm exits Center Stage well what goes on Center Stage is your data the thing that you want to preserve across these software upgrades is obviously your data you you don't want that to uh be tied at the hip uh to uh

73. 36:58

    the foundation models um and uh so some of the research we published uh in this space in in uh previous times um just mentioning those uh there's a whole big story around ontologies which we won't have time for and so that's that's the slides I had prepared for you Joe yeah no I I mean thank you for diving into all that uh really great history and uh made me feel old some of the stuff I

74. 37:23

    remember but um no I there's a lot of things that come to mind here and I think is really relevant to obviously what's happening now and I and I can see why timing wise it makes sense for what you're starting to build here and it fits into that so one of the things I wanted to kind of first start with is the concept of like training versus getting getting data in real time is this something where you know the data that you guys are trying to produce here

75. 37:56

    and and through the knowledge graphs and everything else is that supposed to be for real time or is that supposed to be for training or again is that like kind of depends on the situation of how it works yeah so the the protocol is generic in in terms of um this question okay I I I do anticipate that um um we'll we'll have a lot of basically pre-training kind of data sets high high

76. 38:20

    quality organic and synthetic uh data data sets U that will be utilized in in pre-training but uh this is this is not the only and and the main main use case um in in general what I mentioned earlier about the data model being Universal is really important what it means is that um when you when you want to bring all the data in your life or an organization or or or a project uh all together um by mapping it to knowledge graphs you are

77. 38:51

    then able to unify all of that data you you are able to intersect all of that data and and very importantly also um even Publishers who don't know about each other uh as long as they are publishing data sets in the same um in the same way on the ASO protocol then those data sets can be intersected arbitarily as in um see the thing thing about the graph model is that once your data is a graph which anyway matches how we think about um information on a

78. 39:22

    whiteboard right boxes and arrows that's very intuitive as well but when when it's uh graph in that form well if you have two graphs you can just merge them right so so and you can do that statically as you can do that in in the storage in the database or you can do it at query time on the Fly dynamically and and what that means is that as you execute the graph query and uh we support initially graphql and and and Sparkle are more more powerful standard

79. 39:47

    um as you execute a graph query you can bring in data from any arbitary data set into that same query and you can just uh at at query time you you merge and unify these knowledge graphs including you can um transform them into yet another kind of graph uh so these are these are powerful capabilities that don't exist uh today um in in the form that we're envisioning um and already by themselves these are these are compelling propositions but then uh combine this

80. 40:17

    with a data Marketplace so of course we we are aiming here for um incentivizing uh knowledge creation especially high quality uh very verified knowledge creation um when you combine it with the data Marketplace and you combine it exactly with the utility of this uh high quality data for for agents and and other AI kind of use cases whether it's zero shot use or whether it's pre-training um this is where the the um Network effects really start to kick in

81. 40:48

    uh so um let me let me give you just one one example uh here um here here is a one on tweet I I'll P you also the link on uh on the utilization of knowledge crafts uh in agents to to improve outcomes um and I'll mention I'll mention a kind of related uh uh use case uh so we developing of course a lot of uh demo demo use cases and uh one of the one of the sort of uh really easy ones to explain to people uh in terms of how

82. 41:21

    you can improve uh agents um with with this kind of technology so supposed that you you have um a scenario I'm going to use a really um formulation which is going to be familiar to everybody so movies right if I used books then it would only be a portion of the people but movies everybody is able to um enjoy so um let's say that we have something like um The Matrix as a typical example

83. 41:46

    we all love and love the Matrix um and we we have let's say Ken Reeves Lawrence Fishburn I think is it name and uh we we have of course the vovis as the directors uh and and suppose you just asked an llm you asked the question that which other movies if any have this um set of people this uh crew worked on together um so ANM would actually be in a in principle in a good position to

84. 42:16

    give you an answer it probably has the information in the somewhere the problem is that you won't be able to trust the answer right so you you have you have no way uh to be sure if that's a hallucination or not and and there in lies the rub there in lies the problem so uh how this would differ with an agent that's utilizing the asmo protocol well you would ask the same question

85. 42:39

    everything looks the same but when the answer comes back the answer will have um verification on it as in this this is a verified true answer and how can we verify it well for example uh the Internet Movie Database IMDb that is a high quality data source um about all the casting for all the shows and movies and and so on and uh what we will have done then is we will have turned your your prompt into a query on IMDb uh to check um if you can

86. 43:12

    find an answer for you and if you find that answer uh we know this to be true um not not objectively true we are not making any kind of claims about you know objective truth but be true within a bounded context and bound in this case is for example IMDb but could also be Wikipedia encyclopedia britanica or something like that does that make sense I got yeah no that makes a ton of sense

87. 43:36

    so are you is azimov then correct me if I'm wrong here are you helping support some of those like that connection when you say like you have a piece of data here in this one graph you have another graph over here and they don't know each other is what you're doing with um the neuros symbolic part of this AI combining that on the Fly uh yes uh so the the design of the network uh maybe maybe actually useful for me to pull up um one more slide here just for the architecture of um of this

88. 44:09

    so This is actually sharing now from um the town near near town hall uh deck but we we have in here we have in here the architecture so easier to explain this way um so uh the the architecture here is that data sets are anchored uh to the chain uh so deset domain net in this case um and uh by anchored I mean that the data itself uh could be on chain or it could be offchain because of course some some data sets are just so large that it would be kind of insane to to try to put them on chain right I'm just

89. 44:43

    too expensive so not economically feasible uh but for small data sets there's a lot of benefits for the data being on chain and Nur has a fairly fairly good um uh storage uh costs as in um we we can really put a lot of data on chain as well uh but in in either case whether it's offchain or onchain data it's fully um verified using of course cryptographic Primitives like hashing um

90. 45:08

    and um that means that when something has been uh inserted something has been uploaded into a repository then it can be in fact trusted that that data hasn't changed under you randomly U so then what happens is that the uh as as the tip of the chain moves and um data sets are uploaded data sets are changed um indexers notice those changes and they make available uh the uh current state of the repository or or they may also at

91. 45:39

    their choice make available historical states of the um repositories in question and of course indexers have a full um autonomy on which data sets they wish to index so they don't have to track all data sets they could choose for example a vertical some particular field you know financial data or something else um and uh and the indexes then are um what consumers can query so indexes provide high performance uh query processing initially using the uh

92. 46:09

    graphql language familiar to to most listeners I would imagine and and then also the far more powerful Sparkle query language from w3c that's the the leading standard for advanced graph query languages similar a little bit to SQL in syntax but far more powerful um and and that means that the indexes uh form effectively something like um um a decentralized global graph database where you can where you can query uh these data sets and of course

93. 46:40

    you pay uh for querying them um and the indexers then earn um fees in order to pay pay for their um Hardware costs and and operations but also uh the protocol earns a little small fee and um the data set ERS also earn um for the data sets that are being utilized according to criteria that they able to set makes sense um I want to save this for one second here uh so with all this data coming in where does that consensus on

94. 47:12

    the data quality come from uh so data quality is um judged by validators so that's another note type here uh the the validators um run a a great variety of um checks um and and batch processes on on these data sets um on the low end you have things like um um well this this um data set has declared that it's utilizing these schemas does it in fact conform to the schemas or or does it perhaps not conform to them so in case it conforms to the schemas well the the data set gets a gold St right and as as

95. 47:52

    you get this validation results in from a sufficient number of uh validators then then you you build a sort of consensus around that this is a um data set that passes these validations um but uh you have a whole whole range of validations uh which of course is uh ultimately you know Community contributed and uh extensible um on the high end you have uh you have things like factchecking uh so we are working we are working with uh some European universities uh where the

96. 48:21

    research teams are the leading experts in the world on the automatic factchecking of knowledge cfts and again I want to be always careful to qualify um we are not aiming to to be able to tell if something is objectively true that's way too big of our Burger we are not going to be able to deliver that however what we can certainly um uh do is uh judge if something is true within a certain bounded context so for example

97. 48:48

    is this um information true within the scope of Wikipedia and of course you know Wikipedia as of revision X at this particular date right so so this is the kind of uh work that uh Academia has been uh putting out over the last few years and this is directly applicable to the Asimo protocol use cases very interesting really really really interesting um yeah I I I I do I really I think what uh is really making

98. 49:21

    sense and I love that you told the whole story of kind of how data has evolved especially with you know Ai and everything else because you can see that this is where like everything is going because like I mean anyone who's using these you know like Chad TBT whatever Claude like you C you can see where there is that hesitation of like is this correct like where where did you get that and and if you are not and I saw

99. 49:50

    someone I forget who it was post on Twitter not long ago if like if you are not a domain expert and you're asking a question in that domain it is really hard to tell if it's being correct or not and so like as a developer I can troubleshoot the code it gives me and so that's thing but I'll tell you if you're not a developer it's SP out a lot of bad code still as good as they are yeah so that's where that domain expertise comes in so it's really interesting I can see

100. 50:20

     where this platform starts to alleviate that need to have to be a domain expert and everything that you're asking yeah isn't it ironic Joe that um um you know one one might have expected uh that for example coding assistant based on llms one might have expected that they would benefit Juniors the most but I think it's actually clearly the case that it benefits us seniors the most yes right

101. 50:43

     100% the juniors are not able to tell apart hallucinations and bad code and and so on and for us it's just trivial right discard discard redo redo right so so it means that you know I I certainly perceive that I get a a fairly phenomenal productivity Boost from this assistant yeah yeah yeah yeah no I I it's it's really interesting and I'm I'm really uh excited to see how you guys are to grow this and develop this because I can see where again any

102. 51:16

     service would want to start to have this here because again if I am a data provider I want to make sure that people recognize that my data is accurate for what I'm trying to do as well so I feel like there's the other side of that which we haven't really talked too much about though is like the data providers also want to make sure that like hey my quality data is the one that's being

103. 51:39

     used yes exactly yeah and and of course so data set Publishers need to compete in a in a market just the same as for example the indexers compete right so they all compete as a fee Market but they also compete on on quality and other metrics uh so for indexes it would be about reli ility and latency and and um uh you know connectivity uptime reputation and so on but also dat set Publishers um they they they they might have similar data sets that could be um

104. 52:09

     substituted for each other uh and then it's going to exactly matter who has the higher quality data and what the price point for index or access to it is as in online uh real-time query processing access of course if something is public on a chain then um people can download local anyway but um in terms of the convenience of This decentralized Global um graph database um that's that's where

105. 52:35

     the access fee comes in yeah no I I mean uh it definitely it definitely makes sense and I'll point out it to uh viewers that you know on this show we've had a few different other like I would say AI data providers but I think the difference here that we're seeing is it is about not just like rawad data being fed in and and and a way to monetize that this is much more about for me it seems like it is the validation that the place it's coming from is saying where

106. 53:09

     it's coming from but then also too the the knowledge graph piece to me is super interesting of being able to pull those pieces together that you know I and I submit something you submit something and now the system realizes that this is connected and now it's providing even better experience for everybody else that that part is really interesting to me as well yeah there's a there's a couple of framings I've used uh for that lat part um so in in terms of personal

107. 53:39

     AI that we will all have access to in the future um so one is that um when you break open those data silos that you currently have in your life and you unify that data and your AI has access to it well then that's something like a personal data warehouse and uh for those who know what warehouses that that probably is a compelling value proposition um the uh the other other way that I found

108. 54:05

     um useful in explaining this is that um so Joe intelligence agencies three letto agencies they've had access to this kind of technology for decades um and and just to um to mention this in passing my own introduction to this uh Technologies knowledge graphs and semantic technology as it was known at the time it came on a US Navy project in 2006 I was on project from 2006 to 2008 and um in in in that kind of context um these Technologies

109. 54:34

     were used for a long time and and indeed uh some of the um contributions came from that direction similarly you think about something like Palante uh when Peter teal founded that After exiting PayPal in the same time frame so 20 20 years ago um well for the first decade the technology basis was somewhat secretive but these days if you go on the pal website and you look on what what are they doing well what are they doing they're doing knowledge crafts and doing ontologies which is sort of um you

110. 55:02

     know high high quality scheme out of of knowledge crafts yeah and this is this is the technology basis on which they um helped the three L agencies in America and Beyond uh to to be far more effective because they had this this insights from the intersection of data sets so now in in uh this revolution that's going on um these are the kind of capabilities that I would certainly like for us as individuals to have as in I want to be an Intelligence Agency of one

111. 55:33

     yeah right um so uh Jame James Bond of one right yeah yeah yeah no it it makes sense and uh again you know if you just think back of you know think of your own life how much data is both part of your life and how much data you're trying to consume all that stuff uh I think it's it's really really very interesting here uh I this there's so much more we could go into but uh I do want to in just a couple minutes you have a people don't know you have a fascinating work history

112. 56:06

     and life history I would love to take just like five minutes or so talk a little bit about it's been in the news lately um Silk Road Ross elck um and you actually have a really interesting connection to that story and I think it'd be kind of fun to tell here uh yeah so it's it's not something that um I've spoken that much about over the years but um uh I was of course super pleased that Ross actually Ross actually got out of jail uh this this

113. 56:36

     year because ultimately uh he went to jail for two lifetimes plus 40 years for the crime of running a website um and um it turns out that um um I have a a a connection to all of that which I didn't know about until about Circa 2015 so this already after he was arrested after the Silk Road Show trial all of this um so as as the trial was wrapping up a reporter from The Daily Dot contacted me uh I think it must have been 2015 and um he asked are you perhaps the

114. 57:12

     the arto mentioned in the uh exhibit 270 in the trial and I was uh confused but I went and looked and and and oh yeah that's me um so the earliest um earliest evidence in this Sil Ro trial uh are emails between Ross and myself from 2009 uh so September 2009 and um I don't actually have access to those emails any any longer so I'm unable to add to the historical record here um the government the US government does have access to

115. 57:44

     those emails because they nabbed Ross with his laptop open so even if they were pgp encrypted uh the contents are certainly um somewhere available uh but um um I I had uh not made the connection until the reporter contacted me that U this was actually that guy uh so at the time in 2009 uh I was maintaining as a as a cipher bankk um activist I was maintaining quite an active pgp correspondence I've since then discontinued pgp use for obvious reasons

116. 58:19

     um but um uh at the time um Ross would have been just one of many many people who sought my advice as in young than me and and sought my advice sure and in this case he apparently sought the advice that um I have this idea how do I build it and I pointed him at Bitcoin I pointed him at Thor and I gave him um an important book reference called a lodging of wayfaring men U at the time published anonymously uh but eventually Paul Rosenberg another Cipher Punk um

117. 58:50

     admitted to writing that and uh this apparently uh then inspired Ross to actually proceed with the plan yeah yeah no it's super interesting and uh it just I mean to me I like this story because it just shows how small the you know cyber Punk world really has been and especially back then I mean it really it was such a small uh group of people um we're we're short on time I know there's so many other stores you can go to I know you have a connection to the old room 77 um back in Berlin uh people don't

118. 59:25

     know that term this is is like aging us but that story of that bar um accepted Bitcoin and it would became a really popular place for uh you know cyber punks and just people who I think just interested in you know like uh self- sovereignty and things like that so uh really really interesting stuff definitely and um um a lot of lot of stories to tell from those years before they fade from memory yeah ABS absolutely that's the challenge is getting older for sure uh well I arto I

119. 59:56

     think this was so exciting I I really I have I have so many more questions for next time um I definitely want to have you back especially as you guys launch to mainnet uh what is the approximate road map and timeline here for for you guys at ASM off uh so so currently uh we are not doing anything on mainnet yet um test net only and um um I expect mainnet will be certainly in the Q2 time frame

120. 1:00:21

     um but probably early early Q2 so we pretty far we're pretty far along in the tech uh we've been working on this for a while and yeah Q to awesome that's exciting and people to definitely go to the website there's some information there links to research as well um so if you're interested keep dabbing more in there uh but yeah no thank thanks again for for your time today and I look forward to seeing as my protocol grow thanks Joe pleasure any take care bye

121. 1:00:50

     bye now uh folks uh thanks so much for for joining us today this was really exciting uh so much great stuff there I really encourage you to go to the website go to their Twitter account um take some time and uh dive into the protocol a little bit more there's lot of really great stuff uh in there uh coming up just some news and notes uh we do have a bonus episode tomorrow uh that

122. 1:01:16

     we're going to dive into uh so we have the proximity team uh jumping in we're going to talk a little bit more about some of the stuff that's going on over there related to some of the projects they're working on on we have a couple episodes next week up top of the normal Dev Hub live related to the hackathon uh so look for the announcements around that as well uh if you have not checked out the hackathon all the links have been scrolling down below uh definitely check that out uh the AI agent hackathon

123. 1:01:41

     is really going strong and definitely worth checking out uh but until then uh I'm your host Joe and we will see you the same time next week and I hope to see you also uh tomorrow for our uh bonus episode with proximity thanks all bye