NEAR AI Office Hours #37 ASIMOV - Building Trustworthy AI with Knowledge Graphs

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

1. 0:00

   your own AI explore hi everyone uh this is episode number next of uh a near office hours uh today we have arto from azimo protocol many people in your EOS system have known arto since his work at Aurora arto do you want to say a couple of words about yourself and uh a quick introduction into what aim protocol is doing uh sure so Alex nice to see you it's been a few years and um uh yes indeed so I started at near in 2020 to lead the evm development with sper off as Aurora and after Aurora

2. 0:40

   uh I've been working on trustworthy AI which of course is very much the same thing as the user owned AI vision of near Ai and Asimo protocol is is the latest spin-off from the last couple of years of um work I've been doing uh we we want to specifically focus with Asimo protocol on the problem of uh high quality trustworthy data uh and specifically structured data um which is um I perceive it as an um

3. 1:10

   underserved area of the market and also my own background is in knowledge representation and reasoning so I started in krnr in 2006 so 19 years now nice so I'm I'm also very interested in uh trustworthy AI I never I never thought of it From perspective of using knowledge graphs so that's an interesting that's an interesting perspective I was looking a lot into formal verification uh so which I guess more of a complimentary than

4. 1:41

   Computing technology in this case uh and U yeah so I guess there are several uh things we can start from one one is U why why knowledge graphs right and uh I guess uh with that we can also uh can we give like a a short overview of how you can so so so the the the whole point is is for the knowledge graph to be trustworthy right so I remember like in the past we had some uh there was a I wouldn't remember

5. 2:11

   the name but there was the open- source version of the knowledge of the Google Knowledge Graph right and uh it had a lot of uh there was a a free base if you remember there was a free base yes there was Google Google Google acquired meta web in I think it was 2010 or something and that's actually um why we now call these things knowledge cows because prior to that that wasn't really a a

6. 2:34

   popular term when I got started in 2006 um we called this semantic technology as in part of the semantic web vision and and all of this yeah right but but from my understanding is that part of the asmf protocol is ensuring that the the knowledge graph is correct right uh you know with some definition of correct so so I would be very interested to hear how like how you approach in general and uh you know if you have any metrics or uh benchmarks to verify how like how how

7. 3:07

   what the current Knowledge Graph is uh yes so maybe maybe useful to introduce to people what a knowledge graph is first because uh you know I'm sure but um I I would suspect most of the audience hasn't yet heard this term so um usually uh if you have a whiteboard uh you would be drawing you would be modeling the world in terms of boxes and arrows right so Alex is a box arto is a box or or perhaps a circle

8. 3:33

   depending on style and then we draw some kind of relationship in between them so for example Alex knows Aro right so that's a Knowledge Graph it's it's as simple as that so it's it's super intuitive um it it absolutely fits how you would just um from first principles or or perhaps you know human intuition Draw Something on the Whiteboard and as it happens um it's also a really good

9. 3:55

   way to to just model the world I would I would argue that it's the Universal data model uh all other data models are subsumed by knowledge crafts meaning that anything else can be mapped into a knowledge grafts so whether you have an Excel uh whether you have a relational table whether you have a JPEG EG whatever um Json uh XML everything can be mapped onto a Knowledge Graph so it's

10. 4:20

    sort of the the one graph to to unite and bind them and then of course um once you bring all of your data into this um one graph it means that everything is in interrelation to everything else in in some way even if not directly and that means that you can achieve extremely powerful results with graph queries you can Traverse the graph you can transform the graph uh you can um once you once

11. 4:46

    you have everything mapped to to this data model you can intersect data set at will there's no problem about merging things and this is one of the uh unique propositions of knowledge graphs so U Know gra have been used um in the past decades especially in Enterprise data integration so for example the whole business of Palante is built on all graphs they didn't speak about that much

12. 5:11

    in the I guess first decade of the existence but um these days you go to the Palante website it's all laid out there it's knowledge grafts ontologies everything is there and in fact I point even some of our own recruits at their documentation because it's you know fairly comprehensive and um banteer of has been um building these knowledge craft systems for um especially the three

13. 5:35

    letter agencies in America um and I think they had Once Upon a Time on slogan like um you know actionable insights from bringing all of this data together so very much that is what um we want to help achieve for individuals for all of us so you could consider it from the Palante angle something like um your personal intelligence agency in your pocket will be possible um or if you consider it from the Enterprise

14. 6:04

    angle are familiar with those terms then it would be something like your personal data warehouse that your AI is accessing your trustworthy personal AI is accessing so I think this this is a really important future to build right so so how do we how do we ensure that the knowledge graph is uh uh you know uh at least resemble some correct data yeah so U just like any any any sort of data data system um these these are not really um we don't have truth in the

15. 6:38

    system right epistemically that would be impossible it's very difficult to to to Really nail down objective truth even in even in science we can really the best we can do is falsifiability right uh we can't be sure that our model of the world is correct so we we don't aim for that uh what what the knowledge craft consists of um assertions of facts uh so U Aristotle defined knowledge as Justified belief which uh has been

16. 7:05

    slightly iterated upon since then but uh it's it's still a pretty good definition uh so um a knowledge craft then carries within it um not only the data itself but typically also data about data metadata where did this data come from that's Provence uh why should we believe believe it to be true and and so on and so on so so this means that um you have U you have some basis for why you might might believe that something is true uh in in that knowledge craft but they are fundamentally it consists of asserted

17. 7:39

    facts and and um um if you want to do fact checking of a knowledge cph which definitely we are pursuing uh then that can't be done again epistemically you you you're not aiming at objective truth instead you're aiming at factchecking of this knowledge within some bounded context and a bounded context could be for example Wikipedia certainly but I mean Wikipedia is variable quality right

18. 8:06

    so it could be the encyclopedia britanica um or it could be some other data set that's relevant to to that particular uh thing that you want to validate so we we have a collaboration with some of the leading researchers in the world in in this um focus uh that's the dice research group at paror University in Germany so very interesting research from there you can can check that out

19. 8:30

    mhm so so in in some of the uh of the material uh online about the protocol uh there's a diagram which has you know what so so some people propose new data to be included in the knowledge graph right and then there's a validation step so so you mentioned right now that there are some ways um uh like you know that the data has um information about Provence and whatnot but at the end of the day it it you know is it fully automated is it primarily manual is it like

20. 9:03

    semi-automated what is the exact process of uh insuring that that metadata is actually valid well so very importantly it's open-ended right uh meaning that it's extensible it's um community contributable and and it can improve over time but fundamentally you want automatic validations and you also want human validations right and and all of this can be folded into basically annotations about um the existing knowledge the existing data so so metadata right uh I I believe that in

21. 9:35

    the future it would be rather sensible that most of the data that we consume uh and reaches us would travel with a lot of um cryptographically verified metadata about it and in particular of course we really want to know where information comes from we want to know that um at the very least we want to know that you know this particular cryptographic uh keer is the origin of this claim about the world and then it has been forwarded and and and traveled you know kind of in a peer-to-peer fashion perhaps but in in the end we can

22. 10:07

    we can still see that it came from you know that particular public ke interesting so um I I guess also an interesting I don't know if you guys looking into that but but part of the uh part of the problem here is that in many cases you don't want the data to be associated with you specifically but you do wanted to carry some of your uh expertise right so like you know if I'm a doctor and I want to speak up about you know the importance of covid vaccines but the current you know the current government is conservative and

23. 10:41

    that's not a popular opinion I I want my credential of being a doctor to be attached to the information but not my name right uh so it's that can of I don't if you're looking into that we absolutely are so the let's say I would fold that under the info Finance info uh Finance banner and and there's many ways to achieve that one one would be of course some kind of a blinded uh scheme um perhaps zeron knowledge based uh another one would be uh actually forget

24. 11:13

    cryptography and just allow um allow a a allow somebody to put their money where their mouth is meaning that let's say I'm a whistleblower uh I on that Unown account that N I don't have any reputations so nobody's going to listen to me however um if I have the means I could post something like a like a bond on a on a claim about the world right and this would be a little bit like a either mini prediction Market meaning

25. 11:43

    that uh if somebody can um show um that within the scope of the protocol that this is a false claim about the world then that money is forer and perhaps distributed to to to the people who who show that but it could also just be basically a a a kind of short that um this money this money is locked up forever it's just there to emphasize that this is really important right uh both both models make make sense in in different context yeah the letter just happened recently right with

26. 12:16

    the with someone sending 500 ether into the void yes nice yeah yeah so so yes even ad hoc these kind of things can be done right it makes sense so formalizing that as a protocol um these These are directions that we we we are working on but of course first um first we need to build uh the uh the basic infrastructure here uh the the data market and and um all the publishing and and consuming

27. 12:48

    mechanisms yeah so so from technological perspective I I guess there are a couple things that interesting here one is uh where is the data stored right because presumably those data sets would be quite right and ideally you want them not to disappear over time especially the valuable ones right so so I guess also some some some concept of value needs to be attached here to the data set right

28. 13:13

    because like you know if someone uploads one PAB of data yeah yeah yeah yeah exactly now on this front um I'd say that in the original semantic web vision from Tim bernly the inventor of the web uh so he he was never quite satisfied with the web and he envisioned something more and and in fact I would encourage everybody excited by agents to go and read um his his piece in Scientific American from 2001 uh where

29. 13:43

    he envisioned both the semantic web and and of course the agentic web you might call it now uh so this was uh Tim bernes Lee oraa and Jim hendler sort of the foundational document for um you know the what I began working on as as well in 2006 and and what has inspired thousands of others in the field um so um in in terms of um sorry I think I going on this aggression I slightly lost the thread there can you remind me of your specific

30. 14:14

    question so so my question was how do we not lose the important data right like if I take a dependency as a user on the data set I would like the data set not to disappear I guess I can download it exactly exactly so so this uh first of all we can we can think about how not to do it uh so we have kind of seen in the last 20 years how not to do it and and that is in fact storing on the web is a

31. 14:38

    fairly poor idea just empirically most of these data sets that used to be part of for example that um linked open data Cloud that I showed in a previous presentation um they have disappeared over time they they have four or Force now and uh therefore we I think we we have sufficient evidence that that's not the way to do it right um so I actually think that um this is where the where the connection to

32. 15:04

    blockchains really makes sense so often uh blockchain everything is criticized about being a solution in search for a problem but in this case it really isn't right in this case I'm a database guy I I wrote a distributed graph database in 150,000 lines of C++ in a previous startup I did uh in 2010 I did my first Knowledge Graph startup so um if you think about the fundamental of the

33. 15:28

    blockchain well is a happened only IM mutable transaction log so that is also the basis of modern databases so it it makes a lot of sense to treat the blockchain as though it is the I call it the anchor the anchor for these data sets right over time and it means that as the data sets um uh mutate over time um you still have the possibility to go back to historical revisions by just going to previous block Heights previous

34. 15:56

    block hashes right um so uh in in asmo protocol indexers can um uh make available the current tip the current state material state of a data set or or they can at their choice also you know support the historical revisions and as for where the data actually is stored uh we have h two two methods we we do support onchain data but obviously onchain data is um only sensible uh economically for fairly small data sets

35. 16:27

    um mind you uh near product call is is one of the uh better chains in this Regard in in terms of the storage costs that um uh I'd say that even even mediumsized data sets um might be reasonable on on your protocol depending on depending on the Publishers choices and and means uh and what it will be utilized for so in principle onchain data could be utilized from Smart contracts for example and that gives additional utility to that data being on chain but we also support of course offchain data there there's really no

36. 16:58

    choice about that as you said the data sets could run into the terabytes or petabytes um and and as a result uh it must be supported of the chain because nobody's going to put put that on the Chain not anytime soon and um um in this way if you have offchain data then effectively you have data dumps somewhere that the indexes will um pull and load but they of course cryptographically anchored to the chain so that we can be sure that they they will not change underneath us over time

37. 17:29

    but also um you don't want just a single URL of where these dumps are accessible and stored there's a there's a mirror list right uh and the the mirror list for that um data set dump U is is just a list of URLs and the URLs could be anything uh they can certainly be in an S3 bucket they can certainly be on ipfs they can be on RV Iris or or any other protocol um so so long as at least one

38. 17:59

    of them remains accessible over time uh which of course those protocols could also work hard to guarantee for example in the case of ipfs um then we're good right we we able to pull the dump we able to validate that it hasn't changed and the indexer is able to well index and materialize that into um a current state suitable for query processing y yeah interesting so the um the the problem I guess here is that

39. 18:32

    if the data set is massive then the the replication cost is relatively high right and uh it's hard to um yeah there's definitely quite a few interesting things to consider here starting with the it's hard to verify that those links are actually live right because like you know if it's an S3 bucket the the only real way to verify that it's correct is to fully and load the data set right which could be you

40. 18:57

    know 15 terab say yeah mind you there's a there's an alternative there so so one of the things we are we are encouraging um in in talks with um Iris um 384 which is a stealth startup in the space as well uh from an next goog actually and um uh a number of others um we are encouraging uh them to provide uh Primitives where there would be at least a little bit of computation possible on on their side right directly on on the store storage

41. 19:29

    right and especially hashing as long as we can get yeah if if you can get the hash of of that remote object on a protocol level without having to pull the whole object well that certainly is a is a good idea right yeah yep yeah yeah there there's there's also a possibility that two of the URLs both s the file but from the same underlying storage right so this way there's a there's a sense of redundancy um right might not exist but yeah I guess it's a it's one of the

42. 20:01

    interesting questions which is not which is not specific to Asim of protocol right as data becomes more important figuring out how to store it the the largest dat sets we are currently the largest data we're currently looking at um um in terms of making available in the initial seing of the protocol uh so we are certainly making available open street maps that's about 400 gbt in in the data dumps um

43. 20:27

    and and we have collaborations in the bioinformatics space uh I I worked with bioinformaticians for I guess 10 plus years now and those are some of the very very largest knowledge crafts in existence so data sets like unipro for example or Pub cam I another principles for those uh projects and um these are just gargantuan huge uh data sets um so making this available um for especially the scientific Community uh is is is a goal

44. 20:59

    here yeah great then I guess the next uh again on on the infrastructure side uh there's quite a bit of uh you mentioned you built a big graph database in the past right so this uh uh this product also requires some sort of graph wearing right so so is it uh are you using one like node for J or something similar or is it again uh grassroot uh well so so so in principle uh indexers in the protocol could be using anything that they choose right they could be using any software stack

45. 21:37

    they choose but of course we'll make available a reference implementation in coming months uh the reference implementation is is based on um on a graph database called clever uh from the um from a university in in uh Germany and um this is a very very capable um cph database but over time uh this this implementation could of course evolve or or even be changed for something else um there there's basically the um the graph database World divides into two tribes

46. 22:10

    um there's the ne forj tribe that you mentioned this is usually known as the LPG tribe labeled property graphs tribe and then there's the rdf tribe which I I come from uh so this is the resource description framework uh otherwise semantic technology semantic web uh coming from the w3st standards uh in this space for the last 20 years starting with rdf and uh this this um uh databases um slightly different for slightly different use cases so while you could store um the data from Asim protocol in something like n forj um it

47. 22:45

    wouldn't really have the rich semantics that um that we um think are important to make possible the automatic merging of data sets so this is the this is why this was originally called semantic technology well original of course this is symbolic AI you know way back in the day but um um the the semantic part is very important if you want to merge data right because if you think about um

48. 23:09

    let's say all the applications that you built over yourself uh even among the applications you built yourself you're going to have used very likely in a career spanning decades you're going to have used slightly different semantics for example for name fields or you know first name last name maybe some sometimes middle name right so unless unless there's agreement on even basic things like that which is the realm of

49. 23:32

    semantics it's going to be harder to automatically merge data sets right but when when the semantics have been nailed down on all of these properties then you can just merge it and and moreover you can merge it um dynamically on the fly right so in in query processing on the asmo protocol one of the unique value props here is that you can your query can intersect any data sets on the protocol including the full closure of the whole protocol right and this is

50. 24:05

    this is phenomenal because normally we access data sets you know specifically or data sources right through some API or something and and so one of the uh one of the most important things in all everything we're doing is that um as we build a decentralized data market which could also be considered um a sort of trustless global graph database of all the worlds in information um you can just uh intersect any of these at query time uh and as long as they have well

51. 24:35

    defined semantics which is where ontology is come in um it all checks out yep um so also in some of the demos you were showing uh in the past uh the the DAT is fed into LMS right for uh for processing so is it like is there any science there is it just like a tool use or or do you guys have some uh more sophisticated integration with with L LMS yeah so there's there's multiple

52. 25:09

    phases to this uh so so right now um we are mostly utilizing existing uh open models and and of course especially the Llama models have been a focus in in the last half a year um but uh there's also a longer term Direction so Academia has been uh pursuing the So-Cal neuros symbolic synthesis for the past several years uh neuro for neural networks and symbolic for the symbolic knowledge representation and there there is a lot of research that marries these things

53. 25:42

    deeper than than for example prompt engineering as in deeper than the the Z zero shot approach and um I think there's a lot of promise there which I can describe separately but of course in terms of the current available technology Beyond Academia yeah um we are certainly building it on top of the open models which means um it's ultimately prompt engineering right yeah so for example we've explored in our research papers we've explored also uh fine tuning but uh the the gains

54. 26:14

    over over the zero short approach are not that uh significant mostly with fine tuning you change the interaction patterns not so much for example feeded new data but but the interaction with the with the knowledge graph is similar to interaction with the right like at some point model just chooses to uh to query the uh uh the protocol or or like at which point does it connect to the knowledge graph uh sure so so there there's many many ways this can be

55. 26:44

    achieved where we began from for in terms of the trustworthy personal AI we began from knowledge capture uh this meant that uh currently currently people are basically storing the whole conversation history that you have with JP or whatever ever and uh I don't think that's a very useful or good or scalable approach because you can't fit all of that into the into the context window anyway so so we we extract Knowledge

56. 27:10

    from the interactions with the user my my kind of go-to example is that um because it it actually kind of happened um let's say that it's your wife's birthday um you you want to um tell your personal AI that hey order some roses for tomorrow right tomorrow is the birthday um and uh at that point we can capture from this super tiny interaction we can capture user has a wife we don't

57. 27:34

    know her name we'll fill it in later she's a blank node in in uh norcraft terms um we um we know that her birthday is that particular calendar date we don't necessarily know the year but let's say that you you say you know order 20 roses 30 roses 40 roses whatever we could of course uh you know um infer and and perhaps prompt the user to confirm the birth here from that uh and and we know that she prefers flowers and specifically roses and through the inter Flora integration we can achieve

58. 28:04

    the the order right so so that's an example of knowledge capture so of course all of that will be in the in the conversation history as well but the conversation history is like a chat log it's it's not high quality it's not useful uh you want to actually extract the the useful things from it and then basically archive the the chat log right does that make sense yeah interestingly enough uh from from ordering 4 Roses you

59. 28:29

    can immediately infer the person is not in in in a Russian speaking country because in none of those even number of roses is is an appropriate gift aha yeah than that's that's part of the culture that the number of flowers must be odd okay yeah interest interesting cultural knowledge that might go into specific uh things like this um but so that's one one side of the utilization of knowledge um or in this case the extraction and capture of knowledge um

60. 29:01

    and we published papers on that now the other side though of course is the utilization of knowledge and and here there's there's many many approaches but the the obvious one of course is that um like you said replacing rag with um with a more fine grained approach right so so rag remember rag is often called semantic s similarity but um it's it's not really uh semantic it's just

61. 29:25

    similarity uh so it means that sometimes um you get a rather irrelevant snippet and and also of course often the the snippet that you want to back into the context is actually like mostly irrelevant it's just that you know it's it's in paragraph form so what you're going to do you know you you you kind of pack that that context term so with with the knowledge graph based rag uh so you

62. 29:49

    know of course Microsoft research published graph Rag and and many others have uh pursued this as well uh we get a very very fine grained approach meaning that um ultimately it's may be useful to mention knowledge crafts decimate into something called uh triples so every every Knowledge Graph in the computer is not represented as a graph typically um instead uh it's actually a set of statements or facts so asserted facts and this consists of subject predicate

63. 30:19

    and object um and and and sometimes some more metadata about it but fundamentally those three things and and what it means is that uh um you get a extremely fine Grange ability to pack the context um because you you're ultimately dealing with individual statements individual facts instead of paragraphs of perhaps irrelevant text so you can fit more into the context window in this in this way and and obviously um you have multiple ways to do that uh you could you could

64. 30:52

    render it into natural language uh which is you know fine for a proof of concept or something uh or or you can make sure that the model understands a more compact syntax and you can just fit fit more and and with better specificity yeah yeah so it's like a I mentioned earlier today right the many of them support the concept of tool use where you can specify very specific uh uh format yeah there actually there's actually one one more one more thing

65. 31:22

    that I want to mention there is that um in in conversations with investors over the last few few weeks um it's clear that there's one particular use case that we probably ought to focus on in in the in the beginning and and this is really uh enhancing uh agentic Ai and AI assistance in terms of verified answers right so um my my go-to example I gave on stage uh last week in um at Dena uh

66. 31:53

    suppose that you're a you're a movie buff everybody likes movies and and and The Matrix especially so suppose that you you ask um an agent an assistant a question similar structurally similar to um you know in The Matrix you had Theos directing you had K Reeves you had lawence Fishburn uh which other movies did the same crew work on right so in in principle llms are well position to answer because they've read the web they

67. 32:24

    they have this information in their somewhere but uh the problem is whatever they answer you Alex you you won't be able to trust it right you you'll need you'll feel like you need to double check it and how do you double check it well you go to IMDb right the Internet Movie Database or perhaps you know somebody goes to Wikipedia or the like so this kind of a use case um is is is

68. 32:49

    something where we can provide value and and it it basically then transforms into tool use meaning that as as you identify that this this kind of a question could be answered from verified knowledge then um that activates the tool use of going out on the Asimo protocol to shop for an answer in in a sense literally right uh meaning that this this question gets turned into a graph query and that gra graph query then accesses those highquality um verified and vetted data

69. 33:21

    sources that provide a high quality answer and in the user interface that's rendered similar to perplexity with with far higher quality as in some kind of a check mark that this is a verified answer based on verified data and and citations these data sour were used right so so this this seems like um like a use case that we'll we probably focus on initially quite a quite a lot yeah yeah yeah I think I'm not a big movie person right but you were showing again

70. 33:53

    watching one of the videos you were showing an example with books right and uh uh the the particular query you were using Greg agen right and if I see someone uh mentioning aim and Greg Igan in in two consecutive sentences I immediately know that I want to read books that person read right so like an interesting use case here would be me exporting your good reads temporarily right and saying hey uh what are the books that that person ranked high on their good reads which I haven't read

71. 34:24

    yet right please give me those right so that I think is the so that's an example of a really nice graph query that can be turned into a really nice graph query that um uh basically accesses my published um uh good reads your published good reads bookshelves and yeah it's a fairly simple thing to answer as a graph query yeah well it's even a little more challenging right because I don't have

72. 34:49

    good reads but I do have a text file that has all the books I've read right so it's a which is obviously uh I presume it should be exportable yes absolutely just importable in yeah just absolutely just needs to be then handled a little differently because then then the identifiers don't exist but but of course it can be answered from free form text as well yeah yeah yeah I think I think

73. 35:13

    generally when it comes to assistance there's a lot of uh a lot of structured data uh that you know is involved in our lives and uh there's I think quite a few like you know we already have plenty of Interest products that are assistant like and most of them work in the in the space of language right so so if you say something it put all the data in the context and then you know like like you know it's a deepseeker one doing some reasoning and saying this is the answer to your request it's obviously in the

74. 35:45

    language space 99 I don't know about 99 90% of the time it will do everything correctly 10% of a time it it will miss a critical uh critical part which in the symbolic space would be first of all much cheaper to execute right and and second of all there is no opportunity yeah I actually developed a nice analogy I think you'll you'll appreciate so this might go over some of the listeners heads but in terms of passing Theory um I I think that uh what

75. 36:15

    we you know as an industry are currently doing with llms is very similar to to just having a tokenizer just a Lex right it's it's it's a text Andor tokens in and out and where else in computer science are we really you know inputting plain text and outputting plain text it doesn't make a lot of sense so um I think that the um clearly llms are a good answer to that ageold problem of how do we pass and understand language and and the structure there makes a lot

76. 36:44

    of sense but in terms of the reasoning well it would be good to reason on Concepts instead of tokens right um and and and to this effect um this this's Research into into things like large concept models instead of large language models where in fact the the embeddings in there correspond to to Concepts instead of to tokens right and and uh therefore um I think that hybrid systems are inevitably the future where we have the the front end of everything is the language understanding uh and perhaps

77. 37:17

    that's preceded by the you know handling of U audio as in voice speech um but then on the back end uh we we must uh get to reasoning in in cont Concepts instead of tokens it doesn't make a lot of sense to reason in tokens yeah yeah so I in I think it was over a year ago now what El and I were discussing the U neuros symbolic AI right and the idea was that ideally you want to do not only you want to do it in

78. 37:46

    the concepts but also you want it you want to do it in a symbolic way right where there's no no probabilities involved ideally right and it's a u especially like at least in the places where you can uh do it in the uh you know in in the without involving probabilities the at the end of the day I think the problem is that the incoming request inherently is in natural language which is you know nonspecific right so so you have to involve at least step of conversion it into uh you know

79. 38:23

    into the into the formal language which will always be leaky right and then uh and I think our the way we were looking into it because once it is converted into some sort of formal language right then you can execute it on top of uh you know graph data in a in a you know like some sort of higher order logic uh right and so the way El and I were thinking about it is that you start with the natural language you convert it

80. 38:52

    into into this formal representation and then what you can do is you can convert formal represent ation back into natural language using you know conversion that is not neural network based right and and and say it back to the user and say hey is this what you want from me right so this way you avoid this problem of initial initial conversion but why once you get into symbolic space then it's a then it's interesting right yeah I think I think that's fundamentally right as an

81. 39:23

    approach as in the analogy I I I meant to get to is that um um just like a a Lexa is the input for everything we do with passing well the lexo is is followed by a passer right and the passer operates on on a BNF grammar bakus now form and and it forms a a symbolic structured representation in memory we call the abstract syntax Tre pass three and then abstract syntax three right so uh very similarly I think that inevitably here we will evolve Beyond you know plain text and and

82. 39:54

    tokens and so on mean it's not that they aren't needed they definitely needed in terms of understanding language and handling typos and and and you know ambiguities and so on but then that's just the front end and that's followed then by uh you know as you said uh symbolic space one way or another yeah yeah I think uh the uh I I was thinking quite quite a lot about so I think in this case the knowledge graph right and some sort of uh formal you know like higher order logic they are quite

83. 40:26

    complimentary right and U obviously we can do quite a lot before we get to high order logic right so we have plenty of examples where you can construct a query query the database get the results and that's already solving 90% of the problem right but at some at some point I think miring this technology with some sort of high order logic where we can go and you know do multiple steps of infint of you know reasoning what exactly the user user wants or how to get the data to what

84. 40:58

    they want well well look so so look in in in my field that's called entailment right so that's that's always been that's always been the part of part of the story it's a knowledge representation and reasoning right so the reasoning is Automated machine reasoning and this has you know existed for decades and an example is we we had the earlier knowledge graph of you know Alex NOS arto well the the nose

85. 41:21

    predicate um is um a so-called inverse functional predicate meaning that it impli the other way too if you know me I know you right so it means that as as that first fact is added to the knowledge base the inverse is added to automatically right and and similarly let's say that you have something like a family relationship that you know uh your father is this and his father is this well that implies that your grandfather is this so This has worked great for decades right this this exists

86. 41:51

    and and is a standard part of this kind of a technology stack and uh this can complement um the very different kind of reasoning that llms are are able to do right so basically this is an enrichment of data that happens automatically as um data is inserted into the knowledge base and and by the way uh last last week just last week um what I read on the plane to e Denver a fantastic new paper out of MIT by Professor Boer um

87. 42:22

    which is entirely about U exactly everything we're talking about the neuros approach to scientific discovery automated scientific discovery so he starts from a a Knowledge Graph uh that is basically um how how should I say very shallow in terms of what he wants to accomplish in his case a very cool use case he wants to accomplish um the automated invention of new meta materials that have strong impact resistance right and and so by the end

88. 42:53

    of the paper um the the approach does come up with rather novel meta material that perhaps humans wouldn't have been able to invent but to get there there's iterations where the knowledge graph is is enriched and and expanded and and just you can see the visualization just unfolding the knowledge graph grows grows grows grows grows connects to different domains of Science and and engineering and

89. 43:17

    just so so I think that there's a bit of a preview of the future right there yeah that's like a dream dream application of something like protocol right yes yes exactly yeah yeah okay we're running out of time but uh uh in the uh in some of the discussion prior to this you you had an interesting uh question which is if we discover aliens will they use Knowledge Graph so I'm I'm curious to hear if if you think they would or they wouldn't um yeah so so um look the same

90. 43:52

    way as uh binary data is is the we can reduce everything down to finary data and and it's a little hard to conceive how you would reduce it any any less than or or more than than just two choices uh there's a similar um fundamental aspect to knowledge crafts um so it's not only that they map to our intuition very well uh they map to the structure of the world very well uh it's that if you just conceive of representing any any information about

91. 44:22

    anything about um whether it's some physical Corporal uh thing like a person or house or whether it's an abstract concept um to represent two things in relation to each other you necessarily need three things right you you have you have a entity you have an entity and you have a relationship between them so that's actually what we call a triple in knowledge representation so that subject

92. 44:46

    predicate object and predicate is like an property or attribute or relationship and you can't really reduce it further than that so in in a sense it's it's really pretty fundamental and and so you can think of it as um sort of is a numeric tower that sometimes in programming languages we talk about similar there a semantic Tower right so it begins with uninterpreted bits that we then uh

93. 45:10

    choose for example to group into segments of eight bits and with that interpretation it becomes information and of course that has multiple levels we can say that that 8 Bits is a signed integer so that's another level of interpretation um and and eventually as we transcend the semantic Tower we get to Justified belief as in knowledge and and that's sort of the gold standard of

94. 45:32

    information as in we have given an argument why we should believe this is to be true um potentially potentially there's a additional arguably there's an additional level which would be wisdom but um I'm I'm leaving that out of the picture right now that's a a bit subjective and hard to uh understand but so I think fairly reasonably um if he discovered aliens uh they ought to be familiar with binary and they to be

95. 45:57

    familiar with um you know knowledge representation which we would call Knowledge crafts interesting interesting uh also I found it interesting completely unrelated uh that in all the works of azim of ir readed uh there are never aliens which is you know yes I never thought about it before but that's fascinating it is isn't it yeah he wasn't a big believer in in in that but but he was a big believer in trustworthy Ai and that's that's where the name uh

96. 46:28

    Choice came for for us as well yeah that that's what I figured that it must be related to the three law of Robotics eventually four yeah he was not only um not only the first AI aligner uh who ever lived but he was also of course the very inventor of the term Robotics and uh and his the path dependen is that all of our um future imaginings about Ai and robots ultimately flow through his Works he shaped how we think about this whole uh Universe yeah it's interesting that I

97. 47:01

    I actually obviously it's very hard to live in the modern world and not know about the three laws of robotics similar to how you might never watch Star Wars but you would know that you know Darth is Luke's father right but I only recently read I Robot and I found it fascinating that the entire book is actually entirely about misalignment not right so it's a lot of the a lot of the

98. 47:24

    Asimov stories about this almost like agasa Christi novels about AI alignment right as in we know that the robot done it we just don't know how the hell he did it because three law of Robotics yes yeah it's it's it's the it's the entire series about a lot of aligned robots doing misaligned things so it's it's very interesting in the context of of alignment in 2025 you know to read how

99. 47:48

    someone in what what what would it be 1940 right for for the earliest one right 100 years ago would think about this problem it's it's it's it's an impressive imagination and and by the way maybe I can mention here for those who aren't readers or aren't familiar uh I Robot by Will Smith one of the movies uh based on aimos works most directly uh the Bicentennial Man with Robin Williams

100. 48:13

     uh and then of course the the visually spectacular Foundation series on Apple TV right and uh also for those who watched the the the foundation series and didn't like it it's worth mentioning that it has nothing to do with the actual book right so you know you might still enjoy the the book for sure for sure okay yeah I think uh we almost twice over you know over time which is a but that was a great discussion uh you know it's uh it's definitely in the field that I enjoy a lot lovely to talk to

101. 48:47

     you thanks a lot for for joining us and uh thanks a lot for sharing your wisdom and you know looking forward to aimo protocol launching soon main n right that's the Q2 Q2 yeah that's what we're looking we're expecting yeah yeah great talking and uh yeah you know all the best with with the launch thank you Alex thanks a lot