NEAR AI Office Hours #2 Decentralizing Data: AI, IP, and the Future of Synthetic Data

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

1. 0:27

   okay sounds like we're live hello everyone great to have you here this is our second episode of uh AI office hours and uh excited to dig again I think today we have a really good group uh we're gonna talk a lot about data kind of Licensing IP there's obviously a lot of news happening around this topic uh with everything that's going on in AI so let's start with introdu introductions uh whitne need do want a start then Shu then Leo and then we can dive in this is

2. 1:00

   also opportunity for you guys to ask questions you know to me and uh each other as well great well Ilia thank you for having us I I appreciate the opportunity to be on here my name is Whitney I'm the CEO of ring fence and we're building the data monetization leader for AI uh we've been building for about a year now and our first platform our creator platform eer is set to go live here in about a month

3. 1:25

   nice awesome uh go ahead CH yeah yeah hi thanks for the invitation AIA and I'm glad to be here and meet everyone here so my name is Shu I'm found of Mizu Mizu is actually building the first synthetic open data layer which aims to create the most open and the collab collaborative uh synthetic data layer like to serve for the big mission of big like open source Ai and we are targeting to launch testet in in a so like looking forward to

4. 1:57

   everyone like to join the network and try it out thank you sh and thank you Whitney uh hi I'm Leo great to be here I'm the founder and CEO of superside and essentially what we're building at supersite is a network of decentralized AI oracles and the problem that we're trying to solve is hallucination so uh whenever you actually ask a natural language question of of a large language model um like any

5. 2:22

   of those today such as chat gbt or uh perplexity the problems that you face uh is essentially like you don't get the right answers uh to the questions because these are usually based on semantic search when they do um Implement like modern uh rag and essentially what we do is we use real databases to answer questions so we're building a network of databases that are linked to models uh to answer each question that requires um data and uh we're super excited to be here today and

6. 2:53

   thank you so much for inviting us AA for sure yeah so the idea was to get uh kind of a group of folks so they working on different aspects of data uh kind of in this decentralized space because uh there's a lot of going on right I mean obviously we've seen the Scarlet Johansson thing there's you know Sony going out and suing people for using their data so I'm curious how you guys

7. 3:17

   think about kind of the space uh obviously you know from licensing to like how synthetic data plays their role how you know search plays a role so I'm curious kind of how you um Overlook at this and also what questions do you think are uh important to be answered and so we can actually discuss that I'm happy to jump in on the Scarlet Johansson topic uh because this this problem was essentially what led us to starting ring fence you know you were seeing all these lawsuits pop up with open Ai and stable diffusion getting

8. 3:50

   sued by different uh different content creators or the New York Times um and it just speaks to this challenge that we face right now and and that is that all of these models that have been trained have been trained with other people's data and those people haven't been given attribution or any kind of compensation uh and web 3 is a nice solution to begin to fix this problem moving forward um but the Scarlett Johansson situation was

9. 4:16

   was wild because she directly rejected the request to use her voice and then it was used and then you know Sam Alman tweeted out her as kind of a you know a shot across the bow um and if you've seen her it was it was certainly you know her voice right um but I think that moving forward what's going to happen is that people are going to start paying attention to how they are putting their

10. 4:39

    content and their data out into the world uh and they're going to look at how they can protect it uh and you know we've heard for so many years now that our data is valuable only none of us have ever realized the value of that data right like I can't think of you know one story where someone has said my data paid my rent last month and we need to get to a point where people can realize the monetary value of their data uh and I think that we're three to six

11. 5:04

    months away from starting to open that door to give people the the benefit of the data that they're putting out into the world just by living their life and using the internet those are some great Point Whitney um it's super interesting that you mentioned this actually so I've been thinking about this uh deeply like in terms of a philosophical like concept so essentially like the way that centralized entities are working today is uh they're fully Focus on acceleration and they don't care about

12. 5:31

    the cost of acceleration and uh this is clearly displayed upon even like for example just neglecting Scarlet Johansson's uh basic rights uh of uh like sharing her voice um just neglecting that completely and uh using a model U that has her voice uh despite all of that so it's essentially like um what we're seeing today is like centralized entities have adopted this like U Viewpoint that uh acceleration at

13. 5:57

    no cost is necessary and and um at all costs and essentially like um we don't really know if that's the right way to go about building um like um a system that is able to answer all questions in the world or like uh even agis actually uh is this the right way to go about it we don't know yet and um what crypto essentially offers us is Like A Primitive uh for us to be able to like

14. 6:22

    incentivize um the contributions from people as you rightly mentioned uh to monetize their own data and like uh hook into like different networks to be able to like be involved in the contribution of I guess one of the most greatest Innovations of uh not just our generation but of all time and um so essentially like that's how I've been thinking about it like um as we grow as a space like it's really fundamental for us to contribute to the development of

15. 6:48

    AGI um because like we can't let like a profit-seeking entity own AGI and uh like decide like how the world is going to play out but yeah I would love toar your thoughts on it show as well Andia yeah so we're thinking that in the in a different perspective basically uh I know like the centralized entities they definitely like uh kind of like you know doing this kind of things in their own minds and are not care about the cost uh but we're thinking more about

16. 7:17

    like what if they want to open source the thing like they they they want they built and what the blocker there and the one big problem we're seeing here is that actually a lot of data in current stage cannot be open sourced at all for example like Lama 3 is Trend using like definitely like using some common public data like common C but also use lot like private data right and this kind of private data can never be open sourced

17. 7:41

    it's not like about like if they want to or not because like regulations or user consents all these kind of legal issues so what we thinking is like imagine like we want to achieve to this kind of AGI space and want to build a fully open source AI ecosystem and then like how the open data layer would look like that's why we're so focusing on synthetic data we believe that only synthetic data can be open sourced

18. 8:05

    because it has no like license issue or copyright issue or IP issue actually not a lot of people have linked the synthetic data part with the open data part there is using synthetic data part to solve the data scity problem but we see the potential of that could be openness and eventually I believe imagine we have this kind of know a very large language model which is completely open soured with open data layer and they has to be synthetic or has to like

19. 8:30

    know license free or copyright free risk- free yeah so that's what we're thinking about yeah I think I mean all of those good points and obviously agree a lot on uh kind of you know we need a alternative we need a user owned AI as an alternative for kind of the for-profit generation AI models at optimizing for success of single company and I do want to zoom in into synthetic data a little bit later but I actually want to take a little bit of controversial stance which is uh well let's imagine that they actually did

20. 9:04

    find an actress uh and a voice that sounds similar to uh scy Johansson but is different and they have literally you know paid somebody to record this voice and so now we're actually in a weird situation where uh you know like what is too similar and also like you know there's a bunch of inputs went into this model that actually obviously was trained on scy Johansson at some point because I'm sure they fed a bunch of videos across internet to it and so kind

21. 9:34

    of at which point should they been filter this out right and how uh what is similar like what does two similar even mean uh and so I think there's an interesting I would say failure mode that we see right now in the kind of legal system in general because copyrights were invented back in 1700s they were invented on the kind of Foothills you know time things were slow moving slower back then so but of invention of printing press and a lot of

22. 10:07

    it is kind of designed to be like this is a Content somebody produced now you consume it right and there's a copyright law protecting this but we're now in a generative age right where there's a model that consumes a lot of different content and recombines it and reconstructs it and now uh it's really kind of the standard uh like plagiarized plagiarization systems don't apply anymore um to this models because they

23. 10:34

    can recombine them in such Noble ways so I actually think there is a I mean a massive shift that's happening where the existing copyright law are just like not able to handle and uh as part of this we do need a new system so I actually agree with you whitne that like we need a new IP uh tracking we need a new way of Licensing it we need a new way to uh compensate people for creating new content uh and it needs to be going directly right it cannot go through this

24. 11:07

    existing distribution platforms and Publishers because they just become pretty much data aggregators and resellers uh and kind of taking their own cut for not much additional value so but kind of crypto here is more of a facilitator now instead of the existing legal system that kind of you know relying on this outdate approach of measuring the attribution right now on Hacker News

25. 11:38

    there is another story trending where uh uh there's this singer Tom waights and he had a very similar situation that scar Johansson is in right now where uh they they aired a song which sounded like his song and the voice sounded like his voice but it was a different person so he won that particular case so there's a a precedent yeah there's a couple of these out there I mean Ed Sheran was sued a while back for having a sound that was similar to uh to to other artists and in court he actually played I think it was

26. 12:12

    over 100 songs that Ed the same chords to show how um you know imitation and something sounding similar are very different and you know something that we talk about often is you know what's the difference between AI generating a piece of art that's similar to Pablo Picasso and an artist actually painting something with Picasso as their inspiration you know if if you're able to to figure that out then I think there's there's something to be said but I think that really what what the focus

27. 12:43

    for us has always been is how do we how do we move forward like how from here do we start to set the tone for not only copyright and IP but how your content your data um is able to be protected into the future I don't know that we're ever going to have any way to you know fix the damage that's been done since Pandora's Box has been opened but I think that you know crypto as it is is

28. 13:06

    about self- sovereignty and if we can start to to teach people how to be more self- Sovereign with every aspect of their life their data included then then we're off to a good start I I'd like to uh just like share some thoughts actually uh I I think in my understanding the copyright is actually to protect the data not to protect the person for example like I write a book the book could be trans the ownership of the book of Ip could be transferred from one person to that person right it's not like necessary to

29. 13:38

    be the writer of the book to anyone so ultimate like the copyright is basically protecting the earners and ownership is actually necessary like you know it's very native concept right like read write own right own is like one of the most important part of that and I don't have answer that e web3 is actually facility in the the system or it's kind of like building a brand new system on

30. 14:04

    top of this but let's imagine a very extreme case like the the the I think like you mention this kind of the voice case that scar Johansson like has has this kind of voice right imagine there's another person who has exactly same voice with scar Johnson and they both own the voice right there's two people own the voice like in that case can you buy the voice for another person like skipping scar Johnson like it's it's basically what what is protecting here are you protecting the person are you

31. 14:35

    protecting the voice eventually I think we will need ownership that system that who decide who own the voice maybe they share the voice they share like 50% of this kind of voice right and uh uh then like all you need is basically to define the ownership of the of the data of the voice and then like you know you you you you you make some decision process on top of that I think that's eventually how how it looks like but like uh that's

32. 14:59

    just my sauce I don't have like exact number of that but like like you know some some like this kind of a uh imagination on top of this yeah but but isn't it the case that both in Scarlet Johansson's case and in Tim W's case it was actually a different person who was doing the voice right it is still the case that there was an attempt to impersonate the person uh with the original voice so so

33. 15:23

    it doesn't matter whose voice is actually being played I think the ownership itself is a social thing is a society thing you know it depends on the society so I own a house who decide I own a house actually it's actually the government decides I own a house right I cannot say I own this house because this kind of ownership scene is actually defined by the society structure by the government not by myself yeah so the voice thing I

34. 15:49

    think is the same eventually the society has decideed who own the voice either no one on voice it becomes the public good or because gly has this kind of higher reputation then Society decid okay you own the voice because you actually promote the voice to be like no well known by the by by the word so it may become like super complex you know it's not like native building thing like it's a bwing with you it's like people other

35. 16:13

    people decides who belongs to what you know that kind of thing I mean uh I think I would like to take a moment to contradict you here I think there's two schools of thoughts so essentially one school of thought is um like what we own is dictated by like um the societal rules and the other school of thought is essentially we are sovereign owners of what we own essentially like we should own our voice we should own our idea as

36. 16:37

    we should own our thought so I think these are two different schools of thought and um so I mean uh I completely understand your perspective from the first school of thought but from the second school of thought um again like if uh for example like uh you had bought like the voice that sounds similar U to Scarlet's voice um that and you had actually bought it from a real person then from the second school of thought

37. 17:01

    this would have been a valid way to go about solving the problem but uh yeah I mean I would love to hear you your thoughts on this as well uh like IIA Alex and Whitney like um because I mean it's very fundamentally different perspectives one is um Society should set the rules of ownership and the second is essentially we ourselves should set the rules of ownership I mean I I think that it's it's a combination right because it we

38. 17:28

    are like we do want self- sovereignty but we are living in a society and so although like we're trying to push as much as possible for each individual to be S Sovereign but at the end there is still a consensus around that and like be that you know social consensus or a onchain consensus with you know uh with nodes or or inter subjectivity right which is like economic consensus around

39. 17:53

    uh something that's subjective and so I think the the reality is uh what I'm I'm kind of leaning more is actually we need a different model for rewarding people for their data right because some of this data is just out there like you you cannot like by the fact that you you know spoke on the street you cannot take it back right the fact that somebody recorded it and then you know now you're

40. 18:17

    saying like oh nobody can use it but it was you know they were walking past and they recorded you speaking like there there is kind of all of this conflict and so what we need is a novel model on of going back and actually attributing the challenge is the I mean some of the some of the folks proposing to uh to kind of evaluate based on you know hey this voice is used in chat GPT so we should attribute it to you know scy Johansson and whoever actor that was

41. 18:49

    recorded and the challenge with that is actually it's very computationally intensive right if you actually I mean you don't know who that people is you need to score every single person How likely you know there always how how close it is to matching and so instead and I and I actually think this is very healthy for in general our uh our world is transitioning off this kind of attention economy where you we attribute

42. 19:14

    and reward based on how much attention you know how much times it's used or how how many views the article gets or how many plays it gets we need a new way a new metric to Define how valuable something is based on then later it's been used in machine learning model been recombined and being used in various ways across different applications and so I have some ideas of how that can actually work but that's kind of a Direction Where We we're moving away

43. 19:41

    from again copyright is protecting data it's protecting ownership but it's protecting kind of the artifact that people consume and the problem is the artifact here that people consume are now model output and this model has you know pretty much seen probably millions of voices right and yes it's like they parameterize it to output something specific but it has seen like so much

44. 20:03

    different data and you like going now you know by son and saying like hey you know take everything out that we've copyrighted is like extremely complicated right it's like well what about image where there is a I don't know uh whatever the like their character is like in on top of you know somebody's digital art on top of you know somebody's new creative work like it's it's it's super complicated and so we need I think a different model that's like scoring and rewarding their

45. 20:33

    creators at the kind of at the time of creating and contributing this data and then allowing models to recombine all this data and create kind of noble experiences for people that sounds um really reasonable especially given the world that you are in currently uh what it sounds like is essentially you're proposing a completely new primitive um such that like whenever anything is created like um the uh potential for monetization is already set and the rules are set and uh

46. 21:05

    then like how much uh like the level of monetization is depends on how much attention that piece actually gathers and uh so essentially like if this primitive were uh like assumed to be fully applied then that what that means is for example all of the data that we give Apple for example like uh through our smartphones uh to like train their recommendation engines we would be able to also like um I guess um like earn in

47. 21:29

    contribution u in proportion to our contributions personally uh to the recommendation algorithm for example uh I mean this is super interesting as a concept and uh we'd love to learn more yeah I think Wy you're Pro you know working a lot in a lot of the direction right we are yeah and this is it's a really good point Leo so I think one of the main things here is that the there data pricing is a bit opaque right now because essentially it's just being done busines to business you don't have

48. 22:01

    individuals being able to place value on their own data you essentially have companies that are bundling all this data together and then offering it for sale to different companies somebody who wants to sell you you know provide ads or or apple who wants to use this for uh ux UI improvements um but you know the numbers come down basically to this your data your personal profile data is worth about $600 per year right now sold to just Google and meta right uh when you

49. 22:31

    extrapolate that out over all of the companies and platforms and services that you use on an annual basis U it gets considerably higher but the challenge right now is that there's not a not a place for your data to aggregate in a meaningful way to where some company can find you and get this data in an anonymized fashion to where you feel protected which is you know essentially what we have been working towards and there's there's challenges

50. 22:56

    to that um but web 3 does solve and crypto solve some of these challenges right account abstraction is a very big thing it's an important thing if you're going to be compiling someone's entire data profile for let's say a whole Year's internet usage you want to make sure that you're not attaching your name and address to all of that as well right um but on the other side of this uh it also allows you as the user to determine

51. 23:19

    who can get access to your data like maybe I don't want to see ads so I I don't offer my data for for sale to those kinds of companies um and on the flip side that maybe I do because it's going to afford me an an incremental amount of additional money every year uh that I can put towards something that that I enjoy so uh you know this is a it's a it's a big problem um and I probably better said it's a big

52. 23:43

    opportunity you know the old cliche was that everyone got into crypto you know you got in for the gains and you stayed for the tech and uh I think that now we're we're seeing this wave again um but you know with data it just it offers such a massive opportunity for people to capture additional value value every year just off of living their life um that they've missed up to this point you

53. 24:09

    know and over the course of your adult life your lifetime as an adult you're looking at potentially hundreds of thousands of dollars in missed Revenue directly to your pocket that you're giving up for the data that you're freely providing to a lot of the apps and services that you use so I think there's a huge opportunity to recapture that and to let people see that value so I I actually want to like shift gears

54. 24:33

    a little bit and and kind of take the op take the opposite here which is I mean right now this uh like the langu large language models for example require about a thousand times to be shown some fact to actually remember it effect like uh accurately uh and so we have this kind of this huge data gers right that require 15 trillion tokens for example to be trained and so to give you uh I mean for the audience the idea Wikipedia

55. 25:07

    is about two billion tokens so English Wikipedia is two billion tokens so it's 75,000 times Wikipedia right now shown to this models to for them to actually um remember things and be able to reason and so uh this is where like I think synthetic data is interesting because it's pretty much taking the existing data right for for example Wikipedia and recombining it many times but the other

56. 25:32

    side is can we make this models more effective at training from a small amount of data where you don't need as much and then data users data can stay private and you know potentially combine it in a different way in a more in a more kind of uh specific way versus just like I just throw you know billions of hours of YouTube and you know trillions of tokens of text into this model and

57. 25:57

    Hope hope it learns how to figure everything out I know I mean I'm sure sh you have been thinking a lot about this like how both synthetic data benefit I mean benefits with training right now but also what other the ways this can actually go forward yeah so we are thinking about like U uh how we can generate data by Ai and then feed into AI again so let AI to train AI I think that ultimately that's away it's now it's it's not the case is

58. 26:30

    because AI s is not AGI yet it's not that smart so you like you need people to keep putting intelligence into the system as a positive feedback group I was in the lecture and I was like listening to a speech uh from a professor from MIT who was like the youngest professor in MIT who in in neuroscience and he made a very very good point he says like that actually he believed that AGI will be there in 18

59. 26:56

    months to five years which is quite optimistic yeah but like he made one point that we should do everything now we should do everything just to assume AI is already there you know so that's something like we actually following here we assume AI is already there like assume AI is already there like how you going to do this kind of model training you're going to let AGI to produce new more knowledges there and then like use AI to train AI right you got AI to train

60. 27:20

    AI just like you the people learn and then like teach other people that's how it works so we're thinking like imagine we can for example like generate whatever data we want from AI directly and gener high quality data in that case it'll solve all the problems now what we're doing is literally like primitive Gathering and hunting and in that case it becomes like industrial mass mass in production yeah that's what I thinking about the biggest challenge here is that

61. 27:52

    the AGI data actually like generated data is uh not as high quality as we expect because it has Illusions you don't know what the boundary what the intelligence boundary of these kind of a models are so you will need you will have to for currency inide the workflows you have to add a lot of like this kind validations or checks or even like you know human reviewers necessarily to do

62. 28:16

    this kind of job to ensure like the quality is is is as high as as as high as enough but uh this kind of reviewing is very different from data labeling for data labeling is very like like labor intense you know it hire someone from like like some other country like who have no knowledge of AI like you know do this kind of labeling for you but for this kind of thing you have to really collect the intelligence because this

63. 28:40

    kind of thing they have to they have to understand what AI is saying they have to tell if AI is telling the correct thing or not they have to very Focus even even like sometimes you have have specific domain knowledge to do this kind of things so I think we will H we will have to shift from this kind of Labor intense labeling to intelligence intense labeling plus this kind of s workflow they will help solve this kind

64. 29:02

    of disgus problem yeah that's my uh pun sh I just have a a followup question if if I may right now how are you thinking about the the flipping if you want to call it that of real data versus synthetic data how far away do you think we are from from that point where synthetic becomes the primary source uh sorry I lost you for three seconds can can repeat the question yeah so with with regards to right now real data seems to have more value than synthetic data in training AI how far

65. 29:35

    out do you think we are from that flipping and synthetic data taking the lead spot I would doubt that real world data has like you know more valuable than than than synthtic data because there are tons of different types of real data there's personalized social data this kind of data is actually not contributing to the intelligence level of large language model if you dump that into larg language model it's not helping improving the intelligence level it's not helping the performance what is

66. 29:58

    more used for is basically as a rag for example personalized data you can use as a rag in Dex local to serve as this kind of agent stuff right so different data serve different purposes if you want to improve the intelligence level of this kind of large Foundation model you really have to input this kind of knowledge and knowledge is Wikipedia's knowledge and this kind of paper you

67. 30:21

    know from the website's knowledge social data is not knowledge so so so I think synthetic data will play a critical role in Foundation model training because we hope it can generate some knowledge from that but it's hard for now yes yeah so yeah and real world data is uh that I said it's it cannot be open sourced why we need open source when we want to build open source AI we want knowledge pattern to be open source we don't want

68. 30:45

    the data itself to be open source right so it's like you know not that sensitive at all that's I think the ultimate goal like the foundation model and syntic data they will be serving more this kind of foundation model thing and the personalized social data they will serve more for like personal agent and there is a middleware there some agent there's some agent serving for different

69. 31:02

    purposes if serving for a general purpose like could be like you know more synthetic if we serving more personal user cases that could be more real world that that's what I think about the whole like landscap and space thank you it's great that makes a lot of sense uh what you said actually makes a lot of sense um we found synthetic data to actually be super useful as well like especially

70. 31:26

    like um the way that we use it um at suide right now is essentially like we um like we fine tune smaller models for specific tasks by uh creating data sets from the larger language models so essentially uh what we did was essentially create a large Tera set of SQL codes uh using gp4 uh and uh then we find you Lama 3 for the same task and essentially made it cheaper so um like this whole process of um like uh teaching smaller language models um like

71. 32:00

    a very specific set of Knowledge from bigger language models when you have a very defined like domain space it works really well um in our experience um so I mean uh we definitely agree that there's something like very promising when it comes to synthetic data and um like as we move forward like we're going to move a lot from like um like human data collection to um human data being used for synthetic data and then synthetic data collection so uh it's super

72. 32:27

    interesting that you will bring up these points yeah I think you mentioned very good point that that's called knowledge distillation right it distilling Knowledge from a smarter model with gbd4 to a not that smart model like small language model that's very common this case uh but like a more challenge things like if we can use synthetic data to push the boundary of the intelligent level of the smartest model that that's super hard yeah I think but like it has a lot of these kind of use cases you now

73. 32:52

    the use case you mention is actually the major use case right now yeah Alex you wanted to say something oh yeah I'm I'm I'm researching how legal it is to use output of gb4 to train models uh yeah I found I found it specifically you cannot do it you cannot do it if you compete with open AI exactly that's a terminal service that's why we need open source AI you know this kind of we need the smartest open source AI that can help us to do this kind of things pretty

74. 33:23

    use yeah I mean I I think my my aspect here obviously coming from you know training C models is is just the fact that you need to show thousand times the same thing to me seems you know extremely suboptimal right the fact that you know you take smaller model right what you said and you're training it on output of bigger model uh I mean there's like a few few effects that are happening there but including like you you know if you do it if you have access

75. 33:54

    to actually bigger model like actual execution you can take the logits from a softmax and you can train on that that actually uh provides way more signal than pure like token feeding and so you can actually have like way better uh knowledge distillation that way uh and this is actually how like we trained like way smaller models to serve uh for example at google.com where you have

76. 34:17

    like huge latency requirements back in the day and the challenge here again is like it means that we are really bad at training these models right it means that this smaller model is actually able to to know the to know all of the facts but we've not managed to show it the data in the right way and update its parameters to act for it to actually memorize it right and a really good example actually is the Llama 3 the 8B

77. 34:46

    uh which they train on 15 trillion right which usually people don't do right usually it's trained on like two billion three billion uh sorry on 15 trillion tokens and usually it's trained on two three trillion and the fact that there's like outperforms models that were trained on two three trillion tokens right and so and and the general fact that like seven eight billion uh parameter models have been improving

78. 35:10

    over time even though it's a fixed size of parameters right it's fixed size of capacity but they've been able to kind of improve on all the benchmarks over time it just shows that we're like so bad at training these models and so my actually question is can we leverage uh I mean potentially synthetic data potentially uh better optimizers but also just generally kind of our creativity as crypto community of creating a different model of organizing people to figure out how do we solve

79. 35:42

    this problem which is how do we train models with less data more effectively so that they're smarter and able to memorize and uh kind of uh also reason about this more effectively Alex you wanted to say something I just wanted to say that it's not that surprising that it keeps keeps on improving right like if if I if I learn a piano piece if I do it 20 times I I I get better and better at it so it's not surprising that yeah I me do take 1,000 times right so

80. 36:15

    it is not that sample efficient but it's not surprising that it's h yeah I mean clearly you need to show it few times right and again like if you because you want it in different context you want to understand the relationships so like I yeah I'm not saying it it needs to go down to like from one case you like completely understand everything but just like the kind of the idea here that uh there's this really good paper around

81. 36:37

    studying like uh physics of llm it's like a series of papers actually and they analyze a lot of this like differences and uh yeah I mean that's that's kind of what I'm I'm really excited about like what are the approaches we can take here to uh really really drive down the kind of data like requirements to train this models but with synthetic data is you're thinking like let's keep model inefficient but uh but fix you know like we don't have 1,000 examples we're going

82. 37:09

    to generate 1,000 examples and so instead of fixing the model we will fix the the inputs no I was more thinking with synthetic data one idea would be to generate data that's better that's this models are better at consuming right so this can be like way cleaner so that that that actually paper on physics of LM they generate the data set uh which is like taking two pools of uh facts and pretty much generating them and so like what can you do like let's say you take Wikipedia and you want to build just

83. 37:39

    from Wikipedia the best model that can like answer you know uh most of the questions like what what is that that you can do to make it more effective at at learning from that uh data set with you know generating synthetic data Maybe distilling something maybe PR reprocess in differ uh the way that I've been thinking about it is I mean you're absolutely right um so using synthetic data we can build much cleaner data sets and uh the more specialized your data

84. 38:10

    set is for the task the better the model you can get um so that's one way to go about it the second thing that we I've been thinking deeply about is essentially you don't need to build the world's largest model uh that solves every problem like known to man but instead like if they if we can essentially C orally break down um like uh tasks or like for example segments then we can have domain specific models

85. 38:34

    that um operate in Union and essentially like all you have to build is essentially a matching engine which uh matches uh users's query with the appropriate sub model which will be able to uh generate the answer to the question that you're looking for and this is essentially like what we're trying to do at supersite with u like domain specific models for different kinds of data uh so that like if you have any um questions for like different data uh then you can ask them at

86. 38:58

    superide and then there will be like different sub models working so u i mean the CRA of this is essentially like um firstly we want better training data but secondly also like uh if we can have like thousands of different models um like that are hypers specialize for different tasks uh be it um data related or not then essentially we can um like if we get them to work in communion then uh we can like also like uh generate

87. 39:24

    like I guess clearer answers than centralized models can uh like create so um that's my thinking about it uh in my opinion I feel like the the data quality is definitely like you know way more important than data quantity so all we need is high quality data yeah so I think I read the paper I don't remember the details but like they Trend the model with I think 100 megabyte data

88. 39:51

    the other one high quality 100 megabyte data the other one is like you know X books is all you need yeah so uh the the problem becomes like how we can encourage the community to build high quality data sets and to to train this kind of models right and uh uh I I don't know like personally I I'm kind of like a DA infra background I'm not I don't I'm not like super uh knowledgeable in this kind of machine learning and the AI training part but I think openness definitely can help in

89. 40:20

    this kind of use cases basically if you can open the data and let Community to collect the intelligence to contribute find a mechm to do that and uh that would be super cool because I think web is basically the correct way to do that and another part is very challenging is basically how you can verify the quality of the data is high or not or how or even like how we can verify if the model is performing good or bad if that I think it's still like a big problem like

90. 40:48

    for the industry right I was like trying thinking about like building a very cool data set it's called trick the model if you can trick model to the model to in some like nonsense stuff you win because you actually are finding the boundary of the large language model if you found the boundary of the intelligence right it doesn't know this kind of thing and then if you can connect all these kind of boundaries then you find a way you have a direction to keep improving the model to to the next to the next level

91. 41:13

    yeah to expand the boundaries so that's something like I'm thinking about but you cannot do that by a single team by a centralized team like mattera right you want to encourage the community to involve in this kind of process to help you to do this kind of things because it has be very diverse have to cover all these kind of domains and only the community can do that yeah so that's what I'm thinking how how how this should happen sounds like we should build this on near we actually have we

92. 41:38

    actually have some uh kind of an example of this uh it's called Shield GPT and it's designed for users to trick the model to pretty much sell one token and buy another token and it's an It's actually an infinite game because people just rebalance the all the meme coins in the VA by pretty much tricking the model and like convincing it that it should uh it should pretty much switch so you can

93. 42:03

    have something like this which is fun like economical as infinite game uh and actually generates data on the other side which is useful because you actually see how to uh jailbreak the models or how to find vulnerabilities in their reasoning yeah exactly I think we should build a gamey the process and build a product on on top of that just let people to find the boundary of these kind of models yeah that be like super I

94. 42:27

    think uh meaningful and valuable all right there was a question uh in a in a YouTube live chat which is can crypto somehow help uh in the accurate versus probabilistic LM problem I'm assuming it's it's uh related to Leo you mentioned uh you know um sounds of hallucination uh but also just like you know the llms at the end right now statistical system right it's kind of gives you like an average of you know

95. 43:01

    sometimes just what Reddit thinks is true um so I'm curious how you are how you are thinking about it but also how everybody else uh thinks about this problem that's a great question so I mean the way that we've been thinking about it is essentially like uh if you could posee the same question to like different sub models at the same time um that are like hypers specialized to similar kinds of data and then you ask

96. 43:25

    that question and then essentially uh when these models generate the answer to that question um you come to consensus on the U answers from the sub models and essentially like U you can even like slash their stakes and um like you can um essentially find a mechanism to uh award people for accuracy and then um like U also like um slash their stakes in case they're inaccurate and uh

97. 43:50

    so like different submodels working in communion to generate answers uh is definitely one way to solve the problem uh it does get slightly more expensive because um like you are inferencing from multiple models at the same time but uh you can increase the accuracy um so this is one way that I've been thinking about in terms of like solving this problem um so the other aspect of this is essentially like um a lot of language models they um like hallucinate because they don't actually have the access to

98. 44:21

    the data while um answering the question and uh so for example like if you were to ask um anything about near blockchain data on perplexity then essentially U like it would hallucinate while giving you the answer because it doesn't actually have that blockchain data but instead like if you had a model hooked up to the blockchain data um that was posing the question off of the data then essentially you would get a much more

99. 44:43

    accurate answer for that question so these are the two ways that we've been thinking about it uh at suid so first is like consensus of different models and second is uh like integrating them with the actual underlying data to generate the answer but we love to hear sh um like your thoughts as well like uh how you will think about this space um and uh it would be super helpful for us to

100. 45:15

     learn yeah I'm curious uh what other folks are thinking on this question I personally don't have opinion on this because I don't know how to answer this specific question because uh why why like why we need to use crypto to solop this kind of accurate and problemistic problem at all like I kind of like don't understand the motivation behind this kind of question so I would like leave the questions to others yeah yeah I mean so here here's kind of my

101. 45:45

     thinking here uh few fold right one is the you have an opportunity to use for example crowdsourcing uh to indeed find places where the model is not accurate um and uh I mean in a way like what you were describing show before is uh so-called curriculum learning right where you can actually find the limits of the model and then potentially try to get more training data for those examples for

102. 46:16

     that scope of knowledge um I think the other side of this is indeed where models just making up stuff or kind of averaging out uh the information uh and so that's where things are more interesting uh I think there's few aspects right kind of similar to what Leo is describing There's an opportunity to have lots of different more specialized models um which are maybe more uh either have access to data or

103. 46:46

     for ftuned or whatever kind of they were designed to answer better at specific types of questions um so you know imagine that you take the mixture of experts that's right now in inside one model but actually now you put them on different servers and you incentivize them to actually be distinct in different and specialized in different areas and so you can have a crypto kind of network of experts uh that you know

104. 47:11

     one of them is better at Medical one of them is better at blockchain data one of them is better at um you know math and physics and then you can combine and ask them and kind of they can also communicate with each other now and then again there's like a value function value system that relies on human labeling as well uh to give you feedback I think the going beyond that is uh some of the model architecture that I've been thinking as well is like how do we use more more compute when

105. 47:43

     it's needed right sometimes this models are uh you know like you ask it two plus two and you ask it to you know design a plan to run to start a company and it spends exactly amount of compute exactly the same amount of compute for the two problems s right for 2 plus two it you know uh spins up you know 10 to 2 F floating Point operations and for like a super complex problem that has a lot of

106. 48:07

     different you know questions to ask it'll it'll exactly you know spend the same on compute so how do we uh kind of use different amount of compute depending on the complexity of a task how do we have more ways to uh for model to understand where it's either not confident um and actually being able to uh you know to ask more questions instead of trying to like latch on into the uh into the directly answering and this is again where you know some of the

107. 48:41

     fine tuning right now like all of the fine-tuning is The Confident answer I I don't think fine tuning data has like I don't actually know please you know give me more input um so this is where you know having some of this uh ways to um to do like effective crowdsourcing or engaging Community to give you uh kind of inputs uh will be important but beyond that yeah like we also need just architectural improvements I mean honestly like I'm I'm stunned by the fact that we have I mean pretty much the

108. 49:14

     same model plus minus like regularization and activation functions uh that just kept around for you know so many years and we have not had like a major uh kind of improvements over that it's kind of crazy to think about like um the fact that you mentioned that um even the simplest problems and the most complex problems are going to like rely on um mechanisms that use similar levels of compute uh that seems like a

109. 49:54

     very um like deep structural problem in the space uh and also it seems very wasteful so um I think that I mean Innovation there is going to be super import important as we move forward yeah and I think I mean the Chain of Thought and some other approaches are kind of you know trying to fix that uh as well as you know ways to you know Pipe One agent into another agent and stuff like this like all of this are kind of like patches on top of this idea

110. 50:27

     uh but I think yeah like fundamentally there needs to be um kind of you know some upgrade on the architecture um but yeah like to actually achieve that we need to coordinate a lot of efforts and resources to you know make that uh possible and available for folks but I think generally the uh the adoption of techniques is relatively slow like if you think about mixture of experts it's been around for a while

111. 50:53

     before everybody started using it right and the same like like you know it's been two months since a mixture of depths from Deep mind I just Googled around and it doesn't look like it's making too much uh you know that doesn't doesn't generate too much discussion uh but but like you know that's definitely an approach so so it could be that just it's being adopted slowly yeah no I mean I agree and I mean even for Transformers right it took couple years before like it started kind

112. 51:24

     of to be leaned on it takes like multiple confirmations from different pretty much parties before people actually like start to lean in into some technique which is potentially actually an interesting place for crypto to incentivize where you know as new paper comes out um actually incentivizing different research labs to redo I mean like replicate research and kind of uh you know combine it in some way and confirm indeed that this results are stand right because even though right now well first

113. 52:04

     of all there's a lot less open source AI code right now coming out from this big Labs um and so the the the whole replicable research that uh kind of computer science started is is not going in a good direction but I think that's that's an interesting opportunity as well like we we have you know potentially sounds as decentralized computer marketplaces which can offer some um some compute to to to try the

114. 52:31

     models and uh at the same time have like more you know Academia and projects to replicate some of this research to indeed like validate that this techniques are working or not I have one question actually for for for you for IIA yeah so I when I like kind of like explain an open source AI concept to my friends and and they are kind of like skeptical about the idea I'm fully believ in that but they're skeptical because they believe that hey even like you open source the the Llama

115. 53:05

     three right he opens up the data opens the training process like the people like the community cannot try out on that because they cannot like train this kind of thing from scratch it cost a lot of gpus a lot of resources even you have this kind of insights it's like super like hpal driven basically like you have these kind of Concepts you have these kind of slots you want to try it out you don't have the resource so how to solve

116. 53:29

     this kind of problem like how how crypto can solve this kind of problem better that it can incentive to people to try different things without you know restricted by the resource and even like the community or some like very cool developer know oh I think this kind of neuron structure or I think this kind of data can help to improve the performance of the model I want to try it out like I

117. 53:51

     can get this kind of for example like consensus from the community and get support from the community and then like I can try by myself and let the community to to to to to to decide like if I want to move on or not you know that kind of mechanism is something I'm looking for I feel like very promising that a big thing like web can provide uh just like to hear your thoughts on this do you think that argument valid that

118. 54:13

     you know for now or you think like how can we build a more better mechanism to support this kind of like better future yeah I mean this is actually something we've as Alex been uh discussing uh for a bit now which is um how do we enable kind of faster experimentation and and it comes as a combination of things right one is we need a very effective kind of training open source training source code right right now I mean there's like few versions but uh we've actually been

119. 54:45

     playing with Megatron right and it's like pretty complicated to like do some modifications it's pretty hard to spin up so like you need something that's like really easy to spin up really effective ideally you have a way where you know somebody can just either write a model or even GPU Cals you know click a ship button uh like and then it you know can if the community supports it it you know spins up on ADV adventure machines on the cluster and so so you

120. 55:13

     need some kind of almost like send boxing that we have for smart contracts like some like approach like that then then you need indeed a way where like you know let's say you have an interesting idea maybe you test it out on on a smaller model right on on a very like high quality data so you can see like how it performs compared to other ideas right and then if it works then Community kind of almost like gives you

121. 55:39

     a budget to uh train it on a bigger model and it is actually kind of like how some of this companies you know big companies internally work right you train on you know one GPU or like uh on one uh on one kind of machine with a gpus and then if it works you can go and uh kind of request more resources to to train on a bigger cluster and then you know bigger and bigger right like you prove it every size that that actually

122. 56:05

     works before uh before going to next size um so we can actually you know do all this facilitation we just need some of this core infrastructure to actually be able to do this and make it like somewhat permissionless uh for people to experiment right I mean because you can come in and like rent one GPU for a day uh which is going to be like you know $100 uh and then when you need to scale

123. 56:29

     up that's when you go and kind of like raise funds from Community for that benefit from that model as well be that's through public goods or through kind of maybe more uh you know financing and this is also something we've been thinking about like how to create kind of public goods financing for open source models um with some ideas but we're definitely not ready for like getting that out I don't know Alex

124. 56:54

     if you want to add uh on this sorry I wasn't listening too carefully I was reading chat so well I'm saying more like how how do we how do you have competitive uh open source models like how do we get people to be able to experiment more freely right yeah yeah so as you mentioned the the training infrastructure uh would need to be uh much easier than it is today I think I think the uh like naturally Innovation happens where uh where there is Demand right so like in terms of

125. 57:33

     inference in terms of uh fine-tuning uh we have outstanding tools today uh but training I guess there is much less demand for training because much fewer entities have capacity to train and those who have uh they have resources internally to build training Loops right so so training code is uh in the less um is less developed in the open source right but if we do want to have uh Innovation you know in the open source being pushed forward tring will

126. 58:05

     have to be part of it yep uh and yeah I think I think it it boils down to how how simple it is um like you know like let's consider the example you you were giving earlier with the um you know you you you require 1,000 example examples to see something and they use you know like Triplets of facts I I don't remember if those were if this is the same example or Not Right But like you know like if if I hear this example and I'm like hey uh like get the

127. 58:36

     account for this or get the account for that right like let's say I have an immediate thought that something was not controlled for like how easy it is for me to immediately pull that experiment from a database of experiments and mess around with it right like if I have to rebuild it from scratch I like effectively G given a limited resource of time uh a single person has right if if they have to rebuild every experiment they want to mess around with from scratch right or

128. 59:03

     like if it if it's if it requires a lot of effort to replicate the code base even if it is published they will only do that many even if they pursue each of them right if if that amount of work does not detract them from even pursuing it right but if it's uh if it's somewhat standardized easily accessible right like you know hugging face style but maybe even even easier than that um then

129. 59:26

     of course I think people will Tinker much more with it yeah so there sounds like we've we have like three or four new projects in a web 3 that we discussed in this uh in this office hour so for those who are listening you know if you if you're interested in building any of this apply uh on uh hdn doxyz and uh thanks guys for joining I think it was a great discussion um and excited to work with all of you guys uh going forward in this

130. 59:57

     space I think we all going to make a ton of progress and uh we're going to start showing kind of Skeptics of uh open source and user own AI soon why they are wrong thank you thanks for having us guys thank you