DappCon 2023: AI, Data, And You, Driving AI Precision With Decentralized Access Control - Don Gossen

Transcript

woo packed house really draw a crowd um first off uh thanks to nois for inviting me to speak today so I'm going to talk a bit about uh data uh AI how it relates to all of us um some of the Precision that we can start to generate and I think some of the the influence at web 3 we'll have on all of this uh just to sort of take a step back as I'm sure most of you don't know who I am I kind of lurk in the background um so I'm Don gosson I'm the CEO co-founder of nevermind so we are at we operate at the intersection of data web 3 and AI um my prior experience I'm a subject matter

expert in data and analytics if that really means anything but uh I worked in IT consulting for the better part of a decade and a half and then uh jumped into the crypto space in 20177 co-founded a project called ocean protocol launch V1 um I've got another company called Kiko where we do web three services and like I said more recently been working on never mind for a little over a year and a half now though it's been incubated over the course of the last uh three and a half years so I'm going to talk about the current state of AI what's going on um how we think web 3 can help and then I'm going to leave the conversation and set

it up for my colleague David um with respect to where we kind of think things are headed what we should be looking out for hopefully give you guys something to think about so I think what we've recognized what everybody in this room uh can see is you know there's been an explosion on the AI front in the last year or so this is a field that's been around for the better part of a half century Maybe more um but it's only recently that I think it's really captured uh the Public's attention and and entered into the Zeitgeist and these are some of the big players that are out there that are doing things in the space um but if we take a step back and we look at okay

let's decompose Ai and what's what's critical to the the current state um there's three sort of real constraints at the moment the first is compute so um for the newest generation of AIS and by that I mean like Transformers diffusion models Etc you need a particular type of um Hardware it's gpus or tpus on the GPU s specifically like Nvidia A1 100s h100s Etc um and there's a real constraint in accessing those the next piece is expertise so you know you need people that know what they're doing in order to develop these models train them and then

also deploy and maintain them and the last real constraint is you need ship hous of data the these new architectures require a ton of information be fed in them into them and that's how effectively they learn um so the these lead to some drivers and and some additional constraints um if we look at the computational side and and we want to predict where things are probably going to head in the future and what constraints to really focus on um I'm biased and we'll get to that in a second but if we look at computation the analog that I use here is projecting backwards to you know 20 25 years ago if you wanted to do Advanced analytics back

then um what you needed were appliances so these were like fit for purpose technology racks um that normally you leas from one of the big software providers like like IBM or Oracle or sap um that would handle like one to a handful of use cases and they charge you a few million bucks a year to support these use cases um so what you found was they they were limited in their application usually to spaces like Financial Services things like flash trading Etc and then uh at the beginning of the the 2000s Google released a white paper on a technology they created

called map ruce and the team at LinkedIn led by Doug cting they took that paper and turned it into the first uh open-source scalable multi-parallel processing or massive parallel processing solution called Hadoop and that was the Catalyst that led into big data and the associated or or um adjacent uh machine learning capabilities that emerged over the last 20 years or so so so what we see here is from a compute point of view over time we address this usually through software um and drive the cost down with expertise we also see a similar uh deprecation in in cost so basically you

know if you look at data science when that came out as a field very expensive to find people but over time um those costs diminished what doesn't tend to uh reduce in price over the long term is the data um and that's for a number of reasons the unit cost might go down with like data legs Etc but um the the total cost uh for capex and Opex usually tends to be pretty uh stable or maybe increases because you're adding new use cases Etc to these capabilities so from this you know we've taken the position that in our opinion um access to data is going to be the biggest rate limit SAR

to the advancement of artificial intelligence going forward and so everything that we do is kind of hinges on this thesis and so now we're going to get into the nitty-gritty of it so ai's got a bit of a dirty secret um some people will know this others may not but it needs your data um it needs contextually relevant information in order for it to be useful so what's happened is that these companies that are creating these models um over the last 5 years or so um especially with respect to uh Transformer models and and diffusion model architectures we've seen this move to um

suck up as much data as possible and this this information is being sourced from assets that are freely available from the internet so that sucking noise you kind of hear in the background are these models and the companies behind them hoovering up all of these readily available data sets that they can get access to from the internet um the issue here is that up until recently this wasn't a huge problem because it wasn't on anybody's radar but now it's starting to become an issue due to the fact that these models are incorporating freely available information that isn't really free um you know there there are um

rights Associated to this information and I think the more important aspect is that these data sets are being sourced and used in inferences that then go on to be commercialized and the contributors of these Source assets they receive little to no recognition for those contributions and what we're really experiencing again is you know a proliferation of the predatory um system that that exists uh from a or predatory um um commercial model that exists within in particular big Tech right and this is again uh proliferating itself into the AI space

we're talking about P piracy here at the end of the day um and we're all kind of attuned to it based on what's happened with social media but I think you know this is something that needs to be addressed um in order for us to really Advance Beyond uh the the current state of where we are because what we're looking at now the reaction to all of this is to fall back to Legal uh legal means right litigate the out of all of this every week there seems to be a new case emerging you know you've got Sarah Silverman s suing open Ai and meta for you know stealing her books you've got stability AI being sued by Getty you've got the New York Times

contemplating a lawsuit against open AI you've even got companies like Microsoft last week coming out basically with a strategy for handling all of this litigation that they're anticipating telling their customers actually we'll pay the cost of your litigation if you're sued when you use our models which is like seems pretty irrational uh when you think about the the the State of Affairs there's got to be a better means to actually work with this Tech technology and make it Equitable um the result of all of this is kind of this um you know Divergent state where on the one hand you've got those that are

subject to this predatory uh business model they're being exploited and they don't have the means to prevent that exploitation um I used Sarah Silverman as an example in the lawsuit she would be an example of somebody that doesn't have the technical capability or knoow how to necessarily prevent this exploitation um and that's just you know that's leading to to General Chaos on the other side you've got organizations that do have the wherewithal the technical capability and the competence to realize what's going on and their reaction is to shut everything off so you've got the reddits the stack overflows the Twitter X's of the world and they shut their they've shut their data off to these models um so you've got this dichotomy now on the one hand

it's everything's freely available and uh you know exploit subject to exploitation on the other hand you've got the scenario where everything's locked down siloed isn't interoperable um neither of these outcomes seem great or totally viable over the long term so is there something that we can do that Bridges these two outcomes um the other issue to point out when with respect to AI is that you know the fact of the matter is we're on a trajectory where we're likely to run out of data uh and by this I mean data that's like freely available and scrapable from the

internet there was a report put out in October of last year by some researchers at MIT they predicted that as early as 2026 we could run out of high quality um text based information that's uh publicly available and scrapable from the internet so there's there's issues with this there's also the fact that you know these models right now are effectively trained on the same Corpus right the Corpus being the internet um and the result of that is that regardless of the architecture whether it's llama 2 or Bard or gp4 because they're trained on the same uh core Corpus uh or Source Corpus when asked the question these models

generally all converge to the same response so if you look at this these all all of these uh Confluence of issues coming to a head we're looking at a situation here where models in order to remain relevant are going to become incredibly competitive with one another uh to maintain their usefulness right because the the assumption is they exist to be useful maybe not to us but to one another sure um and so that's going to create a market that's incredibly Fierce um and the question then becomes how do we take advantage or how do we um ingratiate ourselves with this Market in our opinion two of the ways to

to do this and Tackle these issues that I've highlighted is to focus on two sort of um uh adjacent or symbiotic attributes the first being consent the second being attribution um so from a consent point of view what we're really talking about is informed consent uh the ability to elect whether or not an asset is can and should be used um for a particular application by an AI and if you can establish that consent um also establishing fine grained attribution now they're related to one another because if

I if I'm going to receive attribution if I'm going to be recognized for my contribution to an outcome right I'm going to provide a data set to a use case where an AI is going to provide a solution and that solution is then subsequently monetized and part of that recognition is I get rewarded in terms of I don't know payment or some other form of reward um the likelihood that I Will consent to the usage of my data sets for that use case I think increases astronomically so we need to establish these um uh characteristics of of consent and attri distribution and really focus on these

three pillars right uh especially as it pertains to data usage by AI so the three pillars being consent attribution and then sort of the the uh underpinning to all of this which is Providence so if we just take a step back again and look at what are these limitations that we're facing from an AI development point of view or facing issues with lack of access um we're facing issues with lack of transparency where are these assets being sourced from um how do we trust the information that's coming out in these inferences and then the third piece is we we're seeing a lack of monetization or commercialization opportunities Beyond

just the singular organizations that are putting out these large language models and diffusion models just so happens that blockchain provides a pretty elegant solution to each one of these issues now um so you know if we can establish Providence and by virtue of that fine grained attribution and Linkin aspects of of consent management um in particular informed consent we create an environment and a solution that makes uh these pipelines accessible um traceable and with embedded elements of Rewards or monetization

from our point of view what are we doing and never mind uh we're focused on an A type of implementation uh from an AI pipeline development point of view called retrieval augmented generation uh AKA in context learning how does this work what we do is identify a a corpus of data um so this is basically you know a subcorpus of information that you're going to Source um in response to a question that uh pertaining to a particular use case um so effectively what this acts as is like a filter set that we're then going to retrieve our responses from the way that we enable

this filtration is by taking that Source data and running it through the model filter so you take the filtering from a large language model and you create vectors or or embeddings and with those vectors we then append to each Vector a did that links back to the source document and correspondingly the wallet or wallets of the contributors of that Source information we index all of that or put it into a vector DB and then we implement or call upon a large language model or a diffusion model whatever the case may be um to basically orchestrate the solution or inference based on a

series of questions that we ask that llm what we're doing is telling the llm to only source from that subcorpus of information so it's not actually sourcing from the latent space within that model it's only sourcing from this subcorpus of vectors and the dids that we have created um for this particular use case and in that way we can isolate the The Source material we can restrict the surface area for the response and we can actually identify what the source um information was for a particular response or inference that the llm made and then from that you can you know uh enforce all of the aspects of uh

attribution Etc um whether that's royalties or residuals or whatever if in the case that that that result set is actually commercialized how do we do this so we've created a hub um the never mind Hub you can find it at nevermind. apppp um but it gives you the ability to register your assets so this is where you create the dids for data sets or models um you can actually establish the the components for Access Control who what can access um this these assets uh the ability to publish to an ecosystem none of this would be any good if you couldn't find it and then lastly elements of of reward or monetization for the assets that you contribute and

um how that contribution is actually uh reflected in terms of recognition and payment so the last thing I want to leave everybody with is sort of some thoughts on where we see things going um you know there's been a lot of interest in what is going on in the AI front I think what we're seeing are Beyond strictly speaking the development of these models is the incorporation of these models in particular into consumer applications to make our lives easier um I think that's deflecting from where the focus really should be in at least in our opinion and that is on the agents

themselves these AIS that are emerging um these AIS are going to need the means to transact and they're not going to have bank accounts I think that's just reality so how are they going to interface and transact with us or with other AIS or third party Services I mean obviously we're biased but we believe that that's with Wallets on chain um so now the question becomes okay how do we facilitate this emergent uh ecosystem of agents that are going to uh either collaborate or potentially compete with one another what kind of infrastructure do we need in place obviously the means to transact but also the means to Signal

what's valuable or relevant to their particular use case etc etc the analog I use in this case is kind of um you know Richard Nixon opening up China in 1972 um and the effect that that had on the global economy right prior to that China was non-existent in the global economy and then all of a sudden the Machinery started to move on one of the biggest economies in the world which really changed Global GDP so the what I'll leave you with is think about the GDP of these AI agents that are emergent and how that is really going to Trump the human rac's GDP overall if we

see the proliferation of these agents that we kind of expect to happen anyway I think David's going to talk more about the agent side of things but thank you for your time um I'll be around for questions after thank you very much d i I think we would still have some time for question yeah if so if anyone has any questions I see one here on the left side oh the mic is coming thank you for the presentation so you said the quality of data can you give any specifics on what do you mean by not good quality because yeah okay

this is a good question because I don't I think it's horseshit qu trying to qualify what quality means it's more relevance right um so how relevant is a particular data set to a particular use case the example I usually give is like um let's say uh you know you have a call center and somebody and and a user's address gets fat fingered into um your your CRM system uh that's bad data for me is a business right it's not quality but it is valuable to the call center operator the because you know maybe the

ambient noise is is is too much for the call center operator to hear the address properly or maybe ergonomically you know the setups poor so they're you know they're not capturing the right information alternatively maybe it's a shitty interface from a software point of view right so it's a it's more a question of relevance not a question of quality if I can then you would have a knowledge graphs combined with alarms I mean anyway I can ask you knowledge graphs are going to play a big role in all of this thanks all right any other questions in the audience otherwise I'm pretty sure you can find Don around here so thank you very much Don

thanks