Crypto X AI: Fad or the Future | BITKRAFT Summit 2024

right welcome everybody uh thanks for joining us at today's panel we're going to try tackle or answer the question well we're going to look at the intersection of crypto and Ai and try answer the question of whether it's a fad or whether it is the future um I want to give some cont for today's panel about why we're talking about this sort of thing you know being a gaming fund um and it's quite interesting being a gaming fund we have this privileged front row seat um at a lot of gaming Studios and developers uh working at the frontier frontiers of Technology uh typic you know devs and Studios um knew about Nvidia long before it was a household name um gpus were harness for their process you know their multi-core processing power to simulate physical

environments you know long before they were used to train AI models devs typically were early users of these the you know GPT 1's and twos before um it became a bit more mainstream procedural generation uh gave us a hint about what neural Radiance Fields would look like um or you gen uh models whereas even bots in Warcraft farming gold for you just gives you a bit of a taste of what agents could look like um in the CP crypto AI world and having a focus uh on on crypto as a gaming fund we took an early bet on a tech uh on a technology where we've seen behaviors um occurring in the gaming space already uh users were trading digital goods and you know desired ownership of these digital Goods uh these users are typically terminally online users uh that that form part of these kind of digital societies and you users are these users are typically

early adopters um of of just New Frontier Technologies so now we find ourselves at this pretty interesting intersection of crypto and Ai and there's this wonderful Peter teal quote where he he says um you know crypto is is is kind of you can think of it as as capitalist because of its kind of self sovereignity and the nature of this like libertarian nature that it has to it and AI is is kind of communist in ideology you know because it's this collectivist appropriation of data and now here we are on a panel trying to discuss how these ideologies are going to mix together which I think is quite a sort of tricky one so we've got our expert panelists uh here today um I'm going to hand it over to you guys just to introduce yourselves and we're going to try dive into this question so go ahead Trevor hey everybody uh my name is Trevor I'm from render network uh render network is the largest uh Network in the world of

distributed consumer gpus it was founded by artists for artists today it runs thousands of rendering and AI jobs in parallel at a fraction of the time and cost uh we were founded by a company called otoy um that really set the um set the yard stick for the first commercial use case for the GPU in rendering and uh it's been amazing to watch it evolve from rendering to become um rendering and AI a very important second use case for the GPU hey guys uh Karan here um I'd say I'm the co-founder CEO of mea we are one of the younger companies in the crypto AI space but uh I think we're founded on the vision that you know there's this entire centralizing force that exists in Ai and we believe that cryptoeconomic Primitives allow you to build really high quality Sovereign Tech in a space that's becoming a lot more controlled by a few companies our focus is actually on building um interesting tools across the

interesting Primitives across the supply chain that focus on kind of handing power and control of these tools back to the user uh currently we work on bunch of different Frameworks including agents models and workflows and are in the process of you know building really important I'd say evaluation criteria that allow us to be more Sovereign in the future cool yeah thanks for having me um I'm honored I'm an investor focus on early stage Investments mostly almost exclusively focused on this weird intersection awesome yeah so I'll start with you and you've been one of the kind of key evangelists of the space um one of the early investors let's start off just for you know the audience sake why why does crypto need AI why why does AI need crypto you know why why is there this intersection happening yeah it um at some point in the last two years it became just like extremely evident that a lot of the Primitives cuz we you know crypto we

love to talk about infrastructure right like this is everything kind of starts from the ground up and we've been building so much infrastructure for so many years and when you know to to current's point when you start to look at sort of this blast radius of issues that are starting to crop up in the world of centralized AI it felt like there are going to be just natural use cases for all these Primitives that have been built for the last 10 years maybe more so pressing you know like in the last four years if you were look at some of the Technologies has come out it was like all right I think the time is now so it felt like a natural marriage where crypto can support AI I mean just like simplistic things right like we talk a lot about I like I'll pick on just one specific example here I you know like meta's models like llama for example is like it's a it's a very popular Model A lot of developers they love the fact it's open open weights they can use it pervasively and you know you don't have this hangover effect of using um you know using open AI or other sort of centralized entities the reality is that like we actually don't know where this data that llamas models are trained on

come from the natural Evolutions that out coming from Facebook right like or Whatsapp or Instagram which is a lot of our data and uh of course then we start to see deals that are being struck between Google and Reddit open Ai and Reddit and then like then you start think about oh wait a second this's thing called Data provenance that exist in blockchains like this is a PR we've had for a long time can we build infrastructure to support users providing their data more systematically right so this is just like one case in point of like where we can support the AI workflow across the stack but this start to become evident and again it's very very early to current point I think we haven't really gotten there yet we starting to unpack this there's a lot more work to do so this is where crypto can support Ai and of course you know if you just think about like the AI economy and like how it's going to boost every single thing we do I mean just coming over here my way I was reading I was like oh you know we need to we need to get better Radiologists that need to be better supported by AI right every single field is going to get boosted by AI it's inevitable that AI supporting crypto is not going to become a thing so

in both these ways it just felt like you know it's rising tide situation think we're all going to win yeah K we last time we were sping speaking about that um scale AI Story related to data Providence you just want to tell that and kind of add to anoun point yeah I mean if you think of any API it's an arbitrary Endo right you're trusting whoever's on the other side to deliver the information that you need and that can you already see issues with that right some model providers say that they're giving X model but actually they're serving y mod model in making Arbitrage on the on the gains or sometimes even more importantly in case of evaluations which kind of Define your AI system you don't want to be in a place where someone's kind of giving you an evaluation output that's like not true or fraudulent or even just like you know like lazy compared to what you would expect so in scale AI for example last year what they found was when who who scale AI what they do oh sorry my bad scale AI is this big Behemoth in the AI space uh it started as a data labeling company but now it basically does everything data related to training fine tuning models every big AI lab

works with it if You' ever used AI you've come and you've touched scale ai's work in some capacity so they're very entrenched incumbent in terms of what's happening in the AI space but last year it recently came out that while meta was kind of fine-tuning some of their chat Bots built on llama they had a contract with scal saying can you give us annotations on some data sets that need we need to find tune our models and the SLA say of course you need humans to do this right that's why you're paying Scalia money but Scalia was just using GP before to do those annotations so in that case you're basically creeping the bias of one model into the other model which is extremely dangerous because if you can do that inadvertently you can also do that maliciously so what systems exist to prevent that Beyond like a contract SLA I think that's where blockchain Primitives become super useful amazing can I be contrar there it I'm I I don't know if centralized AI really needs crypto um as much as we talk about in fact they've done everything not to use crypto when you look at how models were

trained and and how services are delivered I I think really the battle is more um from a consumer and from a Content owner perspective um and and it's open source versus close Source in that in when when you look at these models once they're trained the Gen is really out there's very little you can do there's billions of parameters you know you you can't pull your data and have a meaningful impact on that um but we're what we are seeing is um a real uh shift in folks wanting better ownership of their data that's used in these models and and to do that it it has to be done in a public decentralized way and when when you think public and decentralized for me some of the best examples are crypto Bitcoin really being one so um I I see it as a um as a movement that has not yet embraced crypto but really needs to if it's going to be done the right way in an open source way why do you think that is is that like a threat to the business model um I I you I speak to researchers and

operators in the non-native sort of crypto field and they see a move from uh stability AI Imad uh most like sort of moving into the crypto space is kind of selling out or kind of looking down upon it do you feel that you know you get a sense of that or why do you think there is this bacation yeah maybe I can start I think uh so I have experienced this because I come from a very AI native web to Native fund and my bosses they still ask me me why do you like why you doing this crypto AI stuff right but uh and they were the lead investors in scale AI last round but that's a different story but uh but I think the interesting question there is yes they do things see crypto's yet to prove itself on a lot of things and U AI also I'd say it has more you know viseral pmf in terms of you can taste and feel a lot of these things but how I think about crypto Primitives is it's more of an insurance policy right you the risk is it it might be janky then it might be more friction but the

risk of not having it in certain use cases a lot higher than the cost of actually putting it in and you realize this like you know people didn't realize stable coins were important till like the Venezuelan government was de banking people right or like stuff like that so you don't realize these things matter until suddenly they do and that's like the I think the view you should take of a lot of these things because open source is it in itself hard and open source in itself is not really capitalistic and to be to be successful in any platform shift you have to be ruthlessly capitalistic crypto for the first time brings capitalist Primitives into the open source world that lets them compete may not be in training models because it has this uh and yeah you know a lot of cool companies like Prime intellect and stuff doing things on training but you know just coordinating resources and giving users the ability to own a lot of the resources that they're kind of contributing is going to be like important yeah you um you touched on decentralized training uh and on you know I know you've looked at the space with the kind of 30,000 foot view um could you talk about some other sectors

uh that are interesting in the space we have there is a market map that Friends of the fund um Casey cuso is sort of pushed out that that is quite uh uh exhaustive uh feel free to pick any interesting sectors over there I know you've invested in a couple of uh really cool companies like I think Jensen and ritell um can you just talk about a little bit about what they do you know we've spoken about data Providence and sort of crypto benefit there I think we we get that um what other sectors are interesting to you yeah the uh the learning for me as a firsttime investor is that um this is this is going to sound ridiculous but I don't invest on things that are on a market map uh you know so like the best things to look for when they're not on a market map uh those have been the best investments there's some companies here that showed up when there was no Market map right so I think that's the way to think about it but um I I think if you look forward I or like the areas that I think are particularly exciting are yes there's still a lot to unpack when it comes to data there's um the uh right

here and now like there's even when we talk about uh the utility of air like going back to your previous question like why are we getting so much you know push back from AI researchers on like the utility of crypto I honestly I think this is this is warranted right like we have kind of dug our own grave here we and I think in some ways it's deserved like in a lot of ways because we haven't really built anything that that's crossed its chasm in terms of mainstream utility right so so just on that point um the I can tell you that top of mind for a lot of companies including like self-driving companies or companies that are doing you know Enterprise search at scale companies like glean uh you know privacy and security is very top of mind for these entities right like to your point the trust Factor when it comes to using um you know a close Source model and what is being done with your data when you have to provide it for inference or when it's being used in a rag likee environment these are like questions that are coming up now that are starting to become very interesting so I'm really excited about that as a as

an opportunity Zone because an opportunity set um but I'm kind of veering from your question yeah there's we've invested across the stack um it it felt like training is is here and now you know this about I was just doing this my own internal sort of like cataloging of where we are but like there's there's 13 different sort of research attempts of like distributed training in The Last 5 Years right and I think we're like sounds like we're inching and getting there like you've probably been closer to this than most onchain inference inference just Define inference for uh yeah it's basically the touch point you provide a prompt it provides a response that's the most tangible experience of inference that you've probably all experienced today um but yeah you know I I'm not trying to shill our companies per se but you know it's been it's been across the stack and but I think there's opportunities in security and privacy and data I do think that the next 6 months we we really don't have any cohesion when it comes to what agents really mean like I think we all kind of like pull things together like oh this is what we think an agent needs to look look like but like the proliferation of Agents hasn't really

happened yet and I'm very excited to see that kind of take off actually quick on agents defining it for the audience again cuz cry has a lot of sort of jarg and lingo it's very you know we're used to it but just for everyone's benefit yeah there's there's again like I think between us we might there might not be consensus here but um agents are uh I mean let's use like a real world analog it's like you have a travel agent right like someone who basically you give a set of instructions to and they front of execute the task for you and then the analog for that with large language models or a mixture of these is that like you don't have to do this in sort of a one-off manner with prompts one at a time it kind of like takes the task and kind of delivers response for you in a very reliable way can so on for me on large language models uh training is is really better done centralized uh you've got billions of parameters you you need to have um available memory and networking between these devices that that makes sense in a centralized infrastructure now now that's started changing you saw

drro paper come out quite recently that had an exponential um reduction in the communication between nerds but just for the audience can you sort of Define like why is this the there this distributed training sort of Paradigm and and quite simply on on training you you have so many parameters you're trying to load into the model to train that it's very hard to do on a decentralized distributed basis right but that actually flips when you think about inference and what inference is is really just running the model yeah and and on the inference side you know when stability AI came out I think it was 33 35 gigs um over a couple of weeks let um to months it became something you could run locally on your cell phone um and so you see a natural progression in these models to where that fits a decentralized architecture really well but probably the the biggest uh recent change was for me when open AI brought out their paper on 01 uh they they showed there there was actually point in learning um on the training side where

they' really maxed out and um what they move towards what they is what they call test based learning which is running inference on a problem in parallel and why that's so interesting to us is because those are problems that you can kind of set it and forget it that work really well on distributed networks where you have a swarm of different people working on the same problem and because it's inference it's not a massive problem for them to download and and too many parameters for them to run on memory it's something that can be done in that distribut swarm way so the way I see this is uh there are many many opportunities here but but really you've got to find the opportunities that fit the current state of the market where where jobs that are being run particularly decentralized need to be done almost on an offline basis like we do rendering yes um instead of real time and if you're looking to do things more real time if you're looking to do things more centralized training then I think you're looking in the wrong spot right now uh and that may well change as as this evolves because it does change so significantly week to week yeah um that

distributed training in terms of the the benefits of it and you know a kind of decentralized Paradigm what are the key like value propositions of that for people working with these things is it speed is it cost is it kind of trustless in some way how do you think about that question open to the floor yeah I I can uh I can take a crack out of this I've been I I've been working on something because I've been just from like my very naive point of view I needed to understand this a lot better so I've been trying to distill this down in the Eli 5 kind of format um and the Crux of it of like but you know just to counter to your point right like when it comes to Frontier models I agree with you I do not think that you know we can beat what centralized entities are going to be able to pull off it's just not going to happen however I think for a vast majority like and that is kind of where we are right we have Frontier models and then we have a long tail of stuff on hugging face um we I we need to be able to get to training fine-tuning uh pre-training and fine tuning in ways in which we just haven't had access before

and there is a compromise that researchers academics startups here maybe are willing to make just to be able to get access to pre-train fine tune without having to commit to seven figures and spend and you know this better than most right so I think that's the yeah go ahead no no that's perfect any other thoughts good um I wanted to touch on another sector very quickly the Rena Network started off in the compute uh space could you just sort of crypto's really good at at coordinating stuff can you kind of Define that give give a bit about your history and then what are you doing now so we we started just to solve a real problem we had uh one of our investors this is at otoy was Madison Square Gardens we were helping them do a render and we were trying to do the render using the public cloud and doing the calculation it would have taken all of Amazon's West Coast GPU computes 6 months to deliver this project and it was just a no no we needed to get it done in 3 months so it it made us step back and and um the real opportunity for us was um the Enterprise model that

Nvidia has had led to a bacation in the market where there's an absolute glut of consumer gpus but an absolute shortage of Enterprise gpus so we turned our attention to the Enterprise side and and the lucky thing for us is is rendering is typically of animations and animations are are in frames per second uh 24 frames per second for a movie so they were kind of naturally set to a distributed model where we send each frame out to different note and and that was the onset of the render Network it scaled fantastically but I'd say what we got wrong at the start was uh crypto models were really not designed to incent users in the way they are today and about a year and a half ago um our community voted to change our model completely to a more modern model uh closer to many of the deep in um projects you see today and and that's allowed us to allocate resources across the network in a much more efficient way uh it's allowed us to open up new business model lines like Ai and

allowing Partners to write to us and has allowed us to to Really scale on the back of crypto as a means of of incenting individuals who want to contribute to a decentralized project in a way that just was not possible 5 years ago when we started this yeah amazing on on the computer side I have like an interesting take here is people always talk about oh how we're reaching like a top and like compute demand Supply and stuff I feel the genie is out of the bottle now like if you assume AI is going to be here computers always always going to be needed like right now we're bottleneck on like we were talking about this right we're bottlenecked on like human in the loop and human minutes and a bunch of stuff you you as those things go away like the velocity of inference is just going to keep rising and now with like o1 saying that inference scaling laws exist and stuff I feel like it's always going to be a shortage and figuring out how to constantly acquire and coordinate you know new sources of compute is going to be a very important business regardless of like whether it's centralized or decentralized and I think decentralized ways are going to be like

very interesting way of doing it well the one stat I had um is there are there are about 6 million gpus in the public cloud with the hyperscalers there are 1 and a half billion gpus worldwide outside of that those 1 and a half billion are not all performant but very much a subset of those and multiples more than 6 million um are available for for AI and for rendering should many of these problems be solved so the opportunity is there and I think the need is there uh it's a question of of of Bridging the chasm amazing um just in the interest of time guys uh just a question what's a view or opinion on the space that you hold that you think a lot of other people disagree with you on and feel free to get spicy spicy and controversial yeah I I'm I'm not really a fan of of zks ZK sync can you def couple other haters here well there are good use Define can you define it C sink is is an encryption technique that's quite popular in crypto at the

moment I I think um for a client side if it were um used for identity and and not needing to update State I think it works really well but where I see crypto trying to use it a lot at the moment it's trying to use it in a different service side way and for me the the cost to perform that is just staggering and and it becomes a Time issue so I'm quite hopeful other Primitives like homomorphic or or others come to the four but I'm definitely antiz what is homomorphic encryption it's it's a different type of encryption one that is end to end so on the data being trained and and what that means is therefore you have state stored you you have the ability to do a lot more than you can with a a very much slice in time that you get from zks um I I do think there are ZK use cases but I'm much lesser on many of them that I see today spicy anyone else yeah I I think I think my heart take is like as a Founder building in the space it's always a struggle for legitimacy I feel like like slapping AI

into anything in crypto and calling it crypto AI is pretty easy to do and it's even easier because nobody really understands AI Beyond something and you can kind of a lot of projects have tried to do that and you can make it sound legitimate but you know at the back like as a builder in the space isn't and it's hard for us as Founders to kind of you know if I was building this as like a YY back company no one would ask you and say are you a scam they'll just be like yeah there's a v company building something interesting so that fight for legit legitimacy is tricky but I'd say you know there are some the the hot take is I feel that there's going to be like a like there's going to be a black swor event in Ai and there's going to be like something in the crypto space that either exists at Sol or you'll realize that this is why it really matters in the next like three years any guesses on what that Black Swan event might be I mean is it like geopolitical sort of I think it's just going to be like you know some sort of you know training data injection that's gone wrong or some sort of evaluation data set that's gone wrong that causes some big issues somewhere in

the world okay just you while we're sticking with you Karan do you just want to give an overview um just quickly on the mirror Network I know you gave an introduction but you know we had an interesting conversation last night about um um you know how how you Benchmark these these models and and all of these kind of open and closed Source models and um how you work with them yeah and you know evaluation doesn't sit on this Market map so keep an eye for that but uh but I think U so the question we keep asking for ourselves at ER is like why is block space important for anyone building an AI like where does extrinsic demand come Beyond like the traditional use cases of smart contracts and speculation and financial services and the real value there for us has been that so when we we've been building AI apps for a while now right and you know we have a lot of users but the issue is models hallucinate and you know you have humans in the loop today to kind of protect against that but to really have like agents at scale and all these things really come to life we really need to figure out how to go from like 80% 90% accuracy like closer to 98

99% right and evaluations play a big part in that like how do you give feedback to your systems to make them more robust more efficient and a bunch of these things and evaluations today are either they all of them are focused on quality measurement and not quality control so the things we we've been up to at mea is basically can we use block space and use consensus systems that you know crypto has kind of built out for llms to actually act act as automated evaluators and we've seen some really interesting research breakthroughs there of you know having llms as judges this is based on a paper that Google Deep Mind put out a year ago but having llms as judges really work together to reach consens consensus and output and that becomes like this trustless auditing system for your AI output because a lot of people focus on input parameters you know like instructions prompts data sets not enough people focus on the output parameters because that itself is like a different baste in itself so at mea we've been focusing on how do you create trustless evals just like you know you

can do evals to a centralized entity but you know just like how you can bank with a bank but also Bank on ethereum and there's value in banking on ethereum because of the trustless nature of the transactions I believe there's also value in doing trustless evals where you have a lot more transparency into what the eval system really looks like and just to sort of play Devil's OC good why trustless evaluation you know I think I'm sure people have seen on on Twitter that I don't know pick a model sort of surpass the the SAT score or did really well in a PhD exam like that seems like a good enough evaluation no so let's let's put this way right uh have you seen swe bench it's like The Benchmark for a like a software agent the ISS like it works like you might have an agent that works performs really well on swe bench but does it work well on your code base does it work well within your code base on a particular repository right you can always go into edge cases into edge cases into edge cases because the design space is infinite so evaluations need to become a lot more like end

entric versus like broad-based because broad based lets you compare But like everyone who's used AI here knows like it's a Vibe check like people still use TR error no matter how many like evals exist right so how do you build more trustless evaluations that can go as narrow as possible and there's a quote here from Brett Taylor Who's like the chairman of open AI he says that software is moving from rule based to agentic systems that are goal and guard rail based right and for that you just need to have good goals and and you need to figure out how to keep narrowing your guards because otherwise you're just going to have super inefficient systems yeah amazing uh and I know investors sort of need to be quite friendly to other people in the ecosystem but any spicy takes or or views you have that others disagree with you on I I don't know if this is one that people disagree on but I I I constantly think about the fact that this might be the last jobs we all will ever have right like if and you know I'm not a Doomer I'm not trying to say that like AGI is gonna kill everyone um but I do think that we need to

operate with this mindset that whatever like at least for me you know and I I've been in this industry for for quite some time I think of this as the last thing that I'll ever do and it it feels like AI is going to rewrite the rules for engagement pretty substantially and using that as sort of a mantra it just forces me for to prioritize the kind of things I look at the kind of things I want to engage in and for Founders and others like those are the kind of people that I want to partner with so it's it's a it's a spicy take in that you know we're we're moving at a at a pace that is sort of unprecedented that I've I've really never seen and I've been in I think I was talking to these guys here like I've been in software for 25 years and this is kind of nuts kind of what we're able to do today so uh trying to stay ahead of this is going to be fascinating I want to open it up to the floor if there are any questions uh feel free Mr carton I know that you're involved in this space so if there any others feel free

crypto the question is how crypto aligns with agents I I don't see it very different from um models it's just a an additional level of functionality on what's going to be run on a GPU um and it's going to need GPU power to run it and that comes down to core incentives on GPU power that I think you see a lot of people out here um chasing and and doing a good job of of of meeting yeah I for me it's a um I think like if you just look at the stack everything from like TCP IP htdp rest you know this has been constructed for um an API driven world that we live in today if I want call if I want to use an Uber I call the app the

app's using you know some Services architecture in in the back end this was meant for like the you know the internet era Circa 2023 right like 2024 2025 onwards like if the goal and by the way you know open a is working on a device now assumptions are that it's going to be largely Voice driven agentic in nature the the modality of communication between what we consider be llms today which are largely non-deterministic they can hallucinate a lot like things like this I first of all that's going to change and underly I don't disagree they're not going to run on gpus but the way we agents interact with one another for us to perform certain tasks I think it's going to look vastly different and I don't even know if it's going to be on the traditional stack that we're used to and we're familiar with where that meets crypto is I I hope it's it's a part of the fabric on which on which agents are built because we're not just talking about naturally just like attracting or extracting information we're talking about intelligence here so making auditable verifiable secure Sovereign I think all these things should first principles that we should start to build with today as opposed to making it an

afterthought so uh more wishful thinking on my part there yeah I think um I think it's mainly about you know having verifiability of actions because these these things are going to like they agents they're going to have agency they can do things on your behalf so how do you verify what they do exist because if they can do things on your behalf they can also off skate what they're doing so you need like Sovereign systems and like I'd say verifiable proof technology to make sure that you know you can actually validate all the stuff and audit all the stuff that they're doing so I think that's where it becomes more important there's this take that you know agents are going to use crypto to pay I don't really buy it like if they can use crypto to pay they can just use a bank account no one's going to like change that but um I feel verifiability of a agentic actions is an open space not too many people talk about kind of a follow related to that is I think an unexplored area is the legal liability of large language models when they do causes damage and the legal liability of the data that went into

that model fact it's the data not the model that caus the Dage and the attrition of all that who isable is a meaning risk and does crypto have a and to but defat and they won and was pissed they were like can't yeah just so for the audience I should have given you the microphone but the question is on the legal liability of LMS and what role crypto has to play in there so I I read the story recently where some um travel company some

airline company put a random you know chatbot agent on their front page Canada yeah in Canada and like basically someone just like and you can always hack the prompts right and you can kind of make them do things you want so someone just made the llm given output that yeah you can get a 100% refund because you missed your flight or something and it was admissible in code because now you the company had to give it a refund so how do you get these agentic actors to operate in a way that that doesn't have any backlash towards you and how do you all audit the stuff that they've done because right now it's still preliminary right it's just chat interfaces how do you track actions and not just one action but the chain of actions that ends up happening so I think just having proof of all those things is super valuable yeah I'll just say this is like a request for startup I think the the the the idea around like what insurance and liability looks like in this world is going to look drastically different I'm just not saying you know ml for insurance companies but like specifically when it comes to inference if a radiologist is going to misread you know report uh it's going to have drastic interesting consequences yeah I

don't know how you do it I really don't cuz there's so many billions of parameters in these to point back to one is just about impossible I mean the model takes so much time to train to actually find how much weight went into that data source is impossible um so I I don't know it's very black box uh I don't know how it's solved good I think we'll end it there guys I think you can see you know we can really go down the rabbit hole here um and I think we definitely can at about 600 p.m. if you want to join us upstairs for a drink um so please give it up for uh our panelist today