BBW24 – Day 2 – Securing Truth in the AI Era: Blockchain's Answer to Digital Deception

good morning ah this is my like my third day in a row second conference all right so let's get started why don't we start with a brief introduction of the stuff that you do in your amazing companies I'll go first hi I'm Ginger from Taiwan and I'm from nbco I'm head of uh e uh ecosystem growth and I help the create content creators and also news Medias to Pro uh protect their copyrights and give them the easier way to monetize their contents and uh personally I have more than 15 years of experience in uh tech industry I you I used to work in tsmc as a semiconductor industry before and I transfer into uh the blockchain industry in numers bro and I hope you guys will enjoy the uh the conversation here awesome well it's great to meet you both my name is Christian uh I'm the co-founder and COO of a company called sanctum AI um we can get into what that is in a second but my background is primarily in blockchain I first got involved in 2016 I spent about six uh three years at consensus leading one of their earlier uh portfolio companies called altho we were building very early ethereum analytics um too early but uh that was exciting and then uh I've also been involved in Defi and very excited to be here today and talk about this uh this topic oh well I mean AI is a is already making like a great impact not only on the different Outlets every sector is being like literally disrupted about all the potential that you have but at the same time if it's not handled properly this can lead to a lot of other type of disruption which will be penalizing from one side you know the professionals from the other side the AI and I think that is inevitable that until we get there there will be a first real merge between two technologies the AI from one side for the generation of the content from the other obviously is going to be what today is guaranteeing the truth which is the decentralization of blockchains and here my first question CH it is uh is the content really truly ours are we still you know owner of what uh we are generating or or no uh thank you uh it's a good question I I'm going to see my script uh I mean in in this ai ai era uh the question about content ownership becomes complex as AI system can generate and also distribute content uh the traditional definition for uh authorship and also ownership are challenge while uh you may create a Content a can produce variations adaptations or even entire new works based on your original input so they blood the Blurred the line who owns the final product and the key to ensure that the Cent remains truly ours uh is to establish and maintain clear Providence and right management uh this is this is where technology like blockchain comes into play by embedded content with metadata that uh Records its origin authorship and other any any other information you want and creators can ensure the uh their rights and uh maintain control of their own work so even in this uh AI continues to evolve so in this ER it's crucial to leverage this technology to protect IPS ensure the content ownership is respected and also preserved yeah in my opinion it's not only about what's the content that is generated that needs to be certified that something I've been preaching already in my my previous panel which is uh but the other thing is also from if the output is important from one end the input so the training models are also important from the other end because um improper training models uh training data can generate bias and uh and you know somehow outputs they can have a bias so that is a is a very a very important uh uh thing to consider when we talk about AI it's not only about what you generate and here is my other question for you Christian and it is what do you think how do we think it's good to where is the good balance between the you know the ation uh and the Integrity of what we are doing with the AI yeah I think just one quick comment on the last piece right up front I think um just to your point on inputs I think it's very simple and it's very similar to a lot of uh search engine technology social media the inputs that you as users share with large language models web-based models these These Are Go essentially improve those models over time and the more s of that information is um the more the less you actually own your information um what I mentioned before so sanct we're essentially building an offline version of chat gbt so you download a quantized version a smaller version of uh an open source language model right now the leading model is llama 3 from meta and you essentially have your own offline version of chat gbt so if you were to go to chat gb.com everything you put on that website is essentially logged against your data profile and is used to either tune that model or train that model in a future iteration for where whereas if you maintain some of these capabilities offline you're more or less able to keep control of your your data and also those those outputs the balance I think is important because web-based language models are very powerful and they can generate responses very quickly and they can handle multimodal inputs and generate movies videos so there's very powerful uh possibilities when it comes to using web-based language models but um The Challenge there becomes these large companies taking advantage of the inputs that are put in whether it's from artists musicians or everyday users so you have to find you know for yourself what information do you want to share with these models think of it like you're sharing it with Google search behind the scenes there's not only a data profile that they're collecting and then selling there's also uh you're collectively improving and enhancing the training of these models for future iterations so it's just the balance really depends on the individual user and how much you want to share um you know right now power is kind of coalescing in a few large companies so you know how they approach balance and Innovation is is you you kind of look to to history to see how that goes um but you know this these tools are very empowering to us as individuals we can we can choose what we share and we can um ultimately decide whether or not we we believe the value of sharing information for a response or or an output um outweighs the the cost to us whether it's privacy or or um some some sensitive information you have to share yeah anyway you this this question was uh this this topic was really raised by some controversial allegations on some of the popular llms that someone have been accusing to be RAC or to be sexist because they just trained with they took the internet that DP in and then it happened that the majority of the content online is is is designed is Britten by certain uh ethn and so and then in the end that's the result so that's why I think in the future especially uh there will be regulation in place pretty sure we will get there and I think the best thing is to provide the right tools to make to support the upcoming regulation is probably the best approach is to anticipate the regulation building models and tools that they can take advantage they can leverage and enforce because that's the the other thing I mean everybody here as children and you know or someone or a friend and with the Deep fakes now it's uh it's incredible what you can do the damages you can do is uh it's a very very important thing and uh what can I tell you I'm going to ask this to you because that's pretty much the core business of uh numbers protocol uh if you believe that blockchain is really going to save us from Deep things uh that's a good question uh actually we can provide uh trustable uh trustworthy tools to prevent uh to help us to fight the defect but I think think the answer is no because uh in this world uh if your assets or contents are valuable people will always uh fake it and and use it to profit right so and you don't have currently you don't have a efficient way to verify those contents on internet so is it's it's not very easy for people to verify uh if the the contents are authentic or is made by uh generative AI or is photoshopped by others so the I think the bet uh the better expectation we will have is to have a tool to help us to uh from the beginning to the end uh for example uh if you embedded those metadata into the the digital content for example the pictures or music files or videos and then uh for the information of the creators creation time or any uh gen AI tools you use on and embedded on blockchain so that will make the Providence immutable and also traceable so those tools might help you to make better decisions when you see uh if the picture is real or not uh for example uh num sco had collaborated with Reuters for the uh United States presidential campaign uh in 2020 after the campaign there a lot of going on happened right so uh we help them to preserve all the evidence on blockchain also on internet so maybe our uh in the future people can revie or see those uh incidents and to make the Judgment themselves because uh you make the pictures uh say uh tell the stories themselves so that's uh we can we can provide a better tool for them to do the decisions what they do as s to AI is pretty much the the answer so you can enforce whatever you want but always be there will always be someone that does a computer a private computer where they can run their AI that is not subject to control my opinion is different and is in in the other direction you need to obviously regulate the AI and enforce some tools on the AI if it's public but then you have to enforce the same also on people if I generate content the tools to generate content should be you know track the content I generate and I should take full responsibility for it it's not something you can control in 100% of cases of course but if you think about today you can Photoshop someone you can already do some some damages it's just that right now we tend not to believe so much to pictures and believe more to videos but we we start seeing even with the current you know video editing to start creating pretty convincing effects even without thei but we'll we'll get there and here is uh the question about how the ethical AI looks like in the the develop sure yeah I think just one quick comment on um deep fakes I think you know it all depends on what the goal is um blockchain cannot do anything around preventing you as an individual from reacting to a picture what it can do is try to establish that via some tagging mechanism this is an original unaltered image or video but most people will not take the effort or time to pursue to that end so I think it's important to realize that deep fakes are here they're going to get more convincing and it's important to keep that in mind as as we think about what are some safeguards we have in place and I think you know the more that these large language models by the bigger companies like open Ai and anthropic and Google are becoming ubiquitous they almost present themselves as public goods but but they're really not because the employees that are in the company policies that are setting the weights Behind These models and how they're trained and tuned are not publicly known there's no transparency so the images or the responses that you receive based on your inputs are entirely determined by waiting that you have no control over so I think open source language models help provide an answer to that because you can see the weights you can adjust the weights you can correct the weights now with that will come perhaps some negative outputs for example but the question becomes how do you define accountability when it comes to outputs from AI now with free speech you you kind of set the bar pretty high you can't say bomb on an airplane for example so how how who determines whether an image is damaging and who determines what the pp the punishment for that image is and I think the challenge is we we don't know what the right answer is to who can DET make that determination so I think deep breaks are here they're going to be influential we should be comfortable with that and and kind of start to make our own decisions around kind of what what that means for us and our decision- making and all that I think ethical AI in practice um I think it's going to be a challenging road ahead because right now you know we are seeing again a concentration of hyperscaling companies that are essentially dominating the public language Model Market dominating R&D and they're determining essentially the responses that you get from these models um so you know I look at the startup ecosystem I look to other competitors of sanctum other companies that are building around open source language models and I and I'm encouraged that this can offer a counterbalance to the convenience of just going to chat GP chat gp.com inputting a question and then basing an action on that response because hallucinations are still real and I think our tolerance or or our ability to second guess something whether it's a deep fake or a response from a large language model may continue diminish as these things get more impressive but it doesn't mean we don't have to do our own due diligence and we can't just rely on this technology to to dictate our lives and our decision-making so ethical AI in practice I think has to incorporate transparency it has to incorporate a wide range of parties Regulators Academia it can't just be a Subs segment of Google or open AI setting the weights on these models and that's all we have to use I I mean I think we're we're heading in a very difficult future if that's the path we we kind of proceed on well I mean the story is a little bit like TV was you know in the in the 60s whatever was said on TV was true was the truth and today it's a fact that people are using CHT or other language model and and essentially because they work most of the times for them that becomes the truth so if you really want to make a comfortable use you don't want to do always all that that that due diligence which I agree and for me for example for the kind of stuff I I do so especially when coding until you stay very generic very superficial you get great answerer is a very good at same time instead of going look at the documentation you can actually just uh pull out some of the of the information you need but stays and remains a probabilistic system probabilistic system means that there always a percentage of error in what is pulled out so the hallucination becomes visible to me because I over a certain threshold start becoming repetitive and doesn't really go where I I need the you know sure the llm to go but for the average user might not be that obvious it's it's a very important thing there are models that are a hybrid model between an llm Transformer and a deterministic system which are semantic engines the combination of the two offer a much more powerful solution because the deterministic system is more like a search engine pulls out some results and then the llm transform them in a human readable conversation so but then you have a base information that is pretty much uh deterministic the only thing you don't have a lose in that case is the reasoning that is the pleasant aspect of the llm but if you do this in kind of becomes an acceptable tradeoff and uh it's something for example I invite you to explore but on blockchain blockchain is a is a a phenomenal technology today for certifying contents it's not the 100% solution in specific use cases can be the 100% solution because if you are narrowing down the application on the I don't know on the healthare or in the professional field when there is one software that is going to be you know the connection between the the language model and the and and the output that you need then you can certify the training set you can certify the answer because otherwise the thei becomes the place where to dump all the liability if the decision was a good decision is my Merit and good but if it's bad I was the was the AI that gave me that the transer so you need to make sure that you protect both party parties you know because you want to protect the professional from one side you want to protect the AI from the other side because if you discredit the AI people lose trust you want this to to keep growing and right now about the all the Investments all these thing that is working around the eye I think it's disruptive but there is in my opinion an over an overpricing and over perception but uh and the real Innovation right now is the Transformer it's all around that I hope that all these big thing will instead of deflate converge toward the new models that can even improve that what's your opinion yeah I mean I I I one thing I'll just mention I think um this idea of kind of convenience for users is is you know among the most important things for how they approach many people approach using different Technologies and I think it's one of the reasons why blockchain has taken so long to gather critical adoption across payments or in a range of use cases that should have already taken off it's convenience because there are tools in place that make your life better faster easier and there are safeguards in place if things go wrong and now whether that's the most efficient or or best path for Society for yourself it doesn't matter it's convenient and I think as the as language models continue to evolve in their capabilities and the responses become better and better and hallucinations become less and less apparent I think we have to be very careful about how people approach these tools and how convenient they are in being able to answer any question you have or do your job for you or do your homework or do your or summarize a book for you I think it's going to potentially erode or risk eroding a lot of the critical thinking capabilities that we have as individuals as students and the outputs that we receive won't always be correct but it's convenient and it's easier for a lot of people so I think end of the day it's a it's an interesting technology I think part of it is kind of a smok and mirrors to a degree um in in how they're architected but I I'm concerned that it become the convenient path often becomes a path that's very heavily traveled whether or not it's the best best path forward remains to be seen um but I I think there are ways to put safeguards in place um and you know I I am a new father I'm concerned about what school looks like for my son I'm concerned about what learning looks like for my son and I think these tools can help professionals today do their job a little more efficient like I wish it was capped at that to a degree until we figured out how to approach young minds in this area um but you know the user base is wide in in demographic in an age and the value it provides to different demographics I think is is is very different and I think for us to over index on the value of this for the younger generation is very risky uh and that's what worries me and that's kind of why we're trying to build a version of this capability that you have a bit more control over in terms of how knowing what goes in to to weighing these models and having a sense of um or at least the ability to build custom plugins that can real time determine the whether the outputs of these these models are are truthful or um potentially hallucination the cutest application I've seen in the academic field is like the possibility to upload a paper and then question the llm about the content of the paper simplifying a lot of the access to maybe very complicated studies or things like this why don't we give some uh uh some time to our public for a for a few questions we have a about six minutes okay hello yes my name is Nicholas I was wondering uh your opinion on the uh latest um developments in terms of blockchain and AI together either with payments or service sharing and um security of course um I was wondering it's just the one that I know that it's like one of the most advance but I'm not sure in terms of you guys' opinions or what's curring in the market because it does change so fast um agx uh I don't know if you're familiar with it it was uh Ben Oro I think is the computer scientist who uh founded that one so just that thank you yeah I mean I'm not familiar with that specific instance um I know a protocol called Olas AAS Network olas o they're essentially uh a blockchain based protocol that's building uh a set of automated agents that operate on chain that can execute simple or complex transactions I think a lot of the use cases that actually have um some development uh behind them that are actually being deployed are around using um automated AI Bots to execute trading strategies on chain now how interesting that is I don't know I think you're trying to combine a lot of pieces and does that result in a better outcome than a trading strategy that's deployed by current Technologies I'm not entirely sure um one thing we're looking at doing at sanctum is fine-tuning models on real-time blockchain data so essentially you can create a version of llama 3 that is uh up to speed on what's happened to the latest Block in ethereum whether that's transaction history price feeds you can use this model and it can interact with either a Trader or an analyst or even one of these automated automated Bots and you have kind of an interaction between those two parties um to inform uh better Alpha better trading uh practices for example but I think to answer your question it's super interesting um from what I've seen though the technology isn't quite there yet on both on both sides it's pretty cumbersome for us to do this it took a while to train the model and then to plug in an agent it's a whole you're better off using a human Trader to kind of interact with with that model via chat chat interface is kind of kind of a novel POC but that's kind of where I'm seeing it it at right now cont uh not not I mean in a different use case potentially could code the smart contract if you train it on solidity but in this case essentially you could predefine a trading strategy across a certain set of digital assets and you could either train a model on that trading strategy or you could code that trading strategy into to Olas or another Network and essentially that would um inform the the trading activities the bot makes if certain market conditions are are met or if certain liquidation thresholds are met for example um so that's one area that's interesting uh there's another one but I want to give you a chance to have some else uh I'm not sure if I fully understand your question but I can share the collaboration we had with the AI companies uh that uh this a combination of blockchain and also service we uh there's a company called uh media called Defiance media uh they produce news by Jif AI so uh we help them to embedded the uh the uh widget in into their uh J AI news that uh in Europe and also in the states the ax uh tells people that you have to reveal uh if it's if your content is generated by AI and what percentage you use is gener by Ai and also uh the AI engine you use and you have to reveal uh to them the the creators and also uh any other information so uh is uh only one uh technology could couldn't save us from defects but you have to do uh several uh combination of uh those Technologies to help you to have uh the to to to to find the truth so yeah well do we have another so I just had a question in regards to compute uh going traditional um companies using them for compute compared to like maybe a blockchain uh like render have you guys done cost comparisons like what the difference is on how much it costs to train the training is very expensive correct training rendering is probably less expensive so I'm curious between traditional versus blockchain like you know are you guys familiar with render protocol yes so so far on the use of the AI for rendering you might not know about that my background is actually in real time rendering and also for uh Global illumination algorithms there are still problems with a little artifacts that sometime emerge especially if they are video sequences but uh in general uh let's say the solutions are trending toward a very high quality so they are becoming a actually a way to integrate uh or improve a traditional red systems so there are a number of of of things that will be addressed ESP in the virtual reality uh realm that will be leveraging generative AI massively it's probably matter of another uh five six years but I think will get there very soon one thing I want to just add to that so I think to train a model using blockchain technology would would be an immense uh inefficient way to do it I think one thing we've looked at um the idea of compute is how do you support inference how you support fine-tuning using decentralized compute so right now a lot of this is done either on web-based models with open AI servers or with with sanctum it's done using your laptop's compute power but there are companies like aash Network and obal network that essentially facilitate access to unused compute elsewhere in the world whether it's a laptop or a workstation and they create incentives to allow you to basically use these to run a query or tune a model using a device that otherwise wouldn't have the capability to do that so I think that's probably where blockchain in the near- term has a really interesting use case around compute because there are there's a ton of underutilized compute out there these new MacBooks and PCs are getting really really powerful and for most people you know they're not using the compute to capacity all the time so if there's a way to let your computer run in the background leave it on kind of automat automatically sell unused compute to someone running an inference in a different country or different state make some money doing that in a very secure way that to me is very interesting and we've looked at that uh comparing um how how the problem is centralized servers are probably gonna win out So eventually we can bring these outside I want to thank you Christian and Chi for being here today you're doing and uh really looking forward to see more developments from your companies which I would be happy to keep an eye on