Why AI Needs Web3 - TOKEN2049 Dubai 2024

okay I'm very excited for this panel and we had a very last minute participant show up which is perfect uh my name is Casey I'm really excited to be moderating this quick background we'll do intros across the board but I started my career as an engineer and was an engineer at Google for a long time doing ML and ux work and then most recently I was an investment partner at Paradigm doing early stage investing and then left to start my own Frontier Tech firm and we're focused a lot on the intersection of AI and crypto and I'll hand it to you uh hey everyone uh my name is bidan uh I'm the founder of bagel network uh my background is a mixture of uh machine learning and blockchain um I used to be the AIML infr lead for Amazon Alexa did a lot of work there um and I us to be the ecosystem lead for a layer one storage protocol called ariv and yeah excited to discuss AI today hey everyone I'm Sean uh I'm the co-founder and CEO of Sahara AI um before starting Sahara last year which we're building for over a year uh I'm uh associate professor at USC computer science department working on AI and LP for almost like seven years so really the inuring things for me to step into using blockchain and all these like tools of web3 to power AI to looking at the sort of broken AI economy uh that we have in the traditional web to community were the participants of AI talk for example the data contributors of Wikipedia who were later the data was used to transen chbd wasn't really properly rewarded so uh with it with Sahara AI we are doing a uh utility uh AI blockchain platform that really trying to enable Providence and attribution and monetizations the AI access including data sment model uh with high performance privacy and security hey everyone I'm Michael I'm the co-founder and CEO of zerog Zerg is the first modular AI chain we help launch high performance D apps through the fastest programmable data availability layer in existence today um I started my career at Microsoft as an engineer and Technical product manager then switched over to the business side worked for a baning company Bridgewater Associates and then then uh during graduate school scaled my first web 2 company to become a unicorn and a top YC company and we used AI a lot for uh data cleaning data labeling um and various data extraction issues and so really excited uh to now be at the intersection of AI and web 3 awesome okay that was a lot of info on the companies I actually just want to do one more round can you say your company name in the oneliner of what you do so everybody knows sure uh zero G Labs were the first modular AI chain so Sahara AI is democratized scouting laws and trying to build the Dune of intelligence that owned by you uh Bagel network is the first privacy preserving machine learning protocol for humans and agents okay great all right so with that I think this panel is titled why AI needs web 3 so let's just start there and uh discuss that so why why do we feel AI really does need web 3 today who wants to take a first step I can go first um I'll probably have the strongest opinion on the panel I'll just say that everything AI should be on chain and my thinking around that is that our mission is to make AI a public good and so how do we make something a public good if a bunch of companies are controlling closed Source models where you have no idea what the training data was uh the model what the weights of the model are what kind of decisions are made with the model what's being changed at any point in time and so uh very much AI needs web 3 if we want to have a truly open AI you know pun intended um where for example you the AI stack is essentially anything from uh data creation to data labeling to data storage to model training to inference to then actually the or fine-tuning and uh to then actually the AI dabs and so if everything is fully on chain you have complete understanding of the uh Provence you have complete understanding of the verification the incentive schemes that were used who is actually contributing to this data and then you have complete uh verification of all of this and so that's something when you think 5 to 15 years out if you have ai models running production system administrative systems uh Logistics systems I would want to know exactly what went into that model and what decisions the model is making at any point in time so let's put models into a container where it's always aligned and they can't cheat and so that's very much why AI needs web 3 yeah um I'm a pragmatic person so I'm actually going to take a little bit different view from my c so I don't think we're going to push the entire AI stack on chend which is I don't think it's actually scatterable or economically feasible in uh short term or even in long term I I feel like it's really challenging so my take on where uh blockchain really have the Synergy with u AI is mainly on um attribution onip and Providence and by saying that I basically imply that the current AI economy is broken right I was giving example like Wikipedia GitHub radit or hugging face which they have sort of a been struggling monetizing themselves and but they are hugely important and sort of important resource for building things like character um Track gbt by open AI but the whole Upstream flow of contribution of data points data sets recipes of building model all the Kno hows of tweaking the model um wasn't properly um attributed and rewarded by the current a dech we only see the majority of these Tech JS which are the last sort of uh um application Builders of this AI they got to have the most Revenue share rather than giving them back to all of the Upstream contributor and I see that's pretty broken where and and I think the right way to do it is to think about what other the key metadata what other the key attributes of the data sets and model can be pushed on Trend so you can sort of sort of have a sort of a verifiable ownership or true ownership and then you can use that to do a Providence and attributions down the road for all of the participants of the AI economy yeah so just on that point so data Providence what do you think or what what is it like in even just a year from now because it's such a broken system and it's going to take so much work because the entire internet is already built without Providence and attribution and so what's the master plan of Sahara to fix that yeah I that's a great question I think Providence it's a really long-term sort of a goal um technically speaking uh right now my view of Providence in terms of implementation there's two things right there's a Course Grand Providence where if someone take a model or data set out of their pocket and then say hey this is uh cases model or data sets how could I verify that how could I prove it or we call it auditing right there should be a sort of a proof of ownership or Integrity sort of mechanisms be developed combined with the blockchains that are able to say hey this data sets has been touched by me Sean and Casey and someone else and so when this data sets was used to Trend a model that later be monetized those profits should be attributed back to this couple person and then in term of how much each of us should take that profit sharing that's probably have to be Community governance for now because the second part of the Providence which is the fine Grand Providence is really hard um for that I was talking about for each data training data points that was used to train a model when this model was making predictions for particular Quarry or prom right how the attributions go back to each of the training data points and therefore they owner this is basically machine learning people been using influence function to do that which has the computational cost which like 5x of the inference cost which is totally not usable right so so I love this because I think personally one of the best use cases of crypto is financialization and creating markets around digital assets and ml is naturally digital but I guess I'm curious from your perspective why hasn't it materialized yet or how are you doing in terms of product Market fit so far like are people actually using the product uh use using the which product sorry the data Providence like do yeah I you talk about cuz you have some really big names right that were working with you I just feel like the infrastructure support is not there yet that's the problem that's what Sahara AI is building because right now if you look at the the entire Web Two sort of a landscape most of the stuff are centralized you don't have the transparency in record IM mutable records about who has touch what data sets and how the data sets was used later and how much money the data set made and how and then you can do a world pass to get the things back to everyone that infrastructure doesn't support so really I think the opportunity right now for the the AI in party within the web three Community is to think about how to build this Providence and attribution infrastructure okay I mean there's some examples of grass for example uh creating this financialization layer where it's pretty clear what the attribution is somebody gets tokens contributes data to it so at least in some cases it should be pretty straightforward to to attribute did so if most of data generation ends up or all of it ends up being on chain like this then it would be a pretty simple world but it's not like that unfortunately so there's quite a few infrastructure pieces that need to be built um and um happy to disagree with Sean that everything can be built on Shane but we can get into that as well um I guess it's my turn talk about why AI needs web3 so I actually agree with both of the folks here that U AI can be on shap and can be offchain um one thing we believe is that all of those things like putting Ai onchain and offchain are not going to be done by humans they're going to be done by autonomous agents and it's very soon like we think of it like very futuristic it's going to be here 12 to 24 months from now so our thesis is that there will be billions of online agents they're taking lot of different kind of actions they're buying stuff they're selling stuff uh they're doing prediction markers they're doing stock markets they're doing trading all of those things and they would like to use the most frictionless infrastructure for everything if they're trying to send money they're going to use the blockchain to send money because they don't have to wait couple of days to for it to settle um the same way if they're doing AI if they're doing machine learning they're not going to go somewhere where they have to sign up wait for being kyc so that they are allowed to use whatever they want to do uh they're going to do and do machine learning in a permissionless way and that's the thesis we're building from that AI needs web 3 because web 3 enables permissionless and that's the least amount of friction okay so behind the scenes we were having a pretty controversial conversation and I want to know if Sean would like to share a controversial opinion here I yeah I think we talk about too many controversial things so I I would just like do my high level summarization of what's my view of the centralized AI status quo right so if you look at the key recipe of AI is three things right data model and compute right and then we can go through these three things right when when we started brand storming what's the funding thesis for Sahara AI we just look at these three piece of things and then trying to debate about where blockchain fits in the most natural and and sort of a Sim this way uh with the first thing we reject is the compute um which I think the first thing we were pretty controversial right I was basically thinking that the idea of building a a war network of gpus is a really sort of uh tricky and I don't think it's a very promising idea because I I think the power of this scouting law really comes from a very streamlined and centralized Cloud let's say Nvidia AWS Azu right that make it possible that you can Trend with 10,000 h100 in two weeks to produce a model like llama 3 for example and if you really have to split this out to like your cell phone as and his laptop and the other guys sort of a little GPU desktop and trying to use the Internet connect this small pieces together and call it a network of GPU computer that is spare and can be taken advantage of I don't think that's usable at all for training any usable AI model of course you can use that to run some tiny little toy model but that's not a story of the really usable AI I would say so that's about a compute right and then we look at the model model is tricky because it has too much knowhow that required to build a reasonable uh architecture and know the hyper parameter configuration and so on so I don't think it makes sense to have a pretty decentralized or crowdsourced way to say hey let's come up a new archit Ure and a recipe that is better than llama 3 and Transformer I I think that's kind of really tricky so really about a model part I think there may be something to do with trust list training of the model and then there's something to do with trust list or verification of model inferences that build the trust uh on a protocol and I think now when look at the last piece of data that's where Sahara AI started right we actually started by building at the daap that trying to create a decentralized data Marketplace and then you can imagine that's a decentralized version of scale AI if I put it simply but really everyone have their own expertise and culture background and language background that can come into contribute to a public good data set and become the owner and shareholder of that data set think about if you redo Wikipedia where now everyone have ownership on Wikipedia and they got paid by open AI that's much better work if we can redo that yeah just to make the panel fun I'm going to disagree on the compute part um so Google just uh released a paper where you can do asynchronous uh batching and you could actually create uh large models utilizing different batches and so it's a step forward into truly doing onchain training of models as well yes it still requires a lot of compute and it's uh not going to be sorry what is the what did the paper it's I I don't know the actual over it but I can find it after the panel it just was literally just recently released um Jason are our team flagged it for me he's our head of AI and so I'll I'll I'll give it to you after but um essentially uh you can't do it on a gaming PC and stringing a bunch of them together will still have to be like larger scale uh type of installations but it's a move in the right direction um and it's not all Rosy on the centralized side e either if you do a training run then a bunch of graphic cards fail then you have have to like redo the Run anyways and so U why not move into that direction as well and then on the model side um there's also a move on the AI side towards modular AI so instead of saying bigger models are better uh what if you have an llm middleware and a lot of smaller specialized models that come together and to create a model and so those can be trained in uh in a different way and then as an example of how to out compete I think my shell AI actually creat created a model that performed better than llama 2 with uh very little uh resources I forgot exactly what it was but maybe like 100k to train this model and so those are just some examples um that I can think of um I'd like to say I also disagree with Sean a little bit on this one um so the decentralization already exists um like look at any big startup they train their Model A little bit on AWS they train their Model A little bit on Google Cloud they kind of like distributed the workload that so uh for us is that like decentralization is not fast enough is not an excuse for us to not do it we have to figure out how to uh how to make it work so um Michael uh gave an example how it might work for training I'll give an example how it might work for inferences how to make inference really fast so that it really works uh even if there's a latency problem right um I was talking to a bunch of researchers in Silicon Valley they have been uh researching burning the weights of a model directly in the gpus and that reduces the overhead of compute by 100x right so something uh there are like lot of ways we can figure out like how to make it work some so our mental model is depending on what's possible today we should think about like what's going to be possible in a year what going to be possible in couple of years uh since Chad GPT was released like more than a year ago all you know inovation have just accelerated massively and and it's going to get accelerated more so we don't know what's possible in a year from now we we can say like yeah it's not possible right now um but the latency the GPU cost all of them will come down 100x THX very soon which will make the decentralization possible and what's the benefit of that is that we can actually run models train models in a crowdsourced way so it's a knowledge um so right now the bottleneck is that like it's a a company who are training some people have some knowledge about how to train a certain kind of model but we can never crowdsource it because that model doesn't exist web enables us decentral uh decentralization enables us to do that okay on the bagel front though while we're here I would like to talk about your user base as well and specifically rumors on the street is that you have a lot of AI native people that are using the protocol can you tell us more about that yeah um so we're still in closed beta but um we have been fortunate to have um good Partnerships with mainstream artificial intelligent Frameworks uh some of them are like very well known among developer uh Community like Lang chain and llama index and we work with them directly and we have more than 25,000 Developers ERS use our protocol at least once 5,000 25 25,000 yeah and we're getting like more than half a million API calls per week at this moment and what's the most popular use case of the developers today inference inference data for now so they want to uh generate inferences on additional information so they use Bagel API to do so and it's primarily because of the data you have behind or is it just you're just doing inference right now we're not even um allowing people to use each other's data they're bringing their own data they're using the bagel uh infrastructure to be able to to generate INF on top of that okay and then the gpus in the back are they data centers or where do they come from um multiple we have partners for example one of our um backers or Falon so we're working with them they have like massive data centers uh and a lot of gpus we're also working with other GPU Network uh both decentralized and centralized they have a lot of gpus they want users for we're working with them and increasing their um utilization rate so it's uh there are multiple ways we gathered them together okay and then you also mentioned this kind of thesis that AI agents are going to become your most popular user in the future yeah and I want to talk just more generally about what you're seeing in AI agents what everybody on the panel thinks about them in terms of you know like you said the whole idea of putting everything on chain and what of the AI agent actually needs to go on chain and what doesn't people who aren't familiar with AI agents they're really just AI apps that are doing autonomous tasks um so the definition of AI agents I'll go first because we uh have very strong opinion about AI agents um it's in your Twitter bio yeah exactly we build our product for agents because we think we believe that that's going to be the majority user base for us and for any online business in the world it can be onchain or offchain but as long as you're online you it might be some Shopify uh Marketplace even then your major majority user base are going to be AI agents and what is AI agents in general like now they're just some apps built on top of like some llms hosted by open AI or anthropic or whatever but my vision for AI agents are the every AI agent are going to be a foundation model of its own and each of them will have their different privacy level like their own proprietary data that they generate themselves online through the transactions they make and um yeah we're basically sold on chain like any business uh we're a little bit early on that like being like 100% into AI agents but we were a little bit early on crypto and AI intersection as well like the when uh we started building um it was like end of 2022 early 2023 um people were like what is this crypto and AI thing like are you just putting two words together doesn't make sense and now it's getting more popularity uh same way um like AI agents being the majority user base for all kind of businesses protocol um it's not the consensus at the moment but we're going to get there soon very soon um yeah I guess I would agree with that I just add the dimension of uh control the AI agent has if it's an AI agent that's just doing some shopping for you and has maybe access to 200 usdc in your wallet like run it on your phone an edge device but if it's an AI agent that can actually control an entire traffic system like I would want that to be fully on chain just just again to have uh full visibility full verification of everything that the AI agent is doing and everything that uh went into the creation of the model that's actually running those agents so that's that's that mention that i' add okay a new category that I'm seeing a lot of is intent llms where we have an intent on chain you utilize an llm to automate it via an agent and I'm personally pretty bearish on this because I feel that we're first going to see agents in e-commerce or things that have lower Stakes than with Finance but I don't know if anybody has a different opinion about that category um so I actually think we already have a lot of agents in finance in centralized Finance like most of the trading is algorithmic like algorithms are running all the trades but diff but not with non-deterministic llms um yeah so the algorithms are deterministic llms are are not deterministic right and I also don't think llms are going to do the uh trading as well like large language models are better at generating language but that doesn't mean they're the best at trading we're just using them nowadays for every use case because it's easier to do open made an API which they can everyone can use but they're not the best for a lot of stuff for statistical stuff you have to do statistical model which doesn't even speak in human language so I agree with you that llms large language models are not going to be the one who will be do who will be doing the financial trading but large models are going to do it it can be LSM like large statistical model right so um and a lot of that happen already right now um the algorithms are deterministic the models are a little bit less deterministic but that also gives them superpower they can learn new things as they go they can learn instantly and they can execute the trades better um Financial use because I see it happening sooner than the other one on um yeah that's also based up based on some of the uh conversations and demand we have seen on the bagal network in general yeah I want to share a little bit about my view of agent because you guys been talking about agent right so I think um agent is different from large language model apparently I think my view of agent is basically eventually going to be your personal replica or your digital TN right writing reports for you 24/7 talking to other people or other people's agents right eventually I think we we're going to have a pretty scary situation where there's different you don't need any Marketplace because information there's no information gaps because everyone's agent is talking to other people's agent instantly 24/7 um you got all the information ex except that you don't want to share your information right so to get to that kind of stage I think what the agent technical need to be done is uh it doesn't it it has to have the language capability found backet by large language model but they also need a lot of like Preparatory data right basically your private knowledge Vote or your private knowledge basis that store all your DMs and text messages emails pictures and so on and then you also have a sort of working memory then the language model will interact with the database and working memory to take actions and you also have to have sort of reward mechanisms for these actions taken so you can choose the best course of actions to execute right so under this kind of a broad architecture if you ask me where would decentralizations and blockchains play a role here I think there's two parts right number one is the decentralizations is going to be uh providing security privacy protections for user data right if I'm using say llama 3 as the background language model how could I have a good personal replica without sharing my personal proprietary data to a cloud that host a centralized crowd that host llama 3 right that I think that's where Federated learning some sort of distributed machine learning inference and training pipelines will play a role uh plus you can apply maybe differential privacy which is empirically strong and you can even try to push for multiparty computation to enforce stronger encryption the theoretical guarantee right those are all could happen I think that's there that's technically how can you make it privacy Pro protect the secure sort of a agent framework and then if you have to talk about blockchain I think we we have to go back to all of the things about Providence and ownership right if let's say I'm going to call your proprietary knowledge bases to collaborate with my own agent so that I can make a more informed decision on something or I can make a better technical reports about a web three project then I have to pay for API access to your proprietary knowledge basis and I mean the jaw breaking and all the other things a separate discussion but I think that we need the infrastructure so that our agents or our knowledge bases can collaborate with Providence and proper attribution and I think blockchain will play a pretty critical role to building towards that yeah for the first you mentioned doing privacy preserving and thinking about that why not just go local though because right now with the advancements in Hardware it's going to be expensive right it's too expensive from what capacity like I don't think um if you only think about pushing a 7 billion model to your local devices that's totally durable nowadays if you have a pretty strong smartphone but you're not going to enjoy the power of 400 billion llama 3 which is going to be open source in a few months right so you want to take a Best of Both words I think that's where you want to do a sort of decentralizations so feder yeah but with quantization I feel like it's getting better and I feel that but I don't think we can push a 400 bit model to your no no I mean definitely not but I still think that there's going to be a lot of use cases where a local like Edge compute makes a ton of sense oh yeah uh there's no conflict about what you said right Federated learning basically saying like if I have a Lura or like low adapter module on my Edge devices and I have a backround model a llama 3 on the far Cloud yeah as long as the communication between the two is streamline low latency I can make inferences and also I'd like to say um something here it's probably controversial kind of that the local training is actually not as private as you think um uh we have written a detailed blog post about that on our blog like blog. b.net if anyone's interested in looking into it but but uh when we say that we want local training uh like Federated learning is an example of that you actually send the model where the data is right and then you actually uh revealing uh sacrificing privacy of the model itself both of the model algorithm and the weights what federal learning do is that you send the model to different location they generate some weights they bring them together then you bring the weights together and create like a aggregated weight of that um so all the privacy of the model owner is sacrificed the data owner yeah they have their uh um privacy in theory even though there have been studies where it has been proved that the data is leaked a lot if you do Federated learning there are uh a lot of uh research papers on that as well so we uh were trying to like go really deep in local training it's just not as private as we think it is yeah I mean I can see that also with oama for example because they're doing the quation they technically at one point could have accessed to your logs even though you're running a local machine I don't think that's the case anymore I think they fixed that because enough of the Discord can complain but I I I hear that yeah um and also like uh there's other privacy preserving Solutions we uh use in web 3 like ZK for example but in that that also have some problem uh in ZK you send the data where the model is so that has the reverse problem of like you reveal the data but uh the model owner is kind of like preserving their privacy to some extent and and while we're on this maybe like HF do you want to comment on if you genuinely think that's going to be possible so can I repeat the question again yeah like fully homographic encryption do you think that's going to be I think we're too far away from that why why do you think we're seeing so many startups retry this right now I I think we definitely try I mean this to me is really um or maybe can you explain to the audience what that is before we dive yeah so so Casey was talking about homomorphic encryption or fully homomorphic encryption which is I think the basic idea is like you can make your input and also output entirely encrypted but you can just like enjoy the computation Logics or the sort of AI Logics of any language models or foundational model right doing chaining or inferences without exposing any of your propor information or privacy and I think that's the that's the dream I want to have because I if we got this today there's many of the problem we're discussing just now which is easily be resolved and I think the curent like for example we don't have to think about other type of privacy protection methods like differential privacy or multipart computation we can just use a very uh efficient fhe kind of approach all right and um yeah so that's many about privacy protection and data security and I think right now the problem is uh morphy encryption doesn't works for anything complex than a linear classifier uh and you if you know about Transformer is basically uh dozens of layers of nonlinear linear layers stacked together so if you have to design a homomorphic encryption technique for that uh it's totally intractable in terms of computation it's overly slow um so I think that's why there's tens of billions of dollars of investment into projects doing that right and I'm really looking forward for the results but I don't think we can enjoy any usable results in a in a few years yeah yeah some of my friends that are working on it in the web 3 field think it's two years away and then my friends in the AI field think it's 10 years away and so maybe the truth is somewhere in the middle but one thing is definitely clear what Sean mentioned is the computational requirements are significantly higher so my friends who are working on it they say it's a THX so if today for example you're trying to run some type of highfrequency defi application or onchain order books you need like a data availability speed of about 3 megabytes per second with fhe multiply that times a thousand and so the type of data availability throughput that you need for that is in the gigabytes not in the megabytes what most people are providing today and so those are some of the things uh that we're fixing and also future proofing the system uh so that when FHA actually becomes a reality it actually even feasible to run on chain as well okay um I'd like to say that um I think or we think at Bagel network uh if is the future um for both machine learning and blockchain and when I say future uh that's the keyword that's the future that's not the present like it cannot be done at the moment like for our like uh another like a very short version of uh FHA definition is that you basically do computation of uh encrypted data and get the same result of doing the computation of raw data that's FHA and that's the end goal for any kind of privacy uh when it works everything will move to FHA machine learning will move to FH web will move to F but right now it is actually very slow as uh uh Sean and uh Michael mentioned and uh but we're getting there um I actually think uh all the acceleration we're talking about that's going to happen way sooner again like the mental model we have is that there's like 100 researchers they're doing research on accelerating F that will take them 10 years I don't think that mental model apply anymore someone uh and there people I actually know who are doing that are going to train a foundation model who is a FHA researcher and that AI model is going to figure out how to accelerate the FHA chips that we need for that like the even most recent chips made by Nvidia they were not designed by humans they were designed by an AI which was developed veled inside Nvidia if they were just relied on the human capital they have it would have taken like another decade to get to where they right now as I said like everything got accelerated our mental model for everything if if our mental model for a timeline of AI getting accelerated was 10 years that got shrunk uh by a lot um so it is the future it's not the present I agree but it's not far in the future it's near term okay all right well we are coming up on time I think let's wrap with doing predictions by end of year so not that far in advance in the intersection of crypto and AI I can start first um in a probability too in terms of how confident you are to materialize I think end of year we will see a lot of economic layers that are proposed by this web three Community some will fail some will works to some extent and my guess is the actual working AI products generated by the protocols will be relatively limited uh that's my view I mean they going to take much longer to explain why but we as s AI are really leaning towards bu AI utility sort of a protocol meaning that the best way to quantify your success is to see whether the AI Builders are using them to actually build the AI products that are going to be monetized and sell and I think uh we're going to see some success promising success but many of the pro protocols going to have pretty hard time proving that um not in the crypto space and I'll get to that in a second but I'm super excited about full self-driving mode from Tesla major breakthrough sorry not in the web 3 space but maybe we can bring that to the web 3 space over time but uh web 3 meets AI specifically just excited about like agent marketplaces uh decentralized Vector databases is um like deep fake uh provenance those types of things I think uh pretty seems like it could be very reasonable to figure all of that out end of the year confidence level 80% um so I I say I have two predictions one more related to the crypto um um ecosystem and one less related to the crypto ecosystem so the first one is that I I think 90% of the onchain trading are going to be done by agents autonomous agents and I'm 90% confident about that um and the non- crypto uh stuff uh is that um actually forgot what I was going to say but uh I'll pass on that yeah my prediction is that okay I like all of yours my prediction is that the llm craze falls down a bit and linear models come back in Vogue and we start hearing more things like yeah just regular models that people have kind of forgotten about that are actually baked into a lot of our products and perhaps less consumer focused but I've totally changed all of Technology fa actually remember what I was going to make the other prediction I think we will start having discussions about discrimination between artificial and human intelligence so right now everyone say like hey we don't want to uh get a text which was uh publish a text which was generated by AI or somehow AI generated content is fake um I think it's going to get uh real very soon where the AI models are going to be as intelligent as humans and then we will start discussing like is does it even make sense to discriminate or is it discrimination at all like intelligence is intelligence like worldcoin recently launched a chain where humans get like less gas fees that's onchain discrimination and we're going to discuss like does it actually make sense or not um we would want all kind of intelligence to have equal rights if you may okay I think that's a great place to end thank you so much everybody we'll wrap