Decentralized AI: The Power of Permissionless Intelligence - TOKEN2049 Singapore 2024

so welcome uh everyone Thanks for uh joining us for the panel uh we have a excellent lineup of speakers I've known most of them um we have one of the earliest employees of open AI he left then went back started near went back to open Ai and then you know left and so this is going to be a really exciting discussion so why don't we start with very brief um intros so just people have some context of what you're doing hi everyone I'm emad moac of uh shelling AI previously I was co founder of stability AI we built models like stable diffusion that had 300 million downloads and now at shelling we're working on the concept of universal basic intelligence for everyone in healthcare education government finance and Beyond hey everyone I'm shanan um CEO and co-founder of Sahara AI also associate professor USC computer science department uh been working on AI over 15 years uh mainly focused on National D processing and starting Sahara a which is uh focusing on uh building a blockchain platform that prioritize user sovereignty and Providence uh infrastructure the simply put we're trying to create a sort of a copyright system to power Revenue sharing mechanisms uh on top of the AI accesss hi everyone my name is Alex I'm one of the founders of nearer near is one of the uh layer one blockchains it's a shed blockchain with an emphasis on usability uh but near Back in the Day started as an AI project uh both of the founders myself and IIA we're coming from AI backgrounds and now we're starting a new lab called near AI uh the goal of which is to make sure that we moving towards a world where open source AI is always competitive with the proprietary solutions from from the world Gardens and the large corporations uh yeah exciting to be here hey uh I'm tun uh founder Gauntlet I also uh investor robot Ventures um probably my main uh reason I've been around decentralized AI is my first job involved a lot of hardware and uh we kind of got replaced by AI so I can talk about that great so I guess when don't we start with uh I'd like to make this more of a discussion um how do you think about um competing in centralized Solutions um like what is the role what is the need for decentralized AI right uh okay I will start uh so the there are several problems I guess one of the big problems is that if you look at any any uh I guess technology that uh uh was created in the past decade uh all of them are today controlled by large corporations uh and uh in every single case what happens is that those corporations uh capture all the data monetize it sell it uh users effectively and not customers they products and artificial intelligence is moving in the same direction uh and this is something that um generally I think in web 3 we are good at fixing uh and uh the big differentiator of we call it user owned AI of user owned AI should be that users are in control of the data users are in control of uh how they use AI powered applications yeah um I think there's a different user groups within this AI ecosystems these days right and uh we really care about uh three groups of people the first one is these AI developers or AI model developers and these days if you think about prototyping a model idea it's so expensive and very timec consuming given that you need to really find vendors of data find vendors of compute and then you need to raise cash up from from VCS this whole process takes at least like three to six months to even start um trying out some of the ideas so I think this model developers facing the limitations on accessing the resource as well as the capital which um and and I think there's a great opportunity that the entire blockchain backet platforms and um D5 Capital markets can provide much fast tracks for them to realize their idea right at the same time um if you think about the AI users right users of the model of the application they are dayby day giving out their feedbacks and sort of preferences uh about themselves right very private personal information and exchange of free access say using chbt but but over time they didn't really realize I think this is a really burning problem that um open for example accumulating uh hundreds of millions of this kind of user preference data as well as the things that they uploaded like the PDFs about your financial situation your text message your emails and then over time they're going to create a better version um um aggregating any all of these individual domain experts that might be able to automate part of your jobs and even replacing your jobs and monetization opportunity and that's where we're going back to this whole problem of privacy protection s and sort of user ownership about what they have given to the central provider right so um and I think these are two major big problems we have to tackle and to see a great opportunity coming from blockchain and crypto uh ecosystem for soling that maybe tun or um um Imad I really care about the cost because in this industry it's very ideological right so yeah you know user privacy and decentralization's great and but the actual implementation of it you you realize users really don't care about that they care about lowest cost provider most efficient provider and you know and so is it realistic um and is there actually a possibility for decentralized AI to compete head-to-head on on feature on like on price relative to a centralized solution um I think that just from personal example when we built stability AI we were running about 10,000 Nvidia n A1 100s which at the time when we switched it on 2 years ago was the 10th fastest supercomputer in the world it was eight times the computer of NASA and that's what we needed to build state-of-the-art image video Audio models so we build some language but again we were experts in the media side and we built stated the open source models that cluster cost $400 million Elon musk's new cluster that he's using to train state-ofthe-art models is the equivalent of 300,000 a100 and costs $3 billion so these are the table Stakes to get in on the training side of the core of these models now the interesting thing is uh you see open ai's 01 model last week it's hit 120 IQ which is crazy but it's moved from requiring the large supercomputer thinking longer which suits the more decentralized parallelized things like I'm on the board of render for example we've got million gpus and the ability then to spread things out even then again there is a very interesting example in web 3 whereby the biggest success story has been Bitcoin Bitcoin uses 160 terawatt hours of energy a year on Specialized computer infrastructure all the data centers in the world use 350 so I think there are some elements there where we can compete but right now it's very difficult because you need billions of dollars of investment to build the state-of-the-art models and the question is will that change so we can use distributed billions of dollars as it were or can we create mechanisms that we can create these clusters and then have an open alternative to these fully centralized solutions that typically only reflect one point of view so that's like the Google image generation where you type in sumo wrestler and you get Indian female sumo wrestlers or whatever which you know Dei on steroids yeah I think there's uh a couple components that so first Edge applications so applications that maybe don't have immediate financial outcomes like in a lot of scientific research actually training a model is extremely cost prohibitive relative to the payout so um you know I think for instance even when Alpha fold was started there was actually a lot of contention within Deep Mind as to whether to give them even like one tenth of the resources they needed um and it took a long time to actually for people to believe that like oh we should actually be trying to build AI models for things that aren't text or video or like IM immediately commercializable so I think the edge assets Edge kind of products that people want to to build models for often times get lost in the the cost discussion because everyone is so focused on like the the what seems like the the right model the second thing to imod point is that the architectural changes of going from a pure hey we want like one big model with a large initial data set to we have a more reinforcement learning type of game theoretic model that might make a different latency vers bandwidth tradeoff in the sense that I'm willing to wait more at inference time but I I don't need as much complexity that architectural shift is actually going to completely change how the the Clusters are developed it's probably going to change how you do sort of inference in the long run right like you you were doing this kind of trade-off between training and inference and I think that the the battle for that is not done and it's not going to be done tomorrow there's an interesting thing that I just like to point out so with alphao I was one of the authors of open fold where we did the open source replication of alpha fold and we provided all the compute for that as well there were a lot of errors in it which we didn't understand because it was a black box so I think open source is very interesting and decentralized because you have so many more people looking at it you can figure out errors in these models that are going to have an increasing part of our lives and be an increasing part of our Lives as well yeah I'd like to shift the discussion to the actual products and so we've seen from you know uh marketplaces for compute um we've seen you know the ability for people to monetize their models um what are the use cases you guys are most excited about and and you think are probably going to hit scale and be most uh you know widely adopted um in the ne you know the near future I can start first um I think this part of these questions also from the um to answer the purose question right I mean if you look at um the whole progress we have made in the past 3 years all these advancements on AI model capability there's a whole lot of things come from uh the success from centralizations especially from uh putting gpus together and then putting machines together into Data Center grouping um a team of very Talent AI researchers and let them really like go well in terms of implementing the ideas a lot of things come from resource allocation put together the the best of the people and best resource to create things and I think this really tells the uh the the the the fact that some parts of the AI development shouldn't be immediately decentralized and if you choose to do that it's going to pretty much just like disrupt the success recipe successful recipe right away right and I personally believe that for example compute I don't think we should immediately go for decentralized computes with um a very heterogenous set of gpus putting trying to put together Channel model or inferencial model right but yeah maybe a GPU Market place is a great idea that you can uh balance out the sort of the imbalance distributions of the GPU resource over the globe but you don't have to have a blockchainbased solutions for that right away uh but right now I'm seeing maybe the best part for us to explore in terms of decentralization is actually the the ownership and governance on the AI assets right the data sets and models that you're producing and what that means is if you trying to put put together resource to build a model that you believe is going to be successful why don't you giving out the shares of your model uh to the people who participate in the process for example someone give you 100 hours of G a100 GPU um and you you you have to use that because you don't have that much of cash to purchase that up front right then why don't you give out like 20% of the model you create in exchange of that right similar things can happen to the data providers because you are touching people's private data and you know those private data really Val able for making your model different from the others so this pretty much ask for a new sort of governance and economic models for coming to get crowdsource a model um and then you are giving back everyone some part of the ownership when the model start monetizing itself there should be a permissionless and trustless manner to attribute all these run new sources back to every single stakeholder and that's exactly where the blockchain can come in to help to make this whole thing in a trustless manner permissionless manner yeah uh I think one of the most interesting uh there's couple use cases that we are actively using today one of them is decentralized inference uh there are several companies uh we internally use a company called hyperbolic uh which is uh uh it's very competitive with u like fireworks or together in terms of price and speed but it is based on top of a large decentralized cluster of uh uh gpus pulled from from various actors uh and and I I think generally inference is one of the more successful businesses in web to as well it's uh the most uh has the highest demand in the space generally uh and the Second Use case which is very interesting is we've been uh internally at near we've been working towards uh building what we call an AI developer um like endtoend development of smart contracts or applications uh since uh at least 2021 uh and we've been annotating very large amounts of data um so we have we have a large pool of developers of people who who annotate logic Etc uh and that that runs on top of blockchain and um without blockchain it's generally a very hard task because people who do data annotations they usually come from markets where payments are very hard right and on blockchain you you do work you get paid right away right and it's uh you don't have to wait like if you work on mechanical torque uh you have to wait until like the week ends to get your payment aggregated because payments are expensive but uh uh if you do anation on blockchain the work is done the money is yours right at the end of the day you can already cash out and uh and for many for many of those people money are needed today right for many of like they leave from the payment to payment uh so that has been working very well and uh uh without blockchain that would be practically impossible uh to build in the same way I think if you think about practical use cases so when we released stable diffusion it used all the gpus across Amazon there weren't any available and now then it came 100 times faster thanks to the open source community so went from 30 seconds an image to almost 300 images a second in some instances if you look at what's coming you'll be able to generate a full length Hollywood movie by the end of next year with the current technology and you can do that on an asynchronous basis because it doesn't make sense to do real time because you have to wait an hour to see how things are and movies are made of 2 and 1 half second shots the transformation of media is one of the biggest use cases of overall inference compute but then the IP attribution of that the ability to remix flows of knowledge we have a system called comfy UI we built where rather than just this uh Goldfish Memory AI that you type a prompt in and it comes out you see the entire logic flow and then if I share that media file with you it reconstructs every decision made these are coming and these will be one of the biggest uses of it in media but then also in areas of knowledge so now medical language models and we're building those you'll have a cancer model that has comprehensive authority to upsate knowledge on cancer and outperform human doctors on empathy so nobody will ever be alone again on their cancer Journey or RMS or Alzheimer's or dementia or Autism and think about the use cases of that it's massive but then we look at distributed Ledges and the verification aspects of that protecting your own Healthcare data and having AI analyze it on the edge that's immensely powerful and you can look at something like apple intelligence which is an on thee Edge model that has has adapters for various knowledge features your smartphone has enough knowledge to analyze all your Healthcare data in place but then how does that go and find more knowledge we can build that on infrastructure like some of the people here Sahara and others are building and that will be a huge unlock in the next few years because finally we'll have holistic Healthcare or personalized education one toone education is a two Sigma Improvement and now like um there's an amazing thing notebook LM that Google's just announced you drop a bit coin white paper or anyone's white paper in it it'll generate a podcast of two people discussing it almost instantly things like that are coming but again we need to have trackability verifiability coordination as we send our AI agents that have memory to go and solve these problems so I know we uh we you know you're kind of discussing privacy is sort of this philosophical thing that people always talk about for why you might want decentralize AI but I'm going to give a more uh concrete version which is you know I would argue that from maybe 2012 to 2019 crypto actually sort of helped AI in the S I like it implicitly helped fund GPU development by creating demand for gpus for a long period of time the interesting thing is I actually think the current set of GPU road maps and also the Asic road maps from from people like grock towards their implementations of trusted execution environments is actually something that crypto can take advantage of so I think the the performance on on tees on gpus has gone up almost like 10x over two years and there wasn't really that much interest in improving high performance parallel tees until uh h100 and one of the reasons for that is that at inference time people actually do want privacy for their weights and a lot of there's a lot of demand that uh the hardware companies are getting for that but that actually creates a ton of excess Hardware capacity for private computation that I think people you know in crypto have been dreaming about for years in terms of further out things like zcan fhe but the idea that the h100 h200 tees are you know only suffer 30% performance degradation on inference for llama 3 kind of is something that I think in you can take advantage of within blockchains especially for co-processing and and things like that so it's not exactly a product per se but it's it's something that has existed for a while and has never people never really had Financial incentive to optimize the hardware as much as they do right now and I kind of think that that's sort of one of the things that will enable a lot of things that people in crypto have very expensive proof of concepts for but have never been able to to make feasible people forget but Nvidia was kind of Saved by crypto in many ways um before obviously you know they got ton of demand for quickly follow up on uh that part I I really like the idea of like uh making things more tangible because I think there's this idea of like doing uh thinking about F and say Zer knowledge proof going to be matriz in a year something which is like totally foring and that like some some of the projects we see in the whole space like trying to make uh things uh uh as like a verifiable Compu and Justice Compu and privacy protective methods they didn't even talk too much about te these days uh as a tangible solution that can build within three months rather they talk about a solution that you have to build for like three years uh I think there some some of us here have to think about like how to make stage byst stage approach uh in terms of using all these like uh potential te Solutions on gpus and CPUs that we already have yeah when we think about permissionless aspect it it can be really good but also it can backfire and can be really bad um there's there are certain people that think that AI will end Humanity at some point uh I don't ascribe to that view but maybe a couple of you guys do or don't I'm curious how you think about the permissionless aspect fits into you know are we getting closer to potentially AI destroying humans if it is run in a permissionless decentralized context or or does it does it not even matter so I was one of the signaries of the six-month pause letter last year which is mainly because I needed a break because AI was going so fast AI time is like five times faster than crypto time and that was already exhausting um but when I kind of look at it so open source is permissionless innovation because you can take this couple of gigabyte file or llamas 3 GB and you can do almost what you want with it I'm not actually that worried about things like that but more about governance and standards because in my mind I have one very concrete way that Humanity gets wiped out for example we have six billion robots in a bad firmware upgrade and that feels like a centralized problem you know um anthropic had a paper called sleeper agents with a few thousand words you can poison the four trillion words or 10 trillion words of a language model that it turns evil with a code word and you can't tune it out and you can't find it ahead of time so I think that the permission side does open up additional attack vectors but it can get more resilient as well and I worry more about centralized systems because we've all seen centralized systems have attack points and now like one of my old tutors from University built an AI that performs every human at pen testing but the opposite of that is the AI that performs every human on penetration of systems and we got a wave coming and I think open infrastructure has been proven to be the most resilient Linux MySQL Etc which is why I'm more in favor of that than the centralized Alternatives that will inevitably have flaws that are dangerous I mean I'm looking at you Alex because you were at open Ai and U you know yeah yeah yeah the yeah open AI one of the things so to give a little bit of background I was at open in 2016 uh that's before even the had a logo uh and I left open a to start near um I was there very very briefly in 2016 and then in 2022 after near was launched I was briefly at open a again uh primarily because uh uh by 2022 they have been uh they have been silent for three years and I I suspected they might have reached the singularity and I just wanted to be to be part of it if that was the case um and uh yeah and one of the things I was doing is I was walking U the corridors of the company and I was looking into the eyes of the people and I was trying to judge are they evil or not and it's very hard to do also evil people are usually good at uh not showing that they evil right so so at the end of the day I wasn't able to convince myself that this group of people uh is uh uh that there's a 100% certainty that that group of people will save us from Annihilation uh if the AI will start running away uh so from this perspective I think that open source AI like at the end of the day if there is a chance that the models will uh do harm uh it is much better if the models are available to everyone because in this case you can uh effectively good actors and the Bad actors will have access to the same resource uh to compete with each other than to have the models only available to two actors and then to find out that those two actors actually happen to be uh malicious uh and have bad intents so I think from this perspective we are way way better off uh if the uh ti1 research is uh owned by the society and not by uh uh a couple of actors and uh like like in us right now uh those actors are trying to you know to introduce regulation that will make it effectively uh impossible for anyone but them uh to a big extent to to train very large model so that would be a disaster if that happens yeah Elon Musk was saying hey guys there's a new law in California that forbids uh people training large language models I think this is a good law uh but he himself is in txas so of course it's good if uh uh in California competitors cannot train big models so it's a you yeah I think u a a maybe a more uh research version of of imods point is you know I think one of the reasons personally I like crypto research a lot better than AI research is you have to walk in with a threat model you have to say like here is the set of actions that can be taken and I assume a worst case outcome like any possible sequence of actions can be taken as input and I need to actually have the system robust to that like it has to be able to resolve itself from that failure or it has to not deviate in price too much like you have all these kind of constraints and in general AI research has been much more of the opposite right it's very empirical very like I take something I see if it works on on average it seems to work well you know I Define something like a a metric like an mlu or something like that those are these average case types of things so you don't really know the worst case except for hey when you search for poisoning a data set right and I generally think the crypto mindset of at least starting with a threat model and having this sort of adversarial belief is sort of something you only have in open source and permissionless environments you don't have them the permission there's really never an incentive for you to to Red Team your own code if it seems to work well enough on average against the benchmarks you choose and the idea that there's only one or two benchmarks that people are using to compare is also sort of unfair right for Edge applications I don't really care how well your language model does on you know translation like maybe I actually only care about for my specific subset of tasks and I think you only get this like multi- kind of metric comparison in the permissionless world you don't have this like Hey we're we're going to standardize on one or two comparisons and I think these are all sort of social reasons in that people in the in those centralized environments don't have the incentives to to kind of think about the adversarial stuff as as as deeply as maybe you do in the public environment but I think you have to have both Alternatives otherwise you get stuck in a bad Minima yeah I want to end on that um if you ask any of the AI researchers at AI conference like would you be comfortable deploying AI in a permissionless manner I think probably 90% of them would say no and I and I think the reason uh being two right one is even us we don't really fully understand what these models are doing and how they make the actions based on right because they are just probabilistic models they are not programs they make um stochastic sort of predictions uh depending on the actual environments they will deploy and also the initial setup that you have so sometimes they just make really funky mistakes that um might be as simple as like they just look at a slight different pixels that than the ones they are familiar with so that's like a one big reason they we don't have full um control and understanding on what these models is behaving like compared to say smart contract or any type of a programs that you have written and number two is being that uh there's so much of human uh social values and Common Sense constraints physical constraints all these things hasn't been fully uh Incorporated by these models in the sense that when you push them to high stake applications say feing resumes looking at um application to your credit card stuff like that I mean the textbooks of AI have been using these examples for decades but still are very difficult for them to solve these days in a very robust and and um WID spreading situation uh before that problem has been understood in a in a much more comprehensive way I don't think we should be thinking ready to deploy AI in a permissionless manner at high stake application but for low stake application I think it's good to take some risk and let it fly yeah yeah so we have 10 minutes left I'd like to just perhaps shift that discussion back a little bit to you know the product sort of see three categories one which is you know crypto um enhanced with AI features like you know a trading bot or some sort of risk uh mitigation kind of solution uh you have ai enhanced with crypto properties perhaps AI agents are using a stable coin and then perhaps the more interesting bucket would be the combination of crypto and Ai and and it's only uniquely possible when you combine the two um maybe if each of you could explore these three categories and where you think um where the most opportunity is um or or the the things you know you're mostly interested in um so from our side we have a piece on our blog how to think about AI we're like there's a new continent AI Atlantis coming with 100 billion graduates and billions of robots and they can code and they can sing and they can do everything and again the gold now but they'll have memory soon uh we think that's going to cause massive changes in the global economy like again how many of you are hiring graduates for knowledge work right now or you have the AI developer how many people write a smart contract in 3 to 5 years so one of our key focuses is you know the original of crypto was Bitcoin and again this permission list distributed money algorithmically and we're thinking what does a post- labor economic system look like and money there and this is something uniquely that I think intelligent money supported by AI with intellig market makers and all of this will be really interesting because the world has to develop uh that's a very hard problem and uh again part of that is building these Healthcare education government and financial fully open systems but the entire economy is about to change and that's where the opportunity comes in and I think that fits within the other buckets too but we haven't seen anyone else looking at a post labor monetary system what does currency look like and more in that space and just on that point like you're just saying crypto is better at getting human to do thing like the coordination layer there's the coordination layer but there's also the question of intelligence will be the reserve currency I think in the future because if you took if you take 10 coders to do something and then you add another 200 it probably won't happen whereas now we have inference time scaling again opening ey is one of the first things whereby you're bounded by the amount of compute that you have to do things and you can out compute people if you have more compute to run these agents as they become mature so we think that'll be the foundation of a new economics system system given again why would you hire any graduate programmers or most content writers when the AIS can do that more you have to have a re alignment of the social structure and that's coming incredibly fast again the average human has an IQ of 90 globally adjusted average open ai's new AI model scored 120 so big token 249 attendees would be way up way up there yeah of course yeah I think I think two um two big um use cases or sort of a directions explor for crypto for AI which is what I'm passionated about uh number one as I explained before um really blockchain can be a good bookkeeping system um a copyright sort of uh um um implementation solution to help you understand who has contribut what in the whole life cycle of developing AI models and applications right there data contributor computer contributor model developer who have the recipe applications who integrate the model end points into serving um really the AI customers who give feedbacks to this model to iterate to the next version which is also equally valuable and then to the final monetizing who are basically the AI consumers paying for the application all of these people here what they have done work can be booki by blockchain and then you can really Empower a lot of new economic mechanisms for building AI model which is much more accelerated compared today's Solutions in terms of building AI model that's number one and number two and I think the whole D5 ecosystem of crypto um has a lot of uh high expectations in terms of finding vehicles of investments into the AI ecosystem and they don't have too much choices today right the best way they can probably do that is either they have ecosystem from going to these early AI project Builders or they look at uh ntoo and try to find some stocks that has some AI relevances right but then why don't we find out a new way of tokenizing the AI access such as data sets and model and then there's probably going to be initial offering of these AI access to Benchmark their fundamental value and then you can have the whole secondary Market from the D5 ecosystem to start taking the upsides of these AI access right and I think the most challenging problem here is how do you define the fundamental values of these AI access you cannot simply just throw a stable diffusion model uh onto the marketplace and haaste I think this was 100 uh dollar per share because it doesn't make sense because it's open source model right it doesn't really come with a revenue model yet so really I think you first need to build a whole uh fundamental value infrastructure for The Priory models how to Benchmark the performance judged by the community how to um find a business usage and then revenue and then check check all of these things in a in a Justice Manner and then you go from there to really find investment vehicles for this access so I don't have a good example of uh Ai and crypto together uh you know enabling a unique use case uh but there are so uh What uh Sean just mentioned uh generally if you go to the Whiteboard and you start designing you know like for example how decentralized storage should uh should work right if you don't want to be using S3 uh how should the storage uh be designed for the future you will notice that there's no way to design it without the blockchain layer where people settle you know like agree to store data settle if uh if something went wrong similarly I think that the only way uh for us as a society to train those extremely large language models as we go you know into the future uh is uh through yeah like pulling resources together on the blockchain and having a smart contract that distributes the value generated by the model uh back to the people uh who contributed towards training it uh so those are case examples where crypto enables Ai and I think a very interesting example of where AI enables uh crypto is uh if you guys saw my talk earlier today I was showing an example a demo of a small application where you can on the Fly generate uh fully functional uh web three applications uh it's like an agent that has access to all the metadata on chain to all the contract abis and then you can explain what you want to do in a completely no technical terms and it will lay out for you a fully functional application that where you can do do things uh and that enables like today if you want to have to to issue any transaction that requires any compossibility you effectively need to know how to send lowlevel transactions yourself right like if you want to take advantage of Unis swap uh and some other contract in the same transaction neither of the applications on their own will give you that opportunity uh but with AI uh with AI agents that build applications on the Fly you can just explain your intent right and the you will get a fully functional app which is tailored towards your particular use case right so now all the people who are who understand well uh defi uh and have you know like interesting ideas of how to uh use compos compability of contracts in interesting way but they're not technical enough uh you know to construct those transactions now they enabled uh will be uh enabled in an interesting way with the I developer to issue any of those transactions using you know without using any technical understanding yeah I think yeah personally from a practical standpoint I'm definitely most interested in the AI impacting crypto usage aspect uh like Alex just talked about one of the main things I think we see complexity within crypto over time is that crypto has gone from generally monolithic systems to extremely asynchronous multi-chain multi- environment not one toone comp not like bjective in terms of like what two contracts are supposed to mean to one another and so because of that there's actually insane amount insane amounts of complexity for the average end user going into a crypto system to say I want to do this trade between a salana asset and ethereum asset there's even just something that's simple figuring out what bridge to use figuring out how to distribute your transactions or Bridge figuring out sort of like some optimal routing all of those things are only getting more complex there's almost like a second law of Thermodynamics for cryptos like every year it's just like insanely more complex than it in some ways than it needs to be so I think those tools that distill that are going to be extremely important and we already see that with a lot of telegram Bots that hide all that complexity for the user make it feel like they're just in a chat room or having fun and I think this idea of making crypto transactions fun to use is actually one of the most important things that AI can bring crypto on the other end of the spectrum of like the crypto and AI at the same time I mean I I just spent a lot of time playing around with 01 also have a lot of friends who work on that team and so I I I actually feel like the Paradigm Shift of moving things towards inference and things being much more adversarial lends itself to crypto systems where I can actually have prediction markets on the reasoning abilities of different models so you know as we were talking about earlier it's probably hard to make a direct Marketplace for a particular model you tell me you find tuned to model with some data how am I going to actually test that your MML type metric is not going to tell me whether there's actually a difference there's never going to be some perfect way of comparing them to figure out how to price them on the other hand for reasoning models I can actually imagine comparing different reasoning algorithms which are effectively adversarial with each other and figuring out how to price them and I think we will likely see that shift towards inference spaced uh models being more friendly to decentralized systems that's fting well we're out of time thanks everyone for listening and thank you for uh fascinating discussion really enjoyed it