Fireside Chat with Juan Benet & Illia Polosukhin - Bangkok 2024

[Music] well uh yeah I mean thanks for uh inviting me here uh hopefully everybody's having a great event so far and uh as always great to see you great to see you too uh uh it's been I'm sure you're like super tired of this but it's been amazing to see what Transformers have done for the world and they're going to continue to for the world so thank you for that um and yeah how are you like yeah maybe a question for you um how has your mind changed over the last year or two in how you think the world is going to unfold over the next two to four years um what did maybe you used to believe that you no longer believe or what do you think it's like accelerating in terms of um yeah like how has it like your your world model changing part of why why I want to ask that is I want to help you and I spend a lot of time thinking about this sort of stuff and whatever we can help distill out for people uh the better people are going to be able to make decisions yeah I think I mean I I actually tweeted this because I've been thinking about AI since I was 10 after I watched like artificial intelligence uh movie and uh for me kind of a lot of the things that are happening are just like it's more like it's kind of happening too slow type thing not uh not just like hey this is you know new new like conceptu new things so um it's really hard to say like hey I changed my mind it's more that actually what did change my mind is crypto when like because I worked kind of on you know full-time on AI effectively for 10 years before going into crypto and uh effectively my model was always well you know you need to build a profitable company to effectively inact change right and so like you know if I imagine okay how does this work and how would this work and how would this come to Market I would actually normally would think more like open AI Google style but as you know as I was learning and uh um when we you know when we're working on neari originally in 17 we needed to engage a lot of people around the world to kind of contribute to our work uh with data labeling and we had challenge paying them and so crypto was like Hey you know this is a kind of way to pay everyone globally you know uh not don't need to bother about borders don't need to bother about like specific restrictions that every country has you know you just pay them and kind of they need to I mean they still need to figure out how to off them but at least like uh that's a solvable problem that people know how to do but as kind of you go deeper down the stack right you realize well and I think I I forgot which year was that but I remember it was in Paris during one of the EC conferences and you had this uh talk about like borrowing from the future and so this idea that like hey actually the capital formation can itself be different because you are able to uh instead of creating this corporation that forever needs to grow or it will die right uh you now have a different model where you have people effectively contributing Capital to build something that they can use later and and it doesn't need to effectively you know absorb the whole planet or die it can actually maintain some status and live as a as a uh as a network and so I think that that's a conceptual change and kind of I now pitch this back to traditional AI people because like hey look what's happening right now is actually very scary if you have you know handful of companies that are controlling effectively the intelligence tooling that we all going to use and I mean literally right now in uh for example um Thailand you know you go across the border you don't have access to open ey anymore right uh they can you know right now shut it down for everyone in this country uh they can decide that you know you you it shouldn't respond with medical answers to you or if they want to change something for whatever reason I'm not even saying it's malicious it's more like business reasons right for example I was I was suggesting this idea that like hey you know closed uh uh you know Corp corporations will actually inject ads into the uh models and then uh like I was suggesting as like you know kind of thought experiment and then I learned somebody actually was doing that I was like hopefully not from my suggestion [Laughter] um but like in general right like we have no idea like how the decisions are made and again working at Google before I know how I was making decisions I I was like well I need to filter out Fox News because they saying Barack Obama was born in Kenya and it's you know reducing quality of my data and so the easiest thing is just you know cut out all of the rightwing media now the model is super liberal like not because I'm liberal but because like this is how I cleaned my data right yeah how do you see um the importance of Open Source training and open source inference shifting with time right so a key point that you're making is you have these Labs controlling the the the build out of the models and who gets to use them uh the open AI example can today be averted thanks to llama so people can download llama and then use it and maybe you get an open AI equivalent uh but it also depends on Facebook's extended generosity in continuing to train um Cutting Edge models um yeah how do you think about that yeah so we're announcing something tomorrow the alpha yeah this is not live streamed right yes it is all right we'll watch tomorrow the redaction talk but I would say kind of high level prediction markets are going crazy right now high level the like this is a big problem and like the problem is it's also nobody actually knows what Facebook train it on um it's also like I mean gbd 4 is actually smaller model than the Llama like 400b and llama 400b is way worse it's like barely better than 70b and worse than like in of 70b models now so like the reality is a we don't know we can't improve on it like effectively because we don't actually know what it was trained on where the issues are Etc so it's not really open source it's open weights y now um if you step back why we want open source we want open source because we want the guarantees that you know it it does what it does this is kind of the bias and the data that went in and so we know how it works we want to be able to permissionless access it and we want to have a way to build on top of it and you know fine-tune it and do all those things now the challenge is open source and this is global challenge with all open source is that it's not monetizable right uh you build something any Corporation can grab it and use it and you know there's no attribution kind of back to this now there is a new somewhat new uh kind of way of doing things which is called trusted execution requirements and specifically this year uh kind of new Intel chips have new kind of model of this as well as the h100s and gpus working together you can create an environment where you can actually execute code that the host machine that runs this code doesn't know what it runs it doesn't know what data it runs on and you you as a external party can have guarantees that this code was ex executed so traditionally this was either impossible or was very complicated the previous generation of tees were completely unusable so this is like a fundamental change that happened like within next within previous like six six8 months what this enables is actually we can train a model inside a te right where you know what code it runs you know what data it runs but the model parameter are not open they're encrypted and they're only available inside tees now if you have your own te you can download it there and run it uh so you have permissionless nness if you have and you can get it from anyone else who has the same te or bring it on the flash drive encrypt it if you have you can fine-tune it in s te as well so you can build on top of it you know what data went in you know what code went in so all of the properties we wanted are there uh and the one extra thing now you can do is you actually can monetize it you can actually charge with a payment channel for every inference for every token uh effectively creating a business model that actually creates Revenue stream back to fund the research and fund a training new model and so now you can pull Capital right going back to what you were just saying pull Capital to train the bigger and bigger models with a business model that it actually will return so that is fundamentally what's needed and this is what blockchain actually uniquely enables right I mean this combination with this Hardware uh to actually create business model and pull resources borrow from the future Revenue stream to actually uh have truly open source approach that's pretty awesome so if we um unclear what's going to get announced tomorrow but if we can ship a Cutting Edge model trained in an in this kind of uh verifiable computing uh style and making a network owned resource that could be transformative for the landscape today so I think that would be uh that would be super valuable um how do you think about uh both deciding what goes into the training set I mean certainly you can train many models over time you know as things get cheaper you can train many of them um and maybe how do you decide the inference part of the equation where uh there are some concrete you know there there are certainly some regulations in specific locations that maybe are warranted maybe shouldn't be warranted right like preventing medical access to medical information inm is a deeply fraud ethical quandry right um if you if you do it and there's a mistake in the LM you could cause a lot of harm if you don't do it then you're also causing harms of omission M especially where people don't have access to doctors like there's actually I was talking with some Kuran assistant and their African user base massively uses as just the doctor because they don't have access to any qualified doctors and that probably has anyone done a graph on like how how that is improving uh what what impact that that is having on the actual health I mean I I don't know if there's like way to track that down to like local data on you know hospital visits and stuff but deepen for uh human metrics right like uh you know deepend for like World Bank data why not let's do it chip to everyone right I'm sorry that's that's from a different different story different story it's deep in it's fine beer to beer no but but so so so two questions right so one is so one actually is interesting because the t is also enforce privacy because now instead of you sending all the data to again one of these companies it actually goes to their databases right this actually goes now into te it's encrypted end to end and so you can always you know keep your data but there's no other third party that ever sees this data and there's no kind of additional enforcement that can be done on that and now you can so now you go back to like what models should be available through this path and this is so there is a question of like kind of safety and ethics that that is there and that is important sadly a more important problem right now for this is how to get a good model again right now open source models are way less effective than closed Source model uh even trained by large centralized lab right spending whatever hundreds of millions of dollars so the and this goes back again to approach like building an out centralized lab will not work right I mean you're not different then from uh what so how do we use there's only so many large tech companies that are going to bankroll you yeah I mean you're going to be pretty much competing for the same talent in the same way like just throwing you know millions of dollars the same four offices to go pitch them for money like yeah and so so how do we actually use the same approach right this you know kind of we've been using blockchain in such a way where we can incentivize people to come up with new ideas come up with new methods come up with new data cleaning procedures these new crowdsourcing procedures uh and effectively at the same time collaborate and compete right this is what kind of blockchain is like this Cooperative competition uh that we create like for Block production for everything we do and so again the the thing we're going to be launching is competition related it creates kind of effective like a model and we're going to be iterating on how it works to pull Talent from anywhere you know it can give them resources to be able to train this models to be able to iterate um you can then again see the results you can apply whatever the set of requirements that you know the community thinks like again somebody is funding this model so for example if Government of some country says like hey we want to fund for example model that speaks Thai really well they can also enforce like and we wanted to not do this and do this cool that's you know they can they can see all of the restrictions and then so you break down the problem of the uh ethics and safety policy space into local jurisdictions deciding or communities again like a a you know crypto Community can we're going to find like the most crypto anarchic model how do you solve the Earth level question of like I it'd be great if these things don't kill us so so I I think actually that don't kill us is a different question right so so so there's like few levels of safety to think through one is first of all I actually fundamentally disagree with like AI will go and think of killing us because this is the most like it's a lot of energy extended and there's no return on this so like from a effective AI decision making there's no like you know economical reason to do that we we could dive deep into that but it probably it could like um we should do it make a bet sometimes this is a grade bat type uh question um uh not not in this to put you in the spot but at some point we should have like a good Fireside diving deep into this with But but so there's like few again few different levels of of kind of the safety and and and and security system but the interesting part is this this is something that like you know I know people from open their position is hey we don't want to open source like they're trying to convince themselves this is not for business reasons so they say hey we don't want to open source because people can use this to now break you know code and you know know do some other malicious things and so the reality is actually yes this model that's also true like yes I understand the business mod so let me let me get there yes so this models are getting better and better at actually breaking code like I actually tested This and like it actually finds existing vulnerabilities pretty like that you know it didn't see in training data um so that's a startup on itself and the question is like okay what do we do with that like how do we actually from base principle solve this problem and this this relates to kind of AI safety in general and the answer is math there's one thing you cannot break which is math everything else code is breakable everything else is breakable math is not and so what you need to what we need is we need to actually have a formal proofs for every single line of code for every single program and for every single interaction that any agent be that human agent or AI agent can take now how does this work well we as humans are pretty bad at form approving so we actually need the to help us so that's a separate uh and we we we'll talk about that in in coming weeks um on how to do a kind of use AI for formal approving and how I think it's a great Direction great Direction a lot but the interesting thing is how do you apply it back to everything we do and the interesting thing is right now on a blockchain when you're sending transaction you actually don't have any guarantees ex like yes it'll execute the smart contract and the smart contract can do it like there's a Code there that may have and can do whatever yeah that's very useful and so what you can do now you can actually have the guarantees on a transaction where you say hey I require that everything I interact with from here on out satisfies that this set of principles and so smart contract need to have a proof that it will not lose your money right and if it doesn't your transaction will literally not compute that's will reject formal verification at at runtime that's a great idea run time now let's extend it go offchain when you interacting with other AI agents they need to also comp you to Pro like provide you a proof that they're not going to do XYZ is that harder it's harder but it's still possible right and again this is like we're using AI to do for modification but again like they still doing a state transition one way or another and this state transition may like it may involve dispute resolution it may involve kind of and so in effectively you're making this intents which are you know having set of restrictions and have set of conditions and they matching with whatever agent solver Services whoever is executing like matching executing them and then you have set process that verifies that is correct so there's a whole framework that we need to build kind of from gra like from the ground up to kind of have a way to specify set of uh mathematical kind of requirements which maybe like lose in some places where you want them to be loosed and very Concrete where you need them to be concrete to actually ensure and then you know if people say like hey this is our definition of safety like it should not do this and this cool write it in math you know on all the tees and now the model will you know be actually restricted from doing that so so that's kind of how I see this is like it's a you know we'll progressively math mathematically Define what we actually mean by all these things and then we can ensure that the Machinery we're using the code we're using is actually satisfying it uh two two questions uh building on this uh and that makes a ton of sense and I think it's going to be a really useful Direction um one question is around economics um and I'll kind of paint the direction and then describe the other one uh you know we can go down each path in whatever order you think is best One path is how do we Orient the economic value that these models and the inference that is going to that they're going to make uh how do we Orient that to align with humans broadly defined both in the short term with you know groups of builders that are building Commons or like you know you know societies at large city states and whatnot uh countries um and then second uh there's about to be a large proliferation of many different fairly intelligent agents with their own access to Capital and resources and that could like that that also has a question of um from one perspective like who are we to tell a new species to not have rights um that that always isn't a good idea um and then separately how do we Orient and align the economic the broad ethical and economic interests of like the community of millions of agents that are going to be appearing over the next two to two to six years um so I don't know I don't know we you want to dive into first what's harder what's easier aligning humans or aligning Ai and humans I mean there's this the joke I use is like blockchain is like ultimate meritocracy because it doesn't matter who's on the other side it can be you know human it can be a cat it can be a ey doesn't matter right like if you have made commitments if you made transaction you're executing it you have an address on chain that's kind of all you need to know to actually uh transact now obviously it's a little bit simplifying thing but broadly speaking what we see now right the CI agents will be indistinguishable from humans on the other side and you know yes there's like we need to apply all this restrictions and proving Etc to like make it more robust and so the question here is how do we um how do we formulate a effectively a value function right a goal that is both makes individuals successful people and AA agents and is kind of globally not trying to destroy the world right and so I think that that is a really interesting research problem I don't like have an immediate answer y but at least like kind of the way I I'm looking into this is to start by inversing the current system which is like corporate focused right you kind of you know and single country Focus right I mean there's a deep geopolitical inequity there that it all works out well you know the borders are free and the Commerce is free but the moment you get shut off from the AI pipe like that sucks for you as as a nation state well yeah or or Internet or yeah everything so um or money for that matter and so so the question is yeah like how do we how do we make it individually focused how do we make models and kind of agents and systems that are focused on how do we make individual more successful more you know monetary successful as well as like General well-being as well be that health or be that mental Health Etc right and so that's kind of I mean that's why I like call it user owned AI like to me that's kind of the first focus and then we do need to Define kind of global function as well that kind of maybe puts constraints on the overall system again I I'm not probably not the one to do that but it but it is something interesting to figure out kind of how do we balance this like individual Community level with kind of global you know we are in one constraint system and so we need to like as same time satisfy set requirements there and and for the how do you see this evolving over time with agents emerging like do do you buy the notion that there will be a ton of these little agents or do you think it's going to be more you know glommed up into larger SLE t or no I think that so the way I I describe it I mean because like agent is such a like wide term and it like covers a lot of things um there is a effectively like a agent Services right so this can be like effectively any store any like any even individual can become an agent service an agent Service is like providing something to everybody else in some specific way and you know it has a natural language interface but you can pay for it you can you know you know it requires set of requirements for example I only work with you know kyc people or whatever right like there's kind of that framework something we're working on there's agents that are effectively just tools for yourself right so this this is less of and then there's autonomous agents which are effectively like manyi not like nonprofits uh that can be generating their own profit and reinvesting it into something right and so those are completely different and kind of the way they work they can raise their own money by issuing a token they can use this money to like invest in do things they can capture this profit and reinvest it Etc yep and so uh so I think it's like really important to distinguish this different types yeah I mean more like the latter case where you have party AI powered entities that are going to start being economically relevant on their own right and they have their own entity yeah yeah so I mean those are really interesting and I think there's like set of use cases where this is extremely uh valuable and important right I mean the like any kind of uh support any kind of uh distribution of uh uh like when when there's a um catastrophic event the war breaks out those kind of things are really interesting as a model where like the CI agent can coordinate relief and even some people on the ground to do things uh and again this is because like it kind of fits this nonprofit structure uh really well that's inter that use case is actually very interesting for the community here so you and I are both very keenly aware of how uh coordination super breaks down in any kind of high conflict scenario any kind of war and while maybe the the core of the conflict gets attention all kinds of other things starts falling apart and we now have the internet connectivity globally and we have good ways of getting information around but we lack good coordination tools to actually get people to operate well um how do you think like these tools can help the people here build these kinds of better coordination systems and Frameworks for either jumping in on like an an emergency scenario or just better governing and growing a commment in general yeah I mean again there's a lot of work to be done but effectively the interest like the part that's interesting is what usually breaks down is when you need to coordinate a lot of people is that one individual can only talk to that many people it's also like you know as as running organizations we know when you do Town Hall types meetings most most people don't listen uh so you to repeat like seven times uh but the interesting thing was was effectively even now with Ani model and like a little bit of configuration to to make it agentic and have some memory and have some kind of tools is that it can effectively manage any number of people right it can know exactly what any individual is tasked to do and it can you know check up on him them and it also can do that across you know 10,000 people 100,000 people without any hierarchy right you don't have any broken telephone anymore kind of through the hierarchical structure of organizations and so now the question is like how does it make decisions is it making decisions effectively those are like research problems on making them this model is more higher quality but the From perspective of like you know you have an individual like an app that actually like hey you know I'm your whatever uh help helping people assistant and you know here's a task for you today and there's you know $5 for you if you do it and there's another person who gets the same to like validate that you did the task for example or whatever and they get paid you know 50 cents for that and so like you can actually coordinate like a pretty big crowd of people who you know each one they may still even have a context of what they're doing they can ask like oh you know why am I doing this they're like well you're doing this because you know at 50 people will come and do this five different things and together will achieve this right so like that's a part of which is right now people like don't fully understand why they're doing some stuff and and it's like because missing context they're missing maybe motivation think it's buildable now like you could probably build this now you can build this now yes AES so if somebody wants to build it let me know you you could solve so many large scale coordination problems in these kinds of way yeah and so I call this a IO and so like I'm most interested because if you think of if you think of a blockchain blockchain cannot have a human to be in charge of it yeah but it can have an AI in charge of it that is fully verifiable you know what come in you have token holders effectively being able to vote on I would call that like like a digital corporation like a like a digital AI powered Corporation yeah and so DS are finally coming it's been like 10 years and and and so yeah and like whoever I mean somebody sets it up either an a corporation or a person and they Define set of mission and goal and how to evaluate its success then it it's really cool that these models you you can get a very strong sense of like impartiality and Effectiveness and so on especially if you combine the the verifiability properties like that's amazing that that could like rewrite how we do all human organization today it would be super super cool um sorry I was going to ask or um are you open to uh audience questions yeah yeah just maybe to mention so one one thing that kind of wasin this like even before AI coord one thing we've been kind of in the ecosystem testing is this idea of small tasks because we're going back to where we started we were doing crowdsourcing data labeling and a lot of it is just a small task that person does and the problem is like people lazy they don't want to do stuff they like skip they want to you know make more money and so for about 3 years this project called near crowd been testing effectively how to build a game theory of people doing work and then other people validating it and this system not collapsing because people come up with all kind of creative ways to collapse the system uh there's honeypots there is effective like game theories there's you know something at stake Etc but this kind of a framework itself can serve now for AI to give tasks because you know people then go do them and then somebody verifies this again some money involved for paying Etc so like a lot of this pieces are coming together and kind of creating this effectively more distributed framework of doing work yeah very exciting all right um let's take some audience questions um I think back there can we get a mic uh running around other than R2 or here take this one and uh I'll get the other one thanks thank you uh so a few questions so one is um with the Advent of you know we have this like new neuronal architecture network uh architectures like um you liquid neural networks and you know some of the you know mamba or uh can or whatever uh where do you see um those going in comparison because you brought up you know like Lama 45b super expensive toin Etc so would do you see like new neural architecture network uh new architectures kind of supplanting the old ones and then potentially reducing the risk of the the necess necessity of centralized training uh and then second question is how do we deal with this kind of like very the the problem of like the really like Rogue malicious actor who in the same way that you know people can create viruses on their own if they have enough resources and you know that that's a massive threat uh like a rogue individual actor with enough resources to create um AIS that can like really do a huge amount of damage uh like and you know how do we mitigate that or can we do anything about that yeah so good questions um I mean answering the second one first our World in general already has some mechanisms around that right and I think my point is actually the more every individual is equipped with AI tools and the more we figure out how to make those tools to coordinate the easier it will be to find if somebody Rogue is trying to do something right it's like the same you know if like because Mo majority of people are not trying to do like to destroy the world right so if you majority of people have like more competition resources to actually you know like just stand and optimize for individual and overall success right they can identify when somebody's trying to like you know zero out the value function so uh I mean there's few things around like like I really don't believe in like restricting compute and doing all this this just like doesn't work and there's always kind of uh but I think like there's some combination of uh like how normal you know police and investigation work if they're equipped with AI if they actually they can you know detect when somebody's moving in that direction um and at the same time kind of uh enforce some of the again more like what is the form formal set of rules right that we actually want to enforce on the whole uh world but it's actually expressed in math and like we all kind of need to agree and commit to it versus right now it's like natural language interpretations cour you know all kinds of stuff right so um so I think that that is a direction kind of I believe in and kind of I'm working on uh and then the first question was uh different architectures oh yeah yeah yeah so I I mean from my perspective like I worked on architectures and you know there were new architectures coming out every like year so the fact that like we kind of actually stayed on Transformers and that is like kind of a shelling point is was actually more surprising that like there should be more stuff coming out I think for specifically I guess attention was all you needed I I mean seems like yeah I mean the the reality is and and this is like kind of more fundamentally the the when you when you go to like trillions of parameters right or hundreds of billions of parameters you can express anything in that model right like the the amount of bits of information you can store in that is is insane and so the question was more like how to train that model we didn't have effective ways to train it and Transformers is just like a very effective architecture for training now to your point is like are the architectures better suited for distributed training the answer is probably yes maybe it's you know some modification Transformer maybe it's something new but this is why we need like open up the research and actually like allow people to uh iterate and and try new things versus like close it up what's happening right now and everybody just kind of zeroing in on one thing and just like Ching on it yeah um so famously okay can you hear me yeah famously it's very expensive to train these models um and then like te it seems would probably add overhead how much overhead does te add to the training costs and to the inference costs and maybe also latency yeah so right now te GPU is like 71% for inference and uh we haven't fully like done a training run to measure that but it's probably actually less because the overhead tees come from the encrypting decrypting data uh the GPU itself is like effectively the PCI bus is blocked only to access through the TDX and so nobody else can just access it and so there's nothing there's no overhead on GPU itself it's all overhead in like data marshalling and encryption decryption and and so uh with b100s new generation the networking itself like interconnect within CL uh cluster of the gpus will also be uh kind of encrypt like support that where like you cannot tap it into it and and see it and so you have kind of pretty good guarantees that if you have a cluster let's say of like 4,000 machines you can spin it up you load all the data once and then you like you know you train for a while and it's just doesn't incur additional overhead so you can amortise one one encryption across all of the well you still need you need to do it in parallel on all the machines to like load different batches because you do data parallel but uh kind of beyond that like it's it's whatever effectively you need to to marshall training data in and and kind of to encrypt it into TDX right then so INF that's why inference cost is actually more because like you need to Marshall millions of tokens back and forth yeah and then for the formal verification stuff want to give other people a chance we can one one last question and I think we're out of time um do you know of any feasible attempts of some work groups of how to collect trillions of dollars for public good AI development I don't know if anyone who's doing trillions of dollars but uh you know we're going to we're going to start doing get there in like 10 15 years yeah I I think so right now just to give you some numbers right gbd4 is roughly like you know if you if you assume something is roughly like $80 million for training if you know what you're doing the problem is obviously you need to research and figure out what you're doing but um and also like obviously gp5 is probably like $200 million but like the point is you know you don't need yet like you need hundreds of millions of dollars which is you know in crypto we actually like very effective acely know how to collect and and and contribute and you have a repayment cycle of like you know 6 months probably at like uh good utilization and so like that mass is is working right now and as this thing kind of becoming bigger you know we like the gpus are fee structures on top of whatever applications emerge right like it's super easy to add good fee structures on top of useful products in the way that currently when you build on top of a normal open AI style LM it's difficult to like connect the flows of of the value creation yeah I mean I think here like you can literally tap into the already payment Channel and just say like well you're going to pay a little bit to GPU a little bit to model Creator a little bit to the Fine tuner a little bit to data owner right and you know it just accumulates over millions and Mill like trillions of tokens and then you distribute it uh awesome so thank you very much one last question for me which is uh earlier we were talking about Al getting crypto networks to allocate 10% or more of their value flows to pgf are you guys going to do that well putting you on the spot if if we call you know training AI models a public good well depends on how they're used but well I mean the goal is as I said we mean also like uh public goods more broadly defined than the direct production utility built into it like you know a a national government with a with a large scale Revenue flow has to take some amount of the proceeds and reinvest those into public goods that are hard to fund otherwise and so crypto networks need to evolve this kind of structure of like creating some value flow and then extracting some amount of that and recycling it back to Broad public goods so that's kind of more what we mean yeah I mean we can discuss more on this right now effectiv near near is functioning like that except public goods are more limited to the near ecosystem right like reinvesting infrastructure reinvesting in kind of R&D uh more broadly uh you know near for example developed borish library that is used by salana and by other people so like we are developing and everything we do open source the te like infrastructure will be open source as well um so you can kind of like the idea you can use this on your machines as well right so not just in kind of shared cluster and so anyway I I would say like right now it's mostly through technocratic lens because like we believe that's the most impact we can do right now but you know as near Network you know grows to bigger and bigger uh I'm sure there will be other opportunities it'll just be easier to do it now while the CTO market caps are smaller than later when justifying allocating an entire trillion is harder you know anyway uh thank you very much everybody uh I really appreciate you being here uh thank you so much for [Applause] [Music]