Future of Onchain AI is Optimistic | ETHKL |

yep you me y okay and hello everyone I'm Conway from AA and today I'm going to introduce our lat research opml optimistic machine learning before we started we are going to talk about the proving system in the blockchain currently there are two types of popular proving systems one is z Pro another is a th proof and many people may heard about the Rob system and indeed correspondently there are two types of Ro up system one is the Z up which will utilizes the zero knowledge proof to guarantee a security such as a stock net and Zing another type of the Dr up system is the optimis up also known as op up such as arbitrum and optimism and for the proof time peretic the Z proof can prove something correct but the FR proof adopt a totally different approach the F Pro cannot prove something correct but it can only prove something it is incorrect therefore in the FR system people often directly upload the result on Trend and optimistically consider the submitted result is correct and there there always exists a challenge period and we're waiting for other validator to check the result and the cost of generating a ZK proof is extremely high due to the cryptographic limitation but the cost of generating the th proof is extremely low and here is a page that I connect from the layer two bit and we can find that most of the layer two blockchain adopt the optimis your up your up this is because the FR system is indeed much cheaper than the Z proof and the F system is much more mature and when coming back to the onchain AI many people might have heard about the zil which will adopt the Z proof to bring the AI result on Trend and guaranteed its coric and we wondering that why not adopt the frer system to make AI own chend with that idea in mind we have proposed opml and we have released the paper opml optimistic machine learning on blockchain on archive and if you are interested in it feelit to explore our white paper and reel coming back to the onchain AI we know that in zml we need to convert a AI model to ZK ml model and we need to and we need a ZK prover to generate ZK Pro for the AI model inference and then we'll send the ZK proof back to the back to the onchain verifier contract to check the ZK proof indeed the main limitation of zml is that the cost of generating a ZK PO is extremely high but in opml we do not need to generate a ZK proof instead we have first convert an a model to the opml model and then we can conduct the exclusion of opml model of train in a native environment and then directly upload the result on Trend and optimistically we consider that the submitted result is correct and the FR Pro is needed only when other validator thinks that the provided result is incorrect and want to challenge it compared to the zml and opml the main difference is that zml needs to generate a z Pro but we know that the cost of generating a z Pro is extremely high which limitated the application of SE ml indeep currently SE ml can only handle some T model but from the opml perspective since we can run the am model in the native environment and then upload the result on Trend and the frut is needed only where other challeng is besides the cost of generating a frut is extremely low therefore opml can handle extremely large model for example in for example currently with opml we can now bring the stable diffusion and llama on chain from the AI perspective there exist an AI M love that is every year the model size will increase by 10 and since opml can exclud it in the nearly native environment and thus opml can catch out with the model growth and bring the latest large model on 10 okay after introducing the background we are now going to look deep into the opml architecture there are three important components in opml that is the frion machine F pvn ml engine and the dispute G for the FR prion machine given a state given a stat program and is improv the FR verion machine can chase any in struction step and prove it on the layer one blockchain as for the machine learning engine we have Implement a lightwe framework for the machine learning that is ml go and for the interactive dispute G which will help to locate a dispute done into a single instruction and then we can res and then we can res this disp this this this story and then we can res res this this builds using on train for Pro veral machine and the opml process go as follow there exist a requestor which will first send an AI INF request on chain and then one of the opml nodes will serve as a submitter and then run the AI inference in his native environment using our ml engine and then the submitter will upload the result on chend and automatically we consider that the submitted result is correct and then and there will be a challenge period during which other validator can check the result and if if one of the validator thinks that the provided result is incorrect the verifier can inde challenge the submitter and generating a FR proof and if the submitter is wrong the result will be replaced by the verifier and in the challenge period other validator can also check and challenge it as for the FR verion machine the FR ver machine is for op exclusion and on train arbitration specifically you can consider that the FR version machine is a state transition function which will take the version machine image as a pre as a previous state as an input and then iscale one instruction and then we'll output a new verion machine image as a new state therefore the trace of a program executed by the prop verion machine is indeed an order set of the verion machine State besides the was machine image will including the memory State regit States and other VM State such as PC in the verion machine and the whole verion machine image will be organized as a Mer tree and only the root of the Merle tree will be upload to the blockchain and the root of the merro tree will representive the viral machine State as for the machine learning engine we have two core desde principle of the ml engine the first one is that we will separate exclusion from the proving part and this can guarantee the high efficiency of the machine learning engine another desde principle is about the deterministic machine learning exclusion which can guarantee if the AI result is consistent and verifiable under this design principle we have Implement a lightweight framework for the machine learning andl go as for the design principle of separate exclusion from proving we have provide two c of implementation of the machine learning engine the first one is for the native execution where we will use multistress CPU and GPU to speed up the computation in machine learning and which can guarantee a high speed exclusion another one is for the FR fruit program this one is needed only when someone start challenge you and you need to provide a FR in this implementation we adopt a cross compilation technology where we are compile the the ml computation into the version machine program instruction this separation guarantee that the ution is fast while the proving is based on the machine independent code as for the deterministic Machining exclusion we know that the native exclusion of the ml computation had not guaranteed the consistencies this is because the parallel computations that involve flowing Point numbers often provide nonidentical outcomes due to the running error for example the result of a plus b plus C is not equal to a plus the sum of B and C due to the running error if a b c are flowing Point besides the hardware platform operating system can also affect the Computing result of the flowing Point number in our opml engine the consistency is guaranteed by two technologies one the the first one is that we have convert the flowing Point number into the fix point this is also known as the compensation technique in the ml in the machine learning the second one is that we adopt the Sol the soft flow for the phing point calculation in the version machine and if you are familiar with jkml jkml you'll know that jkml cannot handle the flowing point and therefore zml always leads to accuracy l however in opml with flowing Point support opml can achieve nearly zero accuracy loss compared to zml as for the interactive dispute G there are in in the interactive dispute G there are two parties namely the Challenger and submitter the Challenger and submitter will run the same ma machine learning flpr program to identify who is strong they will indeed conduct a binary search algorithm to located the dispute point and when the dispute point is form the onestep execution will be conducted in the own train version machine for arbitration and here is a simple example to demonstr the dispute G in the FR system assume that assum assume that there is a submitter who have first submit a result on Trend but there exist another validator who think that the providing result is incorrect and therefore the the Challenger will start the challenge with the submitter and we we assume that in this simple example there are 100 instruction in the machine learning program and initially both the submitter and the Challenger agreed on the initial State s s z but disagree with the final State as 100 and then the submitter and the Challenger will start a binary search they first check the middle point that is state 50 assume that at this time both the submitter and Challenger agree with the state 50 and this means that the dispute Point must located in the state 50 to State 100 again the again the submitter and Challenger will check the Middle Point St 75 assume that at this time they disagree with 75 and this means that the dispute Point must located in the state 50 to State 75 and then the submitter and the Challenger will continue and will stop a binary search again and finally they will find a state where both the submitter and Challenger agree agree on this a state SI but disagree with the next state si+ one at this time the state SI will be sent to the OM train version machine for arbitration and if the output State si+ one does not match the submitter state which means that the result provided by the submitter is incorrect and the Challenger result will be considered to be correct and this result we're also waiting for other validator to check it as for the onchain application the onchain version machine will take the previous day SI as an input and conduct One Step execution to Output the next day SI plus one since the onion machine only is SK one instruction the size of the data that that needs to assess is only 01 therefore we can add the necessary data to on train veral Machine by providing a merro pass however the current prop system has has its own limitations which make it impractical to support very large machine learning program the first limitation is that the current Focus system has a very low exclusion efficiency this is because in the current Focus system adoped by the r up all the computation must be excluded within the verion machine but the verion machine can only be exclusive in a single stretch CPU therefore is it therefore it has a very low exclusion efficiency another limitation is that the F verion machine only has limited memory for example the MS version machine can only support a small 4 GB memory this limitation makes the current F made the current F assistance impratical to to support large machine learning program to overcome this limitation we have propos a noral multiphase proper system there are two design principle in our multiphase proper system the first one is the semtive excusion which can help to achieve a high excusion efficiency another one is the lazy loading design which can resove the virtual machine memory limitation in the semina exclusion with multiface design we decompose the dispute game into multiple phases and conduct the computation in the version machine only only in the finals phase for other phase we excluded the computational heavy part as a host function and which can be conducted in a native environment compared to the current F system adoped by the autom dra up the F system in opml is much more efficiency because this is because instead of running the whole program in the single Strat version machine we can now run in the multiat CPU or GPU or even using TPU to speed it up therefore we can achieve more than 1,000 acceleration in our execution as for the lzy loading design the key idea is that we do not need to load all the data into the ver machine memory at once but we can only keep the keep that can identify the data item when a virion machine needs to accept a specific data item the virtual machine can use it kit to fetch it from the external source and then load it into the memory also when the data is not needed we can swap it up with this lazy loading design we can now run a 26 GB large language model in a virtual machine with only 4 memory generally speaking with our proposed novel multiface proper system we can achieve High exclusion efficiency with semina exclusion and also achieve unlimited memory in the FR version machine with our lazy loading design and we are now able able to run opml with a 7 billion Lama whose model size is around 26 gab on a common PC even without GPU and now we are summarizing the compation between of opml and zml and first of all zml and opml adop a totally different proving system ml using Z knowledge proof to guarantee its correctness but opml using the FR system because of the co because because of the cost of generating a Zer knowledge proof it is extremely high they can only handle some small model compared to the Zer knowledge proof the cost of generating a FR proof is extremely low and we have already proposed a high performance multiface propositions and therefore opml can indeed handle extremely large AI model and as for the hardware requirement zml needs a very large memory for the ZK circuit but opml is much more friendly and opml can even can run on any PC with CPU or GPU therefore the service cost for zml is extremely high well opml is much more lower but opml has a problem of the finality which is also the problem of the optimistic R up opml needs to needs to delay for a challenge period but ZL can achieve instant finality when Z when a z z knowledge proof is generated from the security perspective we know that zml is guaranteed by cryptographic techniques using their knowledge proof but opml isq opml isq using frood and frood is inde is secured by the incentive mechanism and Game Theory specifically opml has great advantages on its performance opml can provide machine learning service with low cost and high efficiency compared to zml and the requirement for opml is very low and we have mentioned that we are now able to run opml with a large language model on a common PC but ZL will take around 80 minutes to generate a the Z Pro for a 1 million parameter small model on a servers with 500 gigabyte RAM therefore currently jkml can only handle some small model but opml can handle it nearly all the AI model for the SEC from the security perspective second canri graphic security and opml can achieve Economic Security under the any CH assumption uh here I going to introduce the antitrust assumption the antitrust assumption means that the opml can maintain the security even when even when there is only one Hest validator in this the security model of opml and ZL is quite similar to that of the Tim R up and Z up beside I want to highlight that the opml can achieve much higher security level compared to the existing Oracle protal that which can only ACH achieve security under the majority chust assumption we have provided some mathmatical calculation in our white paper and this figure shows that the an system can indeed achieve much higher security level compared to the majority trust system and as for the finality we can Define the finalized point of zml and opml as follow the finality of ZL is defined as the time that the Z proof of the machine learning inference is generated and verified and the finality of of opml can be defined as the time when the challenge period of the machinary inflence at end and and we know that generating a Zer knowledge proof for a large model will take extremely long time therefore when the model sign is large opml can achieve the finality even faster than the zml while zml is still generating is the JK proof one of the optimization for the opml is to shorten the challenge period so that we can achieve faster finality the challenge period is related to the execution time in the version machine and the number of interaction in the dispute G our M our novel multiface FR protocol can help to greatly shorten the exclusion time in the verion machine besides we can adopt the multi-step multi-step de appr and multistep challenge points to reduce the number of interaction in the dispute G for the multistep Z FR Pro assume that there is a ZK VN which can generate the ZK FR Pro for M step and then and then with the Z VM the number of intergration can be reduced from o log in to O log n / M besides we can also provide multiple chest point instead of only one check point in our previous demo for example we can we can provide K check point in the dispute game at this time we can further reduce the intergration numbers from log two to log k+ one in this with this two technique we can we can greatly speed up the the finality of the opml and it is possible to achieve almost instant finality after introducing the opml we may find that in opml anything everything El needs to be public for the verification so the problem is that whether it is possible both or how to preserve the improved privacy or the model privacy in opml to answer these questions we have released another research paper Nam opine optimistic privacy preserving AI on blockchain the key idea of op is that instead of running the whole model in opml also we don't want to run the whole model into the ZK mail uh in opai we can run part of the model in the ZK but the remaining name part is in the op in this way we can guarantee the improved privacy and the model privacy for example to achieve the Improv privacy we can we can left the first several layer in we can left the first Z layer in second part while the remaining layers in the op in the op part well this well this approach may not achieve fully Z level of privacy but it is possible to to ensure a secure and non reversable of the useless input as for the model privacy although it is not it is impossible to run the whole model in the Z M we can keep the privacy of the part of the model take the fight turning model as an example Laura is a popular fight turning framework we can keep the fightning park in the ZK where the remaining part is still in opml in this way we can guarantee the model privacy of the fight turnning Park therefore the servers can provide ml inference service without making it without making it f model public and open source here is the performance of opai since in opai we only need to keep part of the model into the ZK we can greatly reduce the ZK MML cost for example the cost of opa is only 20% of running the whole model in the Z M therefore Opa is the best practic for preserving privacy and also achieving high performance with opml and opine we implement the First on TR AI Oracle oo with oo people can use AI model result on Trend currently we have supported St B diffusion and Lama 2 in oo if you are interested in om AI Oracle don't hesitate to have a try the architecture of oo is as follow I'm not going to cover the detail if you are interesting you can visit the our G reel uh in oo oh the users Contra can initiated a AI request by calling oo and O sorry and oo will publish this request to opml node for the exclusion and then opml that then o oo will return the AI result back to the user contract with opml and Oo we can unlock many potential om train AI application and inde believe the future of om train AI is optimistic for example with onchain AI we can decide a intelligent Auto we can decide an intelligent amm where the on AI can improve the liquidity and the profit of the market maker besides we can sell the aigc nft and we can even generate a smart contract on CH as a int engine besides on AI can also help to identify malitia transaction and accounts and then we can also set up a distental AI Marketplace one potential or one potential application of the OM AI is the IMO initial model offering IMO can EP tokenized AI models with a token for IMO the open source AI organization can conduct fundraising for the ongoing development of AI models besides it can also motivate the open source AI contributor to to this open source project the io process go as follow the the AI model first performs IMO and then released IMO token and then anyone can buy the IMO token to own their shareff of the the AI model and the holder of the IMO mod the holder of IO token can earn the revenue of that IO AI model generally speaking with IMO the open source project can get fundrais and the token holder can also get the corresponding Revenue yeah that's all thank you and thank you very much uh for for this presentation it's very insightful um so before we go into the question and answering uh section of this um presentation um can you um I mean usually we'll ask this so um can you tell us a little bit about yourself and how you have come to join the um web tree or and also the opml uh scene and tell us a bit about yourself yeah yeah sure and you can call me Conway and indeed I know researchers in in both Ai and crypto I also do I have tried to build and train some large language model the in our pre in in my previous work and I'm also very interesting in Broad chains and and I inde a broad developer first so when coming to the on train AI we start thinking about whether is is possible to bring the AI on chain and then I explore some projects I find that most of the project will use the ZL to bring the A on Trend but as we know that uh Z knowledge proof is not efficiency enough Z M can only handle some model so I wondering there whether whether there exist another proven scheme there is much that that is much more efficiency therefore we can bring some large model on chain and then uh with this question in mind I I also looking for some AI some some blockchain project and I find that in a drw system there are two types of R up z r up and op R up but in the AI part people are only focusing on zml so I'm wondering that whether it is possible to build a opml like a op Rob system and this is the idea that why I propose opml and I try it and I find it is more efficiency than zml I believe if you want to bring some large model on chend opml may be your only choices at least currently oh thank you thank you yeah I mean it's really interesting how opml has came to be according to you um which is um yeah it's it's the response I believe to um essentially what zkl can get to provide um out of curiosity as well so for you personally um what's your favorite uh models that's recently released so you mentioned you you worked on LM before any favorite models that you really like um like favorite ml models you really like well uh inde I like uh Chinese models as pangu developed by Huawei and I am also one of the contributor of pangu inep andarge language model as well yes it is a large language model developed by by Huawei and okay for for other open source model I really like llama it is small but powerful and in in opml the first large language the first large model I bring it on trend is llama and and now you can use our oo to bring the Llama result on Trend oh great that sounds uh interesting yeah thank you thank you for sharing about it so now I will open the floor for um anyone in the audience to ask um any questions so yeah feel free to open your mic or even type in the chat uh while we are waiting for um our audience to ask some questions I I'll start off with one question yeah yeah go for hello yeah here you can go for it yeah uh hello uh yeah can can you guys hear me yes oh yeah uh nice presentation um I'm I'm still not really understand how any trust assumption works and how it make sure that there's a way uh there's a way as this some owners note to challenge the um yeah to challenge the the fra proof basically oh yeah uh let me showare this the slides and uh so are familiar with optimistic R up in opml workiz optimistic R up and there will be one submitter who will submit a result on Trend and automatically we consider that the provided result is correct and any not in our opml Network can verify the result and if the no thinks that provid result is incorrect they can challenge it and the proper system can guarantee that if the submitter is wrong you can you can definitely successfully challenge it and this is a reason why that we have say opml can be achieve security under the any trust assumption this is this is we can we only require one no in our OPM networ that is honest then we can guarantee the provided result must be correct all right thank you um but I'm still curious that how uh how you guys make sure that there are a notes up to challenge because I believe to to check um before make a challenge they have to run the computation again which in ml case is quite computation heavy so how you guys make sure that there's a way some notes that to run this especially in a very large language model that requires very heavy computational power like yes what the incentive to get to keep them running this discount notes yeah you asked a very interesting questions indeed the in this this is the this is known as The verifi Dilemma this is because the verifier in our noes may be lazy they won't check the result they won't like to they would not like to run the the the air inference in their local again to check the result so to to handle this problem and to tackle the verifi Dilemma we have proposed a attention challenge mechanisms and this mechanism is released in our white paper we have proved that with this mechanisms since uh the valid the validators in opml noes has incentive to check the result and if that and because we are reward we are reward that and we have proved the N equivalent of this uh of this sorry we have proved that the N equivalence is that all the all the submit will provide the same result and all the valid dat have the incentive to check it I see uh do they still get rewards in the case that the challenge is not Val let let say I felt that the there's something wrong but in the end there actually everything is fine is correct uh do they still get any rewards oh yes we had decide a similar mechanism the mention the problem you mentioned about also also exist in the automated R system because you know optimacy R system the verifier may not have motivation to check the result indeed we have designed a similar mechanism for example in the we have we proposed a mechanism that called attention challenge that is when a submit a submit result we require all the validators need to pay attention to the provided result and he need to run the result of CH and if and we will randomly slap one of the validator to check whether they have uh do this verification and if the verif and if the verifier have do the verification we are reward it otherwise this verifier will be punished I see so does that means that every notes who wants to par participate in this opml network have to St some tokens or crypto is that correct in order to to get SL yeah it depends if you want to get some reward because if you join join as opml no you can get some reward of because you provide some M service and if you want to be a submitter that is you want to claim some reward for opml and provide a result as opml then you need to stake some money so so in that case if you provide a wrong result you'll be slash but as for the challenge anyone can be a challenger even as user so if a submit provide a wrong result and you are a user you think the provided result is incorrect you can also challenge it with uh with this mechanism in if uh inde as a users you can always verify and challenge the result even there's no onest no in op in opml you can still guarantee that the provided result is correct I see uh I remember seeing that you said there's a be any trust assumption is a lot better than multitrust something like that uh can I know what's the benefit of using the any trust assumption compared to the normal trust assumption oh well uh sorry I'll find the slides yeah uh we are comparing the any trust assumption with the majority trust for example some some Oracle protocols like chain link the chain link need to get to guarantee the provided result the the result provided by chain link is correct you need to you need to think you you need to guarantee that at least the majority know of the chain link is honest but in opml we only require one honest no and this is reason why this is a reason why that opml can achieve a higher security level and we have do some mathematical calculation and we can we can find that the any CH assumption can always achieve a higher security level than the majority trust okay understand thank you um yeah I I still have another question is in regards Oracle uh how does oracle comes into play in all this machine learning stuff especially like the the one that you mention o o right yeah yeah I I still not understand what's o o about H yes and in this this is a product we released recently oo is on Trend AI Oracle and in this the consensus of the oo is using opml so it is totally different from other orle protal so if you want to here's the o overflow uh so if you want to initiate a AI insance you can initiate AI request using oo functions and and then the O function will emit a event and publish this request to opml noes and we also mentioned that the opml is running under the any trust assumption there will be many opml nodes and then one of the opml node will serve as the submitter and who who will submit the result on Trend and other opml node will serve as the validator they will check the result and if the result is incorrect they will challenge it with this Pro with the opml process the result will be sent back and upload to the oo and oo will invol the C function in the users comat in that way you can use oo in your smart contract therefore you can accept to many AI model for example Lama you can directly using oo to use llama in your smart chat I see can you give me one example of using uh AI Oracle instead of using normal Oracle like let's say if I if I just require price feed uh normal Oracle should should available to do the job already so so what what's the case of using a AI Oracle uh well when you want to bring the air result on trend for example currently it is impossible for you to use a llama or using a sa diffusion on chain but with Oo we currently support sa diffusion and llama you can you can click this link and we have provide a demo on it uh for example we we are not we are not supporting the sa diffusion in oo therefore you can mint a aigc nft using oo St diffusion to generate a a pictures and then we can B it as nft also for example for the Llama uh we have open soft reple we will use the oo llama as the intend engine so people will only uh interact with like people can interact with the smart Contra like I want to stand B 10 is then he can just write this test and then send to the OM train AI Oracle the OM the Lama will handle the handle this request and then we can generate uh transaction for the user to conduct it the the transfer the token transfer let see thanks for answering all my questions yeah thank you thank you um so um there's someone in the chat uh called Zang is he's asking uh what are is he saying what interesting topic what are the data that you need to create for proof and where do you store them yes uh indeed this is a problem about the DA for opml and currently the both the input and output of the AI model for example Lama is St on chain because the input and output is not so large and to to create a frut we need the user input a model and also provideed output there three Al we only need three this three component and the input and output will be stored on train and as for the a model because in opml we need we require that all the AI model is open source so we will not solve the a model on TR of course the Lama is very big so and we only we only use the open source model that is available on sorry yeah uh please I think um I might need to actually set set the room something yeah I think that's that's insane um yeah um continue yeah uh yeah uh so any other questions we do have have some questions sorry um give me a moment I really need to see if I can set this oh yeah uh YH is asking do you provide API or SDK uh yeah PML is open source and I can send the repo to you in the chat box withit a moment yeah thank you um yeah the sorry go ahead okay sorry sorry yeah yeah the G report is there yeah thank you very much so um I have question regarding uh how essentially um what's the some some difference between zkl and opml so uh in zkl U basically the idea is to rely on the cryptographic um security um to ENT make sure that the computation is correct and you run one very heavy computation um which is very limited right now and it can be very everywhere to to know that your your result is correct um yes however um opml um right now from what I understand it requires um essentially a network to have people have the verifier or submit fraud proofs to ensure that the result is correct um is that uh I you're correct yes that that's right so opml we need a network or say we need some R data to check the result and if the result is incorrect you can Pro the r data will provide the flut to correct it got it got it so um would would there be a case in which opml requires a lot more um machine to to maintain in the sense that because you require a set of verifiers to check for fraud proofs whereas in zml require um strong provs so the the the weight of your computation becomes running the models instead of running the provs to prove the models uh is that a problem right uh that's a good point but we want to mention that in the cost of ZL is at least thousands of tens of thousands over opml so in opml we only in opml we can run in the native environment and as for the Trap assumption we have made a simulation inde with 10 with data we can achieve a very high security probability so I think that the cost of opml is low and besides since we need some validators of no in opml we can incourage people to join our opml Network serve as a service as a validators and they can they can earn their revenue from opml thank you thank you um great uh any more questions the audience yeah U Can we have the slides oh yeah I can share the slide later thank you uh do you have any test note now uh sorry currently uh the the opml no is commissioned in it not allow other people to join our opml because our FR is still in testing and auding and and we are make open sour and permission L later I see um yeah so is there so given right now opml is a so solution for um resolving the issue where zkl cannot resolve that that is um it's very hard or very competition expensive to to run a u proof on like a big model like chbt or even llama um when zkl becomes mature enough that um it can do such big computation with a relatively cheaper price um what's the plan of or oo would you switch over to use zml or do you think uh at that time opml still have its merits well that's a good questions we also considering the futures of zml and currently we have do some work combining the zml and opml and because we know that the zml is very limited it can only prove some small model so one comination way is that we can participate if the model into two part the one part is in Z Ma another part is in opml currently because the for example for the seven billion Lama it is impossible for zml to prove it zml can only provide approve for the first for example for the first five layers and the remaining s layer must be run in opl and when the ZL become more much more efficiency we believe that we can put more layer in the ZL to guarantees its private private input and guarantees is is it it efficiency and as for whether it is possible to whether it is possible that zml will totally replace opml well I well I think that may be too optimistic because have a slce here uh this is because the the motor size increase very very very fast which mean that that unless we have proposed a very cheap zkm everything since otherwise because opml can be executed in the native environment so unless ml can also achieve such performance I think uh opml is still the only way they can catch out with the model growth and bring the latest large model on chain thank you thank you yep that sounds it sounds uh like this that's good news right because I think opml definitely have its place um in in the ml world and ZK at least right now it's very difficult to to scale to like very huge models and I think in the future will still be expensive right unless that's new brick to yeah yeah I think that's um any more questions if not then we can oh yeah there's uh questions uh in the comments um so so it's asking about the local Noe uh in in chat if you can take a look at a shot okay yeah currently we we do not publ uh uh uh currently we do not public uh make a public and permission is opml Network so you can not join our oo currently but we are makeing permission L later but using our opml repo you can create your own opml you can run your own op PML no and create your opml sub and currently we do not invol any token in our repo so you do not need to generate any token and all coins sounds sounds interesting y definitely going to try that out um any more questions not I think we can end this session yep I think that's all for now thank you very much Conway uh for thank you for coming here and giving us a um presentation on um opml and it's very insightful um sorry for the disruption in the middle um I did not expect that okay yeah thank you very much um I'm going to stop recording now but for anyone who wants to ask question later you feel feel free to um let me just stop recording now yeah thank you