Key for a Better Artificial System: From Gov to AI (Panel) - Tokyo 2024

all right can we all please give a warm welcome for our next panelist we have here uh rioto Kai of Arya Jimmy Chang of fabd and chuj way of gov and Taiwan and they'll be discussing today key for a better artificial system from governance to AI let's please give a warm welcome [Applause] [Music] so I please my slides thanks okay so uh I will you uh in this panel I will use a very brief time to talk about uh our project especially our projects happening last year uh we received the open a grants and then we will talk move on to the discussions about how to form a better artificial systems these artificial system is not only talking about AI Jingo but also talks about the S talk about organization talk about how can we form a new types of organization like kaisha like uh kak like safe these kind of things so okay let's move on so first of all before we move on to the introduction we need to talk about how this V Tai 1 so V one has three feature first of all it a human Centric public participation mechanism which means that we use the digital tools but also we focus on people we use digital tools to breach different peop together and second things is that we are an open transparent commun community based approach which means that we are actually uh you can find our records on uh websites on our websites and also on Artic Co editing notes everything of B Tai one are all on that notes and finally we have potentials to integrate Innovative tools how can we have these kind of potentials to inter integrate these kind of innovative tools is that we have four stages which means that we don't stick to one tools but we choose the tools that is appropriate for our discussion so in our first stage it will be the proposal stages which means that the civil Society or the governments can bring the issues to the RV Taiwan community and we will help them clarifying the key questions and to do the stakeholder exploration and second stages we will collect opinions from the stakeholders and we will do the online delation and fin in third stages we will do the face-to-face deliberations which will be hybrid mode in live streaming and face Toof face dialogue and finally it can be the realization which means that we can finalize the consensus and du the legislation process so by using these procedures we are able to win the open n grants out of 800 different teams in the road and we also cooperate with a very wellknown s tank in United Kingdom which is chattan house so in with the foundings we are able to hold these kind of democratic input AI civil assembly so you can see that it is actually a lot of people out there uh we have uh successfully accumulate people with different background together with reporters with lawyers with computer Engineers with people from indigenous people with people from uh gender group with people from working in the tech technology companies we also even have um the legislator which is this one so we have legislators in our discussions which means that we are able to provide their viewpoints to the government members and this is our final report so uh this final report is uh cont contribute by a lot of people so if you want to see the final reports you can always scam the QR code here to get access to the final report in this final report we find out that it is actually uh able we are able to form these kind of consensus on AI using these kind of civic participation tools and uh using these kind of civic participation tools to have a better alignment on AI and also we are also using the AI tools such as in a cover of these report you can see this is talk to the city and AI uh integrated tools to clustered public opinions into different groups so this is how Ai and civic participation can interact with each other the civic participation can strengthen Ai and do a better alignment but the AI can also provide a better way to do civic participation so basically what we found is first of all during these kind of discussions uh usually there's a paradox between proficiency and plurality if you want to balance the plurality of the opinion and the Prof uh proficiency you need to design the process in a very cautious way and understand what is going on and potentially to happen and also you need to focusing on the people or Tech which means that besides developing the new tools how to revive ize the own and existing tools is very important to increase the participation especially for those people with digital gaps and finally uh we need to be aware and be cautious about the huge company or the governments are still having these kind of absolute power the tech companies and government agencies still holds the power the tech companies controls the platforms controls the technology and the government agencies controls the Public Power so how does how do we as a civil society member as a organization as a FB as a dow organization can win or can leverage our power leverage our resources on these kind of relationship is also very important and that is also how we are going to talk about it in our panel so basically sorry this is old one so I didn't fix the mistake here so new technology is not only challenges but also this is machine to create uh new spaces for common goods and the V taian example can be proved as an example for providing the new technology a way to do the common goods or also conveying or maybe using the V mechanisms to transmitting this kind of common goods into the new technology so I'm also very curious about what is something happening in Japan and that's will come kisan for uh to uh to introduce how this happening in Japan and also about uh he's organization uh AI alignment okay sure all right yeah thank you um yeah so in my day job um I I run an AI startup called Aria and so in the company we generally provide AI Solutions so in that sense like we have like very close contact to the actual Society in terms of what AI can do and but um but recently uh we uh started a separate organization called um AI alignment Network and the focus is to sort of promote like research and conversation about uh air alignment I don't know if you are familiar with this topic but um but but there are like many aspects in this but uh the idea is uh we want to make sure AI will do what we want them to do so so no we tend to think like you know know AI might be you know just computer programs but but but now know as they become more and more complex it's difficult to uh make sure that you know they their behavior will be aligned with human values so so that's the kind of General topic but but then there are like a lot of political issues around it but uh the thing is um when we think about aligning you know Ai and Technology with human values we have to think about what our human values really are because you know probably everyone has like different ideas and you know know different priorities about you know ethics and so on but um but actually you know people are very diverse and people work in the different areas and have different opinions and so that's where you know maybe we need some sort of like a also like technology assisted mechanism to sort of gather opinions from like a lot of people and and then sort of include their values uh when we deploy AI so so in a way U know people started worrying about AI alignment uh because right now the development of AI is very very fast and so it's getting hard to keep up with everything going on and but yeah but then but but this so the initially the the main thing was uh how to sort of set the goal for AI so know in a way it's difficult right so like you know even when we try to incentivize like employees or you know citizens uh it's very difficult to design like perfect goal for everyone usually you know like when when you set a certain goal know there's always some sort of like uh the game ability so so you can you know maximize you know your Rewards you know if it's actually defined but uh so but AI can like really sort of find some sort of gaps uh where people didn't think about before so um yeah so so that's the kind of concern right so like you know you know we we initially think okay so maybe this is not just about AI but whenever we try to sort of implement some mechanism in the society uh know people find some sort of gaps like you know and then you know that can create all kind of like a fraudulent activities as well so so so that in a way like this is a kind of not just about AI but but but then this issue is enhanced when we think about a because a can be much much faster so know if it's about human activity you know we can sort of change the rules know if something goes wrong but but if they act like you know 100 times faster or 10,000 times faster then know we will be always too late to sort of stop uh know problem problematic Behavior so that's the kind of General topic but um but I yeah sorry May getting long yeah but um yeah the reason why uh we started this AI alignment network was that like this this issue was not well recognized within Japan I think it was mainly discussed in the Bay Area and you know in places in some places in the UK like Oxford and London so so in a way like people from the outside thought you know it's Japanese people uh should be more aware of this kind of issue and should be more involved in research on this topics so actually like know overall this is like more like social design uh kind of issue but but there are also like many technical issues so so like when you design kind of reinforcement learning agent there are uh so we we have to find ways to avoid some sort of cheating by AI so so that's more technical issue that also like we need to understand the actual internal mechanisms uh inside uh NE networks because you know now we have like so many parameters and we don't know what's happening and so so in that sense uh you know there are like technical issues uh which require the involvement of uh AI developers and researchers but also uh know we need to have more conversation about how you know we make AI beneficial for the society so so those are the main issues and so so so within this AI alignment Network we like really try to um sort of promote uh like this kind of uh research and discussion and so so in a way like you know it doesn't have to be just limited to Japan and you know it' be like great know have also network with Taiwan and especially I also find that like a lot of conflict between the Western culture and East Asian culture so so no but that part is not very clearly articulated so so that's something you know we should also you know have a conversation like across different countries within Asia as well yes totally agree so I think uh we can actually move on to talk about how to make people communicate and making discussions on these kind of new technology and I think uh Jimmy from fepow can actually provid their Viewpoint and their experience in Taiwan about how to make discussions like this and how to formulate different people from different backgrounds together on the doubt and also uh doing their projects freely and also working on and dealing with the issues happening under and also uh concerning this emerg technology oh like hi everyone like I actually want to Echo to like uh y yeah y y sounds like uh like idea first I think um there's um topic called Singularity and it's talking about the technology growth is in the exponential way well our like a human brand is go to the linear way so in the like a uh coming future and the AGI sh suace the their capability of the AGI sh uh uh surprise the the human brand and this magic uh in his prediction is in uh 20 45 I believe he have three book to talking about that and I think it's quite interesting because in the general like AI or like an a model if you guys are familiar with like the AI or machine learning and you need to set up objective function and the whole like a training process is like a putting uh optimize like pressure to that objective function and so that will become really really important how we set up the objective function in the first place and that will be the main topic today about like what is like AI alignment allign to wh align to what like objective function who set up the function who can represent s the human being and and yeah that's the part I want to like a echo to Yas and yeah from from our point of view uh we are more into the I think many people already familiarly idea called DAC uh because yeah vitalic write article to try to Advocate this idea so it become popular in the crypto space and you can see many uh people like in their twiter Handler they add like a DCC right after their Handler you know like a suffix but uh yeah change like one's Handler or on name is really easy but what does it even mean about like dacc and uh what is defa like uh we have different like um understanding right uh no matter is a democracy it's about decentralized it's about a deficient or or blah blah blah you can uh interpret uh any way as you like is it's the same idea of like a plalism yeah and but uh for for the AI like specific topic oh yeah by the way so the the really short introduction of dacc is the idea of uh uh the book called power and progress and so the one who have the power will usually have the better technology and when when someone have better technology they can have the better power to control like everything and is the positive like like cycle and how can we break this like a cycle break this like a centralization and the idea is to create uh instead of like linear acceleration we need to have the angular acceleration and put this technology and power into the more like a individual the non power people non power like a individual then we can have the Abundant society and that's the idea of like a DCC but uh back to the AI um what I uh the CES I buy in recently is about we do not need the singular architecture of AI we can treat the whole like a internet or like a as a data available layer a DA layer or all the data and we can have the multiple like a AI agent upon it and it's not only the huge like like a large model language model we can have a small like a r a in in apple like a small expert model and those model should be the public good and those model can talk to each other and represent the different ideology behind and and yeah I think that's our takeaway about this yeah I just wondering uh in the book of the power and progress mentioned about AI have a huge power also it's requireed to a huge money to making the powerful AI so that's the reason why our current situation is only the big tech people could be making the AI so uh I'm just wondering uh what do you guys think about how to regulate the AI from the different layer maybe some of the people could be mentioned from the uh digital develop uh digital Devion to right with Taiwan or something like a alignment as a organization or Alliance kind of so yeah yeah yeah so so maybe like I'm also a little bit triggered by Jimmy's comment as well yeah I I think um like like when people talk about like air alignment or air safety uh I I feel like there's some sort like a hidden sort of motivation uh which is about the concern of Monopoly and um know centralization so uh you know people sometimes worry about uh existential risk uh caused by very powerful AI but uh but but actual concern uh like psychological concern is that like really powerful things are owned by know a few U corporates or or maybe just by a couple of governments so so that's the real concern and so so in that sense you know people talk about okay uh sort of making like big AI models or public available so so that might be good for kind of more inclusive approach but but the thing is you know but at the same time it's a very powerful technology so so that also causes concern so it seems like there's a kind of like a strong strong conflict between people who are more conservative about how we should handle Ai and for other people people actually they seem to be know like know I'm talking about eak kind of approach where know people are like a really like a pushing technological advancement and but but that seems to be partially motivated by know current geopolitical conflict because like if you don't build you know better AI or better technology and someone else will do it and dominate you so so I think you know there's kind of you know tension between these two approaches and I think in that sense thec finds like like really the important common ground and so so I think because it addresses the concerns about like democracy in technological development because you know the future of humanity shouldn't depend on just a handful of people but but we want yeah we want to invol everyone right I I become like a P2P ACC recently like yeah I forget about like a dacc yeah yeah uh from a point of view I want to uh following what Kanan has just say that so yes first of all we do not let the power to be Consolidated on some single players such as the big companies or the governments so that is also and if you notice that the yesterday the Mata which is the company of Facebook just released a news pris saying that they will keep their AI models open sourced and they like the uh very uh famous scientist in uh meta young chin just say that they think that open source is the only way to building the AI systems so it is actually contrary to what open a has approach is so what open a approach is they want to mitigate the risk of uh of AI by strengthen the AI systems by themselves they think that is a way to mitigate the risk uh but for mattera it is actually another way they thinks that they will release all the results open to the public to become the public goods become the commons so that we can uh accum accumulate everybody's effort together to dealing with these kind of risk so there two different kinds of approach um from my point of view as a v taian participant we think that open source is actually very important uh not only because we r one is a open source Community but also is uh the collections of our opinions from last year's pro project the most consensus opinion coming from Taiwanese people is that the the AI assistance the information need to be released the AI assistance information needs to be released to the public and open source so that is also the consensus of people in Taiwan and that comes to another question is that how to regulate it how to use the law enforcement to do that like you know right now in European Union they have qu we call the digital marketing act it will be a little bit low stuff because because I'm a law student sorry for that so the digital marketing act the the critical things is that they do not want the power to be controlling by a single or several companies they want these company to open up their applications open up their ecosystems open up their uh developing developments to everybody and making interoperability making other small players to come in the game and making the games we have more players which means means that we can decentralize the risk we can decentralize these kind of systems instead of being controlled by a single player in the game so this is also another key for a better artificial test which means that we need to have different player in a game to compete with each other with different viewpoints like what we have in V one is that we try to brege and connect each stakeholders together different stakeholders together to do these kind of things or like like maybe P out is also having people from different backgrounds SC And discussing how to achieve a better world how to achieve a better organization new types of organizations to help people and also also from Kanan like that Kanan is building up the community in Japan as Kan say the Eastern values and Western values totally different and I really experienced that when I was in open a headquarter last year talking with open Engineers they thinking that the technology is the best they thinking that te ology is the most important part but however in the some Asian project we think that human should be the core of technology I have a I have a rest question for you guys uh so actually I'm using openi every day like a CH gbt same so same you same so actually how can I contribute like uh anti centu or company or something like that so act activity what can I do I think like uh if you have any uh but do you pay to open the eye yeah sure but you can use it for free yeah and actually like you know if you use their service more like you know they actually they have to spend more money for the user so but I don't know um yeah but but I don't no that's the thing it's very difficult to avoid like know their model is good and their service is very useful so you know there's like a really natural uh sort of economic incentive uh so but maybe I I don't know if we can like really decentralize like this kind of trend easily but but I think uh know like traditional government could play a role so for example just uh like U implementing some sort of auditing system uh be also useful because like in a way like it's indirect but but but the government uh reflects like people's opinions but in a very indirect manner so so their action should be in the sort of benefit of the citizens so so so in that sense uh know we can still uh use the uh like mechanisms of the know current governments um but but I think to do that uh I think we need to do also some lobbying and inform you know policy makers and so on but I think that's probably in practice a very effective way to approach this my one's really short my Approach is like share my account credential like account name and password with several of my friend so many people can use one account so the open I like the revenue will like reduced and they don't know like who is like individual like behind this account yeah that's my Approach I want to share one approach so Canan talk about the Backes we need to do the policy lobbying to the governments to making the AI power decentralized and making forcing the these companies to open up their projects and also open up your contributions but uh I want to talk about one small things is that um AI needs what AI needs so AI needs data and the data is high quality data so right now open eyes approach is that they are using uh Outsourcing these kind of data selection or data purification to the other companies mainly in India so there a lot of news is in that Indian people are actually doing these kind of works for open Ai and uh making a lot of profit like that uh but not really because this is not sustainable overall after all the AI will substitute these kind of jobs and uh making the data generating by themselves so one things is that we need to have our own data but how can we generate these kind of data actually it is not that kind of difficult for example right now I'm as doing a project try uh in the Fair Trade Commission in Taiwan and I'm this is my job so when I was doing these kind of jobs I'm actually using open AI to summarize the document and also doing translation but however you knows that the open AI Chinese version is not really quite well it actually containing a lot of wordings that is inappropriate in Taiwan context which means that they were containing the China context and I know know that maybe in Japan the open air also face and CH gbt also have these kind of questions so when you are doing Corrections for the AI generated results you can actually record down the glossery of your correction like what is AI wrong on the world of translation you can and what is the correct one you can create these kind of glossery by yourself and these grocery will be very sector specific for your needs or for your job or for your training and also you can develop these you can develop your own AI systems using these kind of ways in the future so making your own data and try to think about what is the data that you control and it's is very unique then that is the way to leverage your own advantage and your own feature to these AI systems and uh achieve the goal of decentralizing these kind of systems originally centralized thank you yeah perfect uh just one minute so if you have any want to say something okay I think the technology like Peter mentioned is called IG retrieval like argument generation is kind of really popular in recent like imp implementation and the cor of the like AI aging would be like a data aore and uh the computing power and I see many like crypto company are trying to make those three like Dimension as decentralized as possible I think you can check like a bit t b tension like and IO next and a CH base yeah I think they trying to make those three dimension as decentralized as possible yeah that's it this is all thanks so much thank you very much thank you very much seer pedalist [Applause]