Great. So, this agenda is a very light one. I don’t need to go through it in detail. I just wanted to write down notes on things I wanted to touch about.

So, we – Isabel, good to see you. I’m on the other side of the room.

Hello. Good to see you.

Good to see you.

I’m checking the document that you shared.

Great. There’s very little in there. It’s a list that I want to go through. So, no pressure.

Yeah, so last week I had, I think, seven meetings on different groups that I’ve been in Taiwan, which was incredibly exciting. So, I figured there’s actually a lot of intersections between all of these. So, maybe let’s put our heads together.

I talked with Isabel about AI Academy Summit. We talked about the journalistic project with HC and a couple other projects. Another chat with vTaiwan group. I think a lot of these do intersect with each other. So, I wanted to hear about what kind of goals and resources each of these groups have and kind of play a PM hat between you all for a little bit while I’m here, so that we would be able to be helpful as well.

I think one question that came up maybe would be a good place to start at the vTaiwan meeting was what kinds of projects does moda envision in the near to long-term future? This is mainly so that we can prioritize future development from an angle that is most conducive towards there. Because I think I’d love for you to chime in.

Like, one thing I sensed was people were saying, I’d like to work on things that would be used by moda. So, I wanted to be able to offer that as like, these are seemingly strategic directions that would be good. So yeah, maybe that would be a good place to start.

Yeah. So, is it starting with me?

Yes.

Okay, so in the Wednesday’s meeting, I had this discussion with Deger, with Bruno and other vTaiwan folks [and] what we had is that… we feel that moda and AI Academy and vTaiwan, all of us want to use Talk to the City’s tools to empower the further discussions, no matter it is from vTaiwan perspective. Our next goal, next step will be reaching out to the government agencies and to discuss the AI basic regulation that I think is supposed to be released in the end of this year. So, after the regulation is released, we want to use the vTaiwan mechanisms to have a discussion with the civic community.

Yeah, and probably we’ll have Polis, we have the City Assembly, and all these kinds of outputs can be put into the talk to the cities platform and be generated out, sometimes like output and the analyzation to or these kinds of opinions coming from the civic public. So, this is what we’re going to do next.

And I think AI Academy will have several like civic participation workshop, these kind of things in November and December. I have registered for one and I think they will also generate these kinds of outputs, like the discussions or the transcript of the discussions, and they can also be put into Talk to the City’s platform. I’m not sure, I didn’t know much about the detail about how moda wants to do, but I think moda will have the same way to try to use the talk to the cities platform to empower the further discussions.

So, what I propose to David and this Wednesday is that maybe we can create a working group for everybody to communicate together, like for the Taiwan folks, we have a master student, Eli, who is focusing on trying and generating this kind of content for us. So, these kinds of working groups can have, I think can better or improve the communication, because right now David needs to talk to each team individually, but however, if we can circulate these kinds of events and also creating a massive group that can have us communicate with each other much faster and more efficiently, it will be much more important.

I thought all of this is already in the vTaiwan Slack channel, right?

Yeah, so I’m actually thinking that maybe… I think vTaiwan right now will have a weekly meetup online, so I think it’s also a very good chance to use or take advantage of vTaiwan channel or vTaiwan meetup as a way to do discussions. Like, I think Bruno is also in London and I think Bruno and Shu Yang is also based in London, so maybe they can participate online, and so that we can have online discussions to talk about the progress of these projects and talk about the cooperation process.

Yeah, and I think the folks in the Democracy Network department are actually quite active on the g0v Slack, right?

Yeah.

Even during their working hours, because that’s their job, right? So…

Yes.

So, this is totally fine. I mean, I don’t think we need a like moda-less additional communication channel, since it’s like… explicitly, we set up the Department for Democracy Network to work with the Democracy Network internationally, and so if the existing Democracy Network already have, as you said, GT or Slack or whatever, I think we should just work with, instead of setting up something that work for, the civil society.

Yeah, yeah. Definitely. Is Bruno on the g0v Slack?

I don’t know. We can add me and Bruno on the g0v Slack, if that would be helpful.

One thing that I want to underline, and we talked about this in the meeting, is let’s make good use of RFCs because instead of adding more people to a Slack channel where they have to be present in a temporal fashion, if there are RFCs, they can just actually contribute. This is something Paulist did a really, really good job about, and I would like to be able to grow features and projects this way, so that someone that doesn’t know anyone or is not connected to vTaiwan or g0v will still be able to contribute.

So essentially, GitHub.

Yeah, yeah, yeah.

So yeah, I mean, there are many artifacts already on Slack channel, right? From the vTaiwan conversation and things like that. So, essentially what you’re saying is that if you’re like concrete suggestions to Talk to the City or to the integration, to a deliberative system, it should go to GitHub, not staying only in Slack.

That would be great. I’d love for… if you want to take point on some of these, that would be wonderful. This is a feature group that should… Because I don’t know if everyone realizes that they can contribute to GitHub this way. That’s one thing that I’m…

No, I mean Slack, GitHub integration is one of the first integrations.

If it’s easily adaptable, then that would be great.

That’s one thing for the technical integration, another for people to actually take the lead.

Yeah, I don’t think there’s anyone on the vTaiwan website that doesn’t know GitHub…

It really is……

Yeah, but I told you to introduce your work to jothon because there are similar data from the Slack and the previous discussion in g0v. So, we’re thinking about maybe we can try to include some of these datasets from rand0m or those archived from the g0v Slack into Talk to the City.

Then, I think it would be great that we have a chance to introduce these tools to the general g0v community, not only to the vTaiwan people, because I would say most engineers are not in the vTaiwan chat room.

Indeed.

But engineers who knows about how to contribute to the event might chat like AI learning or general or random. So, I think that would be wonderful if we can set up or maybe just register a short talk in the next g0V capsule, which will happen in December 9th. So, I think we can arrange to share how Talk to the City works to the more broader communities in g0v. We invite more people to contribute.

And also, I think what you can corroborate with vTaiwan is not only about the subject of AI discussion, because there are so many different topics and the declarations. I talked to Ronnie and he said he will try to… He actually has the data sets of join.gov.tw platform. There are so many declaration discussions out there about gay marriage laws two years ago. So, he will try to use these data sets to test how Talk to the City’s report will be like, but he needs some time to understand how it works.

But yeah, but what I want to say is that I think this school is not only for AI discussion, but that is AI Academy focused on. But vTaiwan and also g0v community, they are not only focused on AI.

Yeah.

Have interest on a lot of different topics. I totally agree. I think that would be a very good place for continuing working to collaboration. And I also think we will need a page on HackMD to describe what’s going on if we want to propose this in g0v community.

Yeah.

Isabel, do you think we need to create a new channel on discussing this or using vTaiwan channels as well?

Personally, I would like to maybe after they share and introduce the project, then we set up a channel.

Okay.

Because if you set up the channel now, I think no one knows what that’s about. And also, there are other tools about civic participation. There is a project called Join Plus. There is actually a lot of different tools this year. And we are members of that project. I think they got into the final 20 projects in presidential hackathon this year, but didn’t make it to the…

The top 5. Yeah.

Already have a lot of different tools. Yeah.

Okay. So, what I’m hearing is that with the random channel, the O coming at zero, which is already CC0 by definition. Anyone who speaks in that channel is open for it to be trained as common voice. That’s the Mozilla Project, right? And since CC0 is a Mozilla specific, we can also use it to train Talk to the City or whatever.

And so, the existence of a random channel is kind of a playground. I don’t think you’ll get anything very useful. But anyway, it’s good technical playground for importing to Talk to the City. And whatever GitHub integration can be tested in vTaiwan… But what I’m hearing is that after the hackathon, the large hackathon, maybe you will spread to other channels, migrate to a new channel, whatever. Like whoever appears in the new channel or the cross-channel integration are the right people, you know, open-space technology and all that.

Yeah. Because I think for now, we need more people with AI engineering backgrounds in the g0v community and also who like to play with data. We have these two kinds of people in the community, but I find that they are not aware about this project, Talk to the City tools yet. So, I would like to introduce to more people about this.

Or if vTaiwan want to host the whole discussion, I think that would be good. But for AI project, we are now focusing on created data and also, we are trying to find out what the report would look like. Yeah, so we are testing with the existing data sets for now. And also, I think just one more thing I want to make sure. Does the vTaiwan’s report, vTaiwan with open AI’s report is using traditional Chinese data sets or you translated it to English then put it into Talk to the City? What data sets do you use?

Oh, the data set we use is not produced by the Taiwan’s track because we have different tracks of Polis and vTaiwan process. Actually, in our projects, I think it was like two different tracks.

So, one is focusing on Taiwan’s track and the other one is focusing on international folks. So, I think the data we used to produce the reports for the open AI is from international track. So, it is mainly done by the Chatham House folks and they feed the data into the Talk to the City. [And] because it was for international folks, it was English.

Oh, okay. But here is a Chinese version. So that’s a report that [was] translated into Chinese version?

Um, no. Not yet. We are currently working on that.

No, no. I mean I got a report in traditional Chinese on the website, so I just want to make sure how this come out.

Okay. I think the traditional Chinese… I think I need to make sure this information with Shu Yang because I think as far as I know, we didn’t put any vTaiwan on the results like for the Citizen Assemblies on September 24th or the Polis reports on to the Talk to the City. Right now, Eli is currently doing on that.

But I think what you are… I’m not sure what you are reading is from Eli’s production or from the former ones. There is currently too much version of that.

Yeah. I think that’s one of the very important things that we want to continue to co-work together to improve the tools. I think we need to have a format to record the process we test the records, so we can learn from each other’s experiences.

Yeah. I think I will send you the link later and you can let me know where the Chinese version came from.

Yeah.

So, two quick questions on my side. One, who would be the right person for me to follow up with on the g0v hackathon in December?

Yeah. I will register a talk for you and I think we can, would you mind joining the Slack for now? Or we still continue our discussion in Signal?

Sounds good. Yeah, I can join the Slack too. I already added that as a follow up.

Okay. Then, I think we can talk there and I will send you the forms for you to register. But the thing is, the short talk will be, I think it’s 2pm in Taiwan. So, when is your time zone for…

Um. What date will it be?

December 9th.

I will be in… That is great. 9am my time.

Okay.

Great, good timing. Um, cool. I want to shift gears a little bit. I like putting on product strategy hats. I think that’s exciting for me at least.

Isabel, I’d love for you to share a little bit about what you are envisioning for the AI Academy Summit. Because for that, I think there are specific features that could be interesting. And I’d love to hear from you guys from one side on thinking like what kinds of features would be interesting for the near future deployment, because there will be intersections there.

Yeah. Actually, for now, this week I shared the Talk to the City demo with Ronny and other people in the project. And also, we are discussing about what kinds of data sets we can produce for this tool. So now, we are only thinking about…

Actually, I want to have a report which would reflect and provide information for people who have not participated in the workshop to know what happened on the workshop and to join the continual discussion online. So, we are figuring out how to design this process and how to do this. And to maybe have a homepage or to replicate different kinds of tools into this process.

But actually, we don’t have a concrete conclusion yet. So. we are also discussing about the discourse and also discussing about Polis. And also, there’s someone proposing using Slido. But I think… oh, and also Google platform. So, there’s a lot of possibilities that we’re still discussing. But the main purpose is to use Talk to the City to share this report. And I think you mentioned that people might have an interactive conversation with the model. But this part is not released yet, right?

No, not yet. This is one of the avenues Ashwin is interested in.

Yeah, so if that could happen before December, then that would be wonderful. We would definitely want to try that.

Can you describe what the ideal use case would be? What do you envision here, to be more specific? We may be able to make a prototype by December.

Oh, okay. Because the topic of the first workshop is about large language model in Taiwan, so we have a lot of discussion. And I actually do interview before the workshop already. So, I would like to input all of this transcript of our interviews to these opinion leaders here.

So, I’m not sure if I input that opinion and then maybe this transcript for like four or five-hour long data set. Then could people just ask questions to the model? And, the model can respond to their question according to this people’s opinion. Yeah, I’m not sure if we can do that.

So, do you have the transcripts already for the four or five hours?

I have two interview transcripts, and I’m going to do another one tomorrow afternoon. And I think by the end of next week, I will have maybe four or five.

Okay, that is great. So, when would you like this by is my next question.

You don’t have to finish… But it would be great if we have something to try maybe by early December or even mid-December. That would be wonderful. But I will send you the data set once we finish that.

That sounds good. Turning that data set into a report and having the information of every cluster be exported from the report. That is ready. Feeding that into a language model in a new report, that is what we are working on right now. So, I do think we can get a simple version of this done in a timely fashion.

Okay. I was very excited. Because I actually have a very good result from this interview. So, I’m looking forward to what the model would be like.

Yeah, the point is that the report’s generation will be much higher quality than the talking to a bot representative, just given the state of language models today. With that knowledge, I would like that to be a disclaimer on top of anything we say. Because people will say, I would never say this.

Yes, that makes sense.

Or I said this, but the term is wrong.

Right, right. I think this is a great first prototype… I like there to be a reason for a prototype, rather than just ideation, too. Actually Audrey, I’m curious, because we talked about this a little bit as well for the extras repo. Any kind of feature orientation you’re interested in for this? How can we position it? What I’m thinking is simply export the content of the cluster, have a place for a human to be able to enter further context, and it is ready from there on. So, kind of like a wrapper, which is what we talked about.

Yeah, I think we already had the priorities, which is first about cross-cultural translations. Which is very interesting, because that’s exactly the same thing as the term is wrong. Because it all depends on how you look at it.

For the speaker, they may see their ideas being paraphrased in a way that they can barely live with. Even though semantically it’s the same, but the tone or the way it’s phrased, they barely can live with that. But from the other side of the semantic space, that may be the bridge that we need to reach the other group of people. So, this bug is actually a feature. If you’re thinking about bridge making.

100%. Right. That’s what I wanted this meeting to be about first, and I’m glad it’s coming back today.

Yeah, so like bridge making and meeting in the middle necessarily mean that there’s some misrepresentation going on. There really are just two sides of the same coin. And so, to manage a leap of rhetoric that’s not too long, so as to say, oh I totally didn’t say that, but also in small enough steps, head toward a shared common understanding or shared value. This is the core thing that I would like to… this mission facilitator or whatever you call it, to possess.

In fact, I had a conversation with Colin on the prompt to get the mission to do that. I actually have it running on Faraday. And we actually found out that this is running on Zephyr, so can do very small machines. And the most effective prompt we have seen so far is now pasted in the Google Doc. Which is essentially do what Audrey does, because the large models actually know me.

So, it begins. But of course, with very specific instructions. But that is what I mean by bridge making. It has like six different values that it needs to take care of at the same time, right?

So, my question would be then, I’m shifting the agenda order a little bit, but this is a natural sequence. We have four to five-hour transcripts, say one hour each. So, we have five different people. We have clusters that are produced out of this. We have language models that are out of the box right now that represent and misrepresent these clusters. We do this within a month.

How do we go from here, like say we reach that step, what do you envision to go towards having that be an experiment in the journey towards bridge making and this misrepresentation being a feature in the book? What comes next from there?

So, to me, why the, like the River Chat audience that the Democracy Network department works on, almost always starts from a foreign polity, is because Taiwan simply doesn’t have that many polarized tension points to work with. So, we’re one of the very rare democratic polities where it’s not automatic for one party’s like good doings to be automatically considered bad things by the opposite party, right?

So, we’re not at peak polarization. We have polarization, but we’re not at peak polarization. And so, if you are at peak polarization, any topic is a good topic for bridge making. Because the chasm is just wide. And the challenge becoming finding a narrow enough intersection where you can run this bridge making.

The challenge of starting in Taiwan is that the chasm is just not very wide, right? And so, it feels that you don’t need to construct a bridge if you jump long enough. You can just get to the other side anyway. So, the machine exoskeleton that helps you to build a bridge and so on seems overkill for many of the issues.

I understand. Well, I wish it wasn’t the case. But the world isn’t short of needing for bridge making So, I’m sure we could come up with a case that’s not in Taiwan if necessary.

Yeah, exactly.

So, I’ll take two minutes. If you can click on the Figma of reliability experiments and AI summarization. So, part of our work is… Currently, the first thing that the reports are tackling is the ability to summarize. We are building this app called Open Questions, which should be a simple workflow for people to be able to give feedback to the system. So that we can use this as a research avenue and maybe use this to fine tune if it is working well. So, this is something we will prioritize building next.

And this lets us be able to evaluate and otherwise very qualitative. Like a lot of people shy away from qualitative judgment making mainly because it is hard. I don’t think it is being hard is a good reason to not try it. So, not a lot of people take that leap, which is why I’m like maybe I’m not doing this right. So, I think bridge making is an even harder question. That is why I am interested in this.

Say we make a couple language models that represent different viewpoints. And they will have these kinds of gaps.

Would you like to use the whiteboard?

Sure. I don’t actually have something to draw right now…

But it’s easier to visualize.

Yeah, definitely. Sure.

You can draw the clusters and the LLMs on the whiteboard…

Say we have a couple clusters that comes out of a couple different speakers. And we see that speaker A talked a lot about this. And there is a little bit of speaker B here. This one has more division. This one is actually only one other speaker and it has a little bit.

And then we can represent… We can come up with two different things, right? One is representing the contributors. While the other is representing cluster 1, 2, 3. And that is these are two very different ways to represent. Here we have four LLMs. Here we have three LLMs. But what the LLMs are trying to do is very different.

In some ways this is where I want to switch us… We are just brainstorming here, right? Like, where do we go from these two to be able to come to an exploration or a leap that will give us more insight towards bridge making. That’s what I’m really curious about.

Yeah, so, I’m ideologically challenged. Because I’ve already committed to only think 1, 2, 3 as the units of conversation. Because the star or whatever is extrapolating volition. And the models simply don’t know about such individuals well enough unless they are… There is no pre-trained knowledge in the language model to take those extrapolated volitions and infer their actual preferences in context. Whereas 1, 2, 3 is the context. So, we at least know that the contextual confidence on this is good enough.

Especially if they are the speech came from a face-to-face deliberation. Then we can know that at least in the context of this deliberation, these clusters have coherent volition So, that’s my ideology.

This is what I was texting you yesterday on things I want to touch on. I completely agree which is why these are the things I think are worthwhile to build language models I think this can be potentially helpful if our goal is to be able to do things like… oh, that’s not what I said. If someone is saying my wording is not accurate…

Exactly.

I’m curious what kinds of experiments we can have that go from here towards bridge making. Where we can actually measure the models ability to actually have formed the bridge, rather than find the most watered-down consensus.

There was a project that came out of DeepMind last year. I’m sure you all saw it. And it’s like bike lanes are good. And I’m like yeah okay everyone agrees bike lanes are good. I’m interested in like should we have less road lanes and more bike lanes. Like, can we find the actual source of the conflict.

So, I would love to hear ideas on where we go from this towards…

One thought is just to have a synchronous. Like this speech collection we assumed it to be mostly asynchronous. But as I mentioned if we commit to have synchronous events that went after the initial summary into these clusters, part of those synchronous events which may be face to face or hybrid can be used to confirm these synthetic cluster viewpoints. So, it could be as easy as you know this robot MP doesn’t represent me. Or something like that, right?

So here are the viewpoints we summarized. And as you said, this is too watered down. This is too the tone is wrong or whatever. And so, this is almost like in a deliberative setting. There’s a summary of a poll maybe that we did or a survey that we did before. And maybe the initial route is just clarification. But like even I’m quoted and I understand you point to the timeline. That was the wrong context.

So, just the contextual calibration could be part of the synchronous conversation. Of course, you can also keep going on as people are having synchronous conversation as a sidekick or dashboard or something. But I think this transition between asynchronous and synchronous through clarification - I think this is actually powerful.

Right. This is the graph you drew that could not be translated with the translate button, but I figured that is what it means. And Isabel and I talked about this. This is I would love to have a little more high-definition on what’s the sequence of questions would be between these. So, Isabel, I’ll just go with the context given the academy gathering will happen.

One thing we talked about was we would have an open-ended survey possibly before the first deliberation. Do you think this is feasible or possible?

I think we would do… I discussed with facilitator of the workshop. The conclusion is that we don’t think pre-event survey would be a good thing to do because it’s kind of out of context. And it might interfere the chemistry happening in the workshop.

But what we would like to do is we will come out report opinion conclusion and the cluster worked by the human being… by the facilitator. And we also will use the data sets to Talk to the City to generate the report.

And I think after that, we can host another maybe online meeting to aggregate more opinion. We will release these two different reports. They might be the same. They might be different. But we will have a comparison about the opinions or the reports by the human being and by the AI. Then, we would like to have maybe another event to find out some process to have people to come back to comment on the report.

And I’m not sure if this helps what you are thinking about or how we should design the process after the event. So, it would be more like… the workshop would be the first survey. And then, we will have another survey after we have this more concrete deliberation process.

That makes sense. That makes sense because you guys already have a context on how you want people to enter the debate. So, let’s not produce something ahead of time for it, but once the event happens, then we can start the cycle So, the summaries can be gathered into clusters. And then, these can be provided for the follow up session, to say, you know, this is the… Of the diamonds that happens here, this is the part where I would love some guiding questions with respect to how can we bring How can we ask things to people when they are in person.

That also lets us evaluate the success and accuracy of the model. There is some experiment design work that I would love help with. That is the part that is not clear to me at this point. So… We can brainstorm again.

Similar to the way we did it for accuracy of the summary, like, how would this serve? Because here we are doing accuracy of the summary, right? We are turning async clusters into a representation and we can ask does this represent. Now, if the same thing is happening for people to gather in person and discuss, what is our success metric? I guess number of cruxes that we identified, that people said yes that was contentious. Like, what is there? I would love to hear your thoughts on that.

It depends on whether people get to have conversations one on one to those synthesized LLM.

Yeah. Assuming that will be the case.

Because I pasted this Faraday screenshot, right? I just run Faraday with the prompt. You know, facilitated conversation like Audrey Tang. It is in the Google Doc.

I literally started a minute ago, typing in ‘bike lanes are good’. The facilitator would say can you elaborate, how do they contribute to the overall wellbeing. People will like more. The facilitator would say that is a great point, how do we ensure so on so.

So, with technology such as this that can be hosted or can be even just run. People can just download the model that is representing the cluster to their own machine. One can explore more conversations in their own context that doesn’t take anybody else’s time, and see how calibrated it is.

So, I would argue that in that scenario, the success is rather the amount of further calibration that people did to that cluster model. Instead of whether it is initially very well aligned. But that is a harder scientific problem.

And there is another case where this is just in a synchronous meeting and there is a demo, right? People seeing the interactions with those clusters or even in silico transcripts that is acted out by performance on stage and so on. And that people voice their support or concern, or tuning based on how the speakers, robotic speakers, theater actors are representing them. So, one is async tuning and one is sync tuning, basically.

I am interested in the Faraday example, let’s start with this. Maybe this is is easier to do. You said how well calibrated… And that is a great sentence that has a lot of nuance. If I double click on that, what else is in there? How do we measure this calibration? I don’t think… Maybe… there is out of the box, you know already well explored metrics for bridging.

This is a good first start, right? Something is well calibrated if it is clear, if it is in the original context instead of taking the speeches, the cluster side of context, it retains the nuances, the specifics within the initial speech, that all the diverse compositions are there instead of one dominating and focuses on downstreaming impacts. And it is collaborative. Maybe the last one is impossible to measure but at least the first five are somewhat measurable.

I can imagine even using a separate model to score them, and a human to be able to validate this. Like, those are very measurable. The thing is, if a cluster is… I was going to say, what if a cluster is not specific? Or a cluster is not diverse?

Well then, it is not that cohere, right?

Yeah. That is why I said. That is not a good cluster. Or, it is okay for someone to say, you can be very clear and I am not sure what to do. Clarification is very obvious.

It is very clear that none of us have an idea.

Right. That is a valid cluster. It can be very specific, so I retract that statement. I think it would be very interesting. I would probably make it in Europe’s paper. I am not really a paper person, I am more interested in prototypes, but I am sure someone would want to say, these are metrics, let’s try a bunch of different models… Compare and contrast, how are we able to evaluate them? I like that these metrics are also human. They are not just evaluations for a language model. Okay. This is a very good evaluation system. I think it would be worthwhile.

You can ask this to everyone that is both in and outside the cluster, see how they see it. It may be so that people in the cluster don’t think they’re well-represented but the ones outside the cluster might think that the cluster works well.

I studied this a fair amount in college in the context of why do more liberal political perspectives tend to fracture more. And it ended up you know in the elections in Tunisia. Like, One conservative party and 10 different liberal parties, then the conservative party wins just by having a more big tent.

So, I think this is worthwhile to explore for sure. Anything jumps out to… I’m just shifting gears because I think there is plenty of food to chew on on this avenue, is there anything else that is of interest for moda? We talked and made a priority list last night. I am interested in the switch intersections with other projects cause then we can have more shared task forces.

Yeah, I’m trying figure out what department or democracy network that can help Talk to the City. And I think there are three ways. First of all, we will have a page of moda that maybe aggregate the… — — — — — — — — — — —

Oh, that’s our AA report.

So, I think we’ll do more deliberative workshops this year, so there will be a portal on all the webpage to make more data aggregate into the webpage… So, I think that it’s kind of data…

So, if you can put this data together, then we can share it to Talk to the City or Polis because I talked to Liz Barry yesterday and she showed me the open data of all the famous Polis projects in GitHub. It’s useful for the academic or protectionist to use these data.

Secondly, we have collaborated with Shu Yang, Fang-Jui and Indy of the Dark Matter Labs to develop a prototype about AI, “Say Hi to the River”. And I think they’re already undergoing development and I think there should be some resources that can be shared with Talk to the City, maybe Talk to the River.

That’s probably not in Taiwan but somewhere else more polarization, yeah? Why just Cities, huh? Rivers!

And I think it is in the tool layer, not just the results layer. So, the first topic is data aggregation and shared data, and the second topic is the shared results of the AI chippoff. I think it can be done…

I think it would be great. We can do a follow-up chat with Shannon as well.

And as how they’re doing presently.

Yeah. It has been a while since I last chatted with her when I was in the UK. Basically, we won’t fork the two repos. Talk to the city. We should name them in the right way. But then the problem is people get confused because they are after the same name.

Two questions I have is my understanding last time we chatted was that these might not feel as fast for example. HCS article that he was interested in he wanted there to be a hosted version. Does this get hosted in Taiwan or do we host it for now? Like, if photo wants to have a hosted version for a lot of the projects here, that would be great.

But I don’t know what’s your timeline on that side but you was pretty straightforward at this point because that way there is more of a shelling point for more people to not have to deploy an instance but they can just create…

Basically, like one thing I blow is there’s some features that will come across any of you that want to contribute on developing them or making RFCs, that is also a great way for like… if we have RFs that come from you guys that would be actually wonderful because then more people will be stewards of what’s created, especially around the chat avenue.

Like, one is the transcription and diarization tooling for example, that was relevant for civil as well. I’m curious, Peter, your thoughts on this because you guys already have tried to whisper for…

Well, I think can do test first like test about the AI model summarizations to work these kinds of words and these types of results and with these kinds of features, I think you’ll be more like extension. So, it would be like… it is not a core function of the model but it would be like, OK if you have that it would be more convenient for us to convert the data into the super data security into the Talk to the City or we can generate the better outage whether it’s kind of better inputs.

Yeah so, we actually talk about some features that probably that will have more extension and probably have that can be installed on the top two seats model in Wednesday like how to remove it if we can remove the planks and because we know that if we are currently using the western model right now to do the transcript of the video they would automatically have the time script… timestamps on the paragraphs, which means that the models still rely on the time stamps instead of the real paragraph. And we understand the whole paragraph to understand the content, to make the transcript so…

I think it can be one way to like having… so we can actually separate like these kinds of intentional features into two different parts - one is that ‘OK would be better for if we have that but if we don’t have that it would be OK’, and the other one will be ‘OK it will be very important when we have that because you can empower the model in a very huge way’…

Right, right. That’s why I don’t see that as an immediate thing but once it exists it will enable a lot more people. Yeah, so if you want to make an RFC for it and then say hey here’s an experiment that would also be great. But let’s not start from that immediately.

I think I can do that and also, I can ask some of our partners and… not partners but buddies and folks in g0v community and vTaiwan community to participate into these.

Yeah. Um, cool. Anything else? About five minutes… anything else that jumps out here that we want to touch on while everyone ‘s…

Yeah, we’ve went through the whole agenda for not eating anything from Wendy.

There is only one thing I would like to mention it, that is about the public discussion on the review. The results of the public discussion must be linked back to the government policy, because if it is not linked back to the government policy, the content of the discussion will become an invalid discussion. Therefore, the schedule must be meaningful. The premise is that the purpose and method of discussion must be well designed.

Indeed. After the bridges are made between the same building groups or clustering, if result in governance changes either from any level of governments on regulations or maybe municipal level changes if it’s a river.

So… but some sort of governance institutions need to effect change in order for these bridges to be considered meaningful. Otherwise… Well, we still get something meaningful which is the stakeholder groups come to trust each other more which may be the great thing already, but the point being that if there’s institutionalized support for those kinds of things, that it’s even more meaningful.

Right. I mean that is where I would look to you all for example for maybe the first case in the world for like actually having a government party that is actively listening to this form of input.

I have… I don’t want to say worked with more so dealt with this in the context of US federal agencies where there is regulatory feedback, but the way in which the government signals accountability is actually quite misguided. They have to read the content and the content is not really in a readable amount if you receive you know more than twenty million submissions for net neutrality like… what does it mean for them to have read it and we will take this into account here is the consequence.

That bridge is something I’m very interested in like… how can there be stronger attribution by the central decision-making authority to the original idea. Curious to hear ideas on that front on how we can establish…

That was the core proposal of Join Plus, the presidential hackers on the team, that basically say that if we tally everything in a regulatory pre announcement, which is the Join platform, it’s almost always now at least thirty days. And I think we’re seeing more and more solution to sixty days, so within sixty days, there’s a lot of material to work with.

But again, like for truly controversial regulations, it’s not in a consumable amount, so exactly like regulations I’ve done, right? That’s something that Romney has solved very deeply about so I would consider and be expert in this. But I think regulatory good announcement is certainly very institutionalized in Taiwan.

We signed the twenty first century agreement with US. And the core demand of not just AmCham but the European Chamber is on this, not just AmCham but European chamber of commerce too. So, they want to have a voice even though they’re Mandarin may not be perfect, they may want to express their ideas in English and things like that, which is again, something language models can help. So, this is one possible betting.

Right. I would love to meet Ronnie too and talk more about this too. And I’ll take a look at the Join Plus app.

I guess the last thing here that we didn’t touch on was the decolorization tool kit. I did send… I did put the link of Bruno ‘s project on depolarized, which is…it actually has a lot of the some of the aspects on you know is it establishing common ground. it’s more so thinks about this from are you performing the right steps for it to be a healthy discourse. Is the avenue it takes… Um, if you have any thoughts or feedback on that that would be great I don’t think.

You can just review it with the skin but I would love to share that with you so if you have anything you want to see…

… AI set got rails for a conversation.

Which I think is probably better than AI trying to synthesize is AI guiding humans to say, hey can we find common ground right now.

Alright.

Yeah, Guide rails.

Guide rails, not guard rails.

Right, right. Yeah, I just want to share that with you because I think this will naturally queue into some of the avenues there. Cool. Anything else from anyone?

I’m curious about the roadmap of the local deployment to the Talk to the City. Is it…

Yes. We now finally has access to the repo, and Odie already has the local models deployed in google chat mode and this is not a secret. So, we will see if our local model is good enough to power the Talk to the City. Probably not?

But with some fine-tuning… we will be able to power some sort of Talk to the City with a locally hosted… I mean if you’re willing to throw to your edits and you can run Falcon-180B which is guaranteed to be good enough for this but it’s more expensive. So, we are working to explore ways to run it internally because we also have internal data that can actually be material for Talk to the City. Drinking our own champagne.

So, after we taste our own champagne then we will know whether it’s ready or not for something like polis.tw which is a public facing service. But there’s strict restrictions hardware in the GenAI use in the public sector guideline against directly providing public inputs into a GenAI unless it has been interested and certified and evaluated, which is also our job but…

Anyway, but for internal uses, it’s the AIEC, right? The evaluation center. But we cannot skip our evaluation and testing, so it’s mostly just internal testing.

I’d be very curious if you’d be able to share the the testing steps.

Yeah, of course. It’s public, actually. There’s a PDF about that so I will airdrop it to you.

Because this is something that we got to ask and I said oh great question, like I have thoughts about this but we don’t have like a here is our suggested way in which you can use this in a formal fashion. I think it’s a fair answer to say, you know, use it internally like this is meant for a group to self-sense make, rather than decision maker entity to source information, so they can make sense.

Awesome. Thank you. Live long and prosper.