meeting 2025-11-10

The meeting discussed various topics related to the Swiss Inference Utility and the Public AI Inference Utility project, including:

Usage Trends: The number of active users is currently 1534, but this number is not very reliable. The team plans to use third-party uptime trackers and share internal dashboard screenshots weekly. The API usage has spiked during hackathons in Spain.
Survey Feedback: There is an anonymous feedback form available on the frontend but not shared with API developers yet. The team plans to share feedback through a link in the coming days.
Documentation and Collaboration: The need for a shared, collaborative documentation space, potentially using GitHub or Hugging Face Spaces, was discussed.
Transparency and Traceability: Cole proposed a project inspired by OlmoTrace from AI2 to develop a more generalizable trace mechanism for attributing contributions to the model.
Infrastructure and Permissions: Discussed potential use of sandboxed environments for local development and the need to share permissions for testing.
Feedback Mechanism: The Data Contribution Flywheel from Public AI+ was mentioned as a more sophisticated feedback system.
Next Steps:
Cole will scope and propose a project on trace mechanism.
The team plans to meet again to finalize proposals and discuss infrastructure.
A small proposal website will be created by Monday.

https://chat.publicai.co/c/21c80932-f1ee-4b59-a92f-819b9988ea3d

meetings/2025-11-10/swiu-11102025-notes.txt

@@ -0,0 +1,106 @@
+Swiss Inference Utility
+MEETING NOTES 11.10.2025
+
+PARTICIPANTS:
+
+Oleg
+
+    Effinger, Bern, community, nature game launch (Unpetrified by Dreamhunt Studio)
+
+Vasco
+
+    Bern, Maths to ML, LLM at work, open for questions!
+
+Sabine
+
+    Zürich, Swiss {ai} Weeks, marketing, events, data science (studying), MCP tool for Swiss voting system worked on local
+
+Natalia
+
+    Zürich, project manager, interested in open source, Swiss AI, coordinating, documentation
+
+Cole
+
+    Canada, payments engineering, ICML 2025
+
+Josh
+
+    Happy to join
+
+Joseph
+
+    Late, listening in
+
+
+DISCUSSION
+
+Infrstrastructure of tools and services
+
+Set-up meeting, continuing from the ones before - and quite open to any inputs that you and the others have.
+
+On BBB - Josh can't turn on microphone and video at the same time. In a loud space, keeping words short, looks forward to contributing & participating.
+
+Open questions on status of the Swiss inference
+
+    how is the usage trending?
+
+
+Aktive Benutzer: 1534
+Not a very reliable/exact number.
+We can also use 3rd party uptime trackers
+Hackathons in Spain are resulting in spikes to the developer API - the number is across all models and users.
+There is an internal dashboard that can't be shared yet, but would be open to sharing a screenshot once a week.
+What would be the basic KPI? Basic information about the campaign and strategy. Would like to know who is building Apertus prototype, video series. Referencing Public AI, most people Sabine has talked to did not know about it.
+
+    are we getting survey responses?
+
+
+"How has your experience using Apertus been? Let us know at https://forms.gle/9TGSfTK1RfKj2b6X9 ..."
+
+Anna has prepared a form which is integrated into the interface, and Oleg gave her some feedback. Josh: it is anonymous, and like the dashboard the information should be made public. We don't have the time right now to make that available. Our priority is to not expose the private user information and conversations. Let's just ask :) The link has not yet been sent to the API developers, only in the frontend, and we would probably ask a very different set of questions.
+
+Josh is excited about the future version, an opportunity to change the minds of people in the media and our communities. Sabine agrees that in public events promoting the model is hard, the feedback is strange, sometimes even laughing about it. Keep trying! What would help is to showcase prototypes that work. Maybe even showing limitations. I couldn't find many yet. Iseca (?) is working on something interesting. Start with hackathon participants? Tibetan Chatbot.
+
+Vasco: you can click share the link of a chat, and opening the link opens a public session. Is it possible that people can share conversations? Transparency is the most valuable thing that Apertus offers. Cole's project from earlier this week is interesting and worth pushing (we discuss it further down).
+
+Josh: sharing your chats is different from the feedback. Thumbs up / thumbs down is something we can make public. There were last time about 2'800 feedbacks of this kind. We can probably just share them in the channel. .. How do we get better feedback? Joseph did a presentation about Public AI+, in which there was a presentation about a Data Contribution Flywheel. It is a more sophisticated feedback mechanism that is already online. (Send us a message if you're not already on Public AI+...) This is a step towards making Public AI actually open.
+
+Joseph: if you're having a conversation, users can click a "data flywheel" button which adds a row to a dataset on Hugging Face. It adds context and allows the user to decide how they want to be attributed, how much personal data they want to share (Cole notes: PII sweep). As a developer I'm quite familiar with application crash recorders that send info to developers, sometimes even create an item in an issue tracker. Not production ready, this is a demo but there are students working on productionizing this. We need some mechanisms, marketing about what we are doing here.
+
+Josh: many people are based around Zürich, and if you are there on the 15th let's hang out! We can try to meet on Saturday at 1PM near the AI Centre, Oleg, Sabine, Vasko in the evening?
+Oleg's running DINAcon next week and would like to hear about your own upcoming #events and suggestions to meet up any time.
+
+It's important that we start to document all decisions and action items in a shared, collaborative repo, i.e. not on this Slack, not just a messy Proton or Google doc. Ideally also not on GitHub, though we can start within forpublicai. Going forward, I would like to support a local provider like Codey.
+Proposal to get started with a Hugging Face space linked to GitHub:
+https://huggingface.co/swiu
+
+We can together go through the Excalidraw, moodboard a little more or just pick out the good stuff and start coding a website.
+Editing link https://link.excalidraw.com/l/71IEdlODkrp/6WN4PRESo1A
+
+Cole tells us about the idea that he had from the moodboard: inspired by OlmoTrace, he would like openness to be more of a competitive advantage. Let's think big with the Trace Mechanism concept from Ai2, make it more generalizable and take it further. Could be a good utility, cost being potentially prohibitive at this point, latency can be best effort at this point. We could (Wikipedia like) generate metadata from people's attribution efforts.
+
+Proposal from Cole "I had an idea for a project based off the mood board (XAI related)"
+https://docs.google.com/document/d/12pEUMhPzlKuhQ3PXFU71sfGjaHrZ99LBUJcobxpnbIk/edit?usp=sharing
+
+ACTION ITEMS FOR COLE:
+Trace project go ahead proposal, scoping, HLDs + LLDS (all basics)
+AI2 collab maybe
+
+Oleg has been interested in data traceability for a while and has collected notes and done experiments in the Proxeus project. Open Knowledge has a Frictionless Data project that aims to distribute data authorship, which has been the basis of most of this.
+
+We might do something in the future with Ai2 around multimodal projects. Think about how we could do experiments - quickly, but scientifically. And Stay Tuned! There is a prod database which cannot be exposed, even though everything wants to be as open source as possible. To get around that you could sign a legal agreement, especially if there's a concrete plan to implement something. Truly open way would be to firewall the sensitive stuff. We don't have a staging server, Swiss AI approached us a month and a half before they launched.
+
+Oleg suggests having a shared infrastructured to do things in a safe sandboxed way. Not a bad idea, says Josh, but what would that take in terms of giving permissions? Cole suggests mocking the API, or using a smaller model. For local development there's no point of having many permissions. Let's do it!
+
+Oleg, aren't we talking about this?
+https://github.com/forpublicai/open-webui
+https://github.com/forpublicai/platform.publicai.co
+Josh: there's not enough documentation. Zuplo and some other cloud services are not open source, there are some aspects of the infrastructure that are not available. Open WebUI is not as open source as it was before.
+
+We should give you some permissions, like what you already have in Public AI+, which you can do by using the site locally.
+
+Oleg says that in the context of starting an association, we need to work on the pitch to build bridges between the self-hosted Apertus and public AI service.
+
+Josh says we should work offline to work out these issues and have room for engineering conversation as well :)
+
+Vasco and Oleg stay on for a few more minutes to discuss the moodboard, agreeing to try to meet again in the days ahead, and put together a small proposal website by Monday. 

meetings/2025-11-10/swiu-11102025-transcript.txt

@@ -0,0 +1,785 @@
+[Transcript made with BBB / Parakeet 0.6b-v3]
+
+I don't expect us to do much with these meetings except to keep track of the development of the Swiss public inference utility.
+I'm pretty sure that a couple of other people might be joining over time.
+We have about nine potentially participants initially.
+There is this is basically this meeting is a continuation of a weekly series that uh Joshua set up, which used to be on Wednesdays.
+Um and this uh thank you very much, first of all, for responding to the poll and for committing to uh come here today.
+And um, I think probably because Natalia and Sabine, I'm not sure you've been in the previous calls, we could do a quick introduction round and start off with that.
+Um, I'm gonna put the agenda into the there's a little bit of a there's this thing called geteilte notizen.
+I'm not sure how it is in your in English, shared notes, I guess, in in BigBlue button.
+So that's just like an editable document.
+Um, and so the traktan, then like the the agenda items are there now, and you should all be able to just click and make changes, and um, I'm going to continue writing some notes there.
+But yeah, the the idea is that we will collect them somewhere else, somewhere central later.
+So um yeah, let's let's do a quick intro around.
+So my name is Oleg.
+I'm sitting here in the basement of Effinger in Bern.
+Um, this is a wonderful little co-working space where I've happened to be part of the association, and I've actually helped to set up this little film studio and podcast studio.
+And I'm quite excited today because this the reason why I drew like a little nature thing in the top left corner is because um we are launching a video game today that has been many years in development, and it's called Unpetrified.
+I think I left all my flyers upstairs, but yeah, yeah.
+It's it's it's a kind of it's a it's a game for little kids about nature, and it's very very lovely, and I'm very excited that we're we had the game developers here earlier.
+You know, it's oh I love launch.
+So anyway, but I'm also very excited that we're kicking off these meetings, and I hope this is the start of something long-lasting and very good.
+So over to you.
+Who is who wants to pull?
+How about Vasco?
+You were here in the meeting first, why don't you take over?
+So uh oh yeah, uh, I am Vasco.
+Um I am uh actually a mathematician and then I convert myself to machine learning.
+I did um I just also worked with uh LLMs in my last job actually a lot, and uh yeah, I I saw that uh it can be a problem that all these uh models are in the end of uh just a couple of people or a couple of organizations and uh for that reason I got interested in public AI and uh public AI and uh in also in the Swiss initiative as I live in Berlin actually.
+Um yeah, I think it's pretty much me.
+If you have questions for me, please okay.
+I will uh take take over.
+Thanks, Vasco.
+Nice to meet you, Savini here based in Zurich.
+Um Oleg and myself worked um in the project Swiss Airweeks during the last year, almost a year or nine nine months.
+So I'm one of the co-initiators of the Swiss Airweeks.
+My background is in marketing and events, and soon to be data science.
+So I'm currently writing my thesis or starting to find the topic, let's put it that way.
+And in that uh in that uh environment, I had a contact with um Joshua regarding public AI and was reading a bit about what you are doing, and uh also found this ETH challenge in terms of the MCP tools and was trying to build one of these tools for a Swiss uh voting assistant.
+Uh, and it worked on my local server, it worked, but then it didn't work for for uh I think uh Joseph.
+So this is my connection here, and I'm now joining here to a bit to understand um what what you are planning and if I can help, if it makes sense, or that's from me.
+I'll go next.
+Hi all, nice to meet you.
+I'm Natalia Birdnikov, I'm based in Zurich, and uh I am uh a project manager in IT.
+I currently am not employed, so happy to spend some time on this.
+I learned about Swiss AI project um uh through just a regular news source and was excited.
+Um filed fill out the form, and here this is how I ended up here.
+I'm interested in open source, I'm interested in Swiss, I'm interested in I, I'm interested in um coordinating, cleaning things up, documentation above all, because whenever I go, wherever I go, I find there is always a lack of documentation for um engineers and and users and everybody alike.
+So if I can be of any help, um I'd be thrilled.
+Hey, uh my name is Cole.
+I'm uh currently a software engineer at uh Web Services in the payments organization.
+Um I first heard about this through uh a former professor uh at SFU, uh Nick, and then I went to the ICML 2025, went to a few ICML things, um ICML 2025 or 2025 uh public AI uh get together thing.
+Um a few people there.
+Um pretty interested in public AI as like a you know good example of like what we could an alternative future to what we could have now, so like a good uh digital public common goods kind of idea.
+Um how can it benefit the most people as opposed to benefit what looks like just a few people um to the detriment of everyone else.
+Is that looking to help out like the first, I guess first good example of like public AI and Switzerland and hopefully other places such as Canada where I'm right now can begin to build our own and have some like uh good multilateral development discussions and things like that.
+That's awesome.
+Thank you, Cole.
+And I'm I'm glad that you could be here despite the time difference.
+Um, was that ICM 2025?
+I I couldn't quite catch exactly what you meant as the conference.
+ICM just correct ICML.
+ICML, okay.
+Anyway, feel free as to correct things in your notes and collaboratively, you know, feel free to add links to whatever you're doing and something.
+Um so nobody really takes ownership of their sketches, that's okay.
+We can move on.
+Um, so I would like to um to basically I've suggested a little agenda for us.
+I I suggest to kind of get right into it, so especially with Natalia leaving uh a bit earlier.
+Um so this is a setup meeting, right?
+So the first thing is that we have this regular time slot from four to five, and we have this big blue button platform, which is a Swiss hosted, open source, Canadian-made call, you know, um platform for meetings, right?
+But it's not um it's not like the it's not enough for us to run like a whole community on.
+It's just for the meetings.
+Oh, we have Josh coming in at the last minute.
+Um maybe he can do a I mean, I think Josh doesn't probably need an introduction, but he's maybe needs a bit of help with Big Blue Button.
+So a few people are going to complain that this is not Zoom, this is not Teams, it's not working, it's not my my microphone's not enabling.
+Okay.
+Look, tough.
+I think for some things we just need to be, we just need to set some standards.
+And for me, when it comes to public AI and sovereign AI tools like Aperitus, this is part, it's got to be part of an ecosystem.
+And I think we have to enforce the use of um at least some open platforms.
+So with this meeting, I'm kind of yeah, I'm I'm I'm being a bit naughty, I'm just kind of saying let's just use Big Blue Button.
+We could have had a poll, we could have looked at all the alternatives.
+Um this server is hosted by um CH Open, the Swiss Open Source Users Group, which is one of the oldest associations in Switzerland dedicated to promoting open source.
+Um we are organizing the DINACON conference next week.
+So that's kind of like the event that I've been posting in Slack about it's an annual gathering of the Swiss open source community.
+I hope to take some basically some directions from you today into that conference, which I'm helping to organize and set up.
+And uh, you know, we have international speakers and visitors coming, about 270 people registered.
+Um, plus we have the TAC night where and presentations of like uh startup projects and so on.
+So it'll be lots of opportunity to engage with people.
+Um and the server is uh sponsored by green.ch, which is another interesting Swiss company.
+It's a data center hosting provider, web service provider that is really sustainability focused.
+They're trying to be carbon neutral and they've been doing that for many years, and they're sponsoring everything basically by offering us to use this service.
+Okay, so that's not very consequential unless anyone has like really any problems with this with big blue button or any any anything with do you think this is do you think this is the right attitude?
+Do you think we can be a little bit militaristic when it comes to choice of tools or nothing?
+Yes.
+Um why do I hear myself?
+Jesus Christ.
+Um does it uh integrate with anything?
+Like what does it integrate with?
+Nothing at the moment.
+Um we could recommend, we could recommend it, for example, transcription, right?
+And this is something I discussed in the slides that it would be great to have the meeting notes to transcribe.
+Um they're working on it.
+It's like in the open source project, people are thinking about how to bundle AI into it and so on.
+It's used in big universities and classrooms.
+A lot of people have need for high quality recordings and so on.
+I have uh I'll I'll just stay on top of it and I'll keep and I'll and I'll and I'll ask.
+I'll I'll I'll I'll ask for you know all I can say is that we're running on the very latest release, the the most upgraded version of the software.
+So I think the system administrators are doing a good job of keeping it up to speed, but I couldn't tell you which plugins they have installed and at what point AI transcriptions on the roadmap.
+But we could go on GitHub and start an issue saying, hey, we should enable our pair tools inside of the meeting node, so you know, help of transcriptions.
+We could do that.
+I'm just kind of thinking an infrastructure of um tools and services will emerge that are needed to you know support this, like video calls or places to take notes or where documentation will be stored, like that.
+It's like we don't have we can't plan it all in advance, right?
+Like perfectly sketch it unless like we buy something super expensive that's available on the market, but like kind of just be thinking ahead and mindful of that the need may come later.
+Joseph, Josh, are you guys able to join us on voice and video?
+It'd be great to hear something from you.
+Can you guys hear me?
+Yeah, loud and clear.
+Okay, great.
+Um, I can't seem to turn on my video and my microphone at the same time.
+Uh sounds like you all.
+Good seeing you all.
+Uh, I'm John.
+I think I've met many of you through various emails and calls.
+I apologize, I'm in a very loud uh space at the moment, so I'll try to keep my words short.
+I'm really glad to come back to Big Blue button and look forward to participating and supporting what you guys are all are doing.
+And uh if I figure out how to join the text chat, I shall post some updates, or I'll post some updates to the Slack.
+That's perfect.
+Yeah, I suggest we stay on Slack on the Swiss inference channel for this for the chat, so nothing gets lost.
+And hey Joseph, you've managed to cut your video and web.
+Ah, sorry, that's the other question.
+Um no, no, I just uh we have uh the Swiss inference private channel, and I saw Tat Cole and uh Natalia and Sabine are not in.
+Josh, would you like to answer that?
+Okay, I would suggest we keep everything public as far as much as much as possible.
+I think the private channels should be there kind of like for issues with like codes of conduct or things that are really like a problem for us to solve that may not be ready for public consumption.
+But like these meetings, and unless we're keeping them off the record for whatever reason, they we can go and use the public Slack channel to keep notes.
+Yeah, just start off squeaking say hi as well.
+Um met most of you, everyone here, I guess.
+I think except uh Sabine, we've communicated over Slack and uh commented on your your GitHub PR.
+Uh it's been great.
+Yeah, good to finally meet you over video.
+And yeah, uh I am I'm engineering, I I work on engineering for public AI.
+Yeah, it's kind of late for me now, so I'm just gonna be listening in uh mostly.
+Uh happy to support, I'll probably keep my video off.
+All right, great, guys.
+Thank you so much for doing a quick intro around it.
+Yeah, um Joseph, uh even if it's late, I've just given you access to the whiteboard if you want to doodle on on Apertus.
+Feel free.
+If you think the logo needs uh needs a remake, you know, whatever.
+This is your chance.
+Anyway, um so just moving on from the fact that we have this kind of new format.
+Um I basically would like to just kind of say that I'm committed to supporting these meetings.
+I mean, um I really want to help Joseph and Josh.
+I've been, I mean, Sabine and I work together on the Swiss CI Weeks, um, but I've been um working with these type of projects, you know, digital public infrastructure projects for many years, and and I'm I'm really kind of I feel quite passionate about what we're what we're happening here, and I think I can afford to invest into it, not very much, but at least running some meetings and looking for funding and trying to get things going.
+That's something I'm committed to.
+Um, so I will be like the community manager for the time being, and you can report any problems you have to meet um and any improvement requests I would take very seriously.
+So, what I suggested in the agenda is that we talk about a few open questions in terms of the Swiss inference.
+And I would like to kind of just point out um, Vasco, you pinged me a bit earlier today about this.
+I really liked the discussion last week in the Swiss inference channel on 5th of November, where we talked about what type of government interactions user want and what kind of tools we should develop, right?
+Sabi, you also brought up the fact that you're working on an MCP project.
+For me, that kind of qualifies as part of that discussion, you know, building tools on top of the platform.
+Um, and if we're building tools, right?
+So, yeah, we can do that on our own laptops and our own infrastructure.
+As soon as we start to want to put that online and make it and make it available to other users, you know, then we need an app store.
+We need something out there.
+And um what you know Joseph and team have been working on is this you know public inference uh endpoint.
+Uh there's the there's an there's an API and everything that we've been using for the hackathons.
+And I think that one thing we can sort of at a very fundamental level ask is is it up?
+You know, we're the we're we should be like the the just the on the front line and you know checking out that the servers up and just using public AI's um existing servers every day, being aware of any downtown, being the first to report some bugs and issues as they come up, this kind of thing.
+So I was kind of wondering how you how do you folks feel about this and if anyone has any updates about some of these questions that I've put into the into the notes?
+So, like how's the usage trending?
+And we have we had a few updates over past calls about this.
+So, does anyone know like what are the numbers like this week, for example?
+Like is is the is there a trend at all?
+Is are is the is the usage of the public AI is it kind of going up or down or plateauing?
+Does anyone know?
+I'm I'm just I'm just thinking how we should look a number.
+Is there a dashboard?
+Is there a dashboard or something?
+I think there were there's a proposal to make a dashboard.
+I don't think there is a dashboard yet, but there are tools that can be used to make your own dashboard.
+For example, I have a simple dashboard that I have uh that I run on Cloudflare, which pings all my important services and sends me notification in Slack if they're if they're offline.
+You know, and you can you can use all kinds, like there's up 10 monitoring tools that run on your like your local machine or as a as a service, right?
+But we don't and the the only thing we have access to as just users of the public AI server is that number at the top uh top right or bottom right, which says how many users are online right now.
+But in the last conversations with Joseph, and right now the number is 1534.
+But Joseph did mention last time that um it's not it's not very reliable, does not necessarily represent unique users or does not account for API users, I guess, and so on.
+Any comments on that, Joseph?
+Or do we just move on from that?
+Yeah, uh sorry, the the exact question was like how many users we have?
+Is that on the front end or developer?
+More like developer.
+The more questions, like, yeah, what can we get any other numbers besides the one that we can see in the interface?
+And or and could you at least uh share something about the trends?
+Like how is the usage of the interference endpoint trending?
+Yeah, I think on the on the front end it's generally been uh platooing.
+Uh I mean we had a big spike on on the launch day, but now we are getting around like two to three hundred users uh a day.
+Um on the developer side there are some like hackathons that we've been doing with like Spain um and that is resulting in some spikes in in the developer API.
+Um but yeah, this is not split up in terms of like country, it's more just like across like all the different models and across other users at the moment.
+Great, thanks for thanks for clarifying.
+And and I am correct to think that there is a plan to make a dashboard available with some some more information.
+Is anyone working on something like that already?
+Is that just uh an idea at this point?
+Uh I do have a dashboard of uh that we are using internally, uh, but we don't have an easy way right now to expose this uh uh publicly.
+Um yeah, so the it doesn't make sense to invest into this at this point.
+Yeah, at this point.
+I see.
+What about sharing a screenshot of that dashboard like once a week with us?
+Like during these calls, we just like have a just a quick glance at like the trend line.
+Yeah, I'll be open to doing that.
+Uh yeah, I can see we can share my screen from here.
+Yeah, there's a button at the bottom center, the the central button next to the webcam to emoji.
+That's a screen sharing button, and you should be you should be able to share your screen.
+Oh but no, sorry, I just I just give it commission.
+Oh, okay.
+Yes, it's literally uh literally uh okay.
+I I I don't have it available right now.
+Um, but yeah, I'll I'll send it later on if I if I can find it.
+If I can find it.
+That's perfect.
+You could post it to the Slack channel, not urgently, maybe just like even later this week.
+It could even just be just a couple of numbers, like you know, I don't know, something that we can just keep track of and and and and and look at.
+You know, I mean, I'm kind of thinking that this this user count, that's a really basic KPI.
+That's like our starting point.
+You know, we want to drive usage of public AI.
+Or is I guess that I mean, Sabin, you're a marketing person.
+What do you think?
+What's what should be our basic trend line?
+Um, I think Oleg, there are to me, there are a lot of information around the whole thing missing.
+I don't know the strategy, what kind of campaigns or ideas are running around.
+So I'm really in the beginning to understand what is happening.
+So, what I did is to try to find out who is building uh perturbs, and now I'm seeing the screen screen here as well.
+Who is building appertos prototypes?
+Because I would like to do some video series on that.
+Um, but this is right, this is somehow where I'm thinking about and talking to people, but so far I yeah, didn't talk about much about public um uh AI.
+I was referencing it to some of the people I met during the last weeks, and they all didn't know about it, so it was great to convert some people to uh maybe give it a try.
+But I don't know what yeah, what are the uh goals and so maybe maybe you you can also share a bit what's what's on the paper at the moment to better understand, and then then uh maybe propose some ideas based on that.
+Yeah, that's an extremely good point, I think.
+Like I'm just a data guy, I love to jump straight into the numbers, but you're right.
+Actually, before we have KPIs, we need to have some goal in mind, right?
+And I'm kind of I have to say, for me, I just kind of assume that we're here to help public AI with their goals, but it's not really the case.
+I mean, public AI is a concept as much as anything, you know.
+So I think you know, it's it's a concept that comes out of this group, Metakov, Josh, and Joseph, but it's much it's much bigger than that.
+I mean, it's it's a concept that extends to a lot of other projects as well.
+And I think it's uh it's therefore it deserves.
+I mean, the more we can provide a new interpretation to that concept and to ground it in this kind of practical reality in Switzerland that people don't know about public AI, they might not even care about public AI.
+I mean, this is a little bit of divergence, but I've also been talking to people a lot.
+I talk almost every day to people now about apparatus and public AI, and I'm getting a lot of feedback, like people have basically decided it's not good.
+You know, I was I was in a public conversation yesterday, there was a speaker, and when asked about apparatus, the response was, I have not tried it, I'm not interested really in trying it.
+There was such a negative critical media community response at the beginning of September, I decided to just completely skip this version, right?
+Let's say, like version 0.1.
+I'll come back maybe in the future when there's another release, when somebody does a glowing review of it, that's when I'll test it, right?
+These kinds of things, I I think they they're they're they're they're actually quite reasonable, you know.
+When I uh look at my own behavior when it comes to new products, I I totally agree.
+I I often buy the version two or three of a product, you know, especially if I if I need to use it at work or if my family's safety is on the line or whatever, right?
+So we can count ourselves as pioneers in the Swiss AI weeks.
+We are early adopters, right?
+Um and there's and the and the public AI as a project is just um it's it's it's still something like a tool for for for hackathons, I guess that's kind of what Joseph was just saying.
+Uh but um my next question is about the server responses.
+I mean, someone um not someone, Anna.
+Is that her name?
+Yeah, Anna Sotnikova.
+She's uh a student at the ETH who is studying a pair to I mean from various perspectives.
+I don't know exactly what her background is.
+I would be it would be great if she could join us to call and introduce herself.
+But she wrote a survey, and that survey is now available inside of the public AI user interface.
+I helped her with that survey a little bit.
+I gave her some feedback and she made improvements to it.
+I don't know where those responses are going.
+I don't I don't see any of the responses.
+I don't know if Josh or Joseph, you guys are keeping track of that.
+I would just love to know at what point can we expect to get some um some more feedback from people who are users of public AI's service.
+All right, let me can you guys hear me?
+All right, amazing.
+Uh so uh I I have not seen the results either.
+Um I believe they are shared in some folder somewhere.
+Maybe like the very basic thing you can try to do is just um uh assuming that there's no like super private information in those surveys.
+I don't think so.
+Um to set up some sort of like permission architecture so that everybody just like default has access to as many resources as is feasible.
+Uh yeah, that sounds like just one of the basic things we should do.
+Uh it's the same thing with the dashboard, right?
+Um all this information should technically be public.
+The only reason it's not public is because we literally just don't have the time and resources to figure out how to make them public.
+Um given that like the most important thing we need to be careful about is protecting the private user conversations that are in the in the production database.
+But everything else is pretty much open game.
+It's just a matter of like how to figure out how do we pipe that access.
+Same thing with uh surveys.
+Let me let's ping Anna, uh, who is I believe in the public AI stack also.
+I just asked her and to share that information.
+For sure.
+Yeah.
+But I mean, just everyone here on the call, please also just fill out that survey.
+You are all users of Apparitus, and it would be really great.
+It's a little bit uncomfortable, I have to say, you can't click the link, you have to copy and paste a rather ugly Google Forms link in order to use it.
+Um but yeah, and and and it's a Google form, so which is just another pet peeve of mine.
+But hey, it's a survey, it's the first survey around public AI.
+We should just maybe like even put it out into our communities like Sabin.
+Maybe we could put it into the Swiss AI Discord.
+If anyone's still looking at that, maybe they could respond to the survey.
+Um, or or even email it directly to people who you know have built something on top of a Peritus, on top of the public AI API, especially.
+Um and and Josh, maybe you guys could also reach out to everyone who's registered um as an API user.
+You know, that would that would also just make sense for me to just send it, send a mass mailing notifying people about the yeah, that's actually really fair.
+Um we I don't think it has been, it's only on the front end that the link has been exposed.
+It has not been sent to the API developers.
+Uh that said, I think it's I would ask a very different set of questions to the API users, if it makes any sense.
+Um that one is more, or at least I would ask like some additional questions.
+Maybe we can we can discuss it with Anna and just figure out what the best way of integrating on the um survey is so that we can sort of get the most impact out of it.
+Um by the way, I also just want to say like thank you so much for uh sharing that earlier conversation you mentioned all like about hey, you know, I'm gonna adopt this when the next one comes out or the third version.
+And I maybe I should just say that you know one of the exciting things we can look forward to is that there will be for sure a second and a third version.
+Um, because I mean the models are gonna come out.
+Um we kind of know they're already coming out soon.
+So that'll be kind of exciting, and I think it's an opportunity to maybe change the minds of some of these people, uh both in the media and in our communities.
+Yeah, I have one thought on that.
+Um, because I unfortunately have to share this what uh what Oleg just said.
+So every time when I raise my voice in some kind of uh public events and promoting the model, it's a real strange uh feedback of the people.
+Some sometimes they are laughing about it, and uh so it's some kind of similar reaction, but I keep trying.
+Um, and I think what would really help is to showcase prototypes that work, right?
+So where you can see what is the value, what is the model capable of, maybe also showing some limitations or so, but maybe focus more on what it's working, showing maybe the public A AI environment more.
+So, in terms of yeah, to grab it a bit more, what's what's possible?
+So the focus to focus on the positive things.
+And so far, I couldn't find many prototypes.
+I know that Bizeka is working on one, but they're not ready.
+But they would be a great use case or prototype use case to uh do it publicly.
+But may maybe we can start with hackathon uh participants or so, maybe with the Tibetan chat board, or I don't know, but to to promote it more, or maybe I missed it, and there are some some events ongoing or channels or what whatever.
+But I think Sabine, uh let's let's let Avasco is you've had your hand up for a while.
+Uh yeah, two things.
+The first one is Tabata Data was the discussion of before.
+I thought that so something that could be interesting for us to develop.
+Um I saw that if you go on the public chat with Apertus, then you you can click uh share and you can share or on open web UI or update and copy link.
+And if you open the link, you have like like a small page with your conversation.
+And uh so my question perhaps is for Josh is if it's possible that people can actually share this entire conversation with public AI, and this conversation is safe, and then it's possible for us to um go and see the data and understand which kind of conversation people are doing, which kind of tools they need, what are the problems with the conversation?
+This could be really nice if you can save some data in this direction and we can slice, for example, just for Swiss and not for Singapore in our case, and so on.
+And this is the first thing and the second, it's about that um so you say that there is people is not uh friendly with apertos in general, um, from your experience, what you said.
+Um and I was thinking what does apertus or what is interesting from apartus that other um big company don't have um I think it's transparency is the most valuable thing.
+And uh with uh and I was reading the technical report of apartus this week and uh also with the project that called wrote in the channel.
+I think this is something that perhaps it's worth to push because you can really say, Hey, the model was trained with this data, and you can say why why um I have this kind of answer and so on.
+So I would say an interesting thing to push, it's really that it's transparent, it's we know what is behind, and so on.
+And it's something that no one actually has because OpenAI is not going to public which kind of data set they are using, for example.
+Okay, maybe I can answer quickly the first question.
+I just shared a link into the chat to one of my own public discussions about the inference utility.
+Actually, as far as I know, you it that's absolutely works and it's fine, you just need to have a user account.
+Those are not public shares, like you can't just put them in a blog or something, expect people to be just be able to open them.
+You're required to have a public AI account in order to click on that link and actually see the contents.
+Sorry, I'll put it in the Slack as well.
+Uh Joshua says he's lost a connection.
+Um and we shouldn't have to say, uh so I think the sharing is a slightly different from the feedback, right?
+Um so the yeah, you can you can definitely share your private chats, and the functionality should be yeah, unless something's broken, it should be working.
+Um for the feedback, so when you do the like the thumbs up, thumbs down, and then you send something.
+Uh so that I mean I I'm pretty sure that is something we can actually make public.
+Um and you know what?
+So can we just do like a download of that and put it in the channel?
+It's like uh last I remember it was like 2800 feedbacks uh for a purchase.
+Um and what they look like is just like was this a good or bad conversation?
+Uh we should probably uh actually come to think of it.
+Sometimes people do silly things um and accidentally make things like public that they probably shouldn't.
+So we probably still need to do like a some sort of anonymization scripting.
+Um okay.
+Yeah, we should probably do that.
+Probably important.
+But yeah, usually what it looks like um uh these feedbacks is just like somebody saying, Oh, you know, this this isn't proper Swiss German.
+Or why are why is Chinese coming up in my you know conversation in French?
+Those are things that happen uh with OMs and purchases, slowly gonna get better.
+Uh but these are you know obvious things.
+The um but uh they're they're often less focused on like use cases, they're kind of more okay.
+This is obviously something that's kind of broken here.
+The um I will say I think we uh maybe this is a conversation, actually, this is a thing, but this is actually quite a useful thing that we should be talking about is how do we get better feedback?
+Um so were you guys around for that presentation that Joseph shared about public AI Plus and uh kind of um and uh sort of uh a secondary feedback mechanism data contribution flywheel?
+Well, even if not, um maybe we should do like a quick demo of it because what that is is a more sophisticated feedback mechanism that is already uh online for Public AI Plus, which I assume everybody here is on, and if not, just send us a message on the Slack and we'll add you.
+But the key thing is maybe we can also tailor that to offer better sort of feedback.
+And I think that's really, really important if we want uh you know to make public eye actually open.
+If we wanted to, you know, be open in the sense of like, hey, anybody can contribute.
+Sorry, did that make sense?
+Yeah, I I've I've been in the presentation, I've tested it myself.
+Um, and there are I know there's people coming in with interest and expanding on it.
+So there's a that's there's a tools channel where we've been discussing kind of MCP building and and and this as well.
+But I see Joseph is giving us a quick little demo.
+Yeah, I'll just give a thing like this five seconds.
+I'm not sure I can make these interesting okay.
+Oh, this isn't even a uh what this model was.
+Yes, uh basically the idea here is uh if you're having a conversation uh over here and you want to share the conversation, uh what users can do is they can actually uh click a uh quick data fly view button and which opens a adds a row to a data set.
+Okay, this is thinking for me to open this.
+Okay, so let's say let's say it's uh this one this is using an external model.
+Okay, so let's say this is the conversation you just had and you want to contribute it, you just click this button um and it kind of does some like PII it sensors out like your personal information and you press it again and this just contributed it to uh this data set.
+So over here you can see all the contributions that people are making for this like demo data set, and the whole data set is just like every contribution is just one role in this in this data set, putting like all this um information over there, and all this can be configured uh within the chat interface from here, uh including like oh uh how do you want to be your attribution to be done, uh what's your intent for licensing, and you can write a note here.
+Uh but yeah, the only issue now is like users need to know how to get uh the hugging face like uh API token and go in here.
+So the the interface, the x the user experience isn't like ideal.
+Yeah, this is this is really awesome, and it's very much like what you have um when your operating system detects a crash.
+I don't know how many of how many times a day you experience this.
+For me, it's about two or three times a day.
+I'm doing something, and then bang, and yet this window is saying your app has crashed.
+As a developer, you're very used to these things.
+And then you have, depending on the application, the chance to send an automated report to the developers.
+Sometimes it goes straight into like a GitHub issue, which they can start working on.
+And um, in some cases, it you get like uh memory dumps and things which can potentially have very private information in them.
+And as a user, you're you decide whether or not you want to set the dump, and you can even filter out certain things from the dump of your memory if you if you want to do that.
+So that data flywheel to me is actually quite a familiar mechanism of getting kind of contextualized problem reports.
+I'm wondering how why would why would users use it otherwise?
+What's the incentive of using it in a normal day-to-day usage of of the LLM?
+And Cole, I would also love for you to pipe in a little bit because you've you've mentioned something very relevant in the chat, I think.
+I think there's uh look at some of the code for the uh demifly wheel.
+I'm not sure how robust this week for the data flywheel.
+Yeah, I mean it's definitely uh this is really was really more as of a demo.
+Um and there are like some students trying to work on productionizing this.
+Uh just on incentives, uh yeah, those are we probably need some like mechanisms or even just like marketing around like hey, this thing exists yet.
+Uh officially has not been launched.
+Um, so yeah.
+Um with with uh recognizing that um we might not have too much time left.
+I'm kind of planning to to go until five, maybe five quarter past at most.
+Um and I've um I would I'm gonna just propose that we skip a little bit some of these like other technical things that were there, like about the downtime and upgrades and so on.
+Josh, yes, go ahead.
+Oh just one thing before oh sorry.
+Okay, hopefully I'm coming through now.
+Uh one quick thing before uh we hop into the next thing.
+I just want to mention, uh, I know many people here are based around Zurich.
+And I just want to say if anybody is in Zurich on the 15th, I would love to just hang out because I'm going to be passing through.
+Uh and yeah, just reach out and it'd be nice to kind of link some you know Zoom video or I guess big blue button videos to faces.
+Yeah, that's exactly the kind of thing I actually want to talk about.
+Thank you very much, Josh.
+So Saturday, um, you're saying the 15th, right?
+Yeah.
+I've got I've got that in in in and I will I will be looking forward to um meeting you.
+I propose that we meet at the AI Center at the ETH.
+Um even though I don't actually, I'm not a I don't have an ETH badge, I have no idea if we if anyone can get in there on a Saturday.
+But to me, this is a little bit the ground zero of Apertus in Zurich.
+Um there's some nice cafes in Urbekum, whatever.
+I suggested that as a meeting point at one o'clock on Saturday.
+Yeah, great.
+Okay.
+So maybe I can maybe I can get a XM.
+Sabine, so Alex Sabine, anyone else might might join us on Saturday, Basco, you're round?
+Um I was looking I was looking.
+I was looking at I have something doing today.
+But I can join the evening if Josh is still uh around.
+I think Josh needs to take off by around seven, so it depends on your definition of evening.
+Um I I need to get back to Bern uh by I think I'm gonna stay only until five or so, but you guys can also just coordinate one-to-one.
+Um I'm also kind of yeah, squeezing this into a busy weekend schedule, but it would be really awesome to have a quick, very, very spontaneous meetup, and definitely um we can um we can plan these things more in advance.
+I already have been promoting DinaCon as the first possible event that may be relevant to you folks.
+I would also love to hear if there's any other events coming up, you know, until the end of the year at least.
+Um that would be a good good thing for us to to try to cross paths that but we can kind of do that just in the advance channel on Slack or in the Swiss inference channel.
+I don't think we need to take the time.
+What I wanted to actually do is um just agree with you that we'll just document these decisions and action items that we come up with through these meetings um in a in a repository.
+Um and just are you would you all be okay if you of using GitHub and there's like this wiki function in GitHub that's quite easy to use?
+Um could also edit markdown files if we're very hacky.
+But I would just suggest as a simple thing that we create a Swiss uh project, you know, and use that to track and like put all our documentation in one place.
+And going forward, I would try to get us off GitHub and something locally hosted.
+But right now, with all the activity around public AI being there, um would you all be you all okay with that?
+Um so oh yeah, oh yeah, I was going to propose that and then I shit and then I realized GitHub is the past.
+I mean, Hugging Face is the future.
+So I'm not sure to what extent this is actually reasonable, but at least at least in Hugging Face there's also a Git repository, and you can also edit markdown files and stuff directly there.
+So that's that's why I did it this so last so recently, like minutes before, that I actually forgot about it.
+But yeah, I've I've set up a um a Hugging Face organization.
+I thought there's kind of like a more or less active public AI organization or public AI data organization.
+I wasn't sure what that is and whatever.
+Um, but yeah.
+Let's look we'll we can figure that this out later.
+I think between GitHub or Hugging Face, they're kind of all really connected, but I think between the two of them we should have basically just a single place to store the activities of a NASA organization, right?
+And there's quite a few open questions around like what form this is all going to take.
+And for me, a little bit the reason for discussing these technical questions is still to kind of understand to what extent we're ready to actually support Metagov, Joseph, you, Josh, the others in running this instance, right?
+Um, and what what do we need to like know to prepare to even potentially running it, like costs and parameters and so on, right?
+But in the meantime, I just think you know, just having this kind of interest group and documenting our meetings going forward and um yeah, starting to do the work of putting in fundraising appeals and so on, that should probably happen on the central thing.
+And I've called it SWIU, which is very provocative.
+Yeah, here's the link, I'm putting it in the notes and in the chat.
+Um because um a Swiss interference utility, that's kind of the topic we've been discussing.
+Public AI in Switzerland, a Swiss inference utility.
+Thank you very much to whoever drew that wonderful logo.
+I've just taken it from the Excalibraw and put it on there.
+Provocative because TW is also the name with the Y of course the government's EID data app.
+Um spelled differently, probably pronounced the same.
+We probably don't want to step on toes.
+Um, so very much proposals are open what we should really call this thing.
+But I just kind of just putting it out there and just just as a just to get the ball rolling.
+Any thoughts about that?
+Okay, maybe it's just too much all at the same time.
+Nobody seems to want to chime in.
+Um but basically, yeah.
+So if you haven't already, look at the Excalibur.
+Um the goal of the Excalibur, and I'll put the link here as well, um, is to brainstorm um kind of visions for a website at least, you know, to communicate some of this really good work that's being done around the data flywheel, etc., etc., to make the interpret the public AI and so on in it in a new way.
+So this is an editing link to the to the which um yeah, so you can add your own thoughts and comments.
+Um, like Vasco's done, thank you very much.
+Cole has done, thank you.
+And Cole, you have I I know we're kind of out of time here, but Vasco already mentioned it.
+You've taken that um you've taken that mood board and you've actually come up with a whole concept.
+Do you want to just tell us in a few words what's got what's what's what's happening there?
+Yeah, so yeah, something from AI2 uh Allen Institute for AI.
+Um it's essentially a trace.
+So um, that's what kind of mention this.
+Um you want to leverage what uh capabilities you have.
+Um we have openness.
+So we can essentially create a trace of the using the open data to what could be like uh n grams backed from the output to the you know, what was the input for training and things like that.
+Um it's an interesting thing, yeah.
+Oh they have OMO 2 tray or OMO trace.
+Um it's like one of the maybe the first or the only that does that, but I want to try to have like competitive advantages where we have competitive advantages.
+Openness can be a competitive advantage, and we should do that.
+And so I was just gonna look and say, hey, um let's think big with this uh concept from AI2, the trace mechanism, and we can make it more generalizable.
+Um, right now it's quite coupled to their Infinigram algorithm, which is quite good, but we can take it further and decouple it from that, we can decouple it from any kind of implementation of a model whatsoever, such that when another, you know, nation state or whoever has like an open data um open model, we can then have this kind of trace be deployed against their their um model, or they can host it from themselves, or or there's even a future where we act as like a managed service provider for them, stuff like that, like Swiss inference utility or whatever Swiss AI as like a revenue generating mechanism, um, probably at cost or whatever for that.
+So it could be a good good utility for that.
+Uh I look at costs for it a little bit, it's a little bit expensive just because uh 60 terabyte data set is not great.
+So it's probably like a few thousand a month or something like that.
+Um but we don't really have uh any kind of source level agreement or anything like that, so latency can just be best effort, more or less.
+Um yeah, so that was kind of my concept.
+I'm gonna be working on it for a little bit.
+Um like some design work and stuff like that, try to reduce like because in the future, if we do have like more onboarding of clients and stuff like that, different use cases, operational burden on more or less volunteers right now is gonna be massive.
+And so I can get that.
+There's also what Vasco had mentioned with like um uses and things like that.
+Um, when people go into um and use the explainability tool and they want to drill down on like what the trace is for that, that's also like a signal to be like someone's interested in something that the model had said in a very general way, and they want to link it back to that.
+So by having this kind of mechanism is very useful.
+There could also be a way to generate metadata off that of people like commenting on the training data itself, so a bit like a Wikipedia peer production utility with that, but that's like very, very far off in the future.
+Um because you want you know you have data attribution, but what about what the data that you already have?
+What are we doing with that?
+Um, could be a nice utility from that.
+That's also another big project.
+I don't know how that would look.
+Um, it's also in my head um right now.
+So um, but definitely the expandability thing is like good and useful by itself and opens up a lot of cool doors for like cool citizen science stuff, cool research stuff.
+Um, I'm not sure like if it's necessarily like what customers are as customers or users or whoever are like super excited about, but uh the vibes I get from people that use LLM just like why does it say this to me?
+Why is it saying this to me?
+And this is kind of like a a little bit of an attempt at explanation.
+Um yeah, we can, and it's just like a beginning, like first idea off the top of my head for like developing uh explainability stack as it were, or explainability as a service, we could look into that more.
+I don't know about it.
+That's the end of my monologue.
+Thank you very much for my TED talk.
+All right, thanks.
+Uh Sabine already had to drop.
+I can say like I can save for another quarter hour or so.
+Um thank you, Cole.
+Any responses to that?
+Um I was reading the paper that you link it inside your proposition, your document, and it's pretty interesting.
+And I don't know if you have time this week or next week to speak from the technical point of view, just to have a chat because uh I think it's very interesting from the part.
+And although I don't know if someone would want to join.
+Yeah, it's definitely interesting.
+Um, they do a lot of cool stuff.
+Um, I actually think that could have been better built, in my opinion.
+Um, I guess like a bit of shade, but uh it's mostly just a wrapper for their Infinigram thing, but there's like a lot of caching stuff they can do because it is most likely true that not all of the data is going to be presented to everyone.
+That's like a natural read-only database kind of feature that you could probably align your caches extremely well and just not even have to read do a read from disk.
+That's like a the truth of databases if you do like locality and things like that.
+So there's some things I was thinking about, like hotspot localities, which is also an interesting um signal of like what shows up most frequently, um, what matches most frequently.
+So there's a lot of cool, interesting things you can do with this.
+But yeah, definitely, I'll definitely maybe not this week, but next week.
+Uh because if I need to figure out my schedule and things like that, we can do like a bigger, uh more uh larger deep dive.
+I would also like to have like more of a thing to present and talk about because I haven't like fully I've like glanced through the paper one time, but I would like to have more details and to look through their their they have a GitHub of like a wrapper things like that.
+I would like to have some more um more knowledge.
+So if there's questions I can answer it more or better and stuff like that.
+So I'll definitely look at that, I'll get that set up uh for next week.
+Okay, I I I also I love the data flywheel series, great suggestion to cooperate on putting these concepts out in a in a communal way and so they're not just scattered thoughts.
+Um and data traceability, it's been a topic in my mind for me for for years.
+I've been in my existing association uh proxy, this has been kind of like a topic, and I've done quite a bit of research about yeah, what kind of tools and so on and methods there are, ways of organizing even user interfaces.
+Um, so very much support this.
+I even would even I've been thinking basically that like uh I mean apparently task, I think quite a good um ability when it comes to citation mapping.
+You know, it seems it tries to be kind of scientific.
+I don't know to what extent that is on your system prompt, but I think kind of developing even just a better system prompt that pays that makes the model pay extra attention to highly readable citations, maybe even like the format of those citations being something like it that could be a differentiating factor for people who are used to these kind of academic style notes.
+Um, but yeah, it's really a big deal.
+A model coming out of a research context that is really usable for further research and that tries to be a good citizen in the research world by by double checking everything it cites, right?
+And not never never hallucinating citations because that is just the worst.
+But also kind of like this uh this this this this methods of uh you know tagging individual statements and coming up with a graph of assertions and based on things.
+That's that's something we've we've we've also been working quite a lot at open knowledge about in the frictionless data area.
+This you know, this this idea of making data portals not just something governments run, but also individual people can work with, um, anchored on the premise that we can identify authorship and assert authorship using you know you know, ciphers, you know, cryptographic methods in in the things we put on online.
+Um yeah, and yeah, and AI2 is a really awesome project.
+I mean that's that's been my default until Apertis has come along.
+Almo in the Tulu models, they're really, really brilliant.
+Yeah.
+Um I'm sorry, there's okay, the the echo is gone.
+The um uh I need to jump soon, but there is I am hoping fingers crossed um that we will be able to do some more stuff with AI too.
+Because they originally this isn't super public, but it's not super private either, either way.
+Um they um uh they did reach out to us uh about potentially doing something with respect to their new some sort of like multimodal OML um that is coming out or maybe coming out.
+But my sense is like we had a couple conversations, but they I think they're just focused on their OML 3 release and they're I I I think they're just gonna do it themselves.
+Either way, um uh obviously yeah, AI2, big fans, you know, leading uh leading open people who are actually still building really open source uh AI models.
+And you know, given that how few of them there actually are, you gotta celebrate what you got.
+But hopefully we'll be able to do something with them.
+And I think implementing Omal Trace for ourselves and maybe like getting a little bit better and then pushing that as a kind of like a PR back to them.
+That sounds like hell yes.
+Um and actually it'll be kind of probably pretty scientifically interesting too.
+Um yeah, the way I always sort of think about it is like think about like how we can kind of implement this, like do an experiment.
+You know, we can do it relatively quickly.
+Um by the way, uh one thing okay, this is you know, it's a sensitive topic because you know, access to the production database is like sensitive in all things.
+Um there's a way of sorry, I keep getting echoed.
+Um sorry Josh, I just wanted to interject very quickly.
+Um you went there at the beginning of the call when I said that if you had something sensitive you wish to discuss, I could we can take it off record.
+I can stop the recording and you can't.
+Yeah, it's it's not sensitive discussion, actually.
+It's um like I don't mind just being recorded, it's just more like there's a you guys are all of this.
+There's like a prod database.
+Um uh there's a bunch of private information there.
+I cannot like literally legally just expose that to in an open kind of way, um, even though like everything else about this wants to be as open source as possible.
+Like I just can't do that.
+So we just need to figure out how to get around that.
+Um and there are like you could like sign a legal agreement, um, or we could do something else.
+Um especially if there's like a concrete plan of like, okay, we want to implement this and we need like extra admin permissions, basically.
+But like that's basically the the way I think about it.
+Um and hopefully uh eventually we'll have enough resources and time to be able to implement this in a truly open way so that we firewall the sensitive stuff and the rest can be truly kind of just like more peer productionally.
+Um but until we get there, I'm you know, I'm happy to just like give you guys extra permissions um as long as there's like a concrete thing and we figure it out like what the legal situation is.
+Yeah, I mean this this just reminds me of the fact that you guys actually have a dev server or staging server, whatever you I've I've seen it once or twice.
+No?
+We don't we don't have a staging server, we don't have a staging server.
+Ah okay, my mistake.
+I I thought I maybe I dreamt it.
+No staging server, straight to production, uh cowboy style.
+I mean, literally, this entire thing was a giant rodeo.
+I mean, I think I must have told all of you that Swiss AI literally came to us basically a month and a half before they launched and said, like, could you help us do this?
+And we did it in a month and a half, and still like it magically worked.
+And there's a lesson, there is a more sensitive thing that I'm not gonna say here in the recording, but yeah, it I think it was a really good thing that we did do it.
+Yeah, absolutely.
+But speaking as an engineer, you can't you can't live without staging, right?
+And for us as an asscent uh project, this might actually be a very good first challenge, just to run, you know, even just the 8B model, you know, just in in more or less the same way with the same setup to to try to do some experimental things in a in a sandbox.
+Yeah, I think yeah, maybe like just setting up your own staging server.
+I mean actually, I mean not a bad idea.
+Um we could uh does that mean we have to sort of set you up with permissions and like AWS bedrock and things like that, it's possible.
+Well, actually, call if you work for AWS, so you probably have some permissions already.
+Um the uh you could just write like a uh you could just write a fake for the models.
+So it the the actual output for the models doesn't really matter so long as you fake the API well or mock it pretty well, then you can just have like an isolate, an ISO, or you could just have like a smaller version, just like a crappy local model that just gives you junk back.
+But the point isn't like the you're not working on the model itself, you're just you know wrapping up the API and working it locally.
+So it doesn't matter what it says.
+It could literally just say hello over and over again.
+Like it's not, it doesn't matter what it says as long as there's as long as you're respecting the API contract, it throws an arrow back to you if you don't.
+What you get back isn't important in this case at least.
+And then when you go up to like state, when you develop more stages, then it's like okay, um you know, like what's our other like you know, uh integration tests and stuff like that.
+That's what you have to worry about, and those are other ones.
+But for a local develop development one, there's no point in even like having any kind of permissions.
+You just do like build, test this thing, and that's out that's about it for like your tests and stuff like that.
+You know what?
+Let's um Joseph, I don't know if I'm interested, um, but the uh I think it would be great to maybe just like give you guys as much like set up a like uh basically um a staging server of some sort.
+Uh and obviously the permissions there are much like uh it's easier to give you give permissions there, and then we can just give you all the permissions.
+Um and we just like do whatever we like in the staging server.
+I will say like, you know, technically Public AI Plus is already supposed to be kind of do that.
+It's just like the staging is literally inside our open web UI instance.
+But I can see like for thick, I kind of imagine something like Almost you're gonna need like extra permissions because this is like a pretty big piece of like you need to like probably edit the underlying open web UI code.
+So maybe there's a one yeah, I don't know.
+Uh I think we ought to kind of like play around with what that should look like.
+But okay, I would just love to know what we're talking about exactly because as far as I know, platform.publi.co, that's the open source project you have under GitHub for public AI, as well as uh the open web UI configuration, which I believe you also have in your fork of open web UI.
+Um that's basically it.
+That's the open source, those are the two kind of key tools to get public AI chat running.
+Um, and I've been, you know, before I knew that you existed, I was working on trying to get open web UI running with Apertus, and it was, you know, it took a took a month and a half, but you know, last week we had uh finally the the Lama CPP support we needed to get it to get Aperitus finally running.
+And last last week, uh actually last night I talked to some folks in the in locally who are running the 70B quantized model on very reasonable GPU clusters, 24 48 gig uh yeah, 48 gigabyte video cards that fits into that VRAM.
+Um and the question for me kind of around this was more, you know, what kind of you know, how much do we actually would we need to reproduce to work on interesting problems like this to offer at least a learning opportunity?
+And you know, even like to do this kind of benchmarking, this is what we this was I thought one of the most nice and ambitious hackathon projects we had at the Swiss AWix was to try to benchmark Apertos and reproduce even some benchmarks, you know.
+I mean, do you need the 70 P model for that?
+Probably not, just to just to get the benchmark running and to get some values out of it and to compare it with other small models.
+You don't, but you're you're learning a heck of a lot, right?
+And and but I still I'm still kind of missing some fundamental inputs there because I I've also worked with VLLM.
+I know that's the default, but I don't I haven't managed to find like the Kubernetes configuration somewhere.
+So my my feeling is that probably not everything is open sourced about the the public A platform, or just the documentation is not there, and I just or I'm just too dumb but I can't figure my way around to I think this is one of those things where Natalia is uh very correct in the sense that there's just not enough documentation.
+Um I we definitely have some cloud services that are not necessarily fully open source.
+I think I'm pretty sure Zuplo is not open source.
+Um and kind of other aspects of the underlying infrastructure.
+Uh and to be quite honest with you, uh you know, open UI, open web UI is not necessarily as open source as it was before.
+The um that's like its own separate issue, but anyways, the um uh model API development locally should be sufficient.
+Yeah, I think like for me it's just like what's what's the what's the best way I can the best thing I can do is just give you the ability to deploy and start you know testing stuff at the very least.
+And we have like this is probably gonna have to lean on Joseph to kind of figure out what the what the best way of doing this is.
+Um right now public AI Plus, uh if you're already in that that basically gives you all the permissions you need to deploy tools, functions, and you know, independent models.
+So that's already pretty you know, that's most of what you can do in OpenWeb UI, right?
+That's basically it.
+The other permissions, yeah, as Joseph mentioned, is just like making like much larger front-end adjustments to the underlying OpenWeb UI instance.
+You can probably already do that just by hosting, like downloading Open Web UI and doing it locally, right?
+Because we have like a an up-to-date version.
+Uh pretty sure at all times.
+Yeah, I'm kind of starting to run out of time.
+Just just to be clear here, I think I think the intent is really, I mean, uh really put it well before.
+She had the MCP server she's working on, which is like a local use case that she's exploring.
+Um, she's doing it for her studies.
+It's not like you know, work-related, it's an experiment.
+She has it running locally.
+She wasn't very specific exactly what she means by locally.
+I assume it's Olama and maybe, I don't know, maybe some small model, maybe not even apparentist.
+And she has been working with you through the pull request to try to get that into a production setting so that she can actually interface with the full apparatus 7 TB model.
+So I think that the question we should really be discussing is kind of like the strategy of how do we explain ourselves as an association's initiative, in particular when it comes to getting like funding and support, right?
+That's that's the real key issue here.
+I mean, we can talk about cool ways to do technology hacks and experiments all day long, right?
+But I think that for me, that's the the key point here is that it is fundamental in open project.
+Anyone can download apparentness, and as of you know from the beginning, VLLM, yeah, more or less, but in the last couple of weeks, it's really become really accessible, really, and as accessible as most people need to be.
+I think the Mac users were the first, right?
+They're the always the privileged ones, right?
+They had apparently running.
+Now the rest of us can get on board.
+People can learn about the project, they can they can go on chat uh on chat public AI, they can compare the results just using the interface, and maybe our association initially is there to just explain those two sides of the coin and to create bridges between the two, right?
+To kind of assist with those pull requests, assist with that showcase of use cases, um, and and maybe help to prioritize what kind of development effort, what are those high value MCP servers, what are those high value showcases that that might actually really serve the public good?
+What do you what do you folks think about that?
+I'll just say I 100% agree.
+I will follow up async, but let's work on the pitch, uh, as you as you mentioned, and hopefully we can do more of that uh in person on Saturday.
+And I think we'll come out of it very likely with a stronger narrative that says like this is why what we're doing is important, this is how we can sort of like figure out right usage of time and sort of like so that we can make more forward progress.
+And uh in particular, sort of how that sort of change or shift our technical strategy and our sort of like basically engineering conversations, which I'm very much in favor of, and I'm glad we got to spend some uh of that time on this call doing that.
+As you mentioned, it's very easy to kind of just go back and forth on like ABCDE.
+Uh but then you know, I think to me it just feels better to like ship something.
+Right.
+Um, I uh you guys are thank you, Josh.
+You have a great time, we'll keep keep in touch.
+So Cole also dropped right.
+He said I got some action items to go ahead for trace project, begin building some plans, set up meeting for next week with Vasco.
+Um what do you think?
+How should we what do you think?
+I mean, okay, we're like 20 minutes over time now, but I would love to just uh put some closing thoughts in.
+And I think you've also been very good, Vasco, just in complimenting me so far.
+I just want to say thank you, and to that you've been a really good sparring partner and someone else physically here in Bern that is really supportive of the idea and interested.
+There's so many people yesterday at this AIML meetup, there were so many people.
+But nobody I think would come to a call like this.
+I mean, I don't know.
+It's just really good to have you there and the chief series, and also keeping you accountable.
+You're like, okay, where are we?
+What's the what's the agenda?
+Where are we going here?
+Um, this is this is going to continue to be maybe hard for me.
+So I really I really appreciate it and I hope that you'll you'll you'll stick around and support this for the weeks and months to come.
+This is just call 001.
+No, I think I think it's starting to really be really interesting that we are getting more ideas and more structure from one side.
+Um yeah, I I I I don't know what is your impression, but sometimes I have the feeling I don't know, we should do something, but I have the impression, for example, it is uh traceability, I think it's a super good idea.
+It can be really interesting, or I don't know, MC MPC server and in some way are a little bit going slow, or so building some community.
+I think it's pretty difficult in this moment.
+Um from one from my side, I will try to meet Joshua in Surich.
+He said he's leaving Sunday, so probably I can meet him Sunday.
+I think I think I did he okay.
+To me, he said he's leaving Saturday night.
+Maybe he changed his plans.
+Um he wrote in the chat he's leaving.
+I mean AF2, but he wrote something in the chat.
+Then I leave seven on the 16.
+Okay.
+Yeah, look, um so for me, like the the the thing is like even just just using public AI every day.
+When you said when you said, do we have an agenda for this meeting?
+I immediately started chatting with public AI about an agenda for this meeting, right?
+And I mean I'm just thinking everything to do with this this community, this project we're trying to do, we need to really be those those pioneers, those those really good users.
+We need to be sharing our experiences and not just filling up the survey, but probably just like putting on social media, screenshots, and blogging and so on.
+There's just two things that I haven't mentioned the call.
+I should I should really mention at least to you, and at least for the recording, I'll try to put it in notes.
+Um DinaCon, so one of the discussions I had is that there is a challenge for DinaCon, which is in the parliamentary data space, which is something that Sabine is working on, it's quite close to what her interest is.
+And um we can um we discussed the idea of benchmarking uh comparing to Mistral on a very specific question of understanding proceedings from parliamentary discussions.
+So just see more like this is like just a particular use case that's already been posted as a challenge on this DinaCon hack night, and that's that's what I plan to be working on a bit next week.
+Um, I've also continued working a little bit on the fundraising side.
+I haven't done anything in the doc since last week, except to move that top part out where we had some of these ideas about other foundations, and put that in a separate doc, and it's like a master list.
+And I've combined it with some notes I've been working on a bit earlier, like it's quite a lot.
+I've actually shortlisted like 40 organizations, maybe even 50, that I think could be potentially interested in supporting this project.
+Um, and I'm gonna just keep doing that research and very happy to collaborate on that.
+But I think the goal there would be just very systematically start putting those applications and approaching organizations for fundraising.
+And just, yeah, that document is really still essential.
+We need to figure out what's our value proposition, what are we doing here exactly?
+And that that that mood board is helping to kind of in a different way because of those screenshots and those slogans and logos, it's starting to kind of gel into kind of a communicate, something we can communicate, you know, and that's very synergistic for me.
+Yeah, do you think we should push to have uh at least one landing page where we explain the vision and uh the ideas behind to the public?
+This could be really interesting because a lot of when I ask when I speak with my friends, what are we doing your free timer?
+I am looking around with I am uh doing something with public AI, what is and then uh we are trying to help uh for the Swiss uh Swiss uh country and so on and so on, and it would be nice to give a landing page or something like so.
+Oh, what do you think about it?
+Yeah, yeah.
+Let's let's let's try to white code this landing page in over this over the weekend.
+It could just be something really just as an expression of our personal interest.
+I think the more true to life it is, you know, rather than contrived and trying to, you know.
+My problem is that I just you know, like there's just so much going on.
+And Joshua and Josh, they could talk for hours about all these things.
+It's too much.
+But there are things that are very clear to me, and I think those clear things is those that would need to communicate, and they will help us to even just to articulate our ideas to to potential sponsors or organizations that we partner with.
+And if we can do that by by Tuesday by DinaCon, then then I could kind of like spread the word there.
+I mean, there'll be 270 people I could potentially share that link with, right?
+So that's that's kind of my next deadline.
+Um, yeah.
+So for the first problem, I mean we we could try to make the meeting more structured such that uh we can go what we want, say okay, this is the agenda for today, we want to discuss that, and then we can start about speaking about staging if you are over, then we can push for that.
+And sorry, for the page, you are you going to do this or should we do a draft and put on my GitHub and share the links with you?
+As you want, like I I can I mean that just depends on time, right?
+I mean, I'm not really gonna have much time tomorrow, but Thursday I'm going to try to um I'm doing the presentation about Swiss Aye weeks actually in the morning.
+Um, so I'll be very much thinking about this.
+I'm gonna block off a couple of hours to work on this on Thursday.
+Um yeah, and I'll probably be here at Effinger.
+And if you want, you can come grab lunch or a cup of coffee and work on that together then, or just work on it independently.
+And like I said, there's this talking face space.
+That's really like the first place, which I think is like really our own.
+So I'm if if that sits well with you, I'll be happy to give you permission so you can start putting code snippets or whatever there as well.
+Right.
+For me, like that that um I it just generally think my interest is just getting more into the whole community of ML and you know, could this quantization work and this targeting face as a platform?
+I think it's it's it's it's fantastic, and it's really the home of Apertus more than anything else, for better or worse, you know.
+And both guys will be here.
+They were there last night, they're gonna be at DinaCon, they've been they're part of our local community.
+I've I didn't go to the GitHub universe last week, you know, the big conference for everyone who uses GitHub.
+And I, you know, there's like a if you know everybody uses GitHub, but I didn't like yeah, they're so far away at the same time, right?
+I have like, I don't feel like I have a really strong relationship to GitHub, right?
+And I think it's actually for many reasons we should probably get off GitHub as quick as possible.
+So that that that's my long answer to why you should just try to, if you do do anything, try to use that hugging face space, which I can make you like an administrator co-owner of right away.
+If you're happy with the name, I mean do you are you not you wouldn't be do you think it's a funny joke, or do you think people will be upset?
+It's always like a bottle, right?
+No, no, I think it's good.
+So, what's what is your hugging face username?
+Um I think uh my hugging face username is I will send you into chat.
+Okay, all right, that's my to-do list is to invite you.
+I really have to run now.
+I have people waiting for me for now for half no.
+Um, but yeah, I'll invite you right now and I will be in touch with you, and maybe we can meet up in a couple of days or maybe on Saturday or next week.
+Yeah, of course.
+So thank you for the invitation.
+My pleasure.
+Have a great night.
+Bye, Vasco.
+Bye bye.
+